Battery Modelling

Special Issue Reprint
Battery Modelling,
Applications, and Technology
Edited by
Simone Barcellona
mdpi.com/journal/energies
Battery Modelling, Applications, and
Technology
Battery Modelling, Applications, and
Technology
Editor
Simone Barcellona
Basel • Beijing • Wuhan • Barcelona • Belgrade • Novi Sad • Cluj • Manchester

Editor
Simone Barcellona
Politecnico di Milano
Milan
Italy
Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland
This is a reprint of articles from the Special Issue published online in the open access journal
Energies (ISSN 1996-1073) (available at: https://www.mdpi.com/journal/energies/special issues/
0WS8O92BP5).
For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:
Lastname, A.A.; Lastname, B.B. Article Title. Journal Name Year, Volume Number, Page Range.
ISBN 978-3-7258-0605-8 (Hbk)

ISBN 978-3-7258-0606-5 (PDF)
doi.org/10.3390/books978-3-7258-0606-5
© 2024 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license. The book as a whole is distributed by MDPI under the terms
and conditions of the Creative Commons Attribution-NonCommercial-NoDerivs (CC BY-NC-ND)
license.
Contents
About the Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Emanuele Fedele, Luigi Pio Di Noia and Renato Rizzo

Simple Loss Model of Battery Cables for Fast Transient Thermal Simulation
Reprinted from: Energies 2023, 16, 2963, doi:10.3390/en16072963 . . . . . . . . . . . . . . . . . . . 1
Dapai Shi, Jingyuan Zhao, Zhenghong Wang, Heng Zhao, Chika Eze, Junbin Wang, et al.
Cloud-Based Deep Learning for Co-Estimation of Battery State of Charge and State of Health
Dapai Shi, Jingyuan Zhao, Chika Eze, Zhenghong Wang, Junbin Wang, Yubo Lian
andAndrewF.Burke
Cloud-BasedArtificialIntelligenceFrameworkforBatteryManagementSystem
Iacopo Marri, Emil Petkovski, Loredana Cristaldi and Marco Faifer

Comparing Machine Learning Strategies for SoH Estimation of Lithium-Ion Batteries Using a
Feature-Based Approach
Simone Barcellona, Lorenzo Codecasa, Silvia Colnago and Luigi Piegari

Calendar Aging Effect on the Open Circuit Voltage of Lithium-Ion Battery
Javier Rey and Lázaro V. Cremades

Evaluation of the Power Generation Impact for the Mobility of Battery Electric Vehicles
Rongheng Li, Ali Hassan, Nishad Gupte, Wencong Su and Xuan Zhou
Degradation Prediction and Cost Optimization of Second-Life Battery Used for Energy
Arbitrage and Peak-Shaving in an Electric Grid
Tadeusz Białoń, Roman Niestrój, Wojciech Skarka and Wojciech Korski

HPPC Test Methodology Using LFP Battery Cell Identification Tests as an Example
Matteo Spiller, Giuliano Rancilio, Filippo Bovera, Giacomo Gorni, Stefano Mandelli,
FedericoBrescianiandMarcoMerlo
AModel-AwareComprehensiveToolforBatteryEnergyStorageSystemSizing
SebastianPohlmann,AliMashayekh,ManuelKuder,AntjeNeveandThomasWeyh
Data Augmentation and Feature Selection for the Prediction of the State of Charge of
Lithium-IonBatteriesUsingArtificialNeuralNetworks
MohammadaliMirsalehian,BaharehVossoughi,JörgKaiserandStefanPischinger
3D Heterogeneous Model for Electrodes in Lithium-Ion Batteries to Study Interfacial
DetachmentofActiveMaterialParticlesandCarbon-BinderDomain
v
Ireneusz Pielecha, Filip Szwajca and Kinga Skobiej
Load Capacity of Nickel–Metal Hydride Battery and Proton-Exchange-Membrane Fuel Cells in
the Fuel-Cell-Hybrid-Electric-Vehicle Powertrain
Emil Petkovski, Iacopo Marri, Loredana Cristaldi and Marco Faifer

State of Health Estimation Procedure for Lithium-Ion Batteries Using Partial Discharge Data
and Support Vector Regression
Reprinted from: Energies 2023, 17, 206, doi:10.3390/en17010206 . . . . . . . . . . . . . . . . . . . . 204
Richard Bustos, Stephen Andrew Gadsden, Mohammad Biglarbegian, Mohammad AlShabi

and Shohel Mahmud
Battery State of Health Estimation Using the Sliding Interacting Multiple Model Strategy
Reprinted from: Energies 2024, 17, 536, doi:10.3390/en17020536 . . . . . . . . . . . . . . . . . . . . 218
vi
About the Editor
Simone Barcellona
Simone Barcellona was born in Milan, Italy, in May 1985. He obtained both his M.Sc. and Ph.D.
degrees in electrical engineering from Politecnico di Milano, Milan, Italy, in July 2011 and December
2014, respectively. During his Ph.D. studies, his research focused on the theoretical and mathematical
topology and operation of power static converters, as well as the study of conservative functions
related to the switching networks in which the converters operate. Currently, he is a researcher at
Politecnico di Milano, where his research focuses on power electronic converters and energy storage
systems, with a specific emphasis on modeling and parameter estimation methods for lithium-ion
batteries.
vii
Preface
In this reprint, 14 works are presented on the topic of battery modeling and estimation methods
for their parameters and states. Specifically, they cover battery state of health and state of charge
estimation using different strategies such as sliding interacting multiple model partial discharge
data, support vector regression, and artificial neural networks. Additionally, the reprint includes
studies on 3D models for electrodes in lithium-ion batteries, focusing on the interfacial detachment of
active material particles, as well as evaluations of the power generation impact for mobility in electric
vehicles. Finally, a review article showcases the most recent developments in battery management
systems using cloud-based artificial intelligence.
Simone Barcellona
Editor
ix
energies
Article
Simple Loss Model of Battery Cables for Fast Transient
Thermal Simulation
Emanuele Fedele, Luigi Pio Di Noia and Renato Rizzo *
Department of Electrical Engineering and Information Technology, Università di Napoli Federico II,
Via Claudio 21, 80125 Napoli, Italy
* Correspondence: renato.rizzo@unina.it
Abstract: In electric vehicles, currents with high-frequency ripples flow in the power cabling system
due to the switching operation of power converters. Inside the cables, a strong coupling between the
thermal and electromagnetic phenomena exists, since the temperature and Alternating Current (AC)
density distributions in the strands affect each other. Due to the different time scales of magnetic and
heat flow problems, the computational cost of Finite Element Method (FEM) numeric solvers can
be excessive. This paper derives a simple analytical model to calculate the total losses of a multi-
stranded cable carrying a Direct Current (DC) affected by a high-frequency ripple. The expression of
the equivalent AC cable resistance at a generic frequency and temperature is derived from the general
treatment of multi-stranded multi-layer windings. When employed to predict the temperature
evolution in the cable, the analytical model prevents the use of complex FEM models in which
multiple heat flow and magnetic simulations have to be run iteratively. The results obtained for the
heating curve of a 35 mm2 stranded cable show that the derived model matches the output of the
coupled FEM simulation with an error below 1%, whereas the simple DC loss model of the cable
gives an error of 2.4%. While yielding high accuracy, the proposed model significantly reduces the
computational burden of the thermal simulation by a factor of four with respect to the complete
FEM routine.
Keywords: cable; thermal analysis; skin and proximity effects; battery storage; ampacity
Citation: Fedele, E.; Di Noia, L.P.;

Rizzo, R. Simple Loss Model of 1. Introduction
Battery Cables for Fast Transient Due to the ongoing effects of climate change, a significant reduction in greenhouse gas
Thermal Simulation. Energies 2023,
and pollutant emissions has become crucial. Such reduction demands a rapid transition
16, 2963. https://doi.org/10.3390/
from thermal to electrical transportation [1–7]. Electric vehicles require the use of onboard
en16072963
energy sources and storage systems such as fuel cells [8] and electrochemical batteries [9].
Academic Editor: Simone Barcellona Battery electric vehicles are the most promising technology, although the actual electro-
chemical technology still needs improvements [10]. One of the main limitations to tackle
Received: 22 February 2023
is the reduction in volume and weight of battery cells so as to achieve higher values of
Revised: 15 March 2023
specific power (kWh/kg) and power density (kWh/m3 ). To this aim, abundant research
Accepted: 22 March 2023
on promising chemistries for the electrode and electrolyte materials has been carried on in
Published: 23 March 2023
recent years [11–14]. Besides the inherent tradeoff between weight or volume and vehicle
mileage, other technical issues associated with the correct management of the battery pack
also exist. In fact, the use of a battery management system able to monitor the electrical
Copyright: © 2023 by the authors. and thermal behavior of the cells is mandatory, and optimized thermal management is of
Licensee MDPI, Basel, Switzerland. utmost importance to obtain a lighter and more compact battery pack for a given power
This article is an open access article rating [15]. Indeed, the removal of heat from the cells due to the flow of current and the
distributed under the terms and internal entropy variation is one of the causes that lead to a reduction in the power density
conditions of the Creative Commons of the battery pack. The thermal behavior of electrochemical cells is widely studied in the
Attribution (CC BY) license (https:// literature, and numerical and model-based techniques have been proposed for the correct
creativecommons.org/licenses/by/
evaluation of the thermal state of the cells [16–19]. The correct performance of a battery
4.0/).
Energies 2023, 16, 2963. https://doi.org/10.3390/en16072963 https://www.mdpi.com/journal/energies

1
Energies 2023, 16, 2963
electric vehicle depends also on the proper operation of external devices attached to it, such
as the DC/DC converters used as an interface to the electrical drives, the protection devices,
and the wiring system [20]. The power cables play an important role in guaranteeing the
vehicle’s performance, especially from a reliability point of view [21,22]. In fact, due to
the low volume availability and the non-negligible size of the cables, redundant wiring
systems cannot be implemented onboard electric vehicles. Therefore, a failure of the main
power cables between the battery pack and the propulsion system can rapidly lead to
an unexpected stop of the electric vehicle. The damage and complete failure of power
cables occur due to rapid degradation of the insulation, which can be caused by excessive
electrical stress or overheating [23].
Virtually all the battery packs employed onboard EVs are interfaced with the traction
and auxiliary loads through static power conversion stages. For this reason, the battery
cabling system usually operates with distorted DC currents characterized by high-frequency
ripples. Evaluating the thermal behavior of the cable in this condition is nontrivial, due to
the mutual correlation between heat generation and electromagnetic phenomena, such as
skin and proximity effects occurring in the conductor bundles [24,25]. On the other hand,
the electrical power demanded by electric propulsion usually varies with fast dynamics
determining sudden changes in the cable current, making transient analyses necessary.
Complete numerical approaches based on the Finite Element Method (FEM) are often
employed in the literature [26–30]; however, the high computation time required by the
FEM simulations on a real multi-stranded cable geometry makes the method of application
difficult for transient thermal simulations. The main idea of this paper is to derive a simple
analytical model that provides a good estimation of the losses in high-power battery cables
due to the flow of currents affected by high-frequency ripples. The model builds upon
the premises of analytical equations known in the literature for the AC losses of round
conductors in multi-layer transformer windings [31,32] and matches with good accuracy the
results of electromagnetic and thermal FEM simulations. Thus, it can represent a fast and
affordable means to evaluate the transient thermal behavior of the battery cabling system.
The remainder of this paper is organized as follows: Section 2 briefly introduces
a typical battery EV powertrain architecture and its power cabling systems. Section 3
presents analytical loss models for multi-stranded cables carrying pure AC and distorted
DC currents and provides a relation for the calculation of cable resistance at a generic
temperature and frequency; in Section 4, the accuracy of the analytical loss models is
assessed through the use of magnetic FEM simulations; in Section 5, a transient analysis
of the cable heating is carried out using the simple method proposed in the paper and
compared to that obtained by a full numerical routine comprising magnetic and thermal
FEM simulations. Finally, Section 6 remarks on the methods and the results of the work
and draws conclusions.
2. Battery Electric Vehicles Cabling System

Electric vehicles require an intensive adoption of power electronic converters, as they
enable a flexible exchange of power between the battery pack, the propulsion motor drives,
the auxiliary loads, and the onboard or offboard chargers. Figure 1 gives a sketch of a
typical power circuit scheme of a battery EV, comprising the propulsion system and the
onboard AC charger.
The DC/DC power converters play an important role in interfacing the battery pack
with external power sources and loads. In fact, one DC/DC converter is often connected
between the battery pack and the traction inverter to adapt the battery voltage to the level
of the DC link and provide a stabilized voltage at the inverter input terminals. One more
DC/DC converter is always included in the onboard charger to control the recharge of the
battery pack from the external AC supply as prescribed by the charging profile and BMS
limitations. As is well known, the operation of the power converters inevitably introduces
some harmonic distortion in the currents and voltages of the power system and places extra
stresses on every component of the drivetrain. For its part, the vehicle wiring system must
2
Energies 2023, 16, 2963
be designed to withstand a number of stress factors, namely, steep voltage transients at the
motor terminals, high environmental temperatures, and non-negligible ripple content in the
DC current drawn from the battery. As is clear from the circuit scheme, such high-frequency
ripples are always present in the battery current, since power converters are operating
during both propulsion and vehicle recharging.
GRID AC/DC + DC/DC
MOTOR DC/AC DC/DC BATTERY
Figure 1. Typical EV power circuit including the propulsion system, the onboard charger, and the
power cables.
For high-voltage EV applications, single-core multi-stranded high-temperature ca-

bles with PVC or silicon rubber insulation are usually employed in the DC stages of the
propulsion system. The physical dimensions of the conductor core and insulation, as well
as the nominal number of strands and their diameter for every standard cross-section, are
prescribed by industry standards that cable makers are required to comply with [33,34].
Due to the multi-stranded geometry of power cables and the high-frequency harmonic
of the battery current, additional Joule losses related to skin and proximity effects among
strands can occur. Under particular ambient conditions, these extra Joule losses may cause
excessive overheating, with a consequent fast degradation or even complete failure of the
insulation layer. Therefore, an adequate evaluation of the cable’s thermal behavior is of
primary importance for the design and validation of the battery wiring system. In fact, a
robust design can prevent accelerated aging, which can cause the intervention of protection
devices or the fatal failure of the system, with consequent safety hazards [23].
3. Modeling of Cable Losses

3.1. Sinusoidal Current Waveform
The power dissipation per unit length P0 of a round conductor subjected to an exter-
and carrying a sine current of amplitude I0 at frequency f 0 can be
nal magnetic field H
expressed as [35]
ρI 2 √ πρd0 2
P0 = √ 0 ψ1 (Δ) − 2 H ψ2 (Δ) (1)
2πδd0 δ
where d0 is the conductor diameter, ρ is the material resistivity, and δ is the skin depth,
which is given by
ρ
δ= (2)
πμ0 f 0
In Equation (1), Δ is the normalized diameter with respect to the skin depth, i.e., Δ = d0 /δ,
while ψ1 (Δ) and ψ2 (Δ)) are functions of the real and imaginary parts of the k-th order
Kelvin function bek ( x ) and their derivatives [36].
3
Energies 2023, 16, 2963
The geometry of a multi-stranded DC cable is schematized in Figure 2. Each strand is

subjected to the magnetic field H produced by the currents in all the other strands inside
the cable. This field can be expressed in polar coordinates as
(r, θ ) = I
H rîθ (3)
2πrc2
where I is the total current carried by the cable, and rc is the total radius of the cable.
rc d0
r
q
H(r,q)
Figure 2. Multi-stranded cable geometry.
By denoting with N0 the number of strands, the power loss density per unit volume
of the cable can be expressed as
N0 P0
p= (4)
πrc2
By introducing the packing factor β,
N0 r02
β= (5)
rc2
and using Equations (1) and (4), the power loss density rewrites as
β I02 ρβ √ ρβ 2
p(r, θ ) = P0 = √ ψ1 (Δ) − 2 2 H ψ2 (Δ) (6)
πr02 2 2π 2 δr03 δr0
By integrating Equation (6) over the entire cable cross-section, the total AC power
losses per unit length of the cable can be obtained:
2π rc

I 2 ρβ √ ρβ 2
P= √0 ψ1 ( Δ ) − 2 2 H ψ2 ( Δ ) drdθ
0 0 2 2π 2 δr03 δr0
ρI 2
= √ [ψ1 (Δ) − βN0 ψ2 (Δ)] (7)
2πδN0 d0
The above result differs from the usual expression of AC losses in multi-stranded
multi-layer transformer windings [32,36] and is peculiar to multi-stranded power cables in
which, differently from transformer windings, the external leakage magnetic field can be
considered of negligible magnitude. From Equation (7), the AC equivalent resistance of the
cable can be derived straightforwardly as
√
P 2ρ
RAC = 2 = [ψ (Δ) − βN0 ψ2 (Δ)] (8)
I /2 πδN0 d0 1
4
Energies 2023, 16, 2963
Since the DC resistance per unit length of the cable is simply equal to the ratio between
the resistivity and the cross-section,
4ρ
RDC = (9)
N0 πd20
the AC to DC resistance ratio k R is equal to
RAC Δ
kR = = √ [ψ1 (Δ) − βN0 ψ2 (Δ)] (10)
RDC 2 2
By recalling that the following Taylor series expansions hold true for ψ1 (Δ) and ψ2 (Δ),
√ 1 1 1

ψ1 (Δ) = 2 2 + 8 Δ3 − 14 Δ5 + . . . (11a)
Δ 32 32

1 1 3 1 7
ψ2 (Δ) = √ − 5 Δ + 12 Δ + . . . (11b)
2 2 2
and considering the terms up to the third power of Δ, the ratio k R can be ultimately
rewritten as
√
Δ 1 1 1 1 1 + 6βN0 4
kR = √ 2 2 + 8 Δ3 + βN0 √ 5 Δ3 = 1 + Δ (12)
2 2 Δ 32 22 3 28
Substituting Equation (12) into Equation (10) yields the ultimate expression for the
AC cable resistance at frequency f 0 :

4ρ 1 + 6βN0 4
RAC = 1+ Δ (13)
N0 πd20 3 28
where the dependency on frequency is not explicit but contained in the normalized skin
depth Δ.
3.2. Generic Current Waveform

The above analysis can be extended to a generic periodic current waveform with DC
and AC components, as typically encountered in electrical drive systems such as an EV
powertrain. An arbitrary periodic current waveform of fundamental frequency f 0 can be
expressed by means of its Fourier series expansion as
+∞ √
i (t) = IDC + ∑ 2Ih sin(2πh f 0 t − φh ) (14)
h =1
where the closed-form expression of Ih and φh for numerous waveforms often encoun-
tered in power electronic applications can be found in the literature. The Joule losses
corresponding to such arbitrary current waveforms are then given by

+∞ +∞
(h) (h) 2
P = RDC IDC
2
+ ∑ RAC Ih2 = RDC 2
IDC + ∑ kR Ih (15)
h =1 h =1
(h) (h)
where RAC and k R represent the AC cable resistance at frequency h f 0 and its normalized
value, respectively. The normalized strand diameter at frequency h f 0 is equal to
d0 d0 √
Δh = = = hΔ (16)
δh 1/ρπμ0 h f 0
5
Energies 2023, 16, 2963
(h)
from which the following expression for k R is derived:
√
(h) hΔ √ √
kR = √ ψ1 ( hΔ) − βN0 ψ2 ( hΔ) (17)
2 2
By recalling the Taylor series expansion provided in Equation (11a,b), Equation (17)
rewrites as
(h) 1 + 6βN0 2 4
kR = 1 + h Δ (18)
3 28
The above expression is useful because it only requires the computation of the nor-
malized skin depth at fundamental frequency f 0 . Equation (18) can be substituted into
Equation (15) to calculate the total AC losses in the cable for a generic current waveform:
+∞
1 + 6βN0 4 2 2
P = RDC IDC 2
+ ∑ 1+ Δ h Ih (19)
h =1
3 28
4. Validation with Magnetic FEM Solver

The analytical expression of the AC cable resistance given by Equation (13) is validated
against the results of the 2D Finite Element Method (FEM) software FEMM 4.2. The software
solves the following phasor equation for the complete multi-stranded wire geometry:

1 ω 1
∇× ∇ × A = −j A + J s − ∇V (20)
μ( B) ρ ρ
where A and Js are the phasors of the magnetic induction vector potential and source current
density, respectively. Several simulations are carried out for different frequencies and
temperatures. This is necessary because the temperature affects the conductor resistivity ρ,
which in turn influences the AC resistance, both directly through the resistivity temperature
coefficient and indirectly through the skin depth. The comparison is carried out on a
35 mm2 high-temperature DC cable for automotive applications whose main parameters
are reported in Table 1.
Table 1. Cable parameters.
Parameter Value
Conductor material Copper
Nominal section 35 mm2
Strand diameter 0.41 mm
Number of strands 276
Ampacity 280 A at 125 ◦ C
Insulation material Silicon rubber
Insulation thickness 1.04 mm
Temperature range −40 ◦ C to +180 ◦ C
The current density distribution inside the individual cable strands as computed by
the FEM solver at 280 A, 20 ◦ C, and for a frequency of 5 kHz and 50 kHz is shown in
Figure 3.
The FEM results show how the impact of skin and proximity effects on the current
density distribution differs between the two cases. At a frequency of 5 kHz, the skin depth
of copper at 20 ◦ C is 0.92 mm, approximately two times the diameter of the strands. For
this reason, the current density distribution is not far from being uniform and has peaks of
9 A/mm2 . On the other hand, the skin depth at 50 kHz lowers to 0.29 mm, approximately
0.71 times the strand diameter. Hence, the distribution of J becomes much more nonuniform
with maximum values of 48 A/mm2 , and higher Joule losses occur in the cable.
6
Energies 2023, 16, 2963
(a) (b)
Figure 3. Current density in the cable at 20 ◦ C and at a frequency of 5 kHz (a) and 50 kHz (b).
The AC equivalent resistance of the cable is evaluated in the post-processing phase

based on the overall losses occurring in the conductor volume and the square of the current:

Vc ρ J dV
2
RAC = 2 (21)
1/2 Sc JdS
where Sc and Vc are the cable cross section and volume, respectively. In Figure 4, the results
of the FEM post-processing given by Equation (21) are compared with those given by the
analytical Equation (13) for different frequencies and temperatures.
Figure 4. Comparison between analytic and FEM-calculated AC resistances at different frequencies

and temperatures.
7
Energies 2023, 16, 2963
As can be seen, the analytical AC resistance equation matches with high accuracy the
FEM results and proves effective in modeling the high-frequency skin and proximity effects
in the strands and their complex dependency on temperature. Indeed, it is worth remarking
that a higher temperature determines an increase in resistance at DC and low frequencies,
but a reduction in resistance at high frequencies due to the increase in the skin depth. This
is the reason why the cable has a higher resistance at 20 ◦ C than at 180 ◦ C at 50 kHz. The
analytical model properly captures this opposite behavior at low and high frequencies.
5. Application to Thermal Transient Analysis

Thermal calculations of multi-stranded cables in free air require careful consideration
of the mutual couplings between electromagnetic and thermal field problems [37]. In fact,
the temperature inside the strands influences the resistivity and thus the magnetic field and
current density distributions, which in turn determine the power losses inside the cable and
its temperature evolution. While FEM analysis also proves effective in the thermal domain,
the coupling between electromagnetic and thermal fields adds considerable computational
costs, as the time scales of the two problems differ significantly [38]. If a generic current
waveform is considered with multiple harmonic frequency magnetic problems, the compu-
tational burden of a fully coupled FEM model can become excessive. The practical benefit
of using the analytical model of the total cable losses is therefore evaluated with application
to the transient thermal analysis of the 35 mm2 cable introduced in the previous section.
A triangular current waveform is considered, which corresponds to the operation of a
DC/DC converter attached to the battery. With reference to the Fourier series expansion in
Equation (14), the following relations apply for a triangular shape [39]:
Δi sin(hπd)
Ih = (22a)
d (1 − d ) π 2 h2
φh = − hπd (22b)
where Δi is the peak-to-peak current ripple and d is the duty ratio.

The flowchart of the cross-coupled FEM simulation of the thermal transient is depicted
in Figure 5.
For the triangular current at fundamental frequency f 0 , the first Nh harmonics shall be
considered. At time step k, the temperature distribution
√ inside the strands is used to solve
Nh FEM magnetic problems at peak current 2Ih and frequency h f 0 . The losses due to
the harmonic currents are added to the DC losses and passed as input to the FEM thermal
model that solves the heat flow equations and updates the temperature distribution, which
is then used as the initial state of the subsequent iteration. Given Nt steps over the entire
thermal simulation, Nt × Nh magnetic problems plus Nt thermal problems must be solved
by two dedicated FEM routines.
On the other hand, the analytical loss model does not need Nh magnetic problems to
be solved at each time step of the thermal simulation, reducing the total number of FEM
calculations to Nt . In fact, at each time step, temperature θ is employed to calculate the DC
resistance RDC and normalized strand diameter Δ. These values are used by Equation (19)
to evaluate the total Joule losses through a single computation. The total losses are then
input to the thermal FEM solver to calculate the evolution of the temperature distribution.
The analysis was carried out for a current of 420 A of DC value (+50% with respect
to the nominal ampacity) and 10% ripple. Natural convection on the external insulation
surface at an air temperature of 60 ◦ C is assumed. Matlab was employed to manage the
iterative execution of electromagnetic and thermal simulations performed by FEMM 4.2 and
for all the required pre- and post-processing calculations and plotting. Figure 6 shows the
temperature distribution inside the cable as computed by the cross-coupled FEM routine
after 10, 20, 30, and 40 min. As dictated by the circular symmetry of the problem, the heat
flows from the core to the outer surface of the cable, and a temperature gradient of 10 to
20 ◦ C is observed between the inner strands and the outer surface of the insulation layer.
8
Energies 2023, 16, 2963
After 40 min, the temperature in the silicon rubber has reached 193 ◦ C, which is above the
maximum continuous temperature rating due to a current higher than the rated ampacity.
Definition of cable geometry

Fourier series expansion
and current waveform
of the current
start
q(0)
q(k+1)
continue?
MAGNETIC DOMAIN
THERMAL DOMAIN Current density distribution at
Temperature distribution temperature q(k) and frequency f0
inside the cable
(strands and insulation) Current density distribution at
temperature q(k) and frequency Nh f0
P(k) Calculation of Jh(k)

total Joule losses
Figure 5. Flowchart for cross-coupled numerical evaluation of the cable thermal transient evolution.
(a) (b)
(c) (d)
Figure 6. Temperature distribution inside the cable as computed by the coupled FEM routine after
10 min (a), 20 min (b), 30 min (c), and 40 min (d).
The variations in the average temperature of the strands and insulation as computed
by the coupled FEM routine and analytical loss model are compared in Figure 7. In the
plots, a third curve is also reported which represents the temperature evolution when only
DC losses are considered, i.e., for P = RDC IDC
2 .
9
Energies 2023, 16, 2963
Full FEM
Complete loss model + FEM
DC loss model + FEM
Conductors
Full FEM
Complete loss model + FEM
DC loss model + FEM
Insulation
Figure 7. Temperature evolution in the conductors and insulations computed through different methods.
The comparison shows that the complete loss model accurately estimates the losses in-
side the cable and matches the thermal transient curve yielded by the complete FEM routine.
However, it should be remarked that the coupled FEM simulation lasted approximately
17 h, while the one relying on the total loss analytical model lasted 4 h (both simulations
were run on the same machine with 16 GB RAM and a six-core 3.10 GHz processor). On the
other hand, the temperature profile obtained through the simple DC loss model is less ac-
curate and presents an error that increases with the temperature. This increasing deviation
can be explained as follows: The DC loss model underestimates the total cable resistance,
which results in lower Joule losses and a consequently slower temperature increase. On
the other hand, since the cable resistance increases with temperature (both in the DC and
complete loss model), a slower temperature increase determines a slower increase in time
of the internal losses. This inherent bidirectional coupling between temperature and power
losses results in a discrepancy between the complete and DC loss model that increases
with time. Based on the thermal constant of the cable here considered, this discrepancy
becomes visible after 10 min. At the end of the simulation, the difference in the insulation
temperature yielded by the two models amounts to 4.6 ◦ C (2.4%). This discrepancy, while
not excessive, suggests that the complete loss model given by Equation (19) may be pre-
ferred when a more precise evaluation of the cable temperature evolution under specific
load or ambient conditions has to be performed. It is worth remarking that this higher
accuracy is obtained at no additional cost, as the DC and complete loss models share the
same negligible computational burden.
10
Energies 2023, 16, 2963
6. Conclusions
The DC cables connecting the battery pack to the traction system of an electric vehicle
play an important role in guaranteeing the performance, reliability, and safety of the vehicle.
The lifetime of a wiring system strongly depends on the temperature reached during
operation. Many factors can affect the temperature of the cable, including the ambient
conditions and the harmonic distortion in the current caused by the switching operation of
power electronic converters attached to the battery. To evaluate the temperature evolution
of DC cables under defined ambient and operating conditions, numerical approaches
based on the Finite Element Method (FEM) represent the prevailing technique due to the
versatility and accuracy of FEM solvers. In a multi-stranded cable geometry, which is
usually found in battery power connections, a strong coupling between the thermal and
electromagnetic phenomena exists, that is, they constitute a multi-physics problem. In
fact, the temperature reached by the cable primarily depends on the Joule losses in the
conductors, i.e., on the current density distribution within the strands. On the other hand,
the distribution of harmonic AC current densities is affected by skin and proximity effects,
whose impacts vary with frequency and temperature. Due to the very different time scales
of magnetic and heat flow problems and the many harmonic components found in the
current, a numerical solution to the problem can often be computationally hard.
This paper presented a simple analytical model to calculate the total losses occurring in
a multi-stranded cable carrying a DC current that is affected by a high-frequency ripple. By
considering the stranded geometry of the cable and the magnetic field inside it, the model
can capture the joint effects of temperature and skin and proximity effects on the total Joule
losses per unit length of the cable. The analytical equations describing the equivalent AC
resistance of the cable at a fixed frequency and temperature were derived from the general
treatment of high-frequency effects in multi-stranded multi-layer winding geometries and
were validated against the results of time-harmonic FEM magnetic models. The model
proved effective in matching with high accuracy the resistance values obtained through
FEM analysis in the entire range of variation considered for frequency and temperature.
The effective advantage of using the proposed loss model to predict the tempera-
ture evolution inside the cable with a reduced computational burden was also assessed.
Specifically, the output of a coupled electromagnetic–thermal FEM simulation comprising
electromagnetic and heat flow calculations was compared to a simpler calculation routine,
in which the loss density input to the heat flow FEM model was evaluated by means of the
derived analytical loss model. The results showed that the temperature curve yielded by the
simplified routine almost overlapped with that produced by the coupled FEM simulation.
However, the simplified routine required only one-fourth of the computation time with
respect to the full FEM simulation. On the other hand, the temperature evolution given by
the simpler DC loss model was found to be affected by an error that increased with time
and reached 2.4% at the end of the simulated time window. These results suggest that the
proposed analytical modeling of total Joule losses in the multi-stranded cable can be of use
when a more detailed evaluation of the temperature is to be performed at a largely reduced
computation cost.
Author Contributions: Conceptualization, E.F., L.P.D.N. and R.R.; methodology, E.F., L.P.D.N. and
R.R.; software, E.F. and L.P.D.N.; validation, E.F. and L.P.D.N.; formal analysis, E.F.; investigation, E.F.
and L.P.D.N.; resources, E.F. and L.P.D.N.; data curation, E.F.; writing—original draft preparation,
E.F.; writing—review and editing, L.P.D.N. and R.R.; visualization, E.F.; supervision, L.P.D.N. and
R.R.; project administration, R.R. All authors have read and agreed to the published version of
the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Conflicts of Interest: The authors declare no conflict of interest.
11
Energies 2023, 16, 2963
References
1. Husain, I.; Ozpineci, B.; Islam, M.S.; Gurpinar, E.; Su, G.J.; Yu, W.; Chowdhury, S.; Xue, L.; Rahman, D.; Sahu, R. Electric Drive
Technology Trends, Challenges, and Opportunities for Future Electric Vehicles. Proc. IEEE 2021, 109, 1039–1059. [CrossRef]
2. Sanguesa, J.A.; Torres-Sanz, V.; Garrido, P.; Martinez, F.J.; Marquez-Barja, J.M. A Review on Electric Vehicles: Technologies and
Challenges. Smart Cities 2021, 4, 372–404. [CrossRef]
3. Fedele, E.; Iannuzzi, D.; Del Pizzo, A. Onboard energy storage in rail transport: Review of real applications and techno-economic
assessments. IET Electr. Syst. Transp. 2021, 11, 279–309. [CrossRef]
4. Sun, X.; Li, Z.; Wang, X.; Li, C. Technology Development of Electric Vehicles: A Review. Energies 2020, 13, 90. [CrossRef]
5. Brelje, B.J.; Martins, J.R. Electric, hybrid, and turboelectric fixed-wing aircraft: A review of concepts, models, and design
approaches. Prog. Aerosp. Sci. 2019, 104, 1–19. [CrossRef]
6. Fang, S.; Wang, Y.; Gou, B.; Xu, Y. Toward Future Green Maritime Transportation: An Overview of Seaport Microgrids and
All-Electric Ships. IEEE Trans. Veh. Technol. 2020, 69, 207–219. [CrossRef]
7. Del Pizzo, A.; Polito, R.M.; Rizzo, R.; Tricoli, P. Design criteria of on-board propulsion for hybrid electric boats. In Proceedings of
the The XIX International Conference on Electrical Machines—ICEM 2010, Rome, Italy, 6–8 September 2010; pp. 1–6. [CrossRef]
8. Sorlei, I.S.; Bizon, N.; Thounthong, P.; Varlam, M.; Carcadea, E.; Culcer, M.; Iliescu, M.; Raceanu, M. Fuel Cell Electric Vehicles—A
Brief Review of Current Topologies and Energy Management Strategies. Energies 2021, 14, 252. [CrossRef]
9. Deng, J.; Bae, C.; Denlinger, A.; Miller, T. Electric Vehicles Batteries: Requirements and Challenges. Joule 2020, 4, 511–515.
[CrossRef]
10. Zhao, G.; Wang, X.; Negnevitsky, M. Connecting battery technologies for electric vehicles from battery materials to management.
iScience 2022, 25, 103744. [CrossRef]
11. Nitta, N.; Wu, F.; Lee, J.T.; Yushin, G. Li-ion battery materials: Present and future. Mater. Today 2015, 18, 252–264. [CrossRef]
12. Balasankar, A.; Arthiya, S.E.; Ramasundaram, S.; Sumathi, P.; Arokiyaraj, S.; Oh, T.; Aruchamy, K.; Sriram, G.; Kurkuri, M.D.
Recent Advances in the Preparation and Performance of Porous Titanium-Based Anode Materials for Sodium-Ion Batteries.
Energies 2022, 15, 495. [CrossRef]
13. Julien, C.M.; Mauger, A.; Zaghib, K.; Groult, H. Comparative Issues of Cathode Materials for Li-Ion Batteries. Inorganics 2014,
2, 132–154. [CrossRef]
14. Li, Q.; Xu, C.; Huang, B.; Yin, X. Rhombohedral Li1+xYxZr2-x(PO4)3 Solid Electrolyte Prepared by Hot-Pressing for All-Solid-State
Li-Metal Batteries. Materials 2020, 13, 1719. [CrossRef] [PubMed]
15. Hamed, M.M.; El-Tayeb, A.; Moukhtar, I.; El Dein, A.; Abdelhameed, E.H. A review on recent key technologies of lithium-ion
battery thermal management: External cooling systems. Results Eng. 2022, 16, 100703. [CrossRef]
16. Kleiner, J.; Stuckenberger, M.; Komsiyska, L.; Endisch, C. Advanced Monitoring and Prediction of the Thermal State of Intelligent
Battery Cells in Electric Vehicles by Physics-Based and Data-Driven Modeling. Batteries 2021, 7, 31. [CrossRef]
17. Piccirillo, F.; Bianco, N.; Di Noia, L.P.; Guerriero, P.; Iasiello, M. Analysis of the Thermal Behavior of Li-Ion Pouch Battery Cell—Part
I: Finite Element Simulations Including the Entropic Coefficient. In Proceedings of the 2022 28th International Workshop on
Thermal Investigations of ICs and Systems (THERMINIC), Dublin, Ireland, 28–30 September 2022; pp. 1–6. [CrossRef]
18. Hu, Y.; Yurkovich, S.; Guezennec, Y.; Yurkovich, B. Electro-thermal battery model identification for automotive applications. J.
Power Sources 2011, 196, 449–457. [CrossRef]
19. Morganti, M.V.; Longo, S.; Tirovic, M.; Blaise, C.Y.; Forostovsky, G. Multi-Scale, Electro-Thermal Model of NMC Battery Cell.
IEEE Trans. Veh. Technol. 2019, 68, 10594–10606. [CrossRef]
20. Aghabali, I.; Bauman, J.; Kollmeyer, P.J.; Wang, Y.; Bilgin, B.; Emadi, A. 800-V Electric Vehicle Powertrains: Review and Analysis
of Benefits, Challenges, and Future Trends. IEEE Trans. Transp. Electrif. 2021, 7, 927–948. [CrossRef]
21. Wang, C.; Liu, L.; Hou, X.; Wang, J.; Qiu, B. Failure Analysis and Improvement of High Voltage Cable of Electric Vehicle Based
on Pseudo Damage Equivalence. In Proceedings of the China SAE Congress 2020: Selected Papers; Springer: Singapore, 2022;
pp. 467–480.
22. Grandvuillemin, J.; Chamagne, D.; Tiraby, C.; Glises, R. Ampacity of power bus bars for Hybrid-Electric or Electric Vehicles. In
Proceedings of the 2008 IEEE Vehicle Power and Propulsion Conference, Harbin, China, 3–5 September 2008; pp. 1–6. [CrossRef]
23. Bukya, M.; Kumar, R.; Gupta, R.K. A study on safety issues and analytical evaluation of stresses for HVDC cable in electrical
vehicle. AIP Conf. Proc. 2020, 2294, 040001. [CrossRef]
24. Dias, R.A.; Lira, G.R.S.; Costa, E.G.; Ferreira, R.S.; Andrade, A.F. Skin effect comparative analysis in electric cables using
computational simulations. In Proceedings of the 2018 Simposio Brasileiro de Sistemas Eletricos (SBSE), Niteroi, Brazil, 12–16
May 2018; pp. 1–6. [CrossRef]
25. Rozegnał, B.; Albrechtowicz, P.; Mamcarz, D.; Rerak, M.; Skaza, M. The Power Losses in Cable Lines Supplying Nonlinear Loads.
26. Karahan, M.; Kalenderli, O. Coupled Electrical and Thermal Analysis of Power Cables Using Finite Element Method. In Heat
Transfer; Vikhrenko, V.S., Ed.; IntechOpen: Rijeka, Croatia, 2011; Chapter 9. [CrossRef]
27. Di Noia, L.P.; Rizzo, R. Thermal Analysis of Battery Cables for Electric Vehicles. In Proceedings of the 2020 2nd IEEE International
Conference on Industrial Electronics for Sustainable Energy Systems (IESES), Cagliari, Italy, 1–3 September 2020; Volume 1,
pp. 298–302. [CrossRef]
12
Energies 2023, 16, 2963
28. Enescu, D.; Colella, P.; Russo, A. Thermal Assessment of Power Cables and Impacts on Cable Current Rating: An Overview.
29. Cirino, A.W.; Paula, H.D.; Mesquita, R.C.; Saraiva, E. Cable parameter variation due to skin and proximity effects: Determination
by means of finite element analysis. In Proceedings of the 2009 35th Annual Conference of IEEE Industrial Electronics, Porto,
Portugal, 3–5 November 2009. [CrossRef]
30. Goga, V.; Paulech, J.; Váry, M. Cooling of electrical Cu conductor with PVC insulation - Analytical, numerical and fluid flow
solution. J. Electr. Eng. 2013, 64, 13. [CrossRef]
31. Ferreira, J. Improved analytical modeling of conductive losses in magnetic components. IEEE Trans. Power Electron. 1994,
9, 127–131. [CrossRef]
32. Dowell, P. Effects of eddy currents in transformer windings. Proc. Inst. Electr. Eng. 1966, 113, 1387. [CrossRef]
33. LV 112; Electrical Cables for Motor Vehicles. ISO: Geneva, Switzerland, 2005.
34. ISO 6722; Road Vehicles—60 V and 600 V Single-Core Cables. ISO: Geneva, Switzerland, 2006.
35. Lammeraner, J.; Stafl, M. Eddy Currents; Iliffe Books: London, UK, 1966.
36. Tourkhani, F.; Viarouge, P. Accurate analytical model of winding losses in round Litz wire windings. IEEE Trans. Magn. 2001, 37,
914375. [CrossRef]
37. Cywiński, A.; Chwastek, K. A Multiphysics Analysis of Coupled Electromagnetic-Thermal Phenomena in Cable Lines. Energies
2021, 14, 8. [CrossRef]
38. Hameyer, K.; Driesen, J.; De Gersem, H.; Belmans, R. The classification of coupled field problems. IEEE Trans. Magn. 1999,
35, 1618–1621. [CrossRef]
39. Kondrath, N.; Kazimierczuk, M. Inductor winding loss owing to skin and proximity effects including harmonics in non-isolated
pulse-width modulated dc–dc converters operating in continuous conduction mode. IET Power Electron. 2010, 3, 989. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
13
energies
Article
Cloud-Based Deep Learning for Co-Estimation of Battery State
of Charge and State of Health
Dapai Shi 1,† , Jingyuan Zhao 2, *,† , Zhenghong Wang 1 , Heng Zhao 3 , Chika Eze 4 , Junbin Wang 5 , Yubo Lian 5 and
Andrew F. Burke 2, *
1 Hubei Longzhong Laboratory, Hubei University of Arts and Science, Xiangyang 441053, China
2 Institute of Transportation Studies, University of California-Davis, Davis, CA 95616, USA
3 College of Big Data and Internet, Shenzhen Technology University, Shenzhen 518118, China
4 Department of Mechanical Engineering, University of California, Merced, CA 94720, USA
5 BYD Automotive Engineering Research Institute, Shenzhen 518118, China
* Correspondence: jyzhao@ucdavis.edu (J.Z.); afburke@ucdavis.edu (A.F.B.)
† These authors contributed equally to this work.
Abstract: Rechargeable lithium-ion batteries are currently the most viable option for energy storage
systems in electric vehicle (EV) applications due to their high specific energy, falling costs, and
acceptable cycle life. However, accurately predicting the parameters of complex, nonlinear battery
systems remains challenging, given diverse aging mechanisms, cell-to-cell variations, and dynamic
operating conditions. The states and parameters of batteries are becoming increasingly important in
ubiquitous application scenarios, yet our ability to predict cell performance under realistic conditions
remains limited. To address the challenge of modelling and predicting the evolution of multiphysics
and multiscale battery systems, this study proposes a cloud-based AI-enhanced framework. The
framework aims to achieve practical success in the co-estimation of the state of charge (SOC) and
state of health (SOH) during the system’s operational lifetime. Self-supervised transformer neural
networks offer new opportunities to learn representations of observational data with multiple levels
of abstraction and attention mechanisms. Coupling the cloud-edge computing framework with the
versatility of deep learning can leverage the predictive ability of exploiting long-range spatio-temporal
dependencies across multiple scales.
Citation: Shi, D.; Zhao, J.; Wang, Z.;
Zhao, H.; Eze, C.; Wang, J.; Lian, Y.;
Keywords: lithium-ion battery; state of charge; state of health; deep learning; cloud; field application
Burke, A.F. Cloud-Based Deep
Learning for Co-Estimation of Battery
State of Charge and State of Health.
Energies 2023, 16, 3855. https://
doi.org/10.3390/en16093855 1. Introduction
With increased concerns about global warming, transportation electrification has
Academic Editor: Simone
Barcellona
recently emerged as an important step across the world. In electrified vehicles, rechargeable
lithium-ion batteries are currently the most widely used systems for electrochemical energy
Received: 11 April 2023 storage and powering electric vehicles (EVs) due to their relatively high specific energy,
Revised: 22 April 2023 acceptable cost and cycle life [1]. However, degradation and aging during the system’s
Accepted: 25 April 2023 operational lifetime is still one of the most urgent and inevitable problems, especially under
Published: 30 April 2023 realistic conditions [2]. In field applications, such as an EV, an online battery management
system (BMS) offers tools to monitor cell behavior under dynamic operating conditions.
However, predicting real-life battery performance in field applications only using the online
BMS is either difficult or impossible due to the limited data computing and storage ability
Copyright: © 2023 by the authors.
of the onboard chips.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
Over the past decade, scientists and researchers are increasingly storing and analyzing
distributed under the terms and
their big datasets by using remote ‘cloud’ computing servers [3]. On the cloud, researchers
conditions of the Creative Commons can interact with field data more flexibly and intelligently. Migrating observational data
Attribution (CC BY) license (https:// from custom servers to the cloud opened up a new world of opportunities to both assimilate
creativecommons.org/licenses/by/ the data sensibly and explore it in depth.
4.0/).

14
Energies 2023, 16, 3855
Several international companies have recognized this and have recently launched their
cloud-based software, including Bosch [4], Panasonic [5] and Huawei [6]. Such public-cloud
services are also termed software as a service (SaaS). The SaaS provided by Bosch—battery
in the cloud—claimed that it is possible to improve the cycle life of batteries by 20% through
the development of digital twins by using the big datasets from vehicle fleets. The universal
battery management cloud (UBMC) service developed by Panasonic aims to identify the
cell state and optimal battery operation. The SaaS launched by Huawei aims to provide
a public cloud computing and storage service for EV companies. By learning from the
historical battery data, the purely data-driven model embedded on its cloud monitoring
system is applied to predict cell fault by discovering intricate structure in large EV-battery
datasets. Beyond enterprise-level cloud services, a national-level big-data platform was
built in 2017 in China, named the National Monitoring and Management Platform for New
Energy Vehicles (NMMP-NEV) [7]. Up to now, the NMMP-NEV has provided remote fault
diagnosis for more than six million EVs.
1.1. Literature Review

1.1.1. Modelling and Predicting Battery States
A battery is a sophisticated material system, with its functionality reliant on the
transport of charge and ions through distinct phases and across interfaces, as well as both
reversible and irreversible chemical reactions, among other material-dependent factors.
The performance of a cell can be influenced by variations in components such as electrodes,
electrolytes, interfaces, microstructures, current collectors, separators, binders, and cell or
pack designs, as well as environmental factors and operating conditions. While significant
progress has been made in first-principles, atomistic, and physics-based electrochemical
modeling of battery systems, the absence of comprehensive predictive models remains a
limiting factor for advancement. The battery management system (BMS) plays a pivotal
role in maintaining the safe and reliable operation of battery systems for EV applications.
Battery modelling is the core function of a BMS. Over the past few years, a variety of
estimation techniques have been developed for the determination of the state of batteries in
terms of two important parameters: SOC [8] and SOH [9]. In the literature, the most-studied
methods in this regard for Li-ion batteries are equivalent circuit models (ECMs), physics-
based models (PBMs), the observational filter model, and, more recently, data-driven,
machine learning-based techniques. Each method has its own advantages and challenges.
But there is always a trade-off between model accuracy and computational cost (Figure 1).
For example, ECMs offer an effective tool to identity the cell states with low computational
cost. Such a simple method has been widely used in onboard BMS for the last decade.
However, it cannot provide accurate cell parameter values due to the simplification and
assumptions in battery behaviors. Compared to ECMs, PBMs can approximate the physico-
chemical processes that take place inside the cell during the system’s operation, which
provides accurate and physically consistent predictions. This method requires detailed
information on cell specifications, including the materials and chemistry of the electrode,
electrolyte, separator, current collectors, and so on. However, it is impossible to obtain the
evaluation of these parameters during the operational lifetime under realistic conditions.
In addition, battery problems in this case are governed by highly parameterized partial
differential equations (PDEs). Solving the governing PDEs faces severe challenges and
introduces multiple sources of uncertainty, especially in real-life physical problems with
missing and noisy data and uncertain boundary conditions. Filter-based models are the
most-commonly used methods for battery parameter estimation in the existing studies.
Two issues constrain the wide application of filtering algorithms: (i) model parameters
need to be updated and (ii) the algorithms may have poor generalization performance.
15
Energies 2023, 16, 3855
Figure 1. The trade-off between computational cost and model accuracy.
Conversely, the data-driven approach, especially machine-learning-based techniques,

displays superior advantages in applications to materials and batteries, from the characteri-
zation of the material properties to the non-destructive evaluation of cell performance [10].
Machine learning allows computational models to discover intricate structure in the dataset
and capture the statistics of the observational data [11].
The machine learning techniques used to predict the evolution of the battery can be
classified into two main categories: traditional machine learning such as kernel-based
approaches, and deep learning approaches such as deep neural networks. Conventional
machine-learning techniques can be applied to process observational data in their raw
form. The learning subsystem in widespread use in the machine learning community,
deep or not, is supervised learning—that is, classification and regression. Such practical
applications of machine learning use hand-engineered features or raw data for almost all
recognition and predictive tasks. For example, extreme gradient boosting (XGBoost) was
used to estimate the battery SOC of Li-ion batteries under dynamic loading conditions [12].
The XGboost technique offers a tool to improve the predictive performance by leveraging a
set of weak learners and aggregating the outputs of each base model. The application of the
multi-step forecasting strategy using XGBoost has been demonstrated to work in capturing,
in real-time, the cell dynamics and predicting the terminal voltage and SOC under WLTP
driving cycles. Moreover, gradient boosting also shows good performance in battery
lifespan prediction [13]. Using the openly shared dataset provided by MIT/Toyota [14],
a variety of features including voltage-related, capacity-related and temperature-related
features were extracted and constructed for the gradient-boosting regression tree (GBRT)
model. Gaussian process regression (GPR) is another common machine learning technique
for battery SOC and SOH estimation. For example, battery SOC is identified by using
electrochemical impedance spectroscopy (EIS) measurements based on GPR [15]. Through
the feature selection from EIS data over frequencies of 1 mHz to 6 kHz, the GPR model
can be trained to establish the mapping relationship between the selected features and
the SOC under various temperatures. In another study, the GPR technique has been
demonstrated as an effective tool to learn nonlinear battery systems and predict capacity
fade in a variety of loading scenarios [16]. By introducing a Bayesian non-parametric
transition, the model can incorporate estimates of uncertainty into predictions, allowing the
determination of varying probabilities of the ranges of possible future health values across a
long-term timescale.
16
Energies 2023, 16, 3855
Conventional machine learning offers a straightforward and effective tool for classi-
fication and regression tasks. However, constructing such a machine-learning system in
general requires careful feature engineering and considerable domain-specific expertise to
design a feature extractor that can transform the raw data (such as battery voltage, current,
etc.) into suitable vector representations from which the learning algorithm could classify
or predict patterns in the input.
In recent years, a new learning philosophy is the family of deep learning, which
enables a machine to be fed with raw observations in mathematically useful latent spaces
and to discover intricate structure in datasets automatically. One popular type of deep
learning model is recurrent neural networks (RNNs) and their popular variants, including
long-short term memory (LSTM) and gated recurrent units (GRU).
For example, a single hidden-layer GRU-RNN model was designed to estimate battery
SOC by using the measured voltage and current [17]. In the proposed gradient method, the
weight change direction takes a compromise of the gradient direction at current instant and
at historical time to prevent the oscillation of the weight shift and to improve the training
speed. Moreover, artificial noise was added to the observational data to improve the
generalization and robustness of the neural networks. Recently, a hybrid neural network
model was developed for SOC estimation of batteries at low temperatures by coupling
a convolutional neural network (CNN) and GRU [18]. The CNN module was applied to
learn the feature parameters of the inputs, while the bidirectional weighted GRU offers
tools to improve the fitting performance of the network at low operating temperatures by
tuning the weights.
In application to battery SOH estimation, a dynamic RNN model with good mapping
ability was established for co-estimation of SOC and SOH for a lithium-ion battery [19]. The
dynamic RNN model was suitable for estimating the nonlinear and dynamic cell behaviors.
Meanwhile, self-adaptive weight particle swarm optimization was applied to improve the
performance of the networks. Compared with the traditional gradient descent algorithm,
particle swarm optimization offers an opportunity to improve the error convergence speed
and avoid local optima. In a recent study, an encoder–decoder model based on the GRU
was developed to be suitable for time series prediction of a Li-ion battery. The GRU-based
encoder–decoder model has demonstrated its ability to predict the dynamic cell voltage
response under complex current load profiles. In contrast to a conventional ECM model,
the data-driven deep neural network does not require domain-specific knowledge and
time-consuming tests under a well-controlled laboratorial environment.
Collectively, the results from these works demonstrate that RNNs and their variants
are effective in modelling and predicting nonlinear battery systems [20]. However, they
suffer from limitations due to the sequential processing and challenges related to back-
propagation through time, particularly in the modelling of long-range connections across
multiple timescales. These are manifested as training instabilities leading to vanishing
and exploding back-propagated gradient problems [21]. The transformer model, primarily
utilized for natural language processing, has recently achieved remarkable advancements
in time series forecasting [22]. The transformer model allows for parallel processing,
enabling efficient utilization of computing resources and faster training. Consequently, this
methodology can be a promising option for battery state estimation. For example, one study
proposed Dynaformer, a new deep learning architecture based on a transformer, which
can predict the aging state and full voltage discharge curve for real batteries accurately,
using only a limited number of voltage/current samples [23]. The study shows that
the transformer-based model is effective for different current profiles and is robust to
various degradation levels. Transformers tackle these obstacles by employing self-attention
and positional encoding methods that simultaneously focus on and encode the order
information while analyzing current data points within the sequence. These methods
preserve the sequential information essential for learning while eliminating the traditional
concept of recurrence. Transformers are capable of capturing such information through
17
Energies 2023, 16, 3855
the utilization of multiple attention heads. In addition, the model bridge the gap between
simulations and real data, enabling accurate planning and control over missions.
1.1.2. Cloud-Based Progress on Battery State Prediction

Engineers are increasingly exploring the option of outsourcing computing needs to
the cloud, which has become popular among researchers dealing with large amounts of
data, due to its rapid improvements [24]. For example, one study proposed a cloud-based
approach to estimate battery life by analyzing charging cloud data, which includes capacity
and internal resistance estimations [25]. The capacity estimation relied on the ampere-hour
integral method, which was further improved using temperature data and optimized
with the Kalman filter (KF). To increase the precision of the estimation results, the study
also implemented fuzzy logic (FL) to manage observation noise. Finally, the battery life
was predicted using the Arrhenius empirical model. Experimental results demonstrated
that the proposed method exhibited a low error rate of less than 4% in estimating battery
life. In another study, the authors proposed a cloud-assisted online battery management
method based on AI and edge computing technologies for EVs [26]. A cloud-edge battery
management system (CEBMS) was established to integrate cloud computation and big data
resources into real-time vehicle battery management. The proposed method utilized a deep-
learning-algorithm-based cloud data mining and battery modeling method to estimate the
battery’s voltage and energy state with high accuracy. The effectiveness of the proposed
method was verified by experimental tests, demonstrating its potential for more effective
battery use and management in EVs.
Solving real-life physical problems in practical applications can be a daunting task,
especially when dealing with multiple sources of uncertainty and imperfect data, including
missing or noisy data, and outliers. This study focuses on the potential of using an
AI-powered cloud-based framework to predict a nonlinear multiphysics and multiscale
electrochemical systems’ evolution in real-world applications. The cloud-based, closed-
loop framework utilizes machine learning models that can learn seamlessly from field
battery data in EV applications [27]. The concept of establishing digital twins for battery
systems is an innovative approach to generating longitudinal electronic health records in
cyberspace. By creating a digital replica of the battery system and continually training it on
a stream of field data, it allows for continual lifelong learning. This approach can lead to
significant benefits such as improved robustness, higher accuracy and faster training times.
The continual training on a stream of field data enables the digital twin to adapt to changing
conditions and learn from real-world experiences. This leads to improved accuracy in
predicting the behavior of the battery system, which can help optimize its performance and
extend its lifespan. In summary, the establishment of digital twins for battery systems offers
a powerful approach to achieving continual lifelong learning, improving the robustness
and accuracy of the system and reducing downtime.
1.2. Major Challenges Involved

Onboard BMS has long been an important component for EVs in the monitoring and
controlling of battery systems. Despite relentless progress, solving real-life battery problems
with noisy data and uncertain boundary conditions through traditional approaches remains
challenging. Modelling and predicting multiscale and multi-physics battery problems for
EV applications require further developments. Challenges specific to battery SOC and SOH
estimation will further stimulate the development of new methodologies and frameworks,
and we identify four major issues why collaboration between onboard BMS and cloud BMS
(Figure 2) is of great importance in achieving the task, as follows.
18
Energies 2023, 16, 3855
Figure 2. Synergy between onboard BMS and cloud BMS.
(1) Although data-driven machine learning techniques introduce considerable savings in

computational cost compared with the traditional numerical methods (e.g., solving
PDEs using finite elements), it still requires complex formulations and elaborate com-
puter codes. Performing such tasks requires scheduling the training of computational
algorithms in a more powerful computing environment. This is where cloud systems
have come into play.
(2) Upon identifying cell conditions in real-world applications, there will be cell-to-
cell, pack-to-pack and batch-to-batch variation, even with the most state-of-the-art
manufacturing techniques. These cells would exhibit distinct states after long-term
incubation. The specific approach to the predictive modelling of such battery systems
significantly relies on the amount of data available and on the cell itself.
(3) In field applications such as EVs, the operation of the batteries depends not only on
user driving patterns but also on environmental factors. Lab tests cannot incorporate
diverse driving cycles and resting periods. Uncertainty arising from the random-
ness of high-dimensional parameter spaces make it difficult to perfectly match lab
experiments to field applications.
(4) Last, but perhaps most important: even with open sharing of test data, reproducibility
and generalization issues make it rather challenging to transfer academic progress to
industry. However, the cloud-computing system provides a very flexible platform
for analyzing, training and developing new frameworks and standardized bench-
marks, which can be leveraged to improve our observational, empirical and physical
understanding of real-life battery systems in a more intelligent manner.
1.3. Contributions of the Work

Despite relentless progress, predicting the dynamics of nonlinear battery systems
by using traditional physical models inevitably faces severe challenges and introduces
multiple sources of uncertainty. First-principle, phase-field, atomistic simulations may lead
to insights into fundamental battery charging–discharging mechanisms, but they cannot
truly predict cell performance for real-world applications [28]. We are well aware of the
many benefits from cloud computing and storage. However, obviously, it is not enough to
migrate data to a cloud platform; researchers need to be able to interact with the data more
seamlessly and intelligently. Contributions of this study are as follows:
(a) Field data, which exhibits irregular loading conditions, dynamic operating scenarios,
and path-dependent deterioration processes, is generated and uploaded to the cloud,
reflecting real-world usage and making reliable predictions meaningful.
19
Energies 2023, 16, 3855
(b) A specialized attention-based transformer neural network model is designed to learn

parameters in the high-dimensional stochastic thermodynamic and kinetic battery
system. The proposed transformer model has the advantage of strong generalization
and robustness in the small data regime.
(c) We examine the evolution of batteries using deep learning approaches in the time-
resolved context and demonstrate how transformer neural networks, which automati-
cally extract useful features, have the potential to overcome the limitations that have
hindered the widespread adoption of data-driven machine learning-based techniques
to date.
(d) The designed cloud-based data-driven framework provides a highly flexible digital
solution for a wide range of diverse physical, chemical and electrochemical problems
in a way that produces promising results for the target outputs.
2. Key Components of Cloud BMS

Machine learning has emerged as a promising technique, but training an intelligent
machine requires plentiful, high-quality and relevant training data. Onboard BMS cannot
competently perform this task due to the high computational complexity of the data-
driven models [29]. To make the best use of such a flexible technique, a cloud-based
BMS provides complementary skills and opportunities to improve our understanding and
evaluate comprehensive battery behaviors. Sensor data can be transmitted continuously to
the cloud, where machine learning models can learn seamlessly from labeled samples while
exploiting the wealth of information in the observations (Figure 3). With the advancement
of sensor networks, it is now feasible to monitor the battery system across multiple spatial
and temporal scales.
Figure 3. Cloud-based framework for commercial EV applications.
2.1. Physical Entity

In EV applications, lithium-ion batteries encounter complex operating conditions,
including stochastic discharging processes for driving, dynamic charging processes for
“refilling”, and resting processes when parked. While most existing research in battery
modeling has focused on either one cell or a specific, well-designed test, these efforts pro-
duce insights and improve our understanding of physical systems but cannot fully reflect
the real-life situations with diverse aging mechanisms, significant cell-to-cell variability,
20
Energies 2023, 16, 3855
and complex loading scenarios. The huge gap between lab tests and practical applications
makes it challenging to transfer academic progress to engineering. However, by assimilat-
ing real sensor measurements to optimize computational models, a digital twin can be used
to replicate the behavior of a physical entity in silico. Focusing on analyzing time-resolved
battery data such as voltage, current and temperature can directly contribute to meeting
certain goals. Ultimately, what matters is the predictive ability under realistic conditions.
These three fundamental parameters are the only information that we can obtain from an
operating a battery using the onboard BMS.
2.2. IoT
The widespread use of the internet of things (IoT) in end-use devices such as EVs
enables a wealth of multi-fidelity observations to be explored across several spatial and
temporal scales [30]. There is a growing realization that terminal devices embedded with
electronics and connected to networks play a crucial role in monitoring the evolution
of complex digital and physical systems. With the prospect of trillions of sensors in the
coming decade, it will be possible to seamlessly incorporate multi-fidelity data streams
from real-world cases into physical models. In electric vehicle (EV) applications, battery
performance, states and mechanical properties can vary greatly with dynamic loading con-
ditions such as charging–discharging current rate, operating voltage window, frequency of
usage and temperature. This calls for sophisticated and continuous monitoring throughout
the operational lifetime.
Sensor measurements of battery cells can be transmitted to IoT components by the
onboard BMS using the Controller Area Network (CAN) protocol. A special IoT protocol,
message queuing telemetry transport (MQTT), allows for dual-direction messaging be-
tween the device and cloud and requires minimal resources. A large amount of sequential
data are generated and collected from both private and fleet vehicles, which can be easily
scaled to connect with millions of IoT devices. Data stored in onboard memory can be
seamlessly uploaded to the cloud using TCP/IP protocols. The IoT wireless system in mod-
ern cities provides infrastructure for real-time data transmission using IoT actuators and
onboard sensors.
2.3. Cloud
Cloud storage and computing have been demonstrated as powerful tools for remote
monitoring and diagnosis. For automotive industry uses, researchers and engineers can
configure their cloud environment and infrastructure to suit their requirements. The
cloud-based BMS can seamlessly learn the stream of time-series battery data and produce
electronic health records in the cloud. The most popular programming languages for cloud
development include Java and Go. In addition, PHP offers a simple, effective, and flexible
tool for web developers to create dynamic interfaces and interact with data deluge. The
servers in these systems should have high-performance CPUs, plenty of RAM, and fast
storage such as solid-state drives (SSDs). Additionally, the storage arrays should have high
capacity, high performance and redundancy features such as RAID or replication to ensure
data availability and durability. Backup and recovery systems are also critical for protecting
customer data in case of disasters or system failures, and they should have high capacity
and reliability. Cloud-based digital twins have demonstrated practical value in closed-loop
full-lifespan battery management, including material design, cell performance evaluation
and system optimization [31].
2.4. Modelling
Despite the progress made on the electrochemical modeling of battery systems using
first-principle, atomistic or physics-based methods, the lack of canonical predictive models
that can associate cell properties and mechanisms underlying their behavior with cell states
has been a bottleneck for widespread adoption. Mathematical and computational tools
have been developing rapidly, yet the multiscale and multiphysics battery system behavior
21
Energies 2023, 16, 3855
dominated by the spatial or temporal context underscores the need for a transformative
approach. Machine learning technology is now a successful part of data-driven approaches,
addressing a wide range of problems that have resisted the best efforts of the artificial
intelligence (AI) community for many years [32]. The availability of shared data and
open-source software, along with the ease of automation of materials tools, has brought
machine learning into computational frameworks. Several software libraries, including
TensorFlow, PyTorch and JAX, are contributing to the determination of cell performance
by using various data modalities, such as time series, spectral data, lab tests, field data
and more.
3. Methodologies
A learning algorithm that can seamlessly combine data and abstract mathematical
operators plays a crucial role in discovering the representations needed for regression or
classification. Deep learning techniques, in particular, naturally offer tools for extracting
features and patterns from data automatically. To explore the observational data (which are
uploaded to a private cloud system) that are characterized by multiple spatial and temporal
coverages, a specialized self-attention transformer-based neural network model is designed
in this study. Transformer-based deep learning (bidirectional encoder representations from
transformers, known as BERT) has received a lot of attention since it was proposed in
2017 [33], particularly in natural language processing [34] and computer vision [35]. In
comparison with recurrent neural networks (RNNs), transformer neural networks perform
parallelization and solve the long-term dependencies problem and thus can process the
observations much more quickly. Inspired by the successful operation, recently, various
transformer-based models have been designed in the aspects of time-series prediction
and analysis. The core idea of transformer networks is the self-attention mechanism,
which belongs to a variant of the attention mechanism that can discover intricate structure
in large time-series datasets and reduces dependence on the unimportant information
across multiple timescales. In this study, we investigate the use of a transformer and
design specialized network architectures that automatically satisfy the physical system for
multivariate time series predictive tasks, as shown in Figure 4.
Figure 4. Specialized transformer architecture for the prediction of the battery system.
22
Energies 2023, 16, 3855
3.1. Transformer Neural Networks for Co-Estimation

In this study, the transformer deep learning model includes embedding, dual-encoder
architecture and a gating mechanism. Time-series data are a collection of samples, observa-
tions and features recorded in a sequential manner over time. Self-supervised techniques
are utilized to enhance transformers by providing them with the capability to learn, classify
and forecast unlabeled data. The embedding vector replaces the original time-series data,
i.e., the entire feature vector at a given time step. The specialized dual-encoder architec-
ture is applied to battery prognostic and health management, which shows a higher test
accuracy than a typical encoder–decoder architecture. A gating mechanism is used for
coupling the predictive results of the two encoders. The success of a data-driven approach
for predictive modeling of such real-life battery systems depends heavily on the amount of
available data and the complexity of the model. Therefore, the model can be trained offline,
and during the system’s operational lifetime, online prediction only requires some sampled
data points after data preprocessing (Figure 5).
Figure 5. Architecture of the deep learning for co-estimation of battery states.
3.1.1. Data Processing

Data normalization has been demonstrated as a crucial step to a training process in a
data-driven machine learning manner. A classical method of data normalization, named
Z-score transformation, is used to normalize all the parameters of Li-ion batteries in the
dataset into the vectors, characterized by the standard normal distribution with a mean of
zero and variance of one. The Z-score method for each feature is calculated as:
xobs. − μ
zfeature = (1)
σ
where xobs. is the raw observational data, and μ and σ are the mean and standard deviation
of the complete population, respectively.
23
Energies 2023, 16, 3855
The Transformer is an encoder-decoder structured sequence-to-sequence model, de-

signed to accept a sequence of observational data as input. Transformers capture this
information by employing multiple attention heads. To implement the transformer model
efficiently, time-series data are separated into different segments based on the charging
and discharging processes using an adaptive-length sliding window. Each time, the deep
learning explores the long-range correlations across multiple timescales based on the sensor
data inside the sliding window (Figure 6). It follows two rules: (a) the model outputs the
SOC every 1 sample points (i.e., 10 s) when the direction of the current flow is constant.
(b) the model re-starts the prediction process when the direction of the current flow changes.
The sliding window offers a simple and effective tool to capture the structured relationships
between the input and output under the charging and discharging processes.
Figure 6. Sliding window mechanism of the transformer model.
3.1.2. Embedding
Unlike the LSTM or RNN models, The transformer model has no recurrence and no
convolution. Instead, it models the sequence information using the encoding included in
the input embeddings. The embedding of a typical BERT model includes token and position
embeddings. By embedding time (seconds, minutes, hours, weeks, years, etc.) into the
input, the model can effectively analyze time-series data while utilizing the computational
advantages of modern hardware such as GPUs, TPUs and others. In some ways, our
embedding strategy is analogous to BERT, but it has unique capabilities and merits for
leveraging physical information. The token embedding of the original BERT is a discrete
variable (word), while the observational data of our model is a time-series variable (cell
parameters) with missing data and sensor noise. Moreover, fine-tuning ensures that the
output embedding for each cell condition encodes contextual information that is more
relevant to the multiscale and multiphysics battery system. The positional encoding applied
to model the sequence information of the battery can be expressed as:
⎧ pos ⎫
⎨ Pi = sin( 2i ) ⎬
PE(t)i = 10,000 d
pos (2)
⎩ Pi+1 = cos( 2i ) ⎭
10,000 d
where pos is the position in the time-step of the input, and i is the dimension of the
embedding vector. It allows the learning algorithm to easily learn to attend by relative
positions [33].
3.1.3. Dual Encoder

In a typical transformer model, the encoder block consists of multi-head self-attention
modules and position-wise feedforward neural networks. To meet the needs of the battery
system, a dual-encoder architecture is designed to produce predictions that respect the
24
Energies 2023, 16, 3855
physical invariants and principles. The transformer model has made breakthrough progress
due to the self-attention mechanism, which offers an effective tool for the automatic extrac-
tion of abstract spatio-temporal features automatically. Such new paradigms of pretraining
and fine-tuning enable large-scale scientific computations on long-range correlations across
multiple timescales and thus enhance the generalization of neural network models. The
multi-head attention mechanism allows the transformer model to extract information from
different representation subspaces, which offers new opportunities for capturing the subtle
differences between different battery cells within the pack.
In the self-attention module, multi-head self-attention sublayers simultaneously trans-
form into query, key and value matrices. A sequence of vectors can be generated from the
linear projections of the scaled dot-product attention:
q kT
Attention(qh, kh, vh ) = softmax( √h h )vh (3)
dk
where qh ∈ Rn∗dk , kh ∈ Rm∗dk , vh ∈ Rm∗dv represents the query, key and value matrices,
respectively; n and m denote the lengths of queries and keys/values, respectively; and dk
and dv denote the dimensions of keys/queries and values, respectively. The multi-head
attention mechanism with h ∈ {0, 1, · · · , H} different sets of learned projections can be
expressed as:
Multi-head_Attention(qh, kh, vh ) = Concat(head1 , headh )ωo (4)
where
q
headh = Attention(qh ωh , kh ωkh , vh ωvh ) (5)
3.2. Data Generation

Machine learning techniques encompass a collection of algorithms, techniques, nor-
mative structures and data that enable the derivation of a plausible model directly from
observational data. The battery raw data under realistic conditions has been recorded
and uploaded to a private cloud server, including cell voltage, current and temperature,
which has been used to achieve a number of tasks for prognostics and health management,
such as battery failure diagnosis [36] and battery SOH prediction [37]. In EV applica-
tions, multivariate time series represent the evolution of a group of variables: voltage,
current and temperature over time. Table 1 lists the key cell specifications. The dataset
is divided into training and test sets. The training set is utilized to learn the model for
developing a base model and improve the accuracy and generalizability of predictions by
fine-tuning the model under unseen battery charge-discharge protocols. The test set is used
to quantitatively predict cell states: SOC and SOH.
Table 1. Cell chemistry and operating windows.
Parameter Value
Cell type Nickel Manganese Cobalt (NMC)
Nominal open circuit voltage 3.6 V
Nominal capacity 135 Ah
Operating voltage window 4.2 V to 2.5 V
Up to 1 C during charging vs. up to 6 C during
Operating current density
discharging/driving
Less than 45 ◦ C protected by thermal
Operating temperature window
management
25
Energies 2023, 16, 3855
3.3. Evaluation Criteria

SOC and SOH are the two most important parameters in the prognostics and predictive
health management, which are generally defined as:
Ahcurrent
SOC = × 100% (6)
Ah f ull
Ahfull_pre.
SOH = × 100% (7)
Ahfull_nom.
where Ahcurrent and Ah f ull are the cell capacity in the present state and its full capacity,
respectively, during the specific charging or discharging step, and Ahfull_pre. and Ahfull_nom.
are the full capacity and nominal capacity, respectively.
The main output of the transformer model in this study is the prediction of SOC and
SOH, which is compared with the observed values of the Li-ion cells. Three metrics are
used to evaluate model performance, including root mean square error (RMSE), the mean
absolute percentage error (MAPE) and the maximum absolute error (MAE). The inputs are
the variables that follow a ground truth joint distribution. Specifically, RMSPE is defined as

1 n
n i∑
RMSPE = (ŷi − yi ∗ )2 (8)
=1
where ŷi and are yi ∗ the observed and predicted value of the i-th sample in the
observational data.
MAPE can be expressed as
1 n |ŷi − yi ∗ |
n i∑
MAPE = (9)
=1
yi
MAE can be given by:

MAE = max |ŷi∗ − yi∗ | (10)
1≤i <<n
4. Performance of Cloud-Based BMS

4.1. SOC Estimation Results
The proposed transformer model was utilized to explore the intricate structures in
battery time-series data and identify the representations necessary for predicting cell states.
However, please note that solving real-life physical problems with missing, gappy or noisy
boundary conditions requires pre-training of transformer models. In this study, tens of cells
are randomly collected and used during their operational lifetime to pre-train the model
initially. The observational data were fed into the transformer model, and its output was
the SOC estimations corresponding to the sampling points (10 s sampling frequency using
onboard sensor measurements). Due to the physico-chemical (thermodynamic and kinetic)
principles, the model split the time-series data into several segments based on the charging
and discharging processes. The model can thus discover intricate structures in two distinct
operating conditions and then couple them together.
In the transformer model, multiple self-attention heads are introduced to operate on
the same input in parallel. Each head uses distinct weight matrices to extract various levels
of correlation between the input data. The transformer dual encoder offers predictive tools
to extract vector representations of multivariate time-series, which can be considered as
an autoregressive task of denoising the input [34]. The estimation results are independent
in the charging and discharging conditions since the transformer model maps the input
voltage and current sequences to SOC separately based on the direction of current flow. The
initial SOC is calibrated at the time when the cell is fully charged or fully discharged during
the system’s operational lifetime, which depends on the usage behavior at an uncertain
time. Once a precise value of the initial SOC is obtained, ampere-hour (Ah) counting
26
Energies 2023, 16, 3855
can be directly introduced to provide the ground truth for those observations. Therefore,
the model estimates the SOC of the cell from voltage, current and temperature data by
coupling the transformer model and the Ah counting method. The transformer-based
model is initially trained for Cell_1, and the SOC estimation result is shown in Figure 7. The
data-driven model achieves a MAPE of 0.76% and an RMSPE of 0.68%, with a maximum
absolute error of less than 2%.
Figure 7. (a) SOC estimation results for Cell_1. (b) the prediction errors.
Subsequently, the self-attention transformer model is calibrated using another Cell_2,

under totally different dynamic operating conditions. Regular calibration and maintenance
of machine learning models require significant resources, including specialized personnel
and technicians. Ensuring that the system remains accurate and effective necessitates
constant attention and care, making it a crucial component of successful machine learning
implementation. A good calibration process can be expressed as
P(ŷ = y| p̂ = p) = p (11)
where the probability p ∈ [0, 1] is over the joint distribution, and ŷ and p̂ are the predictions
and the associated confidence (probability of correctness). Let NN be a neural network,
and thus it can be given by NN ( x ) = (ŷ, p̂). As shown in Figure 8, the developed model
can accurately estimate the SOC for the NMC battery (Cell_2) over both the charging and
discharging processes with a MAE of less than 2.5%, a MAPE of 0.96% and an RMSPE
of 0.81%. The proposed transformer approaches, in particular, provide reliable SOC esti-
mations during the plateau in charge–discharge profiles. While accurate SOC estimation
through machine learning modeling is possible, it should also focus on accounting for SOC
errors induced by aging, temperature and hysteresis. Despite these factors, data-driven
estimation remains a reliable SOC reference for other methods.
27
Energies 2023, 16, 3855
Figure 8. (a) SOC estimation results for Cell_2. (b) the prediction errors.
4.2. SOH Estimation Results

The definition, based on the Equation (12), is used to calculate the SOH in its current
state. The estimation of the capacity can be expressed as:

Ĉmn = trans f ormer ([ xmn,t , xmn,t+1 , · · · , xmn,t+1+i ]Cmj ) (12)
where Ĉmn is the estimated capacity of the m-th cell in the n-th cycle, xmn,t is the observa-
tional data in the n-th cycle, and Cmj is the observed value used as the ground truth. In field
applications, the ground truth of the capacity cannot be obtained for every cycle. Therefore,
j << n. The methodologies of the transformer model need to be revisited. For a complete
explanation of the algorithm, refer to [22].
Herein, the loss value used to determine the hyperparameters of the self-attention
transformer model can be given by:
1 n
(ŷi − yi ∗ )2
n i∑
MSE = (13)
=1
Setting the hyperparameters for a transformer model can be a challenging task, which
depends highly on the specific case, including the size and complexity of the training data
and the available hardware. The model processes the encoder block’s outputs for input
into the linear layers. However, concatenation alone may yield poor prediction accuracy.
Thus, a dense interpolation algorithm [38] with tunable hyperparameters is adopted to
enhance performance. The validity of the trained transformer is demonstrated through
the interpolated results in the time-space domain. Despite a decrease in accuracy with
increasing feature differences between the test and training data, the proposed method
still produces reasonable interpolation results. The trained transformer is then employed
to reconstruct dense data with halved trace intervals for the field data. The reconstructed
dense data exhibit greater spatial continuity, and the spatial aliasing effects disappear in the
time domain. These reconstructed dense data hold the potential to enhance the accuracy of
subsequent seismic data processing and inversion.
Hyperparameters are inherent in every machine learning system, and the fundamental
objective of automated machine learning (AutoML) is to optimize performance by auto-
matically setting these hyperparameters. Table 2 summarizes the hyperparameters used in
this study.
28
Energies 2023, 16, 3855
Table 2. Hyperparameters of the transformer model.
Hyperparameter Value/Method
Layers 10 layers with 8 heads per layer
65 k gradient updates, 4096 sequence length, 64 batch
Training
size (262,144 tokens)
Learning Rate Maximum of 1 × 10−3 , linear warmup of 500 steps
Dropout 0.2 rate
Optimizer Adam with starting learning rate of 2
Use TensorFlow and PyTorch for efficient
Implementation
implementation
Layer Normalization Normalize input to each layer of the transformer
Weight Tying Tie decoder and output layer weights
Label Smoothing Apply label smoothing to target labels
Early Stopping Stop training when validation loss stops improving
Figures 9 and 10 illustrates the performance of the transformer model in estimating

the SOH of cell_1 and cell_2, which are used to train and test the model, respectively. In
each group, we randomly selected 50 sampling points, and the results show that the pro-
posed transformer-based model can achieve high predictive accuracy, with MAPE varying
between ±2.5% within a 98% confidence interval during the system’s operational lifetime.
However, the model still needs further development in some field applications. Firstly,
more efforts are required, such as hand-engineered feature extraction and establishing
ground truth, to label the observational data for training. Secondly, more calibration work
is needed to enhance the model’s performance in short-length cycles, such as charging
from 50% to 80% SOC. As the model lacks sufficient information to learn from, it may
fail to provide accurate and physically consistent predictions for each field charging pro-
cess, especially under random usage behaviors (e.g., charging for only a few minutes).
This can lead to extrapolation or observational biases, which can negatively impact the
model’s performance.
(a) ( b)
Figure 9. The training set. (a) SOH estimation for Cell_1. (b) the prediction errors.
29
Energies 2023, 16, 3855
(a) ( b)
Figure 10. The test set. (a) SOH estimation for Cell_2. (b) the prediction errors.
5. Outlook
Developing cloud systems for battery and EV applications can pose challenges for
many practitioners, but a user-friendly and accessible cloud development environment
could help address some key issues. Observational data can be sparse and noisy, and
may comprise vastly heterogeneous data modalities such as images, time series, lab tests,
historical data, records, and more. The data for certain quantities of interest might not be
readily available. To enhance the efficiency and accuracy of these systems, we propose five
major recommendations:
(i) There is a significant opportunity for synergy between onboard-BMS and cloud-BMS
technologies. Urgent and real-time tasks should be allocated to onboard BMS, while
complex tasks that involve multiple scales and temporal dependencies should be
distributed to cloud BMS.
(ii) Machine-learning models rely heavily on observational data, and new algorithms and
mathematics are needed to yield accurate and robust methods that can handle high
signal-to-noise ratios and outliers. These methods should also be able to generalize
well beyond the training data. However, the model requires craftsmanship and
elaborate implementations on different cell chemistries.
(iii) Battery behavior in EV applications is much more complex than in lab tests due to
unprecedented spatial and temporal coverage. Working with noisy data and limited
training sets and dealing with under-constrained battery problems with uncertain
boundary conditions are major challenges that need to be addressed.
(iv) Developing deep learning architectures for modeling multiscale and multiphysics
battery systems is currently done empirically, which is time-consuming. Training and
optimizing deep neural networks can also be expensive. Emerging meta-learning
techniques and transfer learning may offer promising directions to explore.
(v) Battery performance fluctuates unpredictably throughout its operational life. Precise
forecasting and modeling of long-range spatio-temporal dependencies across cell,
pack, and system levels are essential for efficient learning algorithms. A promising
approach might involve hybrid modeling, combining physical process models with
configurable, structured data-driven machine learning.
6. Conclusions
Field data have the potential to enhance the effectiveness of computational techniques
developed for cloud-based battery management systems (BMS). In this study, we propose
a cloud-based data-driven technique that utilizes state-of-the-art computational methods,
specifically transformer neural networks, to accurately model cell behaviors for real-life elec-
tric vehicle (EV) applications. Our prediction model automatically extracts spatio-temporal
30
Energies 2023, 16, 3855
features using an attention-based deep learning approach, without relying on data from
experimental test cycles or prior knowledge of cell chemistry and degradation mechanisms.
By combining IoT devices to generate field data and machine-learning modeling on the
cloud, our work underscores the potential for understanding and forecasting complex
physical systems such as lithium-ion batteries. Overall, modeling and estimation using
cloud-based BMS can complement other approaches based on simplified battery models
(such as equivalent circuit models), physical and semi-empirical models, and specialized
diagnostics embedded in the onboard BMS.
Author Contributions: Methodology, Supervision, A.F.B.; Software, J.Z. and H.Z.; Resources, Project
administration and Funding acquisition, Y.L.; data curation, J.W.; writing—original draft, D.S. and
J.Z.; writing—review & editing, C.E.; visualization, J.Z. and Z.W. All authors have read and agreed to
the published version of the manuscript.
Funding: This research was funded by [Independent Innovation Projects of the Hubei Longzhong
Laboratory] grant number [2022ZZ-24], [Central Government to Guide Local Science and Technology
Development fund Projects of Hubei Province] grant number [2022BGE267], [Basic Research Type of
Science and Technology Planning Projects of Xiangyang City] grant number [2022ABH006759] and
[Hubei Superior and Distinctive Discipline Group of “New Energy Vehicle and Smart Transportation”]
grant number [XKTD072023].
Data Availability Statement: The data could not be shared due to confidentiality.
Acknowledgments: We thank [Independent Innovation Projects of the Hubei Longzhong Labora-
tory], [Central Government to Guide Local Science and Technology Development fund Projects of
Hubei Province], [Basic Research Type of Science and Technology Planning Projects of Xiangyang
City] and [Hubei Superior and Distinctive Discipline Group of “New Energy Vehicle and Smart
Transportation”] for their financial support in this research.
References
1. Li, M.; Lu, J.; Chen, Z.; Amine, K. 30 years of lithium-ion batteries. Adv. Mater. 2018, 30, 1800561. [CrossRef] [PubMed]
2. Sulzer, V.; Mohtat, P.; Aitio, A.; Lee, S.; Yeh, Y.T.; Steinbacher, F.; Khan, M.U.; Lee, J.W.; Siegel, J.B.; Howey, D.A. The challenge and
opportunity of battery lifetime prediction from field data. Joule 2021, 5, 1934–1955. [CrossRef]
3. Gibney, E. Europe sets its sights on the cloud: Three large labs hope to create a giant public–private computing network. Nature
2015, 523, 136–138. [CrossRef] [PubMed]
4. Bosch Mobility Solutions: Battery in the Cloud. Available online: https://www.bosch-mobility-solutions.com/en/solutions/
software-and-services/battery-in-the-cloud/battery-in-the-cloud/ (accessed on 19 August 2022).
5. Panasonic Announces UBMC Service: A Cloud-Based Battery Management Service to Ascertain Battery State in Electric Mobility
Vehicles. Available online: https://news.panasonic.com/global/press/data/2020/12/en201210-1/en201210-1.pdf (accessed on
19 August 2022).
6. HUAWEI: CloudLi. Available online: https://carrier.huawei.com/en/products/digital-power/telecom-energy/Central-Office-
Power (accessed on 19 August 2022).
7. National Monitoring and Management Platform for NEVs. Available online: http://www.bitev.org.cn/a/48.html (accessed on
19 August 2022).
8. Zheng, Y.; Ouyang, M.; Han, X.; Lu, L.; Li, J. Investigating the error sources of the online state of charge estimation methods for
lithium-ion batteries in electric vehicles. J. Power Sources 2018, 377, 161–188. [CrossRef]
9. Hu, X.; Xu, L.; Lin, X.; Pecht, M. Battery lifetime prognostics. Joule 2020, 4, 310–346. [CrossRef]
10. Berecibar, M. Machine-learning techniques used to accurately predict battery life. Nature 2019, 568, 325–326. [CrossRef] [PubMed]
11. Ng, M.F.; Zhao, J.; Yan, Q.; Conduit, G.J.; Seh, Z.W. Predicting the state of charge and health of batteries using data-driven
machine learning. Nat. Mach. Intell. 2020, 2, 161–170. [CrossRef]
12. Dineva, A.; Csomós, B.; Sz, S.K.; Vajda, I. Investigation of the performance of direct forecasting strategy using machine learning
in State-of-Charge prediction of Li-ion batteries exposed to dynamic loads. J. Energy Storage 2021, 36, 102351. [CrossRef]
13. Yang, F.; Wang, D.; Xu, F.; Huang, Z.; Tsui, K.L. Lifespan prediction of lithium-ion batteries based on various extracted features
and gradient boosting regression tree model. J. Power Sources 2020, 476, 228654. [CrossRef]
14. Severson, K.A.; Attia, P.M.; Jin, N.; Perkins, N.; Jiang, B.; Yang, Z.; Chen, M.H.; Aykol, M.; Herring, P.K.; Braatz, R.D. Data-driven
prediction of battery cycle life before capacity degradation. Nat. Energy 2019, 4, 383–391. [CrossRef]
15. Babaeiyazdi, I.; Rezaei-Zare, A.; Shokrzadeh, S. State of charge prediction of EV Li-ion batteries using EIS: A machine learning
approach. Energy 2021, 223, 120116. [CrossRef]
31
Energies 2023, 16, 3855
16. Richardson, R.R.; Osborne, M.A.; Howey, D.A. Battery health prediction under generalized conditions using a Gaussian process
transition model. J. Energy Storage 2019, 23, 320–328. [CrossRef]
17. Jiao, M.; Wang, D.; Qiu, J. A GRU-RNN based momentum optimized algorithm for SOC estimation. J. Power Sources. 2020,
459, 228051. [CrossRef]
18. Cui, Z.; Kang, L.; Li, L.; Wang, L.; Wang, K. A hybrid neural network model with improved input for state of charge estimation of
lithium-ion battery at low temperatures. Renew. Energy 2022, 198, 1328–1340. [CrossRef]
19. Che, Y.; Liu, Y.; Cheng, Z. SOC and SOH identification method of li-ion battery based on SWPSO-DRNN. IEEE J. Emerg. Sel. Top.
Power Electron. 2020, 9, 4050–4061. [CrossRef]
20. Schmitt, J.; Horstkötter, I.; Bäker, B. Electrical lithium-ion battery models based on recurrent neural networks: A holistic approach.
J. Energy Storage 2023, 58, 106461. [CrossRef]
21. Sherstinsky, A. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Phys. D Nonlinear
Phenom. 2020, 404, 132306. [CrossRef]
22. Ahmed, S.; Nielsen, I.E.; Tripathi, A.; Siddiqui, S.; Rasool, G.; Ramachandran, R.P. Transformers in time-series analysis: A tutorial.
arXiv 2022, arXiv:2205.01138.
23. Biggio, L.; Bendinelli, T.; Kulkarni, C.; Fink, O. Dynaformer: A Deep Learning Model for Ageing-aware Battery Discharge
Prediction. arXiv 2022, arXiv:2206.02555.
24. Drake, N. Cloud computing beckons scientists. Nature 2014, 509, 543–544. [CrossRef]
25. Li, K.; Zhou, P.; Lu, Y.; Han, X.; Li, X.; Zheng, Y. Battery life estimation based on cloud data for electric vehicles. J. Power Sources
2020, 468, 228192. [CrossRef]
26. Li, S.; He, H.; Wei, Z.; Zhao, P. Edge computing for vehicle battery management: Cloud-based online state estimation. J. Energy
Storage 2022, 55, 105502. [CrossRef]
27. Zhao, J.; Nan, J.; Wang, J.; Ling, H.; Lian, Y.; Burke, A. Battery Diagnosis: A Lifelong Learning Framework for Electric Vehicles. In
Proceedings of the 2022 IEEE Vehicle Power and Propulsion Conference (VPPC), Merced, CA, USA, 1–4 November 2022; IEEE:
New York, NY, USA, 2022; pp. 1–6.
28. Aykol, M.; Herring, P.; Anapolsky, A. Machine learning for continuous innovation in battery technologies. Nat. Rev. Mater. 2020,
5, 725–727. [CrossRef]
29. Zhao, J.; Burke, A.F. Electric Vehicle Batteries: Status and Perspectives of Data-Driven Diagnosis and Prognosis. Batteries 2022,
8, 142. [CrossRef]
30. Mohammadi, F.; Rashidzadeh, R. An overview of IoT-enabled monitoring and control systems for electric vehicles. IEEE Instru.
Meas Mag. 2021, 24, 91–97. [CrossRef]
31. Yang, S.; He, R.; Zhang, Z.; Cao, Y.; Gao, X.; Liu, X. CHAIN: Cyber hierarchy and interactional network enabling digital solution
for battery full-lifespan management. Matter 2020, 3, 27–41. [CrossRef]
32. Lecun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef]
33. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Llion, J.; Gomez, A.N.; Kaiser, L.; Polosukhin, I. Attention is All you Need.
Adv. Neural. Inf. Process Syst. 2017, 30, 3–5.
34. Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Rush, A.M. Transformers:
State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language
Processing: System Demonstrations, Online, 1 October 2020; EMNLP: New York, NY, USA, 2020; pp. 38–45.
35. Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; Guo, B. Swin transformer: Hierarchical vision transformer using shifted
windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada, 10–17
October 2021; IEEE: New York, NY, USA, 2021; pp. 10012–10022.
36. Zhao, J.; Ling, H.; Wang, J.; Burke, A.F.; Lian, Y. Data-driven prediction of battery failure for electric vehicles. Iscience 2022,
25, 104172. [CrossRef]
37. Zhao, J.; Ling, H.; Liu, J.; Wang, J.; Burke, A.F.; Lian, Y. Machine learning for predicting battery capacity for electric vehicles.
eTransportation 2023, 15, 100214. [CrossRef]
38. Wang, B.; Zhang, N.; Lu, W.; Wang, J. Deep-learning-based seismic data interpolation: A preliminary result. Geophysics 2019, 84,
11–20. [CrossRef]
32
energies
Review
Cloud-Based Artificial Intelligence Framework for Battery
Management System
Dapai Shi 1,2,† , Jingyuan Zhao 3, *,† , Chika Eze 4 , Zhenghong Wang 2 , Junbin Wang 5 , Yubo Lian 5
and Andrew F. Burke 3, *
1 Hubei Longzhong Laboratory, Hubei University of Arts and Science, Xiangyang 441000, China
2 Hubei Key Laboratory of Power System Design and Test for Electrical Vehicle, Hubei University of Arts and
Science, Xiangyang 441053, China
3 Institute of Transportation Studies, University of California-Davis, Davis, CA 95616, USA
4 Department of Mechanical Engineering, University of California, Merced, CA 95343, USA
5 BYD Automotive Engineering Research Institute, Shenzhen 518118, China
* Correspondence: jyzhao@ucdavis.edu (J.Z.); afburke@ucdavis.edu (A.F.B.)
† These authors contributed equally to this work.
Abstract: As the popularity of electric vehicles (EVs) and smart grids continues to rise, so does the
demand for batteries. Within the landscape of battery-powered energy storage systems, the battery
management system (BMS) is crucial. It provides key functions such as battery state estimation
(including state of charge, state of health, battery safety, and thermal management) as well as cell
balancing. Its primary role is to ensure safe battery operation. However, due to the limited memory
and computational capacity of onboard chips, achieving this goal is challenging, as both theory and
practical evidence suggest. Given the immense amount of battery data produced over its operational
life, the scientific community is increasingly turning to cloud computing for data storage and analysis.
This cloud-based digital solution presents a more flexible and efficient alternative to traditional
methods that often require significant hardware investments. The integration of machine learning is
becoming an essential tool for extracting patterns and insights from vast amounts of observational
data. As a result, the future points towards the development of a cloud-based artificial intelligence
(AI)-enhanced BMS. This will notably improve the predictive and modeling capacity for long-range
connections across various timescales, by combining the strength of physical process models with the
Citation: Shi, D.; Zhao, J.; Eze, C.;
versatility of machine learning techniques.
Wang, Z.; Wang, J.; Lian, Y.; Burke,
A.F. Cloud-Based Artificial
Keywords: lithium-ion battery; battery management system; machine learning; cloud; artificial
Intelligence Framework for Battery
intelligence; state of charge; state of health; safety; field; real-world application
Management System. Energies 2023,
16, 4403. https://doi.org/10.3390/
en16114403
Academic Editor: Simone Barcellona

1. Introduction
Received: 4 April 2023 The transportation sector’s shift towards electrification is crucial for reducing carbon
Revised: 16 May 2023 emissions and improving air quality [1]. Improving battery performance will enhance the
Accepted: 26 May 2023 benefits of electrifying transportation. Lithium-ion batteries have undergone significant
Published: 30 May 2023 advancements over the past decade [2], but proper evaluation and management practices
are still lacking [3]. The widespread adoption of battery-powered electric vehicles (EVs)
has been hindered by numerous challenges, including range anxiety [4] and battery ag-
ing [5]. The implementation of onboard battery management systems (BMS) provides
tools to address these issues by determining the state of charge (SOC) and state of health
(SOH) of the battery as well as the thermal management and cell balancing during the
system’s operational lifetime [6–8]. Throughout the last decade, significant strides have
conditions of the Creative Commons
been accomplished in reaching this objective via the evolution of sophisticated learning
Attribution (CC BY) license (https://
algorithms [9]. However, despite advances in multiphysics and multiscale battery mod-
creativecommons.org/licenses/by/ eling, seamless integration of academic progress into existing onboard BMS remains a
4.0/).

33
Energies 2023, 16, 4403
challenge [10]. The computing ability of the onboard BMS is constrained by factors such as
cost, power consumption, and size limitations [11].
The onboard BMS for EV applications requires compact and energy-efficient systems,
limiting the processing power that can be incorporated. Furthermore, the high cost of
advanced processors and components may be a significant hurdle, particularly in cost-
sensitive automotive applications. Consequently, BMS is designed to execute essential tasks
like battery cell monitoring and balancing, which do not demand extensive computing
power. However, the accuracy of predicting battery characteristics under real-life opera-
tional conditions such as aging and dynamic environments is often limited. This is largely
attributed to the calibration of the model under laboratory-controlled conditions, which
may not accurately reflect the complex and varied conditions experienced in the field.
Recent developments in statistical modeling and machine learning present exciting
opportunities for predicting cell behaviors by distilling key characteristics from an im-
mense volume of multi-fidelity observational data [12,13]. Nonetheless, these advanced
learning techniques often necessitate meticulous design and complex execution. Before im-
plementation, it is essential to develop a comprehensive solution, and a cloud-based digital
solution may be a viable option [14,15]. In recent years, general-purpose Central Processing
Units (CPUs) that power cloud server farms have replaced specialized mainframe proces-
sors [16], providing researchers and start-up companies with access to public computing
resources from commercial providers such as Amazon, Google, and Microsoft [17]. The
EV and energy storage industries have also embraced this trend, with companies such as
Bosch [18], Panasonic [19], and Huawei [20] launching cloud-based software, referred to
as software as a service (SaaS). For instance, Bosch’s ‘battery in the cloud’ SaaS offering,
through leveraging vast data from vehicle fleets to create digital twins, promises to enhance
battery life cycles by 20%. Meanwhile, Panasonic’s Universal Battery Management Cloud
(UBMC) service aspires to discern cell state and optimize battery operations. Huawei’s
SaaS, on the other hand, offers a public cloud computing and storage service tailored for
EV companies. This service utilizes a purely data-driven model, embedded in its cloud
monitoring system, aiming to predict cell faults by uncovering complex patterns within
extensive EV battery datasets. On a broader scale, China has established the National
Monitoring and Management Platform for New Energy Vehicles (NMMP-NEV) [21]. This
expansive data platform provides remote fault diagnosis for over 6 million EVs.
In this review, we start by providing an overview of the functions and techniques
utilized for onboard BMS, as discussed in Section 2. We then delve into the key technologies
employed in cloud BMS in Section 3, followed by a comprehensive analysis of artificial
intelligence (AI) and machine learning (ML) applications for battery state prediction in
Section 4. Given the rapidly evolving nature of this field, we also offer insights into its
current limitations and future directions.
2. Onboard BMS
For large-scale EV or grid-scale energy storage applications, BMS is a technology that
monitors the performance of a battery system, which is typically composed of multiple
battery cells arranged in a matrix configuration [22,23]. BMS ensures that the battery system
can reliably work within a targeted range of voltage and current for a specific duration of
time, even under varying load conditions. By monitoring the battery’s system operations,
BMS helps to keep operating conditions under control and stabilize employment. BMS can
process and analyze data from various sensors and control algorithms in real-time and aims
to improve performance and ensure safe operation by adjusting battery parameters [24].
BMS technology is essential for many applications, including EVs, renewable energy
systems, and portable electronics, and is continually evolving to meet the demands of
increasingly sophisticated battery systems. However, BMS systems typically have limited
computing power and data storage capacity. The onboard BMS presently cannot be used as
a specialized technology designed to optimize battery performance but rather a general-
purpose computing system used to manage the battery system under a given program.
34
Energies 2023, 16, 4403
Estimation of SOC and SOH, thermal management, cell balancing, and so on are the main
functions of the onboard BMS (Figure 1). An onboard BMS is a dedicated hardware and
software system installed directly within the battery pack of an EV. It monitors and controls
various parameters such as voltage, current, temperature, and SOC for individual cells
or the entire battery pack. The primary objectives of an onboard BMS are to ensure safe
and efficient operation, optimize battery performance, extend battery life, and prevent
thermal runaway or other hazardous conditions. The onboard BMS communicates with
other vehicle systems and provides real-time information to the driver or user.
Figure 1. Onboard BMS for field applications (abbreviations: CMU, Communication Management
Unit; BMU, Battery Management Unit).
2.1. SOC
A crucial function of an onboard BMS is to precisely ascertain the SOC. Essentially, SOC
represents the comparison of the battery’s current capacity to its fully charged capacity,
serving as an equivalent to the stored charge measured in Coulombs. SOC in battery
management is generally defined as:
Ahcur.
SOC = × 100% (1)
Ahful.
where Ahcur. represents the battery’s capacity in its present state, while Ahful. denotes
the battery’s capacity when fully charged. The Ampere-hour (Ah) counting [25] and
open-circuit voltage (OCV) [26] are commonly used for onboard BMS due to their low
computational complexity. However, it is susceptible to certain limitations that impact
the accuracy of Ah counting, including erroneous SOC initialization, drifts caused by
current sensor noise, and battery capacity variations. Furthermore, the OCV can only
be accurately gauged when the battery is not in use, which hinders its ability to provide
real-time SOC estimates during operation. In BMS for EVs, equivalent circuit models
(ECMs) are chiefly used because of their lower computational demand and fewer input
requirements than electrochemical models. Utilizing networks of resistors and capacitors,
ECMs simulate cell behavior tied to diffusion and charge-transfer processes [27]. Hence,
they serve as a pragmatic approach for real-time operation and management of onboard
battery systems in EVs. Early and typical examples of ECMs are the Rint model, Randles
model, Thevenin model, etc. Despite their computational efficiency, most equivalent circuit
models (ECMs) have limited accuracy in predicting battery characteristics, particularly
during complex loading conditions and cell aging. This limitation is due to the fact
35
Energies 2023, 16, 4403
that model parameters are designed based on laboratory conditions and often lack the
incorporation of multifrequency impedance measurements [28–30].
The SOC of a battery is a crucial parameter for field applications since it signifies
the remaining energy capacity within the battery system. The necessity for precise and
real-time monitoring of SOC is underlined by several reasons:
(a) Range estimation: SOC is a primary factor in determining the remaining driving
range of an EV. By continuously monitoring the SOC, drivers can better plan their
trips and avoid anxiety.
(b) Optimal battery performance: Maintaining the battery within an optimal SOC range
helps preserve its health and prolong its life. Operating the battery at extreme SOC
levels (either too high or too low) can accelerate battery degradation and reduce its
overall lifespan.
(c) Charging management: Knowledge of the current SOC is crucial for optimizing
charging strategies. It allows for better estimation of the required charging time and
enables the use of smart charging algorithms that can balance the charging load on
the grid and minimize charging costs.
(d) Energy management: SOC information is vital for the efficient management of energy
consumption in EVs. The onboard energy management system uses SOC data to
optimize power distribution between various vehicle systems, ensuring efficient use
of energy and enhancing overall performance.
(e) Diagnostics and prognostics: Monitoring SOC over time, along with other battery
parameters, can provide valuable insights into the battery’s health and aid in the early
detection of potential issues. This can help prevent unexpected battery failures and
enable predictive maintenance, minimizing downtime and maintenance costs.
2.2. SOH
The SOH describes the capacity of a fully charged battery relative to its nominal
capacity at the point of manufacture when it was brand new. Upon manufacturing, a
battery’s State of Health (SOH) starts at 100% and diminishes to 80% at its end of life (EOL).
Within the battery manufacturing industry, EOL is typically characterized as the stage
when the actual capacity at full charge dwindles to 80% of its initial nominal value. The
count of charge/discharge cycles left until the battery attains its EOL is denoted as the
battery’s Remaining Useful Life (RUL). Consequently, SOH can be articulated as:
Ahful.
SOH = × 100% (2)
Ahnom.
where Ahnom. represents the nominal capacity of the battery when it is brand new.
Battery degradation is a complex issue that involves numerous electrochemical reac-
tions taking place in the anode, cathode, and electrolyte [31,32]. The operating conditions
have a critical impact on the degradation process and ultimately impact the battery life-
time. Predicting the remaining battery lifespan with precision under a variety of operating
conditions is of utmost importance to ensure reliable performance and timely maintenance,
as well as for battery second-life applications [33]. Onboard SOH estimation is used to
determine the health of a battery system during its operating lifetime. The battery capac-
ity frequently serves as a health indicator, given its association with the energy storage
potential of batteries and its immediate influence on the remaining operational duration
and overall lifespan of the batteries. Computational tools have provided insights into
fundamental battery physics, but despite the advances in first principles and atomistic
calculations, they are unable to accurately predict battery performance under realistic
conditions. As is the case for SOC estimation for online applications, the most commonly
used onboard SOH estimation methods are ECMs with limited accuracy. Data-driven
approaches can provide a better nonlinear fitting capability [34–37]. However, due to
the computational complexity, it is challenging to make most existing advanced methods
widespread and practical. This could potentially be attributed to the substantial computa-
36
Energies 2023, 16, 4403
tional resources required to accurately estimate the SOH of a battery, especially when it is
exposed to various operating and environmental conditions during its lifetime. In addition,
the need to continuously monitor and analyze battery performance can place a significant
burden on the vehicle’s onboard system and affect overall vehicle performance.
In field applications such as EVs, battery SOH provides an indication of the battery’s
overall condition and its remaining useful life. Precise and timely evaluation of SOH is
crucial for various reasons:
(a) Battery life prediction: Monitoring SOH allows for better estimation of the battery’s
RUL, enabling vehicle owners and fleet managers to plan for battery replacements or
upgrades, thus minimizing unexpected downtime and associated costs.
(b) Performance optimization: As a battery degrades, its capacity and power capabilities
decrease, affecting the vehicle’s range, acceleration, and overall performance. By
keeping track of the battery’s SOH, the energy management system can optimize the
power distribution among various vehicle systems, ensuring consistent performance
and preserving battery life.
(c) Safety assurance: A deteriorating battery may pose safety risks, such as an increased
probability of thermal runaway events, which can lead to fires or explosions. Moni-
toring SOH can help identify potential safety hazards early, allowing for preventive
measures to be taken in case of anomalous capacity degradation.
(d) Charging management: Knowledge of the battery’s SOH is vital for adapting charg-
ing strategies that account for its current condition. As battery health declines,
charging algorithms can be adjusted to minimize further degradation and maintain
safe operation.
(e) Warranty management: SOH information can be used by manufacturers to manage
warranty claims more effectively and ensure that battery performance remains within
the specified warranty limits.
(f) Second-life applications: Accurate SOH assessment can facilitate the identification of
batteries suitable for applications in their second life, like stationary energy storage
systems, once their performance in EVs has degraded below acceptable levels.
(g) Residual value estimation: The SOH is a pivotal factor in establishing the residual
value of an EV in the used vehicle market, as it directly impacts the battery’s remaining
useful life and the vehicle’s overall performance.
2.3. Thermal Management

The thermal management system plays a crucial role in ensuring optimal performance
and longevity of the battery system [38]. It aims to maintain an average temperature that
balances performance and life, as determined by the battery manufacturer. The thermal
management system should fulfill the requirements outlined by the vehicle manufacturer,
such as compactness, affordability, reliability, easy installation, low energy consumption,
accessibility for maintenance, compatibility with varying climate conditions, and provision
for ventilation as needed [39]. Various methods can be employed for these processes,
such as the utilization of air for temperature regulation and ventilation, liquid for thermal
control, insulation for maintaining temperature levels, and phase change materials for
thermal storage. Alternatively, a combination of these methods can be used. The approach
can be either passive, where it relies on environmental conditions, or active, where an
internal source is employed for heating or cooling purposes. The battery’s electronic control
unit manages the control strategy.
The battery’s temperature directly affects its discharge power, energy, and charge
acceptance during regenerative braking, which can impact the vehicle’s fuel economy
and driving experience. Moreover, temperature significantly determines the battery’s
lifespan [40]. Hence, batteries ought to function within a temperature spectrum that’s
optimal for their electrochemical processes, as specified by the manufacturer. However, this
range may be narrower than the vehicle’s specified operating range, as determined by the
manufacturer. For instance, the optimal operating temperature window for a lithium-ion
37
Energies 2023, 16, 4403
battery is typically between 20 ◦ C and 40 ◦ C. However, the operating temperature range for
an EV battery system might be much wider, extending from as low as −20 ◦ C to as high as
55 ◦ C. Hence, thermal management is crucial for EV applications, and can be summarized
for the following reasons:
(a) Performance optimization: Maintaining optimal temperature ranges for battery cells
is essential for achieving peak performance levels, ensuring efficient energy utilization,
and extending the driving range of electric vehicles.
(b) Safety: Effective thermal management helps prevent thermal runaway, which can lead
to battery fires or explosions. By closely monitoring and regulating the temperature,
potential hazards can be mitigated.
(c) Battery life extension: Prolonged exposure to extreme temperatures can degrade
battery materials, leading to a reduction in overall battery life. Proper thermal man-
agement helps maintain the battery within its optimal operating temperature range,
thus prolonging its lifespan.
(d) Charging efficiency: Effective thermal management enables faster and more efficient
charging of batteries by minimizing temperature-related inefficiencies and maintain-
ing safe charging conditions.
(e) Consistent performance: By maintaining consistent temperature conditions within
the battery pack, thermal management systems ensure that the battery’s performance
remains stable and predictable, regardless of external environmental factors.
At present, EVs do not have temperature information for every cell within the battery
pack due to practical constraints associated with the large number of required sensors.
Adding more sensors and wiring can increase the battery pack’s weight and complexity,
leading to reduced vehicle efficiency and performance. Additionally, the cost of adding
more sensors and wiring can be prohibitively expensive for mass-produced EVs that
need to be cost-effective for consumers. Hence, EVs typically rely on strategically placed
temperature sensors within the battery pack to provide an overall temperature reading,
rather than individual readings for each cell. To enhance the batteries’ safe operation,
one possible way is to develop advanced data-driven learning algorithms that leverage
time-resolved data (voltage and current). However, this comes at the cost of computing
efficiency losses.
2.4. Cell Balancing

A crucial element of electric vehicles, the battery balancing system (BBS), is composed
of two main components: the balancing circuitry and the control strategy governing the
balancing process [41]. To transfer charge between cells and maintain balance, the balancing
circuit can be either passive or active [42]. Passive balancing relies on resistance to convert
excess energy in a high-charge cell into heat, which is then dissipated until the charge
is equalized with a low-charge cell. Passive balancing is simple, cost-effective, and easy
to implement, but it generates substantial heat and has low balancing efficiency. Unlike
passive balancing, active balancing utilizes energy carriers to transfer energy from the cells
with high SOC to the cells with low SOC inside the battery pack. While active balancing is
the preferred choice for applications that operate at high temperatures and require rapid
balancing, it does increase the complexity of the circuit.
The architectural composition of the BBS holds paramount importance, yet the bal-
ancing control strategy it employs shares an equally critical role in dictating the circuit’s
overall conversion efficiency and speed of balance, as indicated by references [43,44]. These
balancing tactics can be classified according to the control variable they utilize, such as the
state of charge (SOC), cell voltage, or capacity. Creating an optimized balancing control
strategy tailor-made for a specific battery system and its corresponding application is vital
to guaranteeing a balanced operation that is both efficient and effective.
The merits of utilizing cell balancing span several areas, including:
38
Energies 2023, 16, 4403
(a) Capacity utilization: Cell balancing ensures that all cells within a battery pack are
utilized to their full capacity, maximizing the overall energy storage and extraction
capability. This, in turn, optimizes the vehicle’s driving range and performance.
(b) Lifespan extension: Imbalances in cell voltages can lead to some cells aging faster
than others, ultimately reducing the overall battery pack’s lifespan. Cell balancing
helps equalize the charge and discharge cycles across all cells, promoting even wear
and prolonging the battery pack’s life.
(c) Safety enhancement: Unbalanced cells can cause overcharging or over discharging,
which may lead to thermal runaway and other safety risks. Cell balancing prevents
these issues by ensuring that all cells are charged and discharged within their safe
operating limits.
(d) Performance consistency: Cell imbalances can result in inconsistent performance
and reduced efficiency. By maintaining balanced cells, the battery pack can deliver
predictable and stable performance, improving the overall driving experience.
(e) Reduced Maintenance: Employing cell balancing can minimize the frequency of
maintenance checks and services. By ensuring uniformity in cell usage, the system
reduces the possibility of individual cell failures and maintains the overall health of the
battery pack. This, in turn, lowers maintenance costs and offers greater convenience
to the user.
3. Key Components and Technologies of Cloud-BMS

Machine learning has emerged as a powerful instrument; however, it necessitates sub-
stantial quantities of high-quality and pertinent observational samples. The computational
complexity associated with this requirement surpasses the capabilities of onboard Battery
Management Systems (BMS). Cloud-based BMS (Figure 2) provides a brand-new digital
solution, as it can process and analyze data in a more efficient and flexible manner. Sensor
measurements can be uploaded to the cloud, enabling machine learning to continually
learn from these data points while harnessing the vast wealth of information present in
the samples. A cloud BMS enables remote monitoring, diagnostics, and even predictive
maintenance, improving overall battery management and reducing the need for manual
inspections or on-site intervention. The cloud BMS can also facilitate fleet management by
aggregating data from multiple vehicles or energy storage systems, allowing operators to
optimize energy consumption and plan maintenance schedules more efficiently. Addition-
ally, a Cloud BMS can enable over-the-air (OTA) updates to the onboard BMS firmware
and algorithms, further enhancing battery performance and extending its lifespan.
Figure 2. Cloud-based framework for battery management in EV applications.
39
Energies 2023, 16, 4403
3.1. IoT-Devices
Given the extensive embrace of Internet of Things (IoT) technology [45], end-use
devices have gained the ability to collect and analyze vast amounts of data across various
spatial and temporal scales. Equipped with electronics and network connectivity, these
devices hold a key position in monitoring and management. As the number of sensors is
expected to reach trillions in the near future, integrating data streams with diverse levels of
fidelity into real-world applications and battery models becomes increasingly feasible.
The physical, chemical, and electrochemical performance of batteries can exhibit
significant variations due to dynamic loading conditions such as current rate, operating
voltage, temperature, and more. Consequently, continuous monitoring throughout the
operational lifetime is of paramount importance [46]. The onboard Battery Management
System (BMS) enables the transfer of sensor measurements from the battery cells to the IoT
component, employing the Controller Area Network (CAN) protocol for communication.
To optimize resource utilization while efficiently transmitting a substantial volume of
sequential data generated by both private and fleet vehicles, the message queuing telemetry
transport (MQTT) protocol [47] enables bidirectional communication between the device
and the cloud. The infrastructure can effortlessly support millions of IoT devices, seamlessly
accommodating their operations. Moreover, the data stored in the onboard memory can be
efficiently transmitted to the cloud system using TCP/IP protocols, ensuring smooth and
reliable upload processes. Modern cities’ IoT systems provide infrastructure for remote
data transmission through the use of IoT actuators and on-board sensors. For a more
detailed explanation of the next-generation IoT, please refer to [48].
3.2. Cloud Server-Farm

A cloud server farm is a large-scale data center infrastructure that offers remote data
storage and analysis capabilities, including real-time monitoring, early warning systems,
and intelligent diagnosis over the internet. This beckons scientists as data sets continue to
expand [49]. Cloud storage and computing has been widely recognized and acknowledged
as a highly effective and flexible solution for remote monitoring, especially in the context of
large-scale EV applications [50]. In this context, developers have the flexibility to seamlessly
tailor their cloud to meet their specific needs and demands, thereby achieving maximum
efficiency and convenience. The cloud based BMS has the capability to learn and analyze
the continuous flow of the charging and discharging data of battery systems, enabling the
generation of health information.
The cloud BMS can learn and analyze the continuous stream of time-series battery
data and generate electronic health records, which provide insightful information about the
battery’s performance and health status. Java and Go are among the most commonly used
programming languages for cloud development, providing developers with robust and
efficient tools to create sophisticated and reliable cloud applications. Additionally, PHP
offers a flexible and effective solution for web developers to design interactive interfaces
and engage with the vast amount of data generated by the system [51].
In order to implement the battery-cloud system efficiently, it is essential for users to
have some basic computing skills, but more importantly, it requires a deep understand-
ing and knowledge of the learning task at hand, particularly in the context of complex,
nonlinear multiphysics battery systems that exhibit gappy and noisy boundary conditions.
Moreover, modeling of battery systems for field applications, such as prognostics and pre-
dictive health management (PHM), is often prohibitively expensive and requires complex
formulations, new algorithms, and elaborate computer codes.
3.3. Machine Learning

In spite of the progress achieved in forecasting the dynamics of battery systems utiliz-
ing fundamental principles, atomic-level analysis, or methods rooted in physics, a notable
obstacle persists due to the lack of all-encompassing prognostic models capable of establish-
ing robust connections between cell properties, underlying mechanisms, and the states of
40
Energies 2023, 16, 4403
the cell. The prognostication and modeling of battery systems’ multi-dimensional behavior,
influenced by various spatio-temporal factors, emphasize the necessity for a revolution-
ary approach. Deep learning has exhibited extraordinary advancements in addressing
enduring quandaries faced by the artificial intelligence community [52]. The widespread
availability of open-source software and the automation capabilities of material tools have
seamlessly integrated machine learning into computational frameworks. Prominent soft-
ware libraries like TensorFlow [53,54], PyTorch [55,56], and JAX [57] contribute significantly
to the analysis of cell performance by harnessing diverse data modalities encompassing
time series data, spectral data, laboratory tests, field data, and more.
In the realm of predictive modeling of battery systems, there has been a recent push
towards synergistically integrating machine learning tools with cloud computing. In this
context, researchers and engineers can access real-time data streams and perform real-
time analysis and predictions of battery performance, which is pivotal when it comes
to the design and optimization of battery systems. The integration of machine learning
algorithms, cloud computing, and big data analysis has created a powerful ecosystem for
the representation of multiscale and multiphysics battery systems. By incorporating actual
sensor data to calibrate the models, a battery-powered digital twin strives to emulate the
dynamics of the physical entity in a digital environment. Physics-informed learning is
poised to emerge as a driving force in the transformative era of digital twins, thanks to its
innate ability to seamlessly integrate physical models and data.
A recent illustration of this innovative learning approach is Physics-Informed Neural
Networks (PINNs). The integration of data from measurements and partial differential
equations (PDEs) is flawlessly accomplished by PINNs through the incorporation of these
PDEs into the neural networks. This approach exhibits exceptional adaptability, allowing it
to effectively handle a wide range of PDE types, including integer-order PDEs, fractional
PDEs, and stochastic PDEs. To illustrate its effectiveness, the PINN model can be success-
fully employed to solve forward problems utilizing the viscous Burgers’ equation, which
can be represented as:
∂u ∂u ∂2 u
+ρ =ϕ 2 (3)
∂t ∂x ∂x
The physics-uninformed networks act as a surrogate for the PDE solution u(x, t),
whereas the physics-informed networks characterize the PDE residual. The loss function
encompasses both a supervised loss, incorporating data measurements of u obtained
from initial and boundary conditions, and an unsupervised loss, which captures the
PDE discrepancy:
L = ωsample Lsample + ωPDE LPDE (4)
where
1 Nsample
Nsample ∑i=1
Lsample = ( u ( xi , ti ) − ui ) 2 (5)

1 ∂u ∂u ∂2 u
+ ρ − ϕ 2 x j , t j
N
LPDE =
NPDE ∑ j=PDE
1 ∂t ∂x ∂x
(6)
The two sets of points, {(xi , ti )} and {(x j , t j )}, correspond to samples taken from initial
and boundary locations and the complete domain, respectively. To effectively balance the
relationship between the two loss terms, weight, ωsample and ωPDE are utilized. The neural
network undergoes training using gradient-based optimizers like Adam to minimize the
loss until it is below a predefined threshold ε. For a detailed discussion and introduction of
PINN, one can refer to a comprehensive review [58].
4. AI Modelling for Battery State Prediction

Machine learning plays a vital role in AI modeling as it empowers systems to learn
and enhance their performance through data-driven experiences, without the need for
explicit programming. Machine learning algorithms analyze large datasets to identify
patterns and relationships, which are then used to make predictions and decisions. Machine
41
Energies 2023, 16, 4403
learning utilizes a customizable function with adjustable parameters to accurately predict

battery behavior, often through the use of experimental training data [59]. This allows
for generalization to other battery systems. A representative example includes transfer
learning, which leverages data-driven mechanisms to govern personalized health status
prediction tasks [60]. Machine learning has surfaced as a powerful tool for analyzing the
ever-growing amount of time-series data, but it faces limitations when dealing with complex
spatio-temporal systems. To overcome this, deep learning has attracted significant attention
over the past few years, with its ability to automatically extract spatio-temporal features.
Deep learning has made significant progress in addressing challenges that have previously
proven difficult for the artificial intelligence community. It has proven to be particularly
effective in identifying complex structures in large data sets for battery systems [61–64].
This would not only improve the accuracy of models for forecasting and long-range
spatial connections across multiple timescales but also enable a deeper understanding
of complex physical processes. In this section, we provide a concise overview of the
recent advancements in battery state estimation achieved through diverse machine learning
methods (Figure 3).
Figure 3. AI and machine learning for modelling and predicting battery states.
4.1. SOC
Machine learning (ML) methods have exhibited remarkable efficacy in accurately
interpolating between data points, even for high-dimensional tasks. With the ability to
learn complex patterns and relationships within data, ML models can accurately capture
the underlying structure of the data, allowing for effective interpolation and prediction.
For example, a gated recurrent unit (GRU)-based recurrent neural network (RNN) has
shown good performance in estimating the battery SOC using data from varied loading
patterns [65]. Despite the training process demanding several hours in a GPU environ-
ment, the testing phase demonstrated remarkably swift execution, even within a CPU
environment. This underscores the efficiency and efficacy, in precisely estimating SOC,
which serves as a crucial parameter for management and control in diverse applications.
In another study, a stacked bidirectional long-short-term memory (LSTM) neural network
was applied to estimate the cell SOC [66]. The study focuses on three main improvements:
(1) the use of bidirectional LSTM to capture temporal dependencies in both forward and
backward directions within time-series data; (2) the stacking of bidirectional LSTM layers
to create a deep model with increased capacity to process nonlinear and dynamic LiB data;
and (3) a detailed comparison and analysis of multiple parameters that affect the estimation
performance of the proposed method. The results demonstrate the effectiveness of the
approach and its potential to enhance SOC estimation. A single hidden layer GRU-RNN
algorithm with momentum optimization for SOC estimation is proposed [67]. GRU is
a streamlined variant of LSTM that integrates the forget and input gates into a singular
update gate, resulting in reduced parameters and enhanced computational efficiency com-
42
Energies 2023, 16, 4403
pared to LSTM. The algorithm employs the momentum gradient method, which balances
the current gradient direction and historical gradient direction to prevent oscillations in
the weight change and improve the speed of SOC estimation. The performance of the
algorithm is evaluated under varying parameters, including β, noise variances σ, epochs,
and the number of hidden layer neurons. The results of the study provide insights into
the accuracy and efficiency of the GRU-RNN-based momentum algorithm in estimating
the SOC of lithium batteries, demonstrating its potential as a promising approach for
battery management and control in various applications. More recently, a combined SOC
estimation method called gated recurrent unit adaptive Kalman filter (GRU-AKF) was pro-
posed, which is both robust and efficient [68]. The method eliminates the requirement for
developing a complex battery model by employing a GRU-RNN for initial SOC estimation
and establishing a nonlinear relationship between observed data and SOC across the entire
temperature range. Subsequently, the Adaptive Kalman Filter (AKF) is utilized to refine
the SOC estimated by the GRU-RNN, resulting in the final estimated SOC. The proposed
GRU-AKF exhibits enhanced adaptability to practical battery applications, facilitated by
the improved adaptive approach. The design cost of the estimation method is reduced since
the hyperparameters of the network do not need to be carefully designed as the output
SOC is further processed by the AKF. AKF offers an effective tool for estimating the state of
a dynamic system based on noisy measurements. The method is specifically designed to
address the challenge of noisy data in dynamic systems, where conventional data-driven
approaches may fall short in delivering accurate outcomes. The study’s findings showcase
the efficacy of the proposed method in accurately estimating SOC for batteries.
In addition to the RNN model, self-supervised transformer model is another deep
learning method that has attracted a lot of attention for predicting cell SOC. For example,
transformer-based SOC estimation was used to leverage self-supervised learning to achieve
higher accuracy with limited data availability within a constrained timeframe [69]. The
framework additionally integrates cutting-edge deep learning techniques, including the
Ranger optimizer, time series data augmentation, and the Log-Cosh loss function, to
enhance accuracy. The acquired parameters can be efficiently transferred to another cell
by fine-tuning, even with limited data available within a short timeframe. Another study
proposes a hybrid methodology for SOC estimation of batteries by employing a sliding
window to pre-process data, using a Transformer network to capture the relationship
between observational data and SOC, and feeding the result into an adaptive observer [70].
The effectiveness of the proposed method is validated across different temperatures using
US06 data, demonstrating accurate SOC estimation with less than 1% Root Mean Square
Error (RMSE) and maximum error in the majority of temperature scenarios. The proposed
method surpasses LSTM-based approaches and exhibits the ability to provide reliable
predictions even for temperatures not included in the training dataset.
4.2. SOH
In a recent research endeavor, a battery health and uncertainty management pipeline
(BHUMP) is introduced as a machine learning-driven solution, showcasing its adaptabil-
ity to various charging protocols and discharge current rates. Notably, BHUMP excels
at making accurate predictions without the need for specific knowledge about battery
design, chemistry, or operating temperature [71]. The study underscores the significance of
incorporating machine learning techniques in conjunction with charge curve segments to
effectively capture battery degradation within a limited timeframe. However, the authors
stress that even if the algorithm produces low errors, it is crucial to perform uncertainty
quantification tests to ensure its reliability before deploying it in real-world applications.
Differential approaches, namely incremental capacity and differential voltages are
frequently employed to identify causes of deterioration in online applications. One research
study, for example, combines the Support Vector Regression (SVR) algorithm with a multi-
timescale parameter identification approach based on Extended Kalman Filter-Recursive
Least Squares (EKF-RLS) and a known relationship model between representative RC
43
Energies 2023, 16, 4403
(Resistor-Capacitor) parameters and State of Health (SOH) [72]. The study’s results show-
case that the proposed method achieves Mean Absolute Error (MAE) and Root Mean Square
Error (RMSE) values below 3% for SOH prediction, utilizing both static and dynamic ob-
servational data. This suggested technique demonstrates excellent capability in accurately
estimating SOH in complex dynamic environments, offering high accuracy, robustness,
and practicality.
In recent years, some large datasets relating to batteries during the daily operation of
EVs have been collected and analyzed. For example, one study used 147 vehicle data points
from two sources to verify a proposed method for estimating the capacity and internal
resistance of EV batteries [73]. The results demonstrate that the estimation results converge
to the true trend, with a maximum estimation error of less than 4% for the capacity of
sampled real EVs. The proposed method can accurately estimate the battery capacity of
EVs and enable life prediction using current cloud data. Another study proposed a SOH
estimation method for EV batteries based on discrete incremental capacity analysis that
is robust, compatible, computationally efficient, and memory-efficient [74]. The SOH of
EVs does not decrease linearly with mileage but shows stagnation and fluctuations due to
seasonal temperature variations, driving habits, and charging strategies.
To emphasize the importance of cloud-based AI modeling for battery BMS, ensemble
machine learning offers opportunities to accurately predict SOH using only daily operat-
ing charging data (i.e., voltage, current, and temperature) [75]. A two-step approach is
employed to reduce noise in battery data, while domain-specific features derived from IC
(incremental capacity) and DV (differential voltage) analysis offer physically consistent
representations of intricate battery degradation patterns. To enhance prediction accuracy
and model generalization, a stacking technique is adopted, leveraging four base-level mod-
els (linear regression, random forest regression, gaussian process regression, and gradient
boosting regression) along with a meta-learner. The proposed multi-model fusion method
exhibits robustness, stability, and compatibility with diverse usage histories, making it a
valuable tool for forecasting cell capacity and constructing battery pack trajectories. Further-
more, the study indicates that with the advancement of onboard computing capabilities, the
proposed method can be seamlessly migrated from cloud-BMS to onboard-BMS by employ-
ing feature engineering techniques and constructing lookup tables. In summary, this study
demonstrates the potential of integrating onboard observational samples with data-driven
machine learning models to predict the dynamics of complex systems like lithium-ion
batteries, even in the presence of missing/noisy data and uncertain boundary conditions.
Reinforcement learning, which combines machine learning principles with neuro-
scientific approaches, offers a normative framework for agents to learn policies and op-
timize their behavior in response to rewards received from interacting with the environ-
ment [76]. In battery prognostics and health management applications, such as optimizing
fast-charging protocols, the BMS acts as the agent, making decisions (like determining
the applied current) based on rewards for each possible action while interacting with the
environment (the battery) [77]. A pseudo-two-dimensional electrochemical model, Doyle-
Fuller-Newman [78], is employed to predict the evolution of multiphysics battery systems
by capturing macro-scale physics, including lithium concentration in solids and electrolytes,
solid electric potential, electrolyte electric potential, ionic current, molar ion fluxes, and
cell temperature. The Deep Deterministic Policy Gradient (DDPG)-based reinforcement
learning demonstrates a remarkable ability to handle continuous state and action spaces by
updating the control policy in the actor-critic network architectures, thereby reducing the
likelihood of safety hazards during fast-charging protocols.
4.3. Battery Safety and Thermal Management

In addition to SOC and SOH estimation, cloud-based BMS can also be tailored to
a much more complex problem—that is, battery failure. Lithium-ion batteries are mul-
tiphysics and multiscale systems, and their safety and reliability are crucial due to their
widespread adoption in various applications. However, given the intricate nature of battery
44
Energies 2023, 16, 4403
behavior, accurately predicting failures remains a formidable challenge, given the lack of
understanding of the underlying degradation mechanisms. In light of the ever-evolving
cell and battery designs, the multitude of potential failure scenarios and associated risks
make it impractical to comprehensively understand the origins and consequences of each
through laboratory testing alone. While computational modeling can reduce the number of
required experiments, its effective implementation can be limited by rigorous validation
requirements and computational resources.
The establishment of a “safety envelope”, defining the operational range in which
individual cells can function safely, is essential for ensuring the overall safety of electric
vehicle battery packs. However, the challenge lies in acquiring a substantial dataset of
battery failure tests. In a recent study, researchers developed a highly accurate compu-
tational model for lithium-ion pouch cells, incorporating calibrated constitutive models
for each material composing the cell [79]. To construct a data-driven safety envelope,
supervised machine learning techniques were applied to a vast matrix of severe mechanical
loading scenarios. This study demonstrates the synergistic combination of numerical data
generation and machine learning modeling to forecast the safety of battery systems.
Emerging technologies are addressing previously challenging obstacles by providing
accessible and effective solutions, highlighting the significance of cloud-based AI modeling
in battery BMS. Machine learning approaches utilizing data-driven frameworks excel at
accurately forecasting complex nonlinear systems. A specific research study [80] focuses
on the development of a tightly integrated cloud-based machine learning system for
predicting real-life EV battery failure. By leveraging graphite/NMC cells, a data-driven
early-prediction model is created, enabling the generation of longitudinal electronic health
records through digital twins. The proposed hybrid semi-supervised machine learning
model combines observational, empirical, physical, and statistical insights, achieving a
7.7% test error utilizing field data. Cloud-based machine learning approaches exemplify
the significance of adopting a multifaceted strategy for continuous lifelong learning. These
approaches not only provide a novel means of forecasting battery failure but also underscore
the value of incorporating diverse methods to enhance accuracy and robustness.
Thermal management is a critical aspect in the context of battery systems, and a
specific study [81] conducted a comprehensive analysis of the performance of a liquid-
cooled Battery Thermal Management System (BTMS). The study primarily concentrated
on the analysis of experimental data pertaining to air conditioning and the exploration of
design considerations for the liquid-cooled Battery Thermal Management System (BTMS).
By integrating these thermal characteristics, a more accurate and efficient operation of
the liquid cooled BTMS can be achieved, thus contributing to the overall improvement
of the HPACS for EVs. This can be achieved by coupling the battery electrochemical
model with the machine learning model of HPACS and optimizing the liquid cooled BTMS
based on the automatic calibration model and battery electrochemical model, leading to
more efficient system optimization. In another case study, a multiphysics approach was
employed to demonstrate the temperature-position-dependent thermal conductivity of
Heat Pipes (HPs) [82]. By leveraging the multiphysics nature of HPs, which provides
variable thermal conductivity, valuable insights into heat pipe efficiency can be gained.
Increasing the condensation surface area of the heat pipes enables a reduction in the size
and number of heat pipes required for cooling applications. However, it is crucial to
utilize advanced methods to analyze the complex equations, multiphysics phenomena,
and boundary conditions associated with these systems. By employing such advanced
techniques, a deeper understanding of thermal management can be achieved, leading to
improved design and performance of battery systems.
Machine learning techniques, such as physics-informed machine learning [58] offer
a promising direction to follow. Such learning approach blends mathematical models
with noisy data, utilizing neural networks or other kernel-based regression networks. By
incorporating physical invariants into specialized network architectures, this approach
45
Energies 2023, 16, 4403
can improve accuracy, training speed, and generalization. Additionally, this technique can
automate the satisfaction of certain physical invariants for more effective implementation.
5. Current Limitations
5.1. Multiscale and Multiphysics Problems
While physics-informed learning has achieved remarkable success in various applica-
tions, ongoing efforts are being made to address challenges that involve multiple scales and
physics. It is recommended to initially study each physics in isolation before integrating
them, as learning multiple physics concurrently can pose computational challenges. Ad-
ditionally, it is important to utilize fine-scale simulation data selectively to gain a broader
understanding of the physics at a coarser scale. The existing body of research primar-
ily focuses on models that specialize in predicting the SOH and Remaining Useful Life
(RUL) over multiple cycles, as well as the SOC within a single charge/discharge cycle.
However, to achieve a more comprehensive understanding of battery performance, it is
necessary to develop a model that can forecast the long-term SOH from any arbitrary
point in the charge/discharge cycle. This can be accomplished through a hybrid approach
that combines sophisticated models capable of accurately forecasting the SOC up to a
specific point in the cycle, such as a fully charged state, with a SOH model that takes into
account multiple cycles. By integrating both short-term and long-term dynamics models, a
comprehensive model of battery development can be created, enabling more accurate and
reliable predictions of battery performance.
5.2. Gap between Lab Tests and Field Conditions

High-throughput testing offers a valuable means to obtain large and reliable datasets
for machine learning applications. Various electrochemical techniques, such as cyclic
voltammetry, galvanostatic charge/discharge, and electrochemical impedance spectroscopy,
enable precise and reliable measurements of batteries’ lifetime, rate capability, capacity, and
impedance. This comprehensive approach ensures that the acquired data reflects real-world
conditions and provides a solid foundation for machine learning algorithms to analyze and
extract valuable insights from the battery performance characteristics. Large amounts of
meaningful data can be swiftly generated. Machine learning models can then be trained
using this data, and the battery testing process can be expedited further by detecting poorly
performing batteries based on their initial cycles.
One approach to reconciling standard laboratory tests with field data involves labo-
ratory testing of batteries using representative loading patterns, based on characterized
typical user driving patterns. This methodology offers the benefit of a controlled envi-
ronment with high-precision equipment and frequent characterizations. However, it is
crucial to supplement this approach with field data for multiple reasons. Firstly, laboratory
experimentation is limited in scope and cannot encompass all necessary conditions over an
extended duration, particularly as the number of aging parameters grows exponentially.
In contrast, field data is readily accessible and covers the complete range of operating
conditions, offering a relatively cost-effective option as the cells are already in practical
use. Secondly, laboratory testing is artificial and may diverge from real-world usage. Con-
straints in time and equipment within the laboratory often result in extreme conditions
and short resting periods between cycles, potentially leading to an underestimation of
battery lifespan and an excessive design of battery packs. Thirdly, external factors encoun-
tered in real-world environments, such as seasonal temperature variations or mechanical
vibrations that contribute to failure, are not accounted for in laboratory settings. Fourthly,
the accumulation of additional data is always valuable in enhancing statistical confidence
when constructing models for battery lifespan and performance, considering both inherent
factors related to manufacturing variability and external factors associated with usage
patterns. The standardization of methodologies for interpreting not only accelerated cycle
aging data but also accelerated calendar aging and scenarios involving a combination
of cycle and calendar aging is crucial to extending the applicability of models beyond
46
Energies 2023, 16, 4403
laboratory-accelerated aging tests to real-world applications. Lastly, gaining a comprehen-

sive understanding of the influence of cycles and calendar conditions on battery lifespan is
imperative for both laboratory and field applications.
5.3. Data Generation and Model Training

Accurately predicting the state of batteries, both in real-time and offline, is critical to en-
hancing cycle life and ensuring safety by enabling informed engineering and adaptation to
unfavorable conditions. However, due to the wide range of battery options and constantly
evolving pack designs, it is challenging to predict cell behavior under various conditions.
One promising approach to address this is to use finite-element model data to train machine
learning algorithms that can predict cell performance when exposed to different loading
conditions. While this approach offers numerous benefits, it also encounters a common
obstacle faced by data-driven methods, which is the acquisition of trustworthy, abundant,
high-quality, and pertinent experimental data. This requires overcoming obstacles such as
experimental data collection and curation as well as ensuring that the data is representative
of real-world usage scenarios. Overcoming these challenges and obtaining the necessary
data will be crucial in the application of machine learning for battery prediction.
The battery dataset provided by the Prognostics Center of Excellence at NASA Ames
is extensively utilized by researchers [83]. These datasets involve subjecting batteries
to various operational profiles at different temperatures. Impedance measurements are
recorded after each cycle. The NASA battery dataset is a valuable resource for researchers
who are interested in studying battery performance, aging, and prognostics. Making open
data and software available is a promising approach to enhancing the transferability of
models and making them more useful for battery design. This involves the systematic
generation of datasets and their release for reuse by other researchers.
Harnessing data-based methods for discerning materials, estimating the lifespan
and efficiency of lithium-ion batteries, has proven promising, leading to a heightened
interest in utilizing these methodologies to amplify the prognostic abilities related to cell
behaviors. The recent developments in amassing and processing vast quantities of data
have paved the way for real-time learning and forecasting of battery operations. For
example, Severson and colleagues [84] recently furnished an openly accessible dataset
packed with a comprehensive array of battery data. This dataset comprises 124 LFP-
graphite cells, subjected to diverse quick-charging conditions, varying from 3.6 to 6 C, and
evaluated within a temperature-controlled chamber at 30 ◦ C, achieving up to 80% of their
original capacities. The cells underwent one or two charging steps, such as 6 C charging
from 0% to 40% State of Charge (SOC), followed by a 3 C charging process up to 80%
SOC. Additionally, all cells were charged from 80% to 100% SOC using a 1 C Constant
Current-Constant Voltage (CC-CV) phase to 3.6 V and depleted with a 4 C CC-CV phase
down to 2.0 V, with the end current regulated at C/50. During the cyclic evaluation, both
cell temperature and internal resistance were documented at 80% SOC. This dataset offers a
valuable resource for those looking to delve into battery performance, especially pertaining
to rapid charging conditions. The feature-based machine learning model adeptly utilized
voltage and capacity data from the initial 100 cycles (roughly 10% of the overall lifespan)
of equivalent commercial cells to build a straightforward regression model capable of
forecasting the cycle lifespan with approximately 90% precision. Nevertheless, a pressing
query persists: How might data-based methods be employed to anticipate cell behavior
within ever-changing field uses? Additionally, can these strategies provide an efficient
solution to comprehend and predict the reaction of emerging cell and pack configurations
to authentic environmental conditions, thus enhancing the efficacy of battery systems?
In certain scenarios, gathering training data can be a costly and challenging task. In
such cases, there is a growing need to develop high-performance learning models that
can be trained using data from different domains that is more easily accessible. This
technique is commonly known as transfer learning [85]. Transfer learning offers a practical
solution to the challenge of obtaining sufficient training data and has become increasingly
47
Energies 2023, 16, 4403
relevant in a wide range of fields. With the help of transfer learning, it is possible to achieve
high-performing models without incurring excessive costs or resource allocation.
While the exploration of negative transfer remains somewhat sparse, infusing negative
transfer methods into transfer learning frameworks could be a potent path for forthcoming
research. One conceivable trajectory entails the creation of solutions that cater to multi-
ple origin domains, potentially enhancing the filtering out of irrelevant data. A further
promising field is the concept of optimal transfer, a process that aims to selectively con-
vey certain information from a source domain to maximize the performance of the target
learner. Although there is some intersection between negative and optimal transfer, optimal
transfer concentrates on boosting the performance of the target, whereas negative transfer
emphasizes the detrimental influence exerted by the source domain on the target learner.
6. Outlook
6.1. Cloud-End Collaboration
In BMS, a collaboration with cloud computing capitalizes on the substantial compu-
tational power and storage space offered by cloud servers, overcoming the constraints
of traditional BMS and paving the way for the use of advanced algorithms such as deep
learning and reinforcement learning. The BMS’s 5G communication module is used to
capture real-time battery data, which can then be employed to build battery models in the
cloud. This allows for a two-way dynamic correlation between the digital twin model and
the actual battery, enabling detailed and secure battery management throughout its lifespan
through online learning and model updating. The data gathered from the batteries and
their associated digital twin models throughout their full lifespan is used to construct an
optimal performance improvement path via the application of smart OTA remote program
update technology. In order to cater to the escalating needs of battery management, the
immediate processing abilities of the embedded system are integrated with the high-level
intelligence offered by the cloud platform. To enhance the efficiency of the system further,
the notion of a collaborative management model that incorporates cloud, edge, and end
is introduced.
6.2. Digital Twins

Leveraging a digital twin, a virtual counterpart of the physical object, may serve as a
bridge linking laboratory experiments to real-world uses [86]. The digital twin, integrat-
ing sensor readings from real-world scenarios into computational models, can faithfully
simulate the conductivity of lithium-ion batteries under a variety of operational states like
random discharge, dynamic charge, and idle stages. Fundamental aspects like voltage,
current, temperature, and so on can be derived from the onboard BMS and used to optimize
the digital twin. The goal is to improve the predictive ability of the digital twin under
realistic conditions, which can lead to a better understanding and evaluation of battery
behavior. The use of IoT technology, such as the MQTT protocol, enables the collection of
large amounts of sequential data from the ever-increasing running time of EVs. This data
can be seamlessly transmitted to the cloud for analysis, improving our understanding of
complex battery behavior under different operating conditions.
Despite the potential benefits of digital twins, several challenges need to be solved
before their widespread implementation can become a reality. Firstly, acquiring accurate
and comprehensive observational data can be difficult, as it may be scarce and noisy and
can take various forms. Secondly, physics-based computational models can be complex
to set up and calibrate, requiring significant effort in pre-processing and determining
initial and boundary conditions, making their use in real-time applications impractical.
Additionally, the physical models of complex systems are often only partially understood,
with conservation laws that do not provide a complete system of equations without further
assumptions. Physics-informed learning, however, offers a solution to these issues by
seamlessly blending physical models with data and utilizing automatic differentiation to
eliminate the need for mesh generation.
48
Energies 2023, 16, 4403
6.3. Data-Model Fusion

Simulations based on fundamental principles are commonly employed to probe mate-
rial degradation, incorporating techniques like density functional theory and molecular
dynamics. Additionally, the transformation of microstructures and formations, such as
lithium dendrite creation and phase segregation of active electrode materials, is examined
using physics-rooted models like the phase field method. Despite the phase field mod-
eling not yet achieving complete cell simulation, the simulation results align well with
experimental data. In order to carry out in situ computations, a machine learning model
needs to be trained first, followed by the establishment of a database comprising previous
outcomes from multiscale first principles and phase-field simulations. This model serves
as a proxy for the simulations; if machine learning demonstrates significant uncertainty, an
additional simulation is conducted, added to the database, and the machine learning model
is retrained. This iterative process of active learning holds the potential to drastically reduce
the number of simulations required to understand a system. Similar machine learning
applications can be utilized in experimental design and to eliminate costly experiments.
Further studies focusing on the mechanical properties of solid electrolytes and voltage
demonstrate how machine learning can expedite simulations.
As the fusion of physics-based modeling and machine learning progresses, researchers
are likely to frequently encounter situations where multiple models of the same phe-
nomenon are developed using the same training data, or even data that is equally infor-
mative. Even though their predictions based on the training data are nearly identical, this
could lead to differentially trained networks. To address this issue, the construction of
machine learning-based transformations between theories, models of varying complexity,
and predictive models that can be validated is crucial. This will ensure that a phenomenon
retains a unique and clear physical interpretation, even when multiple models are used to
describe it. The merging of data and models can yield enhanced representations of physical
systems by capitalizing on the strengths of each information source.
6.4. Explainable AI
In several scientific disciplines, the prevailing trend is an overflow of observational
data that often surpasses our ability to understand and analyze effectively. Despite ma-
chine learning (ML) methodologies showing substantial promise and early successes, they
continue to face hurdles in deriving significant insights from the wealth of data at hand.
Furthermore, a sole reliance on data-driven models can lead to accurate correlations with
observed data, but such models might produce physically inconsistent or implausible
predictions due to extrapolation or biases inherent in the data, potentially diminishing their
generalizing capabilities. In many instances, AI systems fall short of offering clear explana-
tions of their autonomous actions to human users. While some argue that the emphasis on
explain ability is misguided and unnecessary for specific AI applications, it remains vital
for a number of key applications where users need to understand, trust, and effectively
manage their AI counterparts. Explainable AI (XAI) systems [87], striving to improve
their understandability for human users by delivering explanations of their actions, hold
promise for enriching materials science and battery modeling. They can contribute to a
more thorough understanding of the underlying physics, more effective hypothesis testing,
and a higher level of confidence in learning models. By granting researchers the ability to
interpret and visualize decision-making processes in complex models, XAI can assist in
identifying crucial features and parameters impacting material and cell characteristics. This
understanding can further promote the creation of new materials with superior properties
and deepen our comprehension of their behavior under varying conditions. Furthermore,
the transparency and interpretability provided by XAI methods can foster trust in learning
models, empowering researchers to make well-informed decisions and draw accurate con-
clusions. In designing more effective, user-friendly AI systems, certain basic principles and
domain-specific knowledge must be taken into consideration. Specifically, an XAI system
49
Energies 2023, 16, 4403
should be capable of articulating its capabilities and insights, justifying its past actions,
outlining its current trajectory, and revealing crucial information that steers its decisions.
7. Conclusions
Artificial intelligence and machine learning methods are increasingly being utilized
to reveal patterns and insights from the expanding volume of battery data. However,
these approaches often require craftsmanship and intricate implementations, especially
when system dynamics are predominantly influenced by spatio-temporal context. This is
where cloud-based digital solutions come in. The cloud environment can be configured
by us-ers/developers to meet their specific needs and requirements. Cloud-BMS opens
up a new world for collecting observational data and assimilating it sensibly through the
seamless integration of data and abstract mathematical operators. However, merely moving
data to the cloud isn’t enough. New physics-based learning algorithms and computational
frameworks are vital in addressing the challenges faced by complex battery systems,
especially in real-time EV scenarios. Integrating AI and machine learning into BMS could
boost battery diagnosis and prognosis accuracy. Furthermore, integrating cloud-based
frameworks into the BMS can improve battery monitoring and management efficiency
and scalability. Advanced sensing and monitoring technologies, such as wireless sensor
networks and IoT devices, could allow for real-time data collection and analysis, enhancing
battery management precision. The fusion of data-driven and physics-based modeling
through physics-informed machine learning techniques promises to further boost battery
management performance. The potential to model long-range correlations across multiple
time scales, simulate thermodynamics and kinetics, and explore the dynamics of nonlinear
battery systems holds promise for accelerating technology transfer from academic progress
to real-world applications.
Author Contributions: Conceptualization and methodology, J.Z.; software, J.Z. and J.W.;
writing—original draft preparation, D.S. and J.Z.; writing—review and editing, J.Z., A.F.B. and
C.E.; visualization, J.Z. and Z.W.; supervision, resources, project administration, A.F.B. and Y.L.;
funding acquisition, D.S. and Y.L. All authors have read and agreed to the published version of
the manuscript.
Funding: This study was financially supported by [Independent Innovation Projects of the Hubei
Longzhong Laboratory] grant number [2022ZZ-24], [Central Government to Guide Local Science
and Technology Development fund Projects of Hubei Province] grant number [2022BGE267], [Ba-
sic Research Type of Science and Technology Planning Projects of Xiangyang City] grant number
[2022ABH006759] and [Hubei Superior and Distinctive Discipline Group of “New Energy Vehicle
and Smart Transportation”] grant number [XKTD072023].
Data Availability Statement: The data could not be shared due to confidentiality.
References
1. Liang, X.; Zhang, S.; Wu, Y.; Xing, J.; He, X.; Zhang, K.M.; Wang, S.; Hao, J. Air quality and health benefits from fleet electrification
in China. Nat. Sustain. 2019, 2, 962–971. [CrossRef]
2. Li, M.; Lu, J.; Chen, Z.; Amine, K. 30 years of lithium-ion batteries. Adv. Mater. 2018, 30, 1800561. [CrossRef] [PubMed]
3. Meng, H.; Li, Y.F. A review on prognostics and health management (PHM) methods of lithium-ion batteries. Renew. Sustain.
Energy Rev. 2019, 116, 109405. [CrossRef]
4. Neubauer, J.; Wood, E. The impact of range anxiety and home, workplace, and public charging infrastructure on simulated
battery electric vehicle lifetime utility. J. Power Source 2014, 257, 12–20. [CrossRef]
5. Xiong, R.; Pan, Y.; Shen, W.; Li, H.; Sun, F. Lithium-ion battery aging mechanisms and diagnosis method for automotive
applications: Recent advances and perspectives. Renew. Sustain. Energy Rev. 2020, 131, 110048. [CrossRef]
6. Gabbar, H.A.; Othman, A.M.; Abdussami, M.R. Review of battery management systems (BMS) development and industrial
standards. Technologies 2021, 9, 28. [CrossRef]
7. Vichard, L.; Ravey, A.; Venet, P.; Harel, F.; Pelissier, S.; Hissel, D. A method to estimate battery SOH indicators based on vehicle
operating data only. Energy 2021, 225, 120235. [CrossRef]
50
Energies 2023, 16, 4403
8. Li, D.; Wang, L. Onboard health estimation approach with segment warping and trajectory self-learning for swappable lithium
battery. J. Energy Storage 2022, 55, 105749. [CrossRef]
9. Zhao, J.; Burke, A.F. Electric Vehicle Batteries: Status and Perspectives of Data-Driven Diagnosis and Prognosis. Batteries 2022,
8, 142. [CrossRef]
10. Sulzer, V.; Mohtat, P.; Aitio, A.; Lee, S.; Yeh, Y.T.; Steinbacher, F.; Khan, M.U.; Lee, J.W.; Siegel, J.B.; Howey, D.A. The challenge and
opportunity of battery lifetime prediction from field data. Joule 2021, 5, 1934–1955. [CrossRef]
11. Peng, J.; Meng, J.; Chen, D.; Liu, H.; Hao, S.; Sui, X.; Du, X. A Review of Lithium-Ion Battery Capacity Estimation Methods for
Onboard Battery Management Systems: Recent Progress and Perspectives. Batteries 2022, 8, 229. [CrossRef]
12. Berecibar, M. Machine-learning techniques used to accurately predict battery life. Nature 2019, 568, 325–326. [CrossRef]
13. Ng, M.F.; Zhao, J.; Yan, Q.; Conduit, G.J.; Seh, Z.W. Predicting the state of charge and health of batteries using data-driven
machine learning. Nat. Mach. Intell. 2020, 2, 161–170. [CrossRef]
14. Zhao, J.; Nan, J.; Wang, J.; Ling, H.; Lian, Y.; Burke, A.F. Battery Diagnosis: A Lifelong Learning Framework for Electric Vehicles.
In Proceedings of the 2022 IEEE Vehicle Power and Propulsion Conference (VPPC), Merced, CA, USA, 1–4 November 2022; IEEE:
New York, NY, USA, 2022; pp. 1–6.
15. Zhao, J.; Burke, A.F. Battery prognostics and health management for electric vehicles under industry 4.0. J. Energy Chem. 2023.
[CrossRef]
16. Gibney, E. Europe sets its sights on the cloud: Three large labs hope to create a giant public-private computing network. Nature
2015, 523, 136–138. [CrossRef]
17. Drake, N. How to catch a cloud. Nature 2015, 522, 115–116. [CrossRef]
18. Bosch Mobility Solutions: Battery in the Cloud. Available online: https://www.bosch-mobility-solutions.com/en/solutions/
software-and-services/battery-in-the-cloud/battery-in-the-cloud/ (accessed on 19 August 2022).
19. Panasonic Announces UBMC Service: A Cloud-Based Battery Management Service to Ascertain Battery State in Electric Mobility
Vehicles. Available online: https://news.panasonic.com/global/press/data/2020/12/en201210-1/en201210-1.pdf (accessed on
19 August 2022).
20. HUAWEI: CloudLi. Available online: https://carrier.huawei.com/en/products/digital-power/telecom-energy/Central-Office-
Power (accessed on 19 August 2022).
21. National Monitoring and Management Platform for NEVs. Available online: http://www.bitev.org.cn/a/48.html (accessed on 19
August 2022).
22. Xing, Y.; Ma, E.W.; Tsui, K.L.; Pecht, M. Battery management systems in electric and hybrid vehicles. Energies 2011, 4, 1840–1857.
[CrossRef]
23. Liu, K.; Li, K.; Peng, Q.; Zhang, C. A brief review on key technologies in the battery management system of electric vehicles.
Front. Mech. Eng. 2019, 14, 47–64. [CrossRef]
24. Xiong, R.; Li, L.; Tian, J. Towards a smarter battery management system: A critical review on battery state of health monitoring
methods. J. Power Source 2018, 405, 18–29. [CrossRef]
25. Yang, N.; Zhang, X.; Li, G. State of charge estimation for pulse discharge of a LiFePO4 battery by a revised Ah counting.
Electrochim. Acta 2015, 151, 63–71. [CrossRef]
26. Lee, S.; Kim, J.; Lee, J.; Cho, B.H. State-of-charge and capacity estimation of lithium-ion battery using a new open-circuit voltage
versus state-of-charge. J. Power Source 2008, 185, 1367–1373. [CrossRef]
27. Tran, M.K.; Mathew, M.; Janhunen, S.; Panchal, S.; Raahemifar, K.; Fraser, R.; Fowler, M. A comprehensive equivalent circuit
model for lithium-ion batteries, incorporating the effects of state of health, state of charge, and temperature on model parameters.
28. Tran, M.K.; DaCosta, A.; Mevawalla, A.; Panchal, S.; Fowler, M. Comparative study of equivalent circuit models performance in
four common lithium-ion batteries: LFP, NMC, LMO, NCA. Batteries 2021, 7, 51. [CrossRef]
29. Wei, Y.; Wang, S.; Han, X.; Lu, L.; Li, W.; Zhang, F.; Ouyang, M. Toward more realistic microgrid optimization: Experiment and
high-efficient model of Li-ion battery degradation under dynamic conditions. eTransportation 2022, 14, 100200. [CrossRef]
30. Wei, Y.; Han, T.; Wang, S.; Qin, Y.; Lu, L.; Han, X.; Ouyang, M. An efficient data-driven optimal sizing framework for photovoltaics-
battery-based electric vehicle charging microgrid. J. Energy Storage 2022, 55, 105670. [CrossRef]
31. Wei, Y.; Yao, Y.; Pang, K.; Xu, C.; Han, X.; Lu, L.; Li, Y.; Qin, Y.; Zheng, Y.; Wang, H.; et al. A Comprehensive Study of Degradation
Characteristics and Mechanisms of Commercial Li (NiMnCo) O2 EV Batteries under Vehicle-To-Grid (V2G) Services. Batteries
2022, 8, 188. [CrossRef]
32. Han, X.; Lu, L.; Zheng, Y.; Feng, X.; Li, Z.; Li, J.; Ouyang, M. A review on the key issues of the lithiumion battery degradation
among the whole life cycle. eTransportation 2019, 1, 100005. [CrossRef]
33. Hu, X.; Xu, L.; Lin, X.; Pecht, M. Battery lifetime prognostics. Joule 2020, 4, 310–346. [CrossRef]
34. Deng, Z.; Hu, X.; Li, P.; Lin, X.; Bian, X. Data-driven battery state of health estimation based on random partial charging data.
IEEE Trans. Power Electron. 2021, 37, 5021–5031. [CrossRef]
35. Khaleghi, S.; Hosen, M.S.; Karimi, D.; Behi, H.; Beheshti, S.H.; Van Mierlo, J.; Berecibar, M. Developing an online data-driven
approach for prognostics and health management of lithium-ion batteries. Appl. Energy 2022, 308, 118348. [CrossRef]
36. Gou, B.; Xu, Y.; Feng, X. An ensemble learning-based data-driven method for online state-of-health estimation of lithium-ion
batteries. IEEE Trans. Transp. Electrif. 2020, 7, 422–436. [CrossRef]
51
Energies 2023, 16, 4403
37. Li, R.; Hong, J.; Zhang, H.; Chen, X. Data-driven battery state of health estimation based on interval capacity for real-world
electric vehicles. Energy 2022, 257, 124771. [CrossRef]
38. Wu, W.; Wang, S.; Wu, W.; Chen, K.; Hong, S.; Lai, Y. A critical review of battery thermal performance and liquid based battery
thermal management. Energy Convers. Manag. 2019, 182, 262–281. [CrossRef]
39. Kim, J.; Oh, J.; Lee, H. Review on battery thermal management system for electric vehicles. Appl. Therm. Eng. 2019, 149, 192–212.
[CrossRef]
40. Wang, X.; Wei, X.; Dai, H. Estimation of state of health of lithium-ion batteries based on charge transfer resistance considering
different temperature and state of charge. J. Energy Storage 2019, 21, 618–631. [CrossRef]
41. Barsukov, Y. Battery Cell Balancing: What to Balance and How; Texas Instruments: Dallas, TX, USA, 2009.
42. Daowd, M.; Omar, N.; Van Den Bossche, P.; Van Mierlo, J. Passive and active battery balancing comparison based on MATLAB
simulation. In Proceedings of the 2011 IEEE Vehicle Power and Propulsion Conference, Chicago, IL, USA, 6–9 September 2011;
IEEE: New York, NY, USA, 2011; pp. 1–7.
43. Turksoy, A.; Teke, A.; Alkaya, A. A comprehensive overview of the dc-dc converter-based battery charge balancing methods in
electric vehicles. Renew. Sustain. Energy Rev. 2020, 133, 110274. [CrossRef]
44. Moral, C.G.; Laborda, D.F.; Alonso, L.S.; Guerrero, J.M.; Fernandez, D.; Pereda, C.R.; Reigosa, D.D. Battery internal resistance
estimation using a battery balancing system based on switched capacitors. IEEE Trans. Ind. Appl. 2020, 56, 5363–5374. [CrossRef]
45. Madakam, S.; Lake, V.; Lake, V.; Lake, V. Internet of Things (IoT): A literature review. J. Comput. Commun. 2015, 3, 164. [CrossRef]
46. Mahadik, Y.; Thakre, M.; Kamble, S. Electric Vehicle Monitoring System Based on Internet of Things (IoT) Technologies. In IoT
and Analytics for Sensor Networks: Proceedings of ICWSNUCA 2021; Springer: Singapore, 2022; pp. 311–322.
47. Jun, H.J.; Yang, H.S. Performance of the XMPP and the MQTT protocols on IEC 61850-based micro grid communication
architecture. Energies 2021, 14, 5024. [CrossRef]
48. Zikria, Y.B.; Ali, R.; Afzal, M.K.; Kim, S.W. Next-generation internet of things (iot): Opportunities, challenges, and solutions.
Sensors 2021, 21, 1174. [CrossRef]
49. Drake, N. Cloud computing beckons scientists. Nature 2014, 509, 543–544. [CrossRef] [PubMed]
50. Zhang, Y.; Liu, H.; Zhang, Z.; Luo, Y.; Guo, Q.; Liao, S. Cloud computing-based real-time global optimization of battery aging and
energy consumption for plug-in hybrid electric vehicles. J. Power Source 2020, 479, 229069. [CrossRef]
51. Laaziri, M.; Benmoussa, K.; Khoulji, S.; Larbi, K.M.; Yamami, E.A. A comparative study of laravel and symfony PHP frameworks.
Int. J. Electr. Comput. Eng. 2019, 9, 704. [CrossRef]
52. Lecun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef] [PubMed]
53. Han, D.; sang Uk, K.; Kim, S.; Lim, C.; Kin, J. Tensorflow-based Long-Short term Memory Network for SOH Prediction of
a Lithium-ion Cell. In Proceedings of the 2019 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC
Asia-Pacific), Seogwipo, Republic of Korea, 8–10 May 2019; IEEE: New York, NY, USA, 2019; pp. 1–5.
54. Ulleri, P.; Mukund, S.; Sudarshan, K.; Zenith, K.; NB, S.S. Development of contactless employee management system with mask
detection and body temperature measurement using TensorFlow. In Proceedings of the 2021 Sixth International Conference on
Wireless Communications, Signal Processing and Networking (WiSPNET), Chennai, India, 25–27 March 2021; IEEE: New York,
NY, USA, 2021; pp. 235–240.
55. Xu, Q.; Wu, M.; Khoo, E.; Chen, Z.; Li, X. A Hybrid Ensemble Deep Learning Approach for Early Prediction of Battery Remaining
Useful Life. IEEE CAA J. Autom. Sin. 2023, 10, 177–187. [CrossRef]
56. Xu, W.; Yan, C. Prediction of Lithium-ion Battery Remaining Useful Life Based on Long Short Term Memory. In Proceedings of
the 2022 IEEE International Conference on Advances in Electrical Engineering and Computer Applications (AEECA), Dalian,
China, 20–21 August 2022; IEEE: New York, NY, USA, 2022; pp. 942–948.
57. Sulzer, V.; Marquis, S.G.; Timms, R.; Robinson, M.; Chapman, S.J. Python battery mathematical modelling (PyBaMM). J. Open Res.
Softw. 2021, 9, 14. [CrossRef]
58. Karniadakis, G.E.; Kevrekidis, I.G.; Lu, L.; Perdikaris, P.; Wang, S.; Yang, L. Physics-informed machine learning. Nat. Rev. Phys.
2021, 3, 422–440. [CrossRef]
59. Aykol, M.; Herring, P.; Anapolsky, A. Machine learning for continuous innovation in battery technologies. Nat. Rev. Mater. 2020,
5, 725–727. [CrossRef]
60. Ma, G.; Xu, S.; Jiang, B.; Cheng, C.; Yang, X.; Shen, Y.; Yang, T.; Huang, Y.; Han, D.; Yuan, Y. Real-time personalized health status
prediction of lithium-ion batteries using deep transfer learning. Energy Environ. Sci. 2022, 15, 4083–4094. [CrossRef]
61. Samanta, A.; Chowdhuri, S.; Williamson, S.S. Machine learning-based data-driven fault detection/diagnosis of lithium-ion
battery: A critical review. Electronics 2021, 10, 1309. [CrossRef]
62. Sharma, P.; Bora, B.J. A Review of Modern Machine Learning Techniques in the Prediction of Remaining Useful Life of Lithium-Ion
Batteries. Batteries 2022, 9, 13. [CrossRef]
63. Shi, D.; Zhao, J.; Wang, Z.; Zhao, H.; Eze, C.; Wang, J.; Lian, Y.; Burke, A.F. Cloud-Based Deep Learning for Co-Estimation of
Battery State of Charge and State of Health. Energies 2023, 16, 3855. [CrossRef]
64. Chandran, V.; Patil, C.K.; Karthick, A.; Ganeshaperumal, D.; Rahim, R.; Ghosh, A. State of charge estimation of lithium-ion battery
for electric vehicles using machine learning algorithms. World Electr. Veh. J. 2021, 12, 38. [CrossRef]
65. Yang, F.; Li, W.; Li, C.; Miao, Q. State-of-charge estimation of lithium-ion batteries based on gated recurrent neural network.
Energy 2019, 175, 66–75. [CrossRef]
52
Energies 2023, 16, 4403
66. Bian, C.; He, H.; Yang, S. Stacked bidirectional long short-term memory networks for state-of-charge estimation of lithium-ion
batteries. Energy 2020, 191, 116538. [CrossRef]
67. Jiao, M.; Wang, D.; Qiu, J. A GRU-RNN based momentum optimized algorithm for SOC estimation. J. Power Source 2020,
459, 228051. [CrossRef]
68. Chen, J.; Zhang, Y.; Li, W.; Cheng, W.; Zhu, Q. State of charge estimation for lithium-ion batteries using gated recurrent unit
recurrent neural network and adaptive Kalman filter. J. Energy Storage 2022, 55, 105396. [CrossRef]
69. Hannan, M.A.; How, D.N.; Lipu, M.H.; Mansor, M.; Ker, P.J.; Dong, Z.Y.; Sahari, K.S.M.; Tiong, S.K.M.; Blaabjerg, F. Deep learning
approach towards accurate state of charge estimation for lithium-ion batteries using self-supervised transformer model. Sci. Rep.
2021, 11, 19541. [CrossRef]
70. Shen, H.; Zhou, X.; Wang, Z.; Wang, J. State of charge estimation for lithium-ion battery using Transformer with immersion and
invariance adaptive observer. J. Energy Storage 2022, 45, 103768. [CrossRef]
71. Roman, D.; Saxena, S.; Robu, V.; Pecht, M.; Flynn, D. Machine learning pipeline for battery state-of-health estimation. Nat. Mach.
Intell. 2021, 3, 447–456. [CrossRef]
72. Tan, X.; Zhan, D.; Lyu, P.; Rao, J.; Fan, Y. Online state-of-health estimation of lithium-ion battery based on dynamic parameter
identification at multi timescale and support vector regression. J. Power Source 2021, 484, 229233. [CrossRef]
73. Li, K.; Zhou, P.; Lu, Y.; Han, X.; Li, X.; Zheng, Y. Battery life estimation based on cloud data for electric vehicles. J. Power Source
2020, 468, 228192. [CrossRef]
74. Hong, J.; Wang, Z.; Chen, W.; Wang, L.Y.; Qu, C. Online joint-prediction of multi-forward-step battery SOC using LSTM neural
networks and multiple linear regression for real-world electric vehicles. J. Energy Storage 2020, 30, 101459. [CrossRef]
75. Zhao, J.; Ling, H.; Liu, J.; Wang, J.; Burke, A.F.; Lian, Y. Machine learning for predicting battery capacity for electric vehicles.
eTransportation 2023, 15, 100214. [CrossRef]
76. Sutton, R.S.; Barto, A.G. Reinforcement Learning, 2nd ed.; MIT Press Academic: Cambridge, MA, USA, 2018; pp. 443–480.
77. Park, S.; Pozzi, A.; Whitmeyer, M.; Perez, H.; Kandel, A.; Kim, G.; Choi, Y.; Joe, W.T.; Raimondo, D.M.; Moura, S. A Deep
Reinforcement Learning Framework for Fast Charging of Li-ion Batteries. IEEE Trans. Transp. Electrif. 2022, 8, 2770–2784.
[CrossRef]
78. Drummond, R.; Bizeray, A.M.; Howey, D.A.; Duncan, S.R. A feedback interpretation of the Doyle–Fuller–Newman lithium-ion
battery model. IEEE Trans. Control Syst. Technol. 2019, 24, 1284–1295.
79. Li, W.; Zhu, J.; Xia, Y.; Gorji, M.B.; Wierzbicki, T. Data-driven safety envelope of lithium-ion batteries for electric vehicles. Joule
2019, 3, 2703–2715. [CrossRef]
80. Zhao, J.; Ling, H.; Wang, J.; Burke, A.F.; Lian, Y. Data-driven prediction of battery failure for electric vehicles. Iscience 2022,
25, 104172. [CrossRef]
81. Tang, X.; Guo, Q.; Li, M.; Wei, C.; Pan, Z.; Wang, Y. Performance analysis on liquid-cooled battery thermal management for
electric vehicles based on machine learning. J. Power Source 2021, 494, 229727. [CrossRef]
82. Boonma, K.; Mesgarpour, M.; NajmAbad, J.M.; Alizadeh, R.; Mahian, O.; Dalkılıç, A.S.; Wongwises, S. Prediction of battery
thermal behaviour in the presence of a constructal theory-based heat pipe (CBHP): A multiphysics model and pattern-based
machine learning approach. J. Energy Storage 2022, 48, 103963. [CrossRef]
83. “Battery Data Set”, NASA Prognostics Data Repository, NASA Ames Research Center, Moffett Field, CA. Available online:
https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository (accessed on 19 August 2022).
84. Severson, K.A.; Attia, P.M.; Jin, N.; Perkins, N.; Jiang, B.; Yang, Z.; Braatz, R.D. Data-driven prediction of battery cycle life before
capacity degradation. Nat. Energy 2019, 4, 383–391. [CrossRef]
85. Weiss, K.; Khoshgoftaar, T.M.; Wang, D. A survey of transfer learning. J. Big Data 2016, 3, 9. [CrossRef]
86. Tao, F.; Qi, Q. Make more digital twins. Nature 2019, 573, 490–491. [CrossRef]
87. Carvalho, D.V.; Pereira, E.M.; Cardoso, J.S. Machine learning interpretability: A survey on methods and metrics. Electronics 2019,
8, 832. [CrossRef]
53
energies
Article
Comparing Machine Learning Strategies for SoH Estimation of
Lithium-Ion Batteries Using a Feature-Based Approach †
Iacopo Marri, Emil Petkovski, Loredana Cristaldi * and Marco Faifer
Department of Electronics, Information and Bioengineering, Politecnico di Milano, 20133 Milan, Italy;
iacopomarri@gmail.com (I.M.); emil.petkovski@polimi.it (E.P.); marco.faifer@polimi.it (M.F.)
* Correspondence: loredana.cristaldi@polimi.it
† This paper is an extended version of our paper published in the 18th IMEKO TC10 Conference on Measurement
for Diagnostic, Optimization and Control to Support Sustainability and Resilience 2022, pp. 109–113.
Abstract: Lithium-ion batteries play a vital role in many systems and applications, making them
the most commonly used battery energy storage systems. Optimizing their usage requires accurate
state-of-health (SoH) estimation, which provides insight into the performance level of the battery
and improves the precision of other diagnostic measures, such as state of charge. In this paper, the
classical machine learning (ML) strategies of multiple linear and polynomial regression, support
vector regression (SVR), and random forest are compared for the task of battery SoH estimation. These
ML strategies were selected because they represent a good compromise between light computational
effort, applicability, and accuracy of results. The best results were produced using SVR, followed
closely by multiple linear regression. This paper also discusses the feature selection process based on
the partial charging time between different voltage intervals and shows the linear dependence of
these features with capacity reduction. The feature selection, parameter tuning, and performance
evaluation of all models were completed using a dataset from the Prognostics Center of Excellence at
NASA, considering three batteries in the dataset.
Keywords: lithium-ion battery; machine learning; SoH; battery degradation; prognostics
Citation: Marri, I.; Petkovski, E.;

Cristaldi, L.; Faifer, M. Comparing
1. Introduction
Machine Learning Strategies for SoH
Estimation of Lithium-Ion Batteries One of the primary challenges of modern life is global warming caused by the emission
Using a Feature-Based Approach. of greenhouse gases from burning fossil fuels, as well as the urgency to diminish reliance
Energies 2023, 16, 4423. https:// on non-renewable resources. Renewable energy generation has become a top priority for
doi.org/10.3390/en16114423 governments all around the world. Focusing on photovoltaics (PVs) and wind, which are
generally considered non-dispatchable and only partially participate in maintaining grid
Academic Editor: Djaffar
Ould-Abdeslam
stability [1], makes battery and other energy storage systems essential. Out of the various
battery technologies currently in use, lithium-ion batteries have become the preferred choice,
Received: 20 April 2023 owing to their high power and energy density as well as long service life [2]. For these
Revised: 16 May 2023 reasons, exclusively Li-ion batteries are used in electric vehicles [3], where maximizing
Accepted: 19 May 2023 energy density and minimizing the weight of the battery pack is crucial. Continuous
Published: 30 May 2023 research and investments are focused on this technology to improve its performance,
robustness, and stability.
A battery management system (BMS) is commonly used to ensure safe and efficient
operation of the battery pack by controlling the charge and discharge processes of the
cells and providing cell balancing. To achieve this task, the BMS must accurately estimate
crucial battery parameters, such as state of charge (SoC), state of health (SoH), and the
remaining useful life (RUL) [4]. The SoC is related to the available capacity of the batteries.
conditions of the Creative Commons By knowing this factor, the BMS prevents overcharging or discharging of batteries. The SoH
Attribution (CC BY) license (https:// provides information about the aging status of the battery and is indicated by the rise of
creativecommons.org/licenses/by/ internal resistance or capacity decrease. On the other hand, the goal of the RUL prediction is
4.0/). to understand how long a battery will continue to operate before it fails or has unacceptable

54
Energies 2023, 16, 4423
performance. Battery degradation is a highly variable process, depending on cell chemistry,

the BMS, ambient conditions, and use patterns. For this reason, a considerable amount of
model-based [5–9] and data-driven battery aging methods used for the SoH and end-of-life
predictions of batteries can be found in the literature.
Recently, data-driven methods have gained popularity due to the availability of
vast amounts of data, gathered through sensors and other monitoring devices, and ad-
vancements in the field of machine learning (ML). They do not necessarily rely on prior
knowledge of the particular battery cell and are less expensive to develop compared to the
model-based ones. Data-driven methods use large data sets to identify patterns and rela-
tionships that may not be easily discernible using traditional analytical techniques. Many
different features, otherwise known as health indicators (His), have been used to build
various ML strategies. Apart from capacity and resistance changes, these His are based on
voltage charge–discharge limits, the amount of current, battery temperature [10–14], and
incremental capacitance analysis (ICA) [15–18], as well as features derived from statistical
analysis of the other health indicators [19].
In the literature, many machine learning techniques have been studied and used to
perform SoH estimation [20,21] apply regression to model battery aging behavior and com-
pare the RUL-prediction capabilities of two fitting functions, while in [22], a combination
of an exponential function and regression analysis is used. The authors of [23] discuss a
strategy based on support vector regression (SVR) and ICA curves obtained from partial
charging data. Similarly, [14] uses partial charging segments of voltage under constant
current charging and a support vector machine model. Another SVR strategy is presented
in [24], based on curves of battery voltage as a function of charging capacity (V-Q). Finally,
in [25], a solution is proposed based on the random forest algorithm. In [13], a gaussian
process regression (GPR) model is used with four specific inputs extracted from the charg-
ing curves, and a grey relational analysis method is applied to analyze the relationship
between features and SoH. The authors of [26] apply GPR to discover the relationship
between capacity, storage temperature, and SoC of lithium-ion batteries. By optimizing the
feature selection process with an automatic-relevance-determination (ARD) structure, they
provide predictions for the calendar aging of batteries tested under different conditions.
GPR combined with electrochemical impedance spectroscopy is used in [27], adopting
many wave shapes to obtain an estimation of the capacity of the batteries. Novel health
indicators related to the lithium diffusion coefficient are provided and validated.
In [28], a capacity-estimation method based on back-propagation neural networks
(NN) and partial charging voltage segments, corresponding to 10–50% SOC, has been
developed. Another solution based on recurrent neural networks (RNNSs) is proposed
in [29], while in [30], an echo state network (ESN) has been used together with a model-
based approach to predict the SoH evolution curve of the tested batteries, starting from
cycles 80, 100, or 120. From the generated curves, predictions are made for the RUL. Due
to the problem of vanishing or exploding gradient, traditional RNNs are not capable of
dealing with long sequences in practice. The emergence of long short-term memory (LSTM)
has provided a solution to this problem [31], and [32] utilized LSTM to build a RUL model
of the lithium-ion battery. In [33], another method is proposed based on LSTM NNs and
signal processing methods for SoH monitoring and RUL prediction of lithium-ion batteries.
In [34], the authors proposed an approach for SoH estimation based on SVR and a
feature extraction procedure. In this paper, SVR is compared to other ML approaches,
including multiple linear and polynomial regression and random forest. These classical
ML strategies have been chosen because they offer a good compromise between light com-
putational effort, applicability, and accuracy of results, while also providing higher model
interpretability than complex NNs. The performances of all strategies are compared using
a dataset from the Prognostics Center of Excellence at NASA, considering three batteries
of the dataset. This work differentiates itself from the other aforecited papers, including
the ones employing the same NASA dataset [10–13,18,28,30,33], based on the specific ML
strategies implemented, the features used, and their feature numbers. Discussion is pro-
55
Energies 2023, 16, 4423
vided on the feature selection process based on partial charging times between different
voltage limits, as well as the parameter tuning process of the different strategies. Finally,
this research had the goal of minimizing the necessary number of features, considering
models based on one-to-four features, and achieving optimal results with only two features
for all considered ML strategies.
2. NASA Dataset
The NASA Ames Prognostics Center of Excellence (PCoE) released a data repository
composed of six datasets of aged Li-ion batteries [35]. However, only the first of these
datasets is suitable for prognostic degradation prediction, according to their guidelines.
In this work, batteries 5, 6, and 7 were considered, which were tested until failure. The
charging process follows the constant-current (CC) and constant-voltage (CV) protocol.
More specifically, the cells are charged with a current of 1.5 A until the upper voltage limit
of 4.2 V is met, after which CV charging proceeds until the current drops below 20 mA.
The discharge phase is carried out at 2.7 V, 2.5 V, and 2.2 V, depending on the battery.
Cycles are grouped into charge, discharge, or impedance cycles. For every cycle of every
cell, various quantities are measured, including current, time, temperature, voltage, and
discharge capacity. To control the environmental temperature, the tests were carried out in
a climatic chamber.
3. Considered Machine Learning Strategies

3.1. Multiple Linear Regression and Stepwise Regression
Multiple linear regression (MLR) is a statistical approach for modeling the relationship
between a target variable (y) and two or more available descriptor variables (xi ), otherwise
called features, using a linear equation. Regression models are usually fitted using the
least-squares approach, which minimizes the sum of the squared differences between
the predicted and actual values of the target variable. However, fitting based on other
criteria can be performed, such as least absolute deviations or minimization of a penalized
version of the least-squares function, as in the case of ridge and lasso regression. MLR is a
powerful tool for analyzing complex relationships between variables, but it assumes that
the relationships are linear. When this is not the case, better results could be obtained using
polynomial regression, which is a statistical technique that models the relationship between
xi and y as an n-th degree polynomial, thus fitting a nonlinear relationship. Polynomial
regression utilizing multiple features can have many potential terms resulting from the
features raised to a certain power or their combination.
Stepwise regression can be used to automatically identify the most important terms. It
involves iteratively adding or removing different terms according to a stopping criterion,
which can be based on the p-value, Akaike information criterion (AIC), Bayesian infor-
mation criterion (BIC), value of the coefficient of determination (R2 ), or adjusted R2 . The
most popular stepwise methods are forward selection (FS), backward elimination (BE),
and bidirectional elimination. In FS, the model starts with no terms and iteratively adds
them until a stopping criterion is met. In BE, the model starts with all combinations and
iteratively removes terms until a stopping criterion is met. For both the BE and the FS
methods, the decision regarding a term is final and is not reconsidered. This is not the case
with bidirectional elimination, which is a combination of forward and backward stepwise
regression and starts with no terms. If the adjusted R2 is considered as a stopping criterion,
this method will first add the terms that produce the largest increase in the adjusted R2
value. Eventually, the removal of terms can also occur if this results in maximum increases
of the adjusted R2 .
In this work, models based on MLR, as well as second- and third-degree polyno-
mial terms, have been constructed using bidirectional-elimination stepwise regression.
The generated models based on stepwise regression were limited to second- and third-
degree polynomial terms, including combinational terms. The adjusted R2 was used as the
stopping criterion. In all cases, fitting was performed by using the least-squares method.
56
Energies 2023, 16, 4423
3.2. Support Vector Regression

The support vector machine (SVM), in ML, is a well-known supervised learning
model, used mainly for binary classification tasks. It has been extensively applied in
predictive and diagnostic tasks, such as in [36], where a partial-discharge-curve approach
is combined with the least-squares SVM to estimate the state of health (SoH) of Li-ion
batteries. Similarly, in [37], the SVM is utilized on an electric-vehicle (EV) battery-usage-
profile dataset generated by simulations to determine the SoH. The SVM searches for the
optimal hyperplane that maximizes the distance from each training point, making it not
only effective in classifying points but also in finding the most robust hyperplane. When
the points are not linearly separable and a higher-dimensional feature space is needed, the
kernel trick is used.
SVR is a version of the SVM adapted to perform regression tasks. SVR fits the error of
its predictions within the limit while minimizing the loss function in Equation (1), which
is called the L2 loss:
min 12 β β
(1)
|Yn − ( Xn β + b)| ≤ ∀n
where
β , β—values that weight arrays, normal and transposed
Yn —target values
Xn —transposed descriptor array
b—bias
—maximum allowed error.
The constraint is then relaxed, introducing the slack variables and applying what is
called the soft margin approach.
⎧ N
⎪
⎪ (ξ n + ξ n∗ )
⎪ min 2 β β + C n∑
1
⎪
⎨ =1
Yn − ( Xn β + b) ≤ + ξ n ∀n (2)
⎪
⎪
⎪
⎪
⎩
( Xn β + b) − Yn ≤ + ξ n∗ ∀n
where
ξ n , ξ n∗ —slack variables for positive and negative error
C—weight associated with slack variables.
The prediction is expressed as a function of the training samples in Equation (3), in
particular of those data points with either αi or αi∗ different from 0, which are called support
vectors. ⎧ N
⎪
⎪ β = ∑ (αn − α∗n ) Xn
⎨
n =1 (3)
⎪
⎪
N
⎩ f ( x ) = ∑ (αn − α∗n )( Xn X ) + b
n =1
In this paper, the SVR hyperparameters have been initially tuned with the MATLAB
built-in function for SVR models, using the Bayesian optimization algorithm, and run
for 500 iterations to define a good starting point for the hyperparameters. The tunable
hyperparameters are as listed:
Box constraint: Coefficient C that weights the slack variables in Equation (1) and helps
regulate overfitting.
Epsilon (ε): The value that defines the radius of the epsilon tube where the algorithm
tries to contain the points or, in other words, the maximum error allowed.
Kernel scale: The value that rescales the predictors. Each value in the predictors is
divided by the kernel scale value.
Kernel function: The value used to compute the similarity between data points in a
higher-dimensional feature space.
57
Energies 2023, 16, 4423
Additional tuning of the hyperparameters was carried out during the validation
process. The final values of the hyperparameters are shown in Table 1. The linear kernel
function was selected because the features are quite proportional to the target value to
estimate and working in a higher-dimensional space was unnecessary. In fact, different
kernel functions led to lower validation accuracy.
Table 1. Hyperparameter values tuned for the implemented SVR model.
Hyperparameter Value
1 Box constraint 0.1989
2 Kernel scale 11.55
3 Epsilon 0.030
4 Kernel function Linear
3.3. Random Forest

A random forest (RF) is an ensemble learning method that puts together many decision
trees (DT) as weak learners and is one of the best-known and most used algorithms for
supervised learning tasks. In [38], a RF is used to perform an incremental capacity analysis
to estimate the capacity of lithium batteries by only feeding raw measurements of new data
to the model.
A decision tree is a non-parametric algorithm that develops a tree by splitting the
dataset over the values of its features and associates different subsets of the dataset to
different nodes of the tree. First, the entire dataset is paired with the root of the tree. Next,
the dataset is split into two parts according to a decision made over some of the features,
and each part is associated with a new node child of the root, forming the second level
of the tree. This behavior is recursively iterated until subsets of the dataset contain only
one value or a stop criterion is met, with the final subsets representing the leaves of the
tree. Each splitting is made over the value of typically one feature, and the choice for the
optimal split is made by finding the feature and its splitting value that optimize a given
metric. MSE metric minimization was used in this work:
⎧ N
⎪
⎪ MSE(S) = N1 ∑ (yi − y)2
⎨
i =1
(4)
⎪
⎪ splitMSE( F, V ) = NNL MSE(S L ) + NNR MSE(SR )
⎩
( F ∗ , V ∗ ) = argmin(splitMSE( F, V ))
where y is the mean of the target values in the set S, yi is the i-th target value, and N is the
number of samples in the set. SplitMSE, SL , and SR are the weighted error, left, and right
subsets, respectively, generated by splitting S over the feature F at value V, while NL and
NR are the numbers of data points, respectively, in the left and right subsets. F* and S*
are the optimal feature-value pair to split the set. Other metrics, such as Gini impurity or
information gain, can be used.
However, decision trees are considered weak learners and strongly tend to overfit. A
random forest is an ensemble algorithm whose mechanism consists of combining multiple
decision trees with a bagging technique to provide higher accuracy and robustness than a
single tree, reducing overfitting. Bagging is, in fact, known for reducing the variance of
the model (as opposed to boosting, which reduces bias) by training each tree (or learner in
general) on a randomly selected subset of the training data with replacement (bootstrap-
ping), hence introducing diversity in the training data. What diversifies the random forest
from the standard tree bagging ensemble is the use of subsets of randomly selected features
for each tree in the forest, which helps to reduce correlation between each learner, thus
reducing overfitting. In this work, one-third of the total features were randomly used to
train each single decision tree.
58
Energies 2023, 16, 4423
4. Feature Selection
As aforementioned, the considered approaches were applied by considering a specific
feature of the batteries. In most battery applications, the charging stage is conducted in a
more repeatable way. While different chargers can be used, which will result in different
charging profiles, many charging cycles will be the same or very similar. On the other hand,
the battery discharge cycles vary greatly depending on the application and use patterns.
Even though the charging phase is more similar between different cycles, complete charging
cycles are by no means guaranteed. For this reason, a small portion of the charging curve
of voltage was used to extract useful information. More specifically, the extracted feature is
the partial charging time (PCT) necessary for the battery to charge by some small voltage
range.
In Figure 1, the battery voltage versus time during charging for different cycles is
represented. Unsurprisingly, the charging time decreases as the battery ages and the global
capacity decreases. In fact, the charging time is halved near the final cycles compared to the
initial ones. It is further noted that the beginning of the charging process is characterized
by a high derivative and is therefore difficult to appreciate the time differences between
different cycles. On the other hand, the middle part extends for a longer period of time and
is more suitable for PCT feature extraction. This is why, in this work, the lower voltage
limit of 3.7 V was set for the feature extraction process.
Figure 1. Charging curves at different cycles for battery 5.
5. Results and Discussion

The initial choice of the voltage range and limits was made empirically by computing
four features over the limits of 3.7–4.1 volts with a voltage range of 0.1 V. More specifically,
the first feature represents the evolution of the charging time between 3.7 V and 3.8 V over
the number of cycles, the second feature uses the range of 3.8 V to 3.9 V, etc. In Figure 2,
the value of the considered features as a function of the number of cycles is plotted. The
first PCT feature computed for the lowest voltage values, from 3.7 V to 3.8 V, appeared to
be an almost flat curve, containing no variance and thus very little information regarding
the data. Conversely, the features computed from 3.8 V to 4.1 V have a higher variance and
hence are more descriptive of the aging phenomena.
To find the optimal features and model parameters, from the voltage limits of 3.7–4.1 V,
many feature sets were created. These sets differ from each other depending on the number
of features, the upper and lower voltage limits used, and the voltage range. For each feature
set, the models obtained using the different ML strategies are compared.
59
Energies 2023, 16, 4423
Figure 2. PCT values calculated considering the voltage limits of 3.7 V to 4.1 V, with a voltage range
of 0.1 V, as a function of number of cycles, for battery 5.
The fitting accuracy of the various models was assessed through the value of the
coefficient of determination (R2 ). It is a measure used in statistics, indicating how much a
hypothesis describes the variance of the data. In other words, it is a measure of how well a
model can fit the data. R2 is described as
⎧
⎪
⎪ R2 = 1 − SS r
⎪
⎨
SSt
SSr = ∑(yi − f i )2
(5)
⎪
⎪
i
⎪
⎩ SSt = ∑(yi − y)
2
i
where
SSr —residual sum of squares
SSt —total sum of squares
yi —target value
f i —estimated value
y—mean of the target values.
A three-fold cross-validation (CV) procedure was applied to the three batteries of the
dataset to find the best features and ML strategies. This means the SoH evolution of each
battery was estimated based on the data of the other two batteries. The results are shown
in Tables 2 and 3 for the voltage ranges of 0.1 V and 0.05 V, respectively. Initially, a smaller
voltage range of 0.025 V and a larger voltage range of 0.2 V were also considered. However,
the smaller voltage range resulted in features with low variability for most voltage limits
and produced inferior results compared to the ones presented in Tables 2 and 3. The
larger range of 0.2 V and higher ranges did not improve the SoH-estimation capability of
the models. Since minimizing the voltage range was one of the objectives to ensure that
the features would be available, even in the case of partial charging cycles, the ranges of
0.05 V and 0.1 V were regarded as optimal, and the higher voltage ranges were not further
analyzed or presented.
60
Energies 2023, 16, 4423
Table 2. Three-fold CV results for a voltage range of 0.1 V.
Mean Validation R2
Voltage Number of Second-Degree Third-Degree
Feature Set Linear Random
Range Features Polynomial Polynomial SVR
Regression Forest
Regression Regression
A1 3.7–3.8 V 1 0.538 0.595 0.631 0.613 0.660
A2 3.8–3.9 V 1 0.918 0.916 0.921 0.939 0.902
A3 3.9–4 V 1 0.947 0.927 0.930 0.963 0.904
A4 4–4.1 V 1 0.554 0.535 / 0.759 0.743
A5 3.7–3.9 V 2 0.901 0.897 0.894 0.917 0.838
A6 3.8–4 V 2 0.945 0.939 0.946 0.971 0.909
A7 3.9–4.1 V 2 0.942 0.835 0.652 0.961 0.877
Table 3. Three-fold CV results for a voltage range of 0.05 V.
Mean Validation R2
Voltage Number of Second-Degree Third-Degree
Feature Set Linear Random
Range Features Polynomial Polynomial SVR
Regression Forest
Regression Regression
B1 3.8–3.85 V 1 0.781 0.896 0.897 0.810 0.878
B2 3.85–3.9 V 1 0.939 0.900 0.947 0.947 0.908
B3 3.9–3.95 V 1 0.937 0.918 0.916 0.949 0.896
B4 3.95–4 V 1 0.895 0.900 0.880 0.938 0.898
B5 3.8–3.9 V 2 0.931 0.909 0.928 0.941 0.901
B6 3.85–3.95 V 2 0.934 0.928 0.936 0.947 0.909
B7 3.9–4 V 2 0.950 0.912 0.922 0.968 0.903
B8 3.75–3.9 V 3 0.915 0.885 0.893 0.935 0.883
B9 3.8–3.95 V 3 0.899 0.911 0.895 0.948 0.905
B10 3.85–4 V 3 0.943 0.938 0.896 0.964 0.910
B11 3.9–4.05 V 3 0.939 0.756 0.884 0.962 0.898
B12 3.8–4 V 4 0.936 0.922 0.885 0.966 0.907
B13 3.85–4.05 V 4 0.931 0.864 / 0.958 0.911
B14 3.9–4.1 V 4 0.934 0.775 / 0.972 0.892
Table 2 shows the feature sets of partial charging times obtained for a voltage range
of 0.1 volts. The first four single feature sets (A1–A4) explore the whole voltage range of
3.7 to 4.1 volts. Unsurprisingly, they show that all ML strategies perform better when the
voltage limits of 3.8–3.9 V (A2) or 3.9–4 V (A3) are used as a feature. More specifically,
the best results are obtained for the voltage limits 3.9–4 V when a single feature is used.
Additionally, Table 2 shows that if the feature set is built from two features based on the
limits of 3.8–3.9 V and 3.9–4 V (A6), there is only a marginal improvement in the R2 value.
In any case, the best results for single and double feature sets are A3 and A6.
Table 3 presents the feature sets obtained for a voltage range of 0.05 volts. In this
case, feature sets consisting of one to four features were constructed. For example, B1 is
a feature set of a single feature, which is the PCT between the voltage limits of 3.8 V to
3.85 V, while B6 consists of two features, which are the PCTs between the limits of 3.8 to
3.85 V and 3.85 to 3.9 V. The best results, per number of features, are B3, B7 B10, and B14.
Using a single feature, even for the voltage range of 0.05 V, is sufficient if the voltage limits
are between 3.85 and 4 volts. There is marginal improvement when two features are used;
however, a further increase in the number of features does not lead to any meaningful
increase of R2 . Considering the models of both tables, it can be noted that SVR delivers
slightly better results than the other considered ML strategies for all feature sets. Still,
using MLR also leads to satisfactory results. Furthermore, when comparing the three
strategies based on regression, no significant improvement in the R2 value is observed
when increasing the polynomial order using stepwise regression. That means the PCT
61
Energies 2023, 16, 4423
features and capacity reduction, as functions of the number of cycles, have a strong linear
dependence. Hence, high model complexity will not result in an improvement of the results
if the correct voltage range of 3.8–4 V has been selected. Actually, a drop in the mean
validation R2 value can even be observed in some cases due to overfitting the training data
of the higher-complexity models. This is especially apparent in models built from a higher
number of features (A7 and B12 to B14). However, some improvement when increasing the
polynomial order can also be observed for the voltage range of 3.7–3.8 V, which has low
variance. Finally, the models based on RF demonstrated worse performance than those of
MLR and SVR.
The models based on feature sets A3, A6, B3, B7, B10, and B14 all represent satisfactory
performance. Having the goal of minimizing the number of features and the voltage range,
the authors consider the models based on feature set B7 as the overall best. The plots for the
capacity estimation of all the batteries using MLR, SVR, and RF are plotted in Figures 3–5,
respectively.
Figure 3. SoH estimation of each battery achieved using a model based on MLR and trained on the
other two batteries.
Figure 4. SoH estimation of each battery achieved using a model based on SVR and trained on the
62
Energies 2023, 16, 4423
Figure 5. SoH estimation of each battery achieved using a model based on RF and trained on the
All three figures display the previously mentioned three-fold CV. For example, the SoH
estimation of battery 5 was done with a model trained using the data of the chosen feature
set of batteries 6 and 7. The full lines represent the measured SoH for the batteries, while
the dashed lines represent the estimated SoH over the number of cycles. Figures 3 and 4
show that MLR and SVR accurately model the SoH of the batteries, even registering the
peaks in the SoH function that are due to the rest time of the battery. Likewise, the RF is able
to model batteries 5 and 7 with similar success, but the same cannot be said about battery
6, as is evident in Figure 5. After the SoH of battery 6 falls to around 0.7, the estimation
begins to diverge from the measurement because batteries 5 and 7, which were used for
training, do not contain data with SoH lower than 0.7.
The random forest and decision trees are indeed well known for their inability to
extrapolate, that is, make estimations for predictor values lying outside of the range of
the observed data. From Figure 5, it is clear that the SoH value of battery 6 from cycle
90 onwards is lower than that of any other cycle of the training batteries; hence, the
decision trees will not be able to correctly estimate that target value. Furthermore, Figure 6
shows that also the feature value for battery 6 is lower than that of the other batteries.
Consequently, the branches of the decision trees built on batteries 5 and 7 will “explore”
the features in a range that does not include the values of battery 6 predictors after cycle
90. Hence, after this cycle number, all the decision trees of the random forest will infer the
lowest observed SoH value for battery 6, which will be around 0.7 because the training
data is composed of batteries 5 and 7. This is the reason for the observed flat line output. It
is important to specify that this result does not imply that the RF is not a suitable solution
for the general problem of battery prognostic because this precise case is strictly related to
the dataset distribution and data scarcity.
Figure 6. PCT for voltage range 3.9–3.95 V for all three batteries.
63
Energies 2023, 16, 4423
6. Conclusions
Accurate SoH estimation is essential for the safe and reliable operation of lithium-
ion batteries. This paper compares SoH-estimation models based on the classical ML
strategies of MLR, polynomial regression, SVR, and RF, which offer good trade-offs between
applicability, light computation effort, and accuracy of results. Discussion is provided on
the feature selection process and optimal number of features.
The partial charging time proved to be a good indicator of battery aging as long as
the proper voltage limits were selected, and the partial charging phase was equal at every
cycle. To find the optimal features, 21 feature sets were built considering different voltage
limits and the two voltage ranges of 0.1 and 0.05 V. The best results were obtained when
considering the voltage limits of 3.8 to 4 volts for both ranges of 0.1 V and 0.05 V. The
quality of the features degrades significantly for a minimum voltage of less than 3.7 V due
to small variance. Results showed that models based on one or two features are optimal.
Furthermore, the PCT feature demonstrated a linear dependence with capacity reduc-
tion as a function of number of cycles. Consequently, MLR produced very accurate results,
and the use of polynomial regression was not justified. The overall best performance for all
feature sets was achieved using SVR, especially when slightly lower voltage limits were
considered. Finally, the RF had the worst performance when facing the limited dataset.
Author Contributions: Conceptualization, E.P., I.M. and L.C.; methodology, E.P. and I.M.; software,
I.M. and E.P.; validation, I.M. and M.F.; formal analysis, E.P.; investigation, I.M.; resources, I.M.; data
curation, L.C.; writing—original draft preparation, E.P. and I.M.; writing—review and editing, E.P.
and I.M.; visualization, I.M.; supervision, L.C., M.F. and E.P.; project administration, L.C. and M.F.;
funding acquisition, L.C. and M.F. All authors have read and agreed to the published version of
the manuscript.
Data Availability Statement: Not applicable.
References
1. Cristaldi, L.; Faifer, M.; Laurano, C.; Ottoboni, R.; Petkovski, E.; Toscani, S. Power Generation Control Algorithm for the
Participation of Photovoltaic Panels in Network Stability. IEEE Trans. Instrum. Meas. 2023, 72, 1–9. [CrossRef]
2. Pelletier, S.; Jabali, O.; Laporte, G.; Veneroni, M. Battery Degradation and Behaviour for Electric Vehicles: Review and Numerical
Analyses of Several Models. Transp. Res. Part B Methodol. 2017, 103, 158–187. [CrossRef]
3. Werling, T.; Geuting, P.; Höschele, P.; Ellersdorfer, C.; Sinz, W. Investigation of the Electro-Mechanical Behavior of Automotive
High Voltage Busbars under Combined Electrical Load with Varying Indenter Geometry and Environmental Conditions. J. Energy
4. Campestrini, C.; Horsche, M.F.; Zilberman, I.; Heil, T.; Zimmermann, T.; Jossen, A. Validation and Benchmark Methods for Battery
Management System Functionalities: State of Charge Estimation Algorithms. J. Energy Storage 2016, 7, 38–51. [CrossRef]
5. Tanim, T.R.; Rahn, C.D.; Wang, C.Y. A Temperature Dependent, Single Particle, Lithium Ion Cell Model Including Electrolyte
Diffusion. J. Dyn. Syst. Meas. Control Trans. ASME 2015, 137, 011005. [CrossRef]
6. Li, X.; Fan, G.; Rizzoni, G.; Canova, M.; Zhu, C.; Wei, G. A Simplified Multi-Particle Model for Lithium Ion Batteries via a
Predictor-Corrector Strategy and Quasi-Linearization. Energy 2016, 116, 154–169. [CrossRef]
7. Petit, M.; Prada, E.; Sauvant-Moynot, V. Development of an Empirical Aging Model for Li-Ion Batteries and Application to Assess
the Impact of Vehicle-to-Grid Strategies on Battery Lifetime. Appl. Energy 2016, 172, 398–407. [CrossRef]
8. Barcellona, S.; Colnago, S.; Dotelli, G.; Latorrata, S.; Piegari, L. Aging Effect on the Variation of Li-Ion Battery Resistance as
Function of Temperature and State of Charge. J. Energy Storage 2022, 50, 104658. [CrossRef]
9. Xu, B.; Oudalov, A.; Ulbig, A.; Andersson, G.; Kirschen, D.S. Modeling of Lithium-Ion Battery Degradation for Cell Life
Assessment. IEEE Trans. Smart Grid 2018, 9, 1131–1140. [CrossRef]
10. Cui, Z.; Wang, C.; Gao, X.; Tian, S. State of Health Estimation for Lithium-Ion Battery Based on the Coupling-Loop Nonlinear
Autoregressive with Exogenous Inputs Neural Network. Electrochim. Acta 2021, 393, 139047. [CrossRef]
11. Liu, D.; Zhou, J.; Liao, H.; Peng, Y.; Peng, X. A Health Indicator Extraction and Optimization Framework for Lithium-Ion Battery
Degradation Modeling and Prognostics. IEEE Trans. Syst. Man. Cybern. Syst. 2015, 45, 915–928. [CrossRef]
64
Energies 2023, 16, 4423
12. Cao, M.; Zhang, T.; Wang, J.; Liu, Y. A Deep Belief Network Approach to Remaining Capacity Estimation for Lithium-Ion Batteries
Based on Charging Process Features. J. Energy Storage 2022, 48, 103825. [CrossRef]
13. Yang, D.; Zhang, X.; Pan, R.; Wang, Y.; Chen, Z. A Novel Gaussian Process Regression Model for State-of-Health Estimation of
Lithium-Ion Battery Using Charging Curve. J. Power Sources 2018, 384, 387–395. [CrossRef]
14. Feng, X.; Weng, C.; He, X.; Han, X.; Lu, L.; Ren, D.; Ouyang, M. Online State-of-Health Estimation for Li-Ion Battery Using Partial
Charging Segment Based on Support Vector Machine. IEEE Trans. Veh. Technol. 2019, 68, 8583–8592. [CrossRef]
15. Tian, J.; Xiong, R.; Yu, Q. Fractional-Order Model-Based Incremental Capacity Analysis for Degradation State Recognition of
Lithium-Ion Batteries. IEEE Trans. Ind. Electron. 2019, 66, 1576–1584. [CrossRef]
16. Ansean, D.; Garcia, V.M.; Gonzalez, M.; Blanco-Viejo, C.; Viera, J.C.; Pulido, Y.F.; Sanchez, L. Lithium-Ion Battery Degradation
Indicators Via Incremental Capacity Analysis. IEEE Trans. Ind. Appl. 2019, 55, 2992–3002. [CrossRef]
17. Stroe, D.I.; Schaltz, E. SOH Estimation of LMO/NMC-Based Electric Vehicle Lithium-Ion Batteries Using the Incremental Capacity
Analysis Technique. In Proceedings of the 2018 IEEE Energy Conversion Congress and Exposition (ECCE), Portland, OR, USA,
23–27 September 2018; pp. 2720–2725. [CrossRef]
18. He, J.; Wei, Z.; Bian, X.; Yan, F. State-of-Health Estimation of Lithium-Ion Batteries Using Incremental Capacity Analysis Based on
Voltage-Capacity Model. IEEE Trans. Transp. Electrif. 2020, 6, 417–426. [CrossRef]
19. Severson, K.A.; Attia, P.M.; Jin, N.; Perkins, N.; Jiang, B.; Yang, Z.; Chen, M.H.; Aykol, M.; Herring, P.K.; Fraggedakis, D.; et al.
Data-Driven Prediction of Battery Cycle Life before Capacity Degradation. Nat. Energy 2019, 4, 383–391. [CrossRef]
20. Barcellona, S.; Cristaldi, L.; Faifer, M.; Petkovski, E.; Piegari, L.; Toscani, S. State of Health Prediction of Lithium-Ion Batteries. In
Proceedings of the 2021 IEEE International Workshop on Metrology for Industry 4.0 & IoT (MetroInd4.0&IoT), Virtual Event, 7–9
June 2021; pp. 12–17.
21. Lashgari, F.; Petkovski, E.; Cristaldi, L. State of Health Analysis for Lithium-Ion Batteries Considering Temperature Effect. In
Proceedings of the 2022 IEEE International Workshop on Metrology for Extended Reality, Artificial Intelligence and Neural
Engineering, MetroXRAINE 2022—Proceedings, Rome, Italy, 26–28 October 2022; pp. 40–45. [CrossRef]
22. Chen, C.; Pecht, M. Prognostics of Lithium-Ion Batteries Using Model-Based and Data-Driven Methods. In Proceedings of the
IEEE 2012 Prognostics and System Health Management Conference, PHM-2012, Beijing, China, 23–25 May 2012; pp. 12–17.
[CrossRef]
23. Weng, C.; Cui, Y.; Sun, J.; Peng, H. On-Board State of Health Monitoring of Lithium-Ion Batteries Using Incremental Capacity
Analysis with Support Vector Regression. J. Power Sources 2013, 235, 36–44. [CrossRef]
24. Weng, C.; Sun, J.; Peng, H. Model Parametrization and Adaptation Based on the Invariance of Support Vectors with Applications
to Battery State-of-Health Monitoring. IEEE Trans. Veh. Technol. 2015, 64, 3908–3917. [CrossRef]
25. Chen, Z.; Sun, M.; Shu, X.; Shen, J.; Xiao, R. On-Board State of Health Estimation for Lithium-Ion Batteries Based on Random
Forest. In Proceedings of the 2018 IEEE International Conference on Industrial Technology (ICIT), Lyon, France, 19–22 February
2018; pp. 1754–1759.
26. Liu, K.; Li, Y.; Hu, X.; Lucu, M.; Widanage, W.D. Gaussian Process Regression with Automatic Relevance Determination Kernel
for Calendar Aging Prediction of Lithium-Ion Batteries. IEEE Trans. Industr. Inform. 2020, 16, 3767–3777. [CrossRef]
27. Su, X.; Sun, B.; Wang, J.; Zhang, W.; Ma, S.; He, X.; Ruan, H. Fast Capacity Estimation for Lithium-Ion Battery Based on Online
Identification of Low-Frequency Electrochemical Impedance Spectroscopy and Gaussian Process Regression. Appl. Energy 2022,
322, 119516. [CrossRef]
28. Tian, Y.; Dong, Q.; Tian, J.; Li, X.; Kukkapalli, V.K.; Kim, S.; Thomas, S.A. Capacity Estimation of Lithium-Ion Batteries Based on
Multiple Small Voltage Sections and BP Neural Networks. Energies 2023, 16, 674. [CrossRef]
29. Eddahech, A.; Briat, O.; Bertrand, N.; Delétage, J.Y.; Vinassa, J.M. Behavior and State-of-Health Monitoring of Li-Ion Batteries
Using Impedance Spectroscopy and Recurrent Neural Networks. Int. J. Electr. Power Energy Syst. 2012, 42, 487–494. [CrossRef]
30. Catelani, M.; Ciani, L.; Fantacci, R.; Patrizi, G.; Picano, B. Remaining Useful Life Estimation for Prognostics of Lithium-Ion
Batteries Based on Recurrent Neural Network. IEEE Trans. Instrum. Meas. 2021, 70, 3524611. [CrossRef]
31. Zhang, Y.; Xiong, R.; He, H.; Pecht, M.G. Long Short-Term Memory Recurrent Neural Network for Remaining Useful Life
Prediction of Lithium-Ion Batteries. IEEE Trans. Veh. Technol. 2018, 67, 5695–5705. [CrossRef]
32. Marri, I.; Petkovski, E.; Cristaldi, L.; Faifer, M. Battery Remaining Useful Life Prediction Supported by Long Short-Term Memory
Neural Network. In Proceedings of the IEEE International Instrumentation and Measurement Technology Conference (I2MTC),
Kuala Lumpur, Malaysia, 22–25 May 2023; pp. 1–6.
33. Qu, J.; Liu, F.; Ma, Y.; Fan, J. A Neural-Network-Based Method for RUL Prediction and SOH Monitoring of Lithium-Ion Battery.
IEEE Access 2019, 7, 87178–87191. [CrossRef]
34. Marri, I.; Petkovski, E.; Cristaldi, L.; Faifer, M. Lithium-Ion Batteries Soh Estimation, Based on Support-Vector Regression and a
Feature-Based Approach. In Proceedings of the 18th IMEKO TC10 Conference on Measurement for Diagnostic, Optimisation and
Control to Support Sustainability and Resilience, Warsaw, Poland, 26–27 September 2022; pp. 109–113.
35. Saha, B.; Goebel, K. Nasa Ames Prognostic Data Repository. NASA Ames Moffet Field, CA, USA. 2007. Available online:
https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository (accessed on 15 January 2022).
36. Chen, Z.; Xia, X.; Sun, M.; Shen, J.; Xiao, R. State of Health Estimation of Lithium-Ion Batteries Based on Fixed Size LS-SVM. In
Proceedings of the 2018 IEEE Vehicle Power and Propulsion Conference (VPPC), Chicago, IL, USA, 27–30 August 2018; pp. 1–6.
65
Energies 2023, 16, 4423
37. Klass, V.; Behm, M.; Lindbergh, G. A Support Vector Machine-Based State-of-Health Estimation Method for Lithium-Ion Batteries
under Electric Vehicle Operation. J. Power Sources 2014, 270, 262–272. [CrossRef]
38. Li, Y.; Zou, C.; Berecibar, M.; Nanini-Maury, E.; Chan, J.C.W.; van den Bossche, P.; Van Mierlo, J.; Omar, N. Random Forest
Regression for Online Capacity Estimation of Lithium-Ion Batteries. Appl. Energy 2018, 232, 197–210. [CrossRef]
66
energies
Article
Calendar Aging Effect on the Open Circuit Voltage of
Lithium-Ion Battery
Simone Barcellona *, Lorenzo Codecasa, Silvia Colnago and Luigi Piegari
lorenzo.codecasa@polimi.it (L.C.); silvia.colnago@polimi.it (S.C.); luigi.piegari@polimi.it (L.P.)
* Correspondence: simone.barcellona@polimi.it
Abstract: In recent years, lithium-ion batteries (LiBs) have gained a lot of importance due to the
increasing use of renewable energy sources and electric vehicles. To ensure that batteries work
properly and limit their degradation, the battery management system needs accurate battery models
capable of precisely predicting their parameters. Among them, the state of charge (SOC) estimation is
one of the most important, as it enables the prediction of the battery’s available energy and prevents
it from operating beyond its safety limits. A common method for SOC estimation involves utilizing
the relationship between the state of charge and the open circuit voltage (OCV). On the other hand,
the latter changes with battery aging. In a previous work, the authors studied a simple function to
model the OCV curve, which was expressed as a function of the absolute state of discharge, q, instead
of SOC. They also analyzed how the parameters of such a curve changed with the cycle aging. In the
present work, a similar analysis was carried out considering the calendar aging effect. Three different
LiB cells were stored at three different SOC levels (low, medium, and high levels) for around 1000
days, and an analysis of the change in the OCV-q curve model parameters with the calendar aging
was performed.
Keywords: lithium-ion batteries; calendar aging; OCV curve; state of charge estimation
Citation: Barcellona, S.; Codecasa, L.; 1. Introduction

Colnago, S.; Piegari, L. Calendar The recent energy crisis and the issue of climate change are spurring governments
Aging Effect on the Open Circuit to incentivize the production of energy through renewable energy sources (RESs) and
Voltage of Lithium-Ion Battery. electric mobility to reduce gas emissions and reliance on fossil fuels. Unfortunately, the
Energies 2023, 16, 4869. https://
power produced by RESs is intermittent and unpredictable, which means that they may
doi.org/10.3390/en16134869
not provide energy when needed or, conversely, produce it when not required. Therefore,
Academic Editor: to make the most of them, it is convenient to install energy storage systems. Additionally,
Djaffar Ould-Abdeslam energy storage systems are also necessary for electric mobility, preferably with high energy
density, high efficiency, and a long lifecycle. In this scenario, lithium-ion batteries (LiBs)
Received: 15 May 2023
are currently the preferred technology due to their high energy density, relatively high
Revised: 17 June 2023
power density, and low self-discharge, making them suitable for both mobile and stationary
Accepted: 19 June 2023
Published: 22 June 2023
applications [1].
LiBs are subjected to various degradation mechanisms that limit their lifespan and
degrade their performance. These degradation mechanisms are caused by storage and
operating conditions. The former is known as calendar aging and mainly depends on
Copyright: © 2023 by the authors. the temperature and state of charge (SOC) as a function of the time for which the battery
Licensee MDPI, Basel, Switzerland. has been stored under those conditions. The latter, commonly referred to as cycle aging,
This article is an open access article mainly depends on the temperature, current rate, SOC, and charging and discharging
distributed under the terms and cut-off voltages as a function of the total charge exchanged with the battery [2,3]. To
conditions of the Creative Commons reduce the degradation and extend the lifespan of batteries, it is essential to control battery
Attribution (CC BY) license (https:// operations properly and store them under optimal conditions. The former is ensured by
creativecommons.org/licenses/by/ the battery management system (BMS), which predicts battery parameters such as the
4.0/).

67
Energies 2023, 16, 4869
SOC, temperature, and state of health (SOH), using appropriate models and controlling the
battery accordingly. Therefore, it is crucial for the aforementioned battery models to be as
accurate as possible [4].
The estimation of the SOC is of paramount importance as it allows the determination
of the remaining available battery capacity and ensures that the battery operates within its
safety limits. In fact, properly controlling the SOC of the battery helps avoid working in an
operating area where its degradation increases. Many different SOC estimation methods
can be found in the literature. In recent years, some filter-based methods (Kalman filter [5],
unscented Kalman filter [6], etc.) together with some observer methods (sliding mode
observer [7], nonlinear observer [8], etc.) and intelligent algorithms (fuzzy logic [9], neural
networks [10], etc.) have been proposed. However, traditional methods are still preferred
in BMSs [11]. Among them, Coulomb counting remains the most widely used method due
to its simplicity and execution speed [12]. Unfortunately, this method has some drawbacks,
including heavy reliance on initial SOC values, the need for highly accurate current sensors,
and susceptibility to cumulative errors. Other traditional methods are the model-based
approaches which are quite accurate and simple [13]. On the other hand, they require a
good knowledge of the open circuit voltage (OCV) curve [14], which is typically obtained
through lengthy tests involving full charge or discharge of the battery at very low current
rates or through the hybrid pulse power characterization method [15]. Furthermore, the
model-based approaches are generally used to correct the Coulomb counting estimation
method [11].
Different models can be found in the literature that are able to describe the dependence
of the OCV curve as a function of the SOC. They can be mainly divided into table-based
models and analytical models. In the former, pairs of OCV and SOC values are stored in a
table, and interpolation is performed between the stored values [16,17]. The main advantage
of this method is its low computational requirements, but it may require significant memory
if high precision is desired. The analytical models, instead, use mathematical functions
to describe the OCV curve and can be further classified into linear regression models
and nonlinear regression models. Linear regression models consist of a sum of products
between parameters and linear or nonlinear functions of SOC, while nonlinear regression
models involve a general function of SOC and parameters. In linear regression models,
the parameters can be determined using linear regression methods such as the linear least
square method. In contrast, nonlinear regression methods like the nonlinear least square
method are required for parameter estimation in nonlinear regression models. In some
cases, it is also possible to linearize the nonlinear regression model and employ a linear
regression method.
Among the mathematical expressions used in the OCV-SOC curve modeling, poly-
nomial functions are the most commonly employed. These polynomials can range from
the second-degree [18] to the twelfth-degree [19,20], depending on the desired accuracy
and complexity. Lower-degree polynomials are simpler but offer limited precision and
can accurately represent only a small portion of the OCV-SOC curve. Conversely, the
higher-degree polynomials can provide excellent precision and fit the entire OCV curve
well. However, they require many parameters to be fitted and may exhibit incorrect trends
outside the range or between the experimental points. Other analytical models, which are
possible to find in the literature, are based on logarithmic functions, also called Nernst mod-
els [15,21,22] or exponential functions [23]. Logarithmic functions offer a good accuracy
with only three parameters to be fitted but cannot be defined for an SOC equal to 0 or 1.
Moreover, there are a lot of different combinations of the aforementioned functions [24–28],
some of which yield higher accuracy than others. Among them, the model proposed
in [23] demonstrates high accuracy and low complexity. The latter is composed of two
exponentials and a quadratic term, with a total of five parameters.
The OCV-SOC curve changes with the battery temperature and aging. Therefore,
accurately estimating this curve and understanding its dependence on these factors is
crucial for building a reliable battery model and, consequently, a reliable BMS. In [29], the
68
Energies 2023, 16, 4869
authors corrected the SOC estimation with the value of the actual battery capacity and
stated that the OCV curves were the same with this correction, despite different aging
conditions. A similar procedure was employed in [30], where the SOC was defined as a
function of the SOH and then used inside the OCV-SOC relation. The changes in the OCV-
SOC curve and the consequent variations in the incremental capacity of a LiB were analyzed
in [25], in which an extended Kalman filter was used for the parameter estimation. In [31], a
correction of the OCV-SOC curve based on the SOH and temperature was proposed, while
the authors in [32] used a convolutional neural network to estimate the electrode aging
parameters, which were then used for the OCV-SOC curve estimation.
In all cases, there seems to be a lack of understanding regarding how the parameters
of an OCV-SOC curve change as a function of aging. In a previous work [33], the authors
used a double exponential function to model the discharge OCV curve as a function of the
absolute state of discharge, q, instead of the SOC, and investigated the dependency of its
parameters on cycle aging at a fixed temperature.
In light of the above, the focus of the present paper was to examine the variation of
the discharge OCV-q curve as a function of the calendar aging at a fixed temperature. The
same procedure and OCV-q curve model adopted in [33] to analyze the cycle aging of a
LiB were applied to three different LiB cells of the same type. The latter were stored at
three different SOC levels (low, medium, and high levels) for a period of almost three years
under the same temperature conditions. Therefore, the parameter variation of the OCV-q
curve model was studied for different calendar aging levels (i.e., different storage times)
and three different SOC levels, developing a calendar aging model. Finally, the proposed
model was validated through a wide campaign of experimental tests.
2. Battery Model
To develop an aging model that considers how the discharge OCV-q curve changes as
a function of calendar aging, we needed to start from an equivalent electric circuit battery
model. This allowed us to model the discharge OCV-q curve and develop a test procedure
to extract it. In the literature, it is possible to retrieve many equivalent electric circuit models
able to predict both the static and dynamic behavior of batteries. They can be very simple,
as the ones reported in [34–36], or much more complex, as the ones reported in [37–39]. The
choice of the appropriate equivalent electric circuit model depends on the desired accuracy
and the specific aspects to be described for the application.
For the scope of the present paper, the simplest model, shown in Figure 1, was suitable
for characterizing the discharge OCV-q curves. This model consists of a voltage source, E, as
a function of the absolute state of discharge, q, which models the OCV of the battery, and a
series resistor Rin , which represents the total internal resistance of the battery. This resistance
is related to the ohmic resistance of the electrode and electrolyte, charge transfer chemical
reaction resistance, solid electrolyte interface (SEI) resistance, and diffusion resistance.
Rin I
E(q) +
о V
Figure 1. Equivalent electric circuit model.
According to this model, it is possible to express the battery terminal voltage as follows:
V = E − Rin · I (1)
where I is the battery current.

The analytical expression chosen to model the discharge OCV-q curve was the same as
that reported in [33], which, in that case, was used to model the changes in the discharge
69
Energies 2023, 16, 4869
OCV-q curve as a function of the cycle aging. This analytical expression consists of two
exponential terms and a constant one:
E ( q ) = a · eb·q + c · ed·q + f (2)
where a, b, c, d, and f are five parameters that, in our case, are functions of the calendar
aging, e is the Euler’s number, while q is the absolute state of discharge in Ah and defined
as follows:
t
1
q= I · dτ + q(0) (3)
3600
0
where q(0) is the initial state of discharge. In this way, when q is null, the battery is fully
charged, i.e., SOC = 100%; conversely, when q equates the actual capacity of the battery,
Ca , the latter is fully discharged, i.e., SOC = 0%. Therefore, the SOC can be expressed as a
function of the absolute state of discharge, as follows:

q
SOC = 1 − · 100. (4)
Ca
To quantify calendar aging, the calendar time taging expressed in days was used. Thus,
the five parameters of (2) could be expressed as a function of this calendar time.
In this work, the procedure to obtain the discharge OCV-q curve was performed by
discharging the battery at the nominal current rate (1C) to speed up the tests, as performed
in [33]. Therefore, it was essential to correct the discharge voltage curve by eliminating
the voltage drop over the total internal resistance of the battery. Thus, it was necessary to
estimate the value of such an internal resistance. For the sake of simplicity, as assumed
in [33], the battery’s internal resistance was considered to be quite constant as a function
of the absolute state of discharge. Moreover, assuming the battery was fully charged,
when the discharge at constant current started, different phenomena, with different time
constants, led to an electrical transient. To obtain the total battery’s internal resistance, the
entire electric transient had to be extinguished. To do this, we had to consider a proper time
interval related to the largest time constant τ, which corresponds to the diffusion process
of the lithium/lithium ions into the electrodes and electrolyte. Through the procedure
reported in [26,27], it was possible to identify that time interval, considered to be five
times the largest time constant τ, in the first part of the discharge voltage curve (Figure 2).
Therefore, the resulting voltage variation related to that time interval was estimated. By
calculating the ratio between that voltage variation and the current step, the battery’s
internal resistance was calculated, and the discharge voltage curve was corrected, obtaining
the discharge OCV-q curve. Finally, the part of the OCV-q curve corresponding to the
considered electric transient was eliminated to avoid errors in the following analysis.
Figure 2. Electric transient at the beginning of the battery discharge.
70
Energies 2023, 16, 4869
3. Experimental Setup and Test Procedure

In the present work, three LiCoO2 8773160K pouch cells of the same batch, manufac-
tured by General Electronics Battery Co. Ltd. (Shenzhen, China), were employed for the
experimental tests. The main specifications of this type of cell are reported in Table 1.
Table 1. Battery cell specifications.
Parameter Value
Nominal capacity 10 Ah
Maximum voltage 4.2 V
Discharge cut-off voltage 2.75 V
Maximum continuous discharge current 100 A (10C)
The three battery cells were used to test how the discharge OCV-q curve changed as a
function of the calendar aging under three different storage conditions related to low SOC
(about 7%), medium SOC (about 50%), and high SOC (about 93%).
3.1. Experimental Setup

Figure 3 shows the experimental setup that was composed of a potentiostat (SP-150)
connected to a 100 A booster (VMP3B-100). Both instruments are manufactured by Biologic
Science Instrument, and they are connected to a PC via an ethernet cable. The PC controlled
them through the EC-Lab software. The 100 A booster was connected to the battery cell
under test through a power cable.

-ŗśŖ

ř-ŗŖŖ ȱ ȱ
¢
ȱ

¡ȱŘŞŖŜşȱ
Ƹȱ
Figure 3. Experimental setup.
The battery cell arrangement consisted of a heatsink, two fans, three Peltier cells, a dc
voltage source, a temperature probe, and a Texas Instrument DRV8303 inverter controlled
by an F28069M controller board. The three Peltier cells were put between the battery cell
under test and the heatsink to maintain the battery temperature as uniform and constant
as possible at 25 ◦ C during the discharge voltage curve measurements. This was carried
out to avoid changes on the OCV-q curve due to temperature variation. To do this, they
were connected in series and powered by the inverter, which was, in turn, controlled by the
F28069M with a PI controller. Finally, the inverter was supplied using a dc voltage source.
Figure 4 shows the schematics of the temperature control.
71
Energies 2023, 16, 4869
ŘŞŖŜş
ŞřŘřȱ
Tref
+о
Tmea s
ȱ
řŜȱ + ȱ

Figure 4. Temperature control scheme.
3.2. Test Procedure

The test procedure used for the analysis consisted of two main phases: the discharge
voltage curve measurement phase and the calendar aging phase. The former phase was
conducted at the beginning of the test and after each calendar aging phase for all three
battery cells. In this way, it was possible to assess how the discharge OCV-q curve changed
as a function of the calendar aging at different SOC levels.
3.2.1. Voltage Discharge Curve Measurement Phase

This phase consisted of fully charging and discharging the battery cell under test at
the nominal current (1C) using the well-known constant current–constant voltage (CC-CV)
protocol for the charging phase and the constant current protocol for the discharging phase.
In particular, the battery cell was charged with a constant current of 10 A (1C) until the
battery terminal voltage reached 4.2 V. After that, this voltage was applied until the battery
current dropped by 100 mA (0.01C). At that stage, the battery cell could be considered
charged at 100% of SOC. After completing the previous step, the battery cell was discharged
at a constant current of 10 A (1C) until the battery terminal voltage reached 2.75 V. This
entire procedure was repeated seven times for each of the three battery cells. In fact, for the
first use of the batteries and after each calendar aging phase, the battery cells needed to
be activated.
Finally, after each voltage discharge curve measurement phase, the three battery cells
were brought to about 7%, 50%, and 93% of SOC for the low, medium, and high SOC tests,
respectively. In particular, for the low and medium SOC tests, the cells were initially fully
discharged with the CC-CV protocol, and then a charge current of 10 A (1C) was applied
to move 0.7 Ah (7% of the nominal capacity) and 5 Ah (50% of the nominal capacity),
respectively. On the other hand, for the high SOC test, the battery cell was first fully
charged with the CC-CV protocol, and then a discharge current of 10 A (1C) was applied
to move 0.7 Ah (7% of the nominal capacity). It is worth noting that since, during the
aging, the actual capacity of the battery cells decreases, the moved charge (0.7 Ah and 5 Ah)
corresponds to different percentages of SOC. In any case, these differences are minimal
(about 5%) and do not change the SOC ranges at which the three cells were stored.
3.2.2. Calendar Aging Phase

This phase consisted of storing the three battery cells in a cabinet at a temperature
between 20 ◦ C and 30 ◦ C for a certain time interval. The duration of this time interval was
not constant throughout the entire analysis, but it varied according to the availability of the
experimental setup.
4. Results and Model Validation

According to the test procedure reported in the previous section, a total of twelve
calendar aging phases were evaluated at specific time intervals: 0th, 35th, 70th, 99th, 136th,
72
Energies 2023, 16, 4869
205th, 273rd, 344th, 437th, 682nd, 743rd, and 997th day. The entire duration of the testing
period was about 1000 days. Among these twelve tests, a subset of four (at the 0th, 205th,
437th, and 997th day) was selected and used to tune the proposed aging model. As stated
before, for each calendar aging phase and battery cell under test, seven voltage discharge
curves were performed to activate the battery cells and stabilize their capacities. The latter
were estimated in Ah at the rated nominal current of 10 A (1C) by integrating the current
over the entire discharging time. Figure 5a shows an example of the battery capacity trend,
as a function of the seven voltage discharge curve measurements, for the high SOC test
battery cell on the 0th day. From this figure, it is possible to recognize that the values of
the battery capacity become quite stable after at least five full charge–discharge cycles.
Therefore, the last two voltage discharge curves were averaged and used for the analysis.
Figure 5b illustrates the related temperature profile during the entire discharge curve
measurement. Finally, according to the procedure proposed in [26,27], the time interval
related to the lithium-ion diffusion was estimated to be approximately 50 s for all discharge
curves. Therefore, for each of them, the total battery’s internal resistance was evaluated
along with the related voltage drops, and the initial section of the curve was eliminated. In
this way, the OCV-q curves for each calendar aging phase and SOC level were obtained.
Figure 5. (a) battery capacity; (b) battery temperature.
4.1. Model Characterization

In the discharge voltage curve measurements, each battery cell under test was initially
fully charged up to 4.2 V using the CC-CV protocol and then fully discharged with the
CC protocol to obtain the OCV-q curves. Consequently, we can assume that all the OCV-q
curves started from the maximum battery voltage of 4.2 V, and the parameter f of (2) can
be rearranged as follows:
f = 4.2 − a − c. (5)
Therefore, the total number of parameters to be tuned in (2) reduces to just four. At
this stage, the OCV-q curves were fitted using the proposed OCV-q analytical expression
through the nonlinear least square regression method, and the coefficients of determination
(R2 ) were evaluated.
Figure 6 illustrates the experimental OCV-q curves along with the corresponding
fitting functions for the four tests chosen for the characterization of the model at each SOC
level, and Figure 7 shows the related R2 . From these figures, it is possible to recognize a
very good agreement between the experimental and modeled data for the low, medium,
and high SOC tests. Moreover, the R2 is greater than 0.995 for all tests.
73
Energies 2023, 16, 4869
Low SOC Medium SOC High SOC

4.2 4.2 4.2
3.7 3.7 3.7
3.2 3.2 3.2
2.75 2.75 2.75

0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
q [Ah] q [Ah] q [Ah]
0 th day 205 th day 437 th day 997 th day Experimental Fitting
Figure 6. Experimental OCV-q curves and their fitting functions obtained through the nonlinear least
square method for the characterization subset.
Figure 7. Coefficients of determinations R2 for the characterization subset.
Furthermore, it is worth noting that the capacity fade experienced by the battery
cells becomes more pronounced as the storage SOC level increases. In fact, for the battery
cell stored at a low SOC, the capacity fade is about 5%, while for the medium SOC, it is
about 13%, and for the high SOC, it is about 25%. This can be attributed to a significant
potential disequilibrium at the electrode/electrolyte interface resulting from a high SOC.
Therefore, secondary chemical reactions such as corrosion, electrolyte decomposition, and
SEI decomposition occur, leading to the loss of lithium inventory and active material, thus
resulting in faster capacity fade [40,41].
Figure 8 shows the behavior of the parameters a, b, c, and d of (2) as a function of the
calendar time for the three SOC levels. From this figure, it is possible to note that, similar
to the results obtained in [33] for cycle aging, the parameters a, b, and d do not present a
well-defined trend as a function of the calendar time. Therefore, following the analysis
performed in [33], we opted to consider, for each SOC level, the values of the parameters b
and d constant at their mean values obtained previously and reported in Table 2. In this
way, the function of the discharge OCV-q curve becomes linear in the parameters a and c;
thus, the linear least square regression method could be applied.
Figure 9 illustrates the comparison between the experimental OCV-q curves and the
corresponding fitting functions for the four tests chosen for the characterization of the
model and for each SOC level. Additionally, Figure 10 shows the related R2 values obtained
by maintaining parameters b and d constant. From these figures, it is possible to recognize
a good agreement between the experimental and modeled data for the low and medium
SOC tests, while the agreement for the high SOC tests is a little worse. In any case, the R2 is
greater than 0.985 for all the tests.
74
Energies 2023, 16, 4869

0.59 0.62 0.6
0.58 0.6 0.57
0.57 0.58 0.54
0.56 0.56 0.51
0.55 0.54 0.48
ƺ0.23 ƺ0.23 ƺ0.15

ƺ0.235 ƺ0.25 ƺ0.25
ƺ0.24 ƺ0.27 ƺ0.35
ƺ0.245 ƺ0.29 ƺ0.45
x10 ƺ10 x10 ƺ7 x10 ƺ4

0 0 0
ƺ1 ƺ2 ƺ2
ƺ2
ƺ3 ƺ4 ƺ4
ƺ4 ƺ6 ƺ6
2.5 2.5 2.5

2.45 2.25 2
2.4 2
2.35 1.75 1.5
2.3 1.5 1
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
taging [days] taging [days] taging [days]
Figure 8. Parameters a, b, c, and d as a function of the calendar time for the characterization subset.
Table 2. Mean values of parameters b and d.
SOC Level b [Ah−1 ] d [Ah−1 ]

Low SOC −0.2393 2.411
Medium SOC −0.2635 2.183
High SOC −0.2856 2.000

4.2 4.2 4.2
3.7 3.7 3.7
3.2 3.2 3.2
2.75 2.75 2.75

0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
q [Ah] q [Ah] q [Ah]
0 th day 205 th day 437 th day 997 th day Experimental Fitting
Figure 9. Experimental OCV-q curves and their fitting functions obtained by fixing the parameters b
and d and through the linear least square method for the characterization subset.
75
Energies 2023, 16, 4869
Figure 10. Coefficients of determination R2 obtained by fixing the parameters b and d for the
characterization subset.
In all cases, the new values of parameters a and c exhibited a well-defined behavior,
enabling them to be fitted as a function of the calendar time. For the parameter a, a linear
least square regression was performed using the following fitting function:
a = α a · t aging + β a (6)
where αa and βa are the coefficients of the fitting function of the parameter a. For the
parameter c, a nonlinear least square regression was performed instead, using the follow-
ing expression:
c = αc · tδaging
c
+ βc . (7)
where αc , βc , and δc are the coefficients of the fitting function of the parameter c. The choice
to use the expression (7) for parameter c was made to find an analytical expression that was
suitable for all three SOC levels. Finally, Figure 11 shows the behavior of parameters a and
c of (2) as a function of the calendar time for the three SOC levels, along with their related
fitting functions. Table 3 reports the coefficients of (6) and (7). In this way, an aging model
with only two parameters that depend on calendar aging was derived.
Figure 11. Parameters a and c as a function of the calendar time (with b and d fixed) and their fitting
functions for the characterization subset.
76
Energies 2023, 16, 4869
Table 3. Coefficients of the parameters a and c.
SOC Level αa [V/Days] βa [V] αc [V/Days] βc [V] δc

−6.017 × −7.362 ×
Low SOC 2.580 × 10−5 0.5580 1.1
10−14 10−11
−1.420 × −9.443 ×
Medium SOC 1.072 × 10−4 0.5258 1.3
10−12 10−10
−4.808 × −4.551 ×
High SOC 1.833 × 10−4 0.4875 4.0
10−19 10−9
4.2. Model Validation

The validation of the model was assessed by evaluating the absolute error (errOCV,abs )
and relative percentage error (errOCV,rel% ) between the estimated discharge OCV-q curves
and the corresponding experimental curves, using all twelve experimental tests at different
calendar times for each SOC level. They were evaluated as follows:

errOCV,abs = Emod (q) − Eexp (q) (8)
errOCV,abs
errOCV,rel% = · 100 (9)
Eexp (q)
where Eexp (q) and Emod (q) are, respectively, the experimental and modeled values of the
OCV-q curves. Moreover, to have an overall indicator of the goodness of fit, the R2 was
also calculated.
Figure 12 shows the maximum and mean values of the relative percentage error of the
OCV-q curves as a function of calendar aging. From this figure, it is possible to note that
the highest maximum relative percentage errors are about 5%, 7%, and 6.5% for the low,
medium, and high SOC tests, respectively. Instead, the mean relative percentage errors for
the low, medium, and high SOC tests are lower than 0.45%, 0.65%, and 0.9%, respectively,
across all calendar aging levels. On the other hand, the relative percentage error can have
different weights for different states of discharge because the OCV curve changes as a
function of q. For this reason, Figure 13 shows the maximum and mean values of the
absolute error. The maximum value of the latter is about 165 mV for the low SOC tests and
220 mV for the medium and high SOC tests, while the mean values of the absolute error
are lower than 15 mV, 22 mV, 30 mV for the low, medium, and high SOC tests, respectively,
across all aging levels. Finally, Figure 14 reports the experimental OCV-q curves compared
with the ones obtained through the proposed model for all tests and SOC levels, while
Figure 15 shows the related R2 . The latter are larger than 0.9763, 0.9542, and 0.9490 for the
low, medium, and high SOC tests, respectively.
The knowledge of the variation law of the OCV-q curve as a function of calendar aging
can be employed to estimate the actual battery capacity. This, in turn, allows the correction
of the SOC estimation and updating of the SOH in terms of capacity fade. In particular, the
actual battery capacity, Ca , can be calculated by equating the (2) to the minimum cut-off
voltage Emin of the battery. This is due to the fact that when the open circuit voltage E of
the battery equals its minimum cut-off voltage, the absolute state of discharge q is equal
to the actual battery capacity. The percentage relative error of the capacity was evaluated
as follows:
qmod ( Emin ) − qexp ( Emin )
errcapacity,rel% = · 100 (10)
qexp ( Emin )
where qmod (Emin ) and qexp (Emin ) are the estimated and experimental actual battery capaci-
ties, respectively.
77
Energies 2023, 16, 4869
Figure 12. Mean and maximum percentage relative error as a function of the calendar time.
Figure 13. Mean and maximum absolute error as a function of the calendar time.
Figure 14. Experimental OCV-q curves and modeled ones for all the experimental tests.
78
Energies 2023, 16, 4869
Figure 15. Coefficients of determination R2 for all the experimental tests.
Theoretically, the minimum cut-off voltage for the battery used in this work would
be 2.75 V. However, due to the elimination of the voltage drop over the internal resistance,
the actual cut-off voltage was higher and depended on the amplitude of that voltage drop
itself. Nonetheless, since the last experimental point of each discharge OCV-q curve is
close to the minimum cut-off voltage, this experimental point was considered as Emin , and
the related value of the actual battery capacity, Ca , was compared with the modeled one.
Figure 16 shows the capacity percentage errors for the three SOC levels. From this figure,
it is possible to recognize that the capacity error is lower than 1.5% for all the aging and
SOC levels.
Figure 16. Capacity estimation error as a function of the calendar time.
Finally, once the actual capacity of the battery is obtained, it becomes possible to correct
the SOC estimation according to (4) and update the value of SOH in terms of capacity fade
as follows:
Ca
SOH = 1 − · 100 (11)
Ci
where Ci is the initial battery capacity. The results highlight the significant impact of storing
the battery at a high SOC, as it degrades much faster. Specifically, the battery stored at a
high SOC reached the end of its life in less than three years, whereas the one stored at a low
SOC experienced minimal capacity fade.
5. Conclusions
In this work, a simple analytical function composed of two exponential terms and a
constant one with five parameters was used to model the behavior of the discharge OCV-q
curves of three LiCoO2 batteries of the same batch stored at three different levels of SOC
(low, medium, and high levels) for different calendar aging levels. Twelve discharge OCV-q
experimental curves were performed over about 1000 days at different calendar times.
The proposed model was characterized using a subset of four discharge OCV-q experi-
mental curves at the 0th, 205th, 437th, and 997th day. Firstly, the proposed analytical model,
being nonlinear in its parameters, was fitted using the nonlinear regression least square
method. Moreover, one of the parameters, namely f, was expressed as a function of the
79
Energies 2023, 16, 4869
other two parameters, a and c, with the constraint that all the OCV-q curves started from
the maximum cut-off voltage of 4.2 V. This reduced the number of parameters to be found
to four. The fitting procedure yielded modeled data for the three SOC levels that showed
good agreement with the experimental ones. This agreement was further confirmed by the
high value of the R2 for all the tests. The four parameters a, b, c, and d were reported as a
function of the calendar time. On the other hand, the parameters a, b, and d did not exhibit
a well-defined trend. Following the analysis performed in [33], we considered the value of
the parameters b and d fixed at their mean values obtained previously for all the SOC levels.
Therefore, the analytical model became linear in its parameters, and the fitting procedure
was performed again using the linear least square method. In this way, the parameters a
and c could be fitted using simple analytical expressions for the three SOC levels, and the
OCV-q model considering the calendar aging was obtained.
Afterward, the model was validated through all twelve experimental OCV-q curves,
and the absolute error, percentage relative error, and coefficients of determination R2 were
evaluated as indicators of the goodness of fit. Through the analysis of the results, it is
possible to recognize that the mean relative percentage errors of the low, medium, and high
SOC tests are, respectively, lower than 0.45%, 0.65%, and 0.9% for all the calendar aging
levels. Additionally, the R2 values are greater than 0.9763, 0.9542, and 0.9490 for the low,
medium, and high SOC tests, respectively.
Finally, the proposed model can also be used to estimate the battery capacity for
all storage SOC and aging levels. The estimated battery capacity values were compared
with the corresponding experimental ones, and the capacity relative percentage error was
calculated. This error is lower than 1.5% for all the tests. The battery capacity obtained
can then be used to correct the SOC estimation and evaluate the capacity fade. Based on
the results, it is possible to confirm that the worst storage condition for these kinds of
LiBs occurs at high SOCs. Indeed, the battery stored at a high SOC reached its end of life
before three years. On the other hand, the batteries stored at lower SOCs are still usable
considering the same time span. This can be due to a notable imbalance that arises between
the electrode and electrolyte interface caused by a high SOC. As a consequence, secondary
chemical reactions such as corrosion, electrolyte decomposition, and SEI decomposition
take place, resulting in the depletion of lithium inventory and active material.
In light of the above, it is possible to claim that the proposed calendar aging model is
simple yet quite accurate in modeling the OCV-q curve. Moreover, it can be valuable for
correcting SOC estimations and evaluating the actual battery capacity as an indicator of
the SOH.
Author Contributions: Conceptualization, S.B. and S.C.; methodology, S.B. and S.C.; validation,
S.B. and S.C.; writing—original draft preparation, S.B. and S.C.; writing—review and editing, S.B.
and S.C.; supervision, L.P. and L.C. All authors have read and agreed to the published version of
the manuscript.
Data Availability Statement: Data sharing is not applicable.
References
1. Divakaran, M.A.; Minakshi, M.; Bahri, P.A.; Paul, S.; Kumari, P.; Divakaran, A.M.; Manjunatha, K.N. Rational Design on Materials
for Developing next Generation Lithium-Ion Secondary Battery. Prog. Solid State Chem. 2021, 62, 100298. [CrossRef]
2. Nunotani, K.; Yoshida, F.; Kamiya, Y.; Daisho, Y.; Abe, K.; Kono, M.; Matsuo, H. Development and Performance Evaluation of
Lithium Iron Phosphate Battery with Superior Rapid Charging Performance—Second Report: Evaluation of Battery Capacity
Loss Characteristics. In Proceedings of the 2011 IEEE Vehicle Power and Propulsion Conference, Chicago, IL, USA, 6–9 September
2011; pp. 1–4.
3. Barcellona, S.; Brenna, M.; Foiadelli, F.; Longo, M.; Piegari, L. Analysis of Ageing Effect on Li-Polymer Batteries. Sci. World J. 2015,
2015, 979321. [CrossRef] [PubMed]
80
Energies 2023, 16, 4869
4. Barcellona, S.; Brenna, M.; Foiadelli, F.; Longo, M.; Piegari, L. Battery Lifetime for Different Driving Cycles of EVs. In Proceedings
of the 2015 IEEE 1st International Forum on Research and Technologies for Society and Industry Leveraging a better Tomorrow
(RTSI), Torino, Italy, 16–18 September 2015; pp. 446–450.
5. Mawonou, K.S.R.; Eddahech, A.; Dumur, D.; Beauvois, D.; Godoy, E. Improved State of Charge Estimation for Li-Ion Batteries
Using Fractional Order Extended Kalman Filter. J. Power Sources 2019, 435, 226710. [CrossRef]
6. Li, Y.; Wang, C.; Gong, J. A Multi-Model Probability SOC Fusion Estimation Approach Using an Improved Adaptive Unscented
Kalman Filter Technique. Energy 2017, 141, 1402–1415. [CrossRef]
7. Ning, B.; Xu, J.; Cao, B.; Wang, B.; Xu, G. A Sliding Mode Observer SOC Estimation Method Based on Parameter Adaptive Battery
Model. Energy Procedia 2016, 88, 619–626. [CrossRef]
8. Ouyang, Q.; Chen, J.; Wang, F.; Su, H. Nonlinear Observer Design for the State of Charge of Lithium-Ion Batteries. IFAC Proc. Vol.
2014, 47, 2794–2799. [CrossRef]
9. Gan, L.; Yang, F.; Shi, Y.F.; He, H.L. Lithium-Ion Battery State of Function Estimation Based on Fuzzy Logic Algorithm with
Associated Variables. IOP Conf. Ser. Earth Environ. Sci. 2017, 94, 012133. [CrossRef]
10. Jiao, M.; Wang, D.; Qiu, J. A GRU-RNN Based Momentum Optimized Algorithm for SOC Estimation. J. Power Sources 2020, 459,
228051. [CrossRef]
11. Zhou, W.; Zheng, Y.; Pan, Z.; Lu, Q. Review on the Battery Model and SOC Estimation Method. Processes 2021, 9, 1685. [CrossRef]
12. Ren, X.; Liu, S.; Yu, X.; Dong, X. A Method for State-of-Charge Estimation of Lithium-Ion Batteries Based on PSO-LSTM. Energy
2021, 234, 121236. [CrossRef]
13. Li, N.; Zhang, Y.; He, F.; Zhu, L.; Zhang, X.; Ma, Y.; Wang, S. Review of Lithium-Ion Battery State of Charge Estimation. Glob.
Energy Interconnect. 2021, 4, 619–630. [CrossRef]
14. Xiong, R.; Yu, Q.; Wang, L.Y.; Lin, C. A Novel Method to Obtain the Open Circuit Voltage for the State of Charge of Lithium Ion
Batteries in Electric Vehicles by Using H Infinity Filter. Appl. Energy 2017, 207, 346–353. [CrossRef]
15. Xiong, R.; He, H.; Guo, H.; Ding, Y. Modeling for Lithium-Ion Battery Used in Electric Vehicles. Procedia Eng. 2011, 15, 2869–2874.
[CrossRef]
16. Sundaresan, S.; Devabattini, B.; Kumar, P.; Pattipati, K.; Balasingam, B. Tabular Open Circuit Voltage Modelling of Li-Ion Batteries
for Robust SOC Estimation. Energies 2022, 15, 9142. [CrossRef]
17. Pillai, P.; Sundaresan, S.; Kumar, P.; Pattipati, K.R.; Balasingam, B. Open-Circuit Voltage Models for Battery Management Systems:
A Review. Energies 2022, 15, 6803. [CrossRef]
18. He, H.; Xiong, R.; Zhang, X.; Sun, F.; Fan, J. State-of-Charge Estimation of the Lithium-Ion Battery Using an Adaptive Extended
Kalman Filter Based on an Improved Thevenin Model. IEEE Trans. Veh. Technol. 2011, 60, 1461–1469. [CrossRef]
19. Sun, D.; Yu, X.; Wang, C.; Zhang, C.; Huang, R.; Zhou, Q.; Amietszajew, T.; Bhagat, R. State of Charge Estimation for Lithium-Ion
Battery Based on an Intelligent Adaptive Extended Kalman Filter with Improved Noise Estimator. Energy 2021, 214, 119025.
[CrossRef]
20. Zhao, S.; Duncan, S.R.; Howey, D.A. Observability Analysis and State Estimation of Lithium-Ion Batteries in the Presence of
Sensor Biases. IEEE Trans. Control. Syst. Technol. 2017, 25, 326–333. [CrossRef]
21. Xiong, R.; He, H.; Sun, F.; Zhao, K. Online Estimation of Peak Power Capability of Li-Ion Batteries in Electric Vehicles by a
Hardware-in-Loop Approach. Energies 2012, 5, 1455–1469. [CrossRef]
22. Ta, K.P.; Newman, J. Proton Intercalation Hysteresis in Charging and Discharging Nickel Hydroxide Electrodes. J. Electrochem.
Soc. 1999, 146, 2769–2779. [CrossRef]
23. Baccouche, I.; Jemmali, S.; Manai, B.; Omar, N.; Amara, N. Improved OCV Model of a Li-Ion NMC Battery for Online SOC
Estimation Using the Extended Kalman Filter. Energies 2017, 10, 764. [CrossRef]
24. Zhang, C.; Jiang, J.; Zhang, L.; Liu, S.; Wang, L.; Loh, P. A Generalized SOC-OCV Model for Lithium-Ion Batteries and the SOC
Estimation for LNMCO Battery. Energies 2016, 9, 900. [CrossRef]
25. Weng, C.; Sun, J.; Peng, H. A Unified Open-Circuit-Voltage Model of Lithium-Ion Batteries for State-of-Charge Estimation and
State-of-Health Monitoring. J. Power Sources 2014, 258, 228–237. [CrossRef]
26. Zhang, R.; Xia, B.; Li, B.; Cao, L.; Lai, Y.; Zheng, W.; Wang, H.; Wang, W.; Wang, M. A Study on the Open Circuit Voltage and
State of Charge Characterization of High Capacity Lithium-Ion Battery Under Different Temperature. Energies 2018, 11, 2408.
[CrossRef]
27. Dong, G.; Wei, J.; Zhang, C.; Chen, Z. Online State of Charge Estimation and Open Circuit Voltage Hysteresis Modeling of
LiFePO4 Battery Using Invariant Imbedding Method. Appl. Energy 2016, 162, 163–171. [CrossRef]
28. Hu, Y.; Yurkovich, S.; Guezennec, Y.; Yurkovich, B.J. Electro-Thermal Battery Model Identification for Automotive Applications. J.
29. Wang, L.; Lu, D.; Liu, Q.; Liu, L.; Zhao, X. State of Charge Estimation for LiFePO4 Battery via Dual Extended Kalman Filter and
Charging Voltage Curve. Electrochim. Acta 2019, 296, 1009–1017. [CrossRef]
30. Tong, S.; Klein, M.P.; Park, J.W. On-Line Optimization of Battery Open Circuit Voltage for Improved State-of-Charge and
State-of-Health Estimation. J. Power Sources 2015, 293, 416–428. [CrossRef]
31. Farmann, A.; Sauer, D.U. A Study on the Dependency of the Open-Circuit Voltage on Temperature and Actual Aging State of
Lithium-Ion Batteries. J. Power Sources 2017, 347, 1–13. [CrossRef]
81
Energies 2023, 16, 4869
32. Tian, J.; Xiong, R.; Shen, W.; Sun, F. Electrode Ageing Estimation and Open Circuit Voltage Reconstruction for Lithium Ion
Batteries. Energy Storage Mater. 2021, 37, 283–295. [CrossRef]
33. Barcellona, S.; Codecasa, L.; Colnago, S.; Piegari, L. Cycle Aging Effect on the Open Circuit Voltage of Lithium-Ion Battery. In
Proceedings of the International Conference on Electrical Systems for Aircraft, Railway, Ship Propulsion and Road Vehicles &
International Transportation Electrification Conference (ESARS-ITEC), Venice, Italy, 28–31 March 2023.
34. Kim, Y.H.; Ha, H.D. Design of Interface Circuits with Electrical Battery Models. IEEE Trans. Ind. Electron. 1997, 44, 81–86.
[CrossRef]
35. Dürr, M.; Cruden, A.; Gair, S.; McDonald, J.R. Dynamic Model of a Lead Acid Battery for Use in a Domestic Fuel Cell System. J.
36. Chan, H.L. A New Battery Model for Use with Battery Energy Storage Systems and Electric Vehicles Power Systems. In
Proceedings of the 2000 IEEE Power Engineering Society Winter Meeting. Conference Proceedings (Cat. No.00CH37077),
Singapore, 23–27 January 2000; Volume 1, pp. 470–475.
37. Randles, J.E.B. Kinetics of Rapid Electrode Reactions. Discuss. Faraday Soc. 1947, 1, 11. [CrossRef]
38. Zhang, S.S.; Xu, K.; Jow, T.R. Electrochemical Impedance Study on the Low Temperature of Li-Ion Batteries. Electrochim. Acta
2004, 49, 1057–1061. [CrossRef]
39. Buller, S.; Thele, M.; DeDoncker, R.W.A.A.; Karden, E. Impedance-Based Simulation Models of Supercapacitors and Li-Ion
Batteries for Power Electronic Applications. IEEE Trans. Ind. Appl. 2005, 41, 742–747. [CrossRef]
40. Barré, A.; Deguilhem, B.; Grolleau, S.; Gérard, M.; Suard, F.; Riu, D. A Review on Lithium-Ion Battery Ageing Mechanisms and
Estimations for Automotive Applications. J. Power Sources 2013, 241, 680–689. [CrossRef]
41. Birkl, C.R.; Roberts, M.R.; McTurk, E.; Bruce, P.G.; Howey, D.A. Degradation Diagnostics for Lithium Ion Cells. J. Power Sources
2017, 341, 373–386. [CrossRef]
82
energies
Article
Evaluation of the Power Generation Impact for the Mobility of
Battery Electric Vehicles
Javier Rey and Lázaro V. Cremades *
Department of Project and Construction Engineering, Universitat Politècnica de Catalunya,

08028 Barcelona, Spain
* Correspondence: lazaro.cremades@upc.edu
Abstract: European institutions have decided to ban the sale of Internal Combustion Vehicles (ICEVs)
in the EU from 2035. This opens a possible scenario in which, in the not-too-distant future, all vehicles
circulating in Europe are likely to be Battery Electric Vehicles (BEVs). The Spanish vehicle fleet
is one of the oldest and has the lowest percentage of BEVs in Europe. The aim of this study is to
evaluate the hypothetical scenario in which the current mobility of ICEVs is transformed into BEVs,
in the geographical area of the province of Barcelona and in Spain in general. The daily electricity
consumption, the required installation capacity of wind and solar photovoltaic energies, and the
potential reduction of NOx and particulate matter (PM) emissions are estimated. The daily emission
reduction would be about 314 tons of NOx and 17 tons of PM in Spain. However, the estimated
investment required in Spain to generate the additional electricity from renewable sources would
be enormous (over EUR 25.4 billion), representing, for example, 5.5% of the total national budget
in 2022.
Keywords: ICEVs; BEVs; mobility; electricity generation; wind energy; solar photovoltaic energy;
renewable energy sources; vehicle fleet; pollutant emissions
1. Introduction
Citation: Rey, J.; Cremades, L.V. The 2015 Paris summit, COP21, ended with an agreement between all industrialized
Evaluation of the Power Generation nations [1]. This agreement calls on the European Union (EU) to try to avoid a global
Impact for the Mobility of Battery temperature increase of more than 2 ◦ C compared to pre-industrial levels. To achieve
Electric Vehicles. Energies 2023, 16, these goals, the European institutions have proposed to achieve climate neutrality, i.e., zero
5006. https://doi.org/10.3390/ carbon dioxide (CO2 ) emissions, so that no more greenhouse gases (GHG) are produced by
en16135006
human activity.
Academic Editor: Simone In order to contribute to these objectives, it is necessary to regulate vehicle transport.
Barcellona Member States have committed themselves to taking appropriate measures to achieve this
objective, which is why the successive environmental permits for vehicle transport in the
Received: 23 May 2023
EU are becoming increasingly restrictive for internal combustion engine vehicles (ICEVs).
Revised: 24 June 2023
However, the ultimate goal is to ban ICEVs in the medium and long term. Currently, in the
Accepted: 26 June 2023
absence of a final agreement with all Member States, European institutions have decided to
Published: 28 June 2023
ban the sale of internal combustion vehicles in the EU from 2035 [2]. Although this only
affects the sale of new vehicles, major European cities are also restricting the circulation of
ICEVs in low emission zones [3].
Copyright: © 2023 by the authors. It is therefore foreseeable that the current vehicle fleet will be transformed at high
Licensee MDPI, Basel, Switzerland. speed in the coming years, and it is therefore important to study and analyze the follow-
This article is an open access article ing questions: Is it possible to transform the current vehicle fleet into a battery electric
distributed under the terms and vehicle fleet? And if it is possible, how can the transition be made and what are the
conditions of the Creative Commons socio-economic costs?
Attribution (CC BY) license (https:// Virtually since the popularization of the automobile at the beginning of the 20th cen-
creativecommons.org/licenses/by/ tury, all ground transportation has been powered by internal combustion engines. However,
4.0/).

83
Energies 2023, 16, 5006
this form of mobility generates negative externalities that were not considered until the
last few decades. ICEVs emit gaseous and particulate pollutants from the combustion
process [4].
Successive environmental regulations have restricted the maximum value of each of
the emissions or introduced a new restriction that was not foreseen in the
previous regulations.
The EU has adopted and applied a series of environmental regulations, known as
EURO regulations, which affect the homologation of vehicles, in an attempt to reduce
the pollutant emissions of ICEVs. The EURO 6d regulation is currently in effect [5–7].
Each revision of these regulations has further restricted the maximum level of pollutant
emissions and introduced new, more demanding type-approval tests based on real-world
driving, such as the WLTP procedure or the Real Driving Emissions tests [5].
In order to comply with these regulations, vehicle manufacturers have developed
technological improvements to meet the EURO regulations, such as particulate filters or
Selective Catalytic Reduction (SCR) catalysts using AdBlue [8]. Now, with the foreseeable
introduction of the next update of the EURO regulation, EURO 7 [5], European manufac-
turers are declaring that it will be impossible to comply with the new restrictions and make
the necessary investments to adapt the existing technology while keeping it profitable, even
more so when the sale of internal combustion vehicles is finally banned in the EU from
2035. It is therefore foreseeable that the introduction of electric vehicles will accelerate in
the coming years.
There are several types of electric vehicle (Figure 1):
• Battery electric vehicle (BEV): They store chemical energy in a battery, which pro-
vides the electrical energy for consumption by the electric motor, which converts the
electrical energy into mechanical energy. It is currently the main alternative to the
conventional ICEV and is the main pillar of the new mobility. This is due to the fact
that the use of the battery electric vehicles is considered to produce no polluting gases
or particles, although this depends on the origin of the electrical energy and the entire
life cycle of the battery electric vehicle.
• Hybrid electric vehicle (HEV): They combine an internal combustion engine with
an electric motor. The internal combustion engine can operate in two ways. The
combustion engine provides mechanical power directly to the transmission or to an
electric generator. The generator feeds a battery, which feeds an electric motor. The
electric motor provides mechanical power to the transmission. HEVs have seen signif-
icant development and uptake in recent years due to their lower fuel consumption,
combined with a competitive price compared to the ICEV.
• Plug-in hybrid electric vehicle (PHEV): They are essentially HEVs that allow electricity
to be supplied to the battery directly from the grid. The battery is smaller than that
of a BEV. This makes it possible to provide driving modes using only the electric
motor supplied by the battery, and therefore driving modes that are a priori free of
polluting emissions. The main problem with PHEVs is that there is no guarantee that
the user will recharge the battery and not constantly use the vehicle powered by the
combustion engine, which is effectively equivalent to an ICEV.
• Fuel cell electric vehicle (FCEV): They use the electrical energy generated in a fuel cell,
which uses the chemical energy stored in a pressurized tank, mainly from hydrogen.
This electrical energy powers a smaller battery than in the BEV, to ultimately provide
mechanical power to the drivetrain through the electric motor. FCEVs can become
another alternative for zero-emission mobility and coexist with the BEV in the medium
and long term. Currently, the purchase price of existing FCEVs, such as the Toyota
Mirai [9] or the Hyundai NEXO [10], is high and the technology to produce green
hydrogen from clean electricity is not yet developed enough to compete with ICEVs
and BEVs. Finally, the refueling infrastructure is very poorly developed and free
mobility with an FCEV is practically impossible [11].
84
Energies 2023, 16, 5006
Figure 1. Types of vehicles (internal combustion and electric) and their main components. Modified
from [12].
For the end user, the main difference between an ICEV and a BEV is the provision and
storage of the energy needed to operate the vehicle. This difference is currently the main
advantage of the ICEV, and therefore the main disadvantage of the BEV. In order to drive
the BEV, the vehicle’s battery needs to be charged. There are currently four possible modes
of charging from the grid [13,14]:
• Mode 1. The vehicle is directly connected to the conventional grid without the need
for any additional special equipment or systems. This mode is very practical for small
vehicles such as bicycles or mopeds, but is not recommended for commercial vehicles.
• Mode 2. This mode provides a slow charge. This type of charging is single-phase
with a voltage of 230 V and a maximum power of 3.7 kW. The BEV is connected to the
mains via the appropriate plug/adapter to ensure the safety of the charging process.
• Mode 3. This mode provides semi-fast charging. The electric vehicle is connected to
the alternating current grid via a dedicated BEV charging outlet. The most commonly
used plug for this type of charging is Type 2 [15]. This mode allows single-phase or
three-phase charging. Single-phase connections charge at 7.4 kW and three-phase
connections charge at 22 kW.
• Mode 4. Its charging power is equal to or greater than 50 kW, allowing “super fast”
and “ultra fast” charging. The latter is not recommended for daily charging, as it can
damage the battery if used regularly. It is specifically designed for outdoor public
use stations and could be similar to a gas station, where the vehicle can be recharged
during long trips or in specific situations where passengers are short of time.
Mode 4 charging uses direct current (DC), as opposed to the previous modes that
use alternating current (AC). The most commonly used plug for this type of charging is
currently the CCS Combo, which combines a Type 2 plug with two extra terminals to allow
DC power to pass through [15].
The main objective of this paper is to study the feasibility of transforming the current
mobility of the vehicle fleet in Spain based on the use of fossil fuels (gasoline and diesel)
in internal combustion engines (ICEVs) to a mobility based on the use of electricity from
85
Energies 2023, 16, 5006
renewable sources for battery electric vehicles (BEVs). This feasibility is analyzed from a
techno-economic and environmental point of view.
There are several papers in the literature that have examined the impact on the
reduction of GHG and/or non-GHG emissions by substituting ICEVs with BEVs in specific
regions, such as [16,17]. There are also those that performed life cycle analyses comparing
ICEVs with BEVs, such as [18–20]. But none of them explored the feasibility of transforming
the actual mobility of ICEVs into BEVs and its impact on electricity generation in Spain, as
presented in this paper.
2. Method
In order to study the feasibility of this change, an in-depth analysis of daily mobility
at different geographical levels and of the type of vehicle used for each trip was carried out,
making it possible to estimate the amount of additional electrical energy that would need
to be generated to make a fleet of vehicles made up entirely of BEVs feasible. However,
to achieve a true environmental transition in vehicle transportation, the electricity needed
to power a fleet of BEVs must come from renewable sources. This would require the
installation of new solar and wind power plants, which would need to be located and
installed somewhere in the territory, along the associated transportation network.
The following steps were taken to conduct the study (Figure 2):
(1) Collect available information on charging stations, vehicle fleet, actual mobility, certi-
fied electricity consumption of BEVs, electricity generation, and pollutant emissions
of ICEVs according to vehicle age.
(2) Estimate the daily electricity consumption that would be required for the mobility of
the vehicles if they were all BEVs.
(3) Determine the hourly availability for charging these vehicles based on the daily
mobility data and energy source. Since photovoltaic (PV) solar energy can only be
generated during daylight hours, the percentage of daily use of this source can be
defined. The rest of the energy is assumed to come from wind.
(4) Estimate the required installation capacity of wind turbines and PV panels to provide
the energy needed to charge the BEVs. Also estimate the investment required.
(5) Estimate the reduction in pollutant emissions from replacing ICEVs with BEVs.
Figure 2. Schematic of the methodological steps followed in this study.
The fleet information was used to estimate the emission reduction of the ICEVs that
would be replaced by BEVs, based on their age. However, the number of ICEVs to be
replaced by BEVs was determined by the actual mobility. Vehicle mobility trips are assumed
86
Energies 2023, 16, 5006
to correspond to the proportion of such vehicles in the fleet, both in terms of category
and age.
The study, analysis and proposal to be carried out in this work mainly focused on the
province of Barcelona (PB) (Spain), as there is a survey of actual mobility that is regularly
carried out for this region [21]. However, the composition and age of the vehicle fleet have
been obtained from data for Spain as a whole.
The results obtained for PB can be extrapolated to the whole of Spain, assuming that
the mobility pattern is similar.
The PB covers mobility that takes place in the region bordering the city of Barcelona.
At the end of 2022, the PB had a population of 5,727,615 inhabitants [22] in an area of
7726 km2 , which corresponds to a population density of 741.2 inhabitants/km2 , of which
1,636,193 inhabitants correspond to the city of Barcelona, with a population density of
15,992.2 inhabitants/km2 , which acts as a dynamic center of socio-economic activity.
In this paper, clean electrical energy is defined as energy produced by renewable
energy sources. Renewable energy is defined as energy that is consumed more slowly
than it is produced. Therefore, in this study, only energy from solar and wind sources is
considered to be clean electricity.
Available data from actual wind and PV solar farms in Spain were extrapolated to
estimate the area and investment costs required to meet the estimated energy demand.
The optimal use of a BEV implies a usage model with recharging at a dedicated
charging point and at a recharging rate that does not imply a sudden degradation of
the battery system. For this reason, in the hypothetical case study of a ground transport
BEV, it was assumed that the vehicle would not be charged in the usual way at a fast
charging station.
3. Basic Data
3.1. Charging Points
In order to increase the use of BEVs, it is necessary to develop a network of charging
stations with public access, in addition to the private charging points that each user can
install at home.
Currently, the two most powerful public charging stations in Spain are located in
the Basque Country and have been installed by Repsol [23]. The fact that a fuel company
is installing charging stations shows a clear commitment to electrification. These four-
terminal installations have a capacity of 400 kW, which corresponds to a charging time of
between 5 and 10 min, similar to the refueling time of an ICEV.
However, according to the Spanish Association of Automobile and Truck Manufactur-
ers (ANFAC) [24], the public access charging network in Spain is growing very slowly and
is poorly distributed throughout the country. At the beginning of 2021, there were 11,517
public access charging points in Spain, 83% of which had power of less than 22 kW, which
does not allow fast charging. While in the EU as a whole the average number of charging
points per million inhabitants is 573, in Spain it is only 245. This is in line with the current
low presence of BEVs in Spain.
3.2. Vehicle Fleet Composition

From the studies and statistics published by the Dirección General de Tráfico (DGT in
Spanish) [25], it is possible to know the composition of the current vehicle fleet in each of
the Spanish municipalities and, therefore, in the country as a whole.
On 31 December 2022, the total number of main categories of motor vehicles in Spain
was 34,304,426, divided as follows:
• Passenger cars: 25,222,554 (73.5%);
• Motorcycles: 4,006,804 (11.7%);
• Vans: 2,617,145 (7.6%);
• Trucks: 2,457,923 (7.2%).
87
Energies 2023, 16, 5006
The DGT statistics also allowed us to classify each vehicle type according to its age,
from 2022 backwards (Table 1), which allowed us to know its pollution potential according
to the existing environmental approval in place at the time of its registration.
Table 1. Age of the Spanish vehicle fleet by vehicle type in 2022 [26].
%
Vehicle Age Gasoline % 1 Diesel %1 Others % 1 Total 2
<5 years 3,242,568 62.7 1,739,699 33.6 190,474 3.7 5,172,741 20.5
5–9 years 1,813,269 38.6 2,859,712 60.8 26,896 0.6 4,699,877 18.6
Passenger
10–14 years 1,150,761 29.1 2,802,644 70.8 3542 0.1 3,956,947 15.7
cars
15–19 years 1,854,694 32.0 3,944,656 68.0 2511 0.0 5,801,861 23.0
>19 years 3,689,507 66.0 1,899,161 34.0 2460 0.0 5,591,128 22.2
<5 years 817,978 95.9 1398 0.2 33,988 4.0 853,364 21.3
5–9 years 612,144 99.0 1853 0.3 4625 0.7 618,622 15.4
Motorcycles 10–14 years 644,651 99.5 922 0.1 2478 0.4 648,051 16.2
15–19 years 810,876 99.8 1186 0.1 118 0.0 812,180 20.3
>19 years 1,073,538 99.9 645 0.1 404 0.0 1,074,587 26.8
<5 years 49,376 9.2 470,241 87.1 19,986 3.7 539,603 20.6
5–9 years 13,425 3.4 373,018 95.4 4400 1.1 390,843 14.9
Vans 10–14 years 13,616 5.2 249,505 94.5 795 0.3 263,916 10.1
15–19 years 36,964 7.9 430,543 91.9 778 0.2 468,285 17.9
>19 years 302,203 31.7 651,928 68.3 367 0.0 954,498 36.5
<5 years 11,108 3.7 275,904 92.8 10,405 3.5 297,417 12.1
5–9 years 5511 2.1 252,210 97.1 2106 0.8 259,827 10.6
Trucks 10–14 years 3995 1.3 293,101 98.2 1331 0.4 298,427 12.1
15–19 years 18,085 2.3 777,168 97.7 402 0.1 795,655 32.4
>19 years 38,597 4.8 767,888 95.2 112 0.0 806,597 32.8
Total 16,202,866 47.2 17,793,382 51.9 308,178 0.9 34,304,426
1 Percentage of each fuel type in its age group; 2 percentage contribution of each age for the type of vehicle
Table 1 shows how the share of diesel passenger cars has fallen significantly in recent
years, from 60 to 70% of the fleet to just 34% for cars less than 5 years old. This is largely
due to the reputational crisis suffered by the diesel engine as a result of the so-called
“dieselgate” in September 2015. It should also be noted that the sales of alternative vehicles
(“Others” column in Table 1) compared to conventional vehicles have started to increase in
recent years, but still represent a very small number. Furthermore, the age of the fleet is
remarkable, as the number of vehicles 10 years old or more accounts for more than 60% in
all categories.
Within PB, the total fleet of main vehicles in 2022 was 3,560,977, as follows [25]:
• Passenger cars: 2,417,620 (67.9%);
• Motorcycles: 696,678 (19.6%);
• Vans: 241,125 (6.8%);
• Trucks: 205,554 (5.8%).
Age and energy type distribution can be assumed to be similar to that of the
Spanish fleet.
3.3. Mobility in the Barcelona Province

According to the latest published survey on mobility in the Barcelona metropolitan
area in 2021, there were 16,909,491 trips per working day in PB, which corresponds to an
average of 3.5 trips per inhabitant per day [21,27]. These trips were made on foot (46.5%)
or by bicycle/scooter (2.5%), public transport (14%), or private transport (37%), the latter
including cars (86%), motorcycles (10.6%), and vans/trucks (3.4%). The total number of
trips in 2021 was lower than it was in 2019 (19,259,471), in the pre-pandemic COVID-19
situation, but the distribution of trips by mode was similar [21].
88
Energies 2023, 16, 5006
On the other hand, the average distance of each trip by private vehicle in PB was as
follows [27]: 8.9 km by car; 5.8 km by motorcycle, and 15.2 km by van/truck.
Multiplying each trip by the average distance per trip gives the total distance traveled
per working day by each type of vehicle (Table 2).
Table 2. Total daily distance traveled by type of vehicle in Barcelona province.
Private Transport Trips km/Trip Distance (km)

Car 5,370,628 8.9 47,798,589
Motorcycle 664,715 5.8 3,855,347
Van/Truck 214,745 15.2 3,264,124
Total 6,250,088 54,918,060
The total distance traveled by private vehicles was 54,918,060 km, which corresponds
to a total of 9.6 km traveled per inhabitant per working day in a private vehicle.
Figure 3 shows the hourly distribution of private transport mobility during a weekday
in PB. It follows a pattern of occupational mobility (work and/or study), as the largest
number of trips was concentrated between 7 and 9 a.m. and 5 and 7 p.m., coinciding with
the departure and return home.
Figure 3. Private transportation trips in PB during the course of a workday. Based on [27].
3.4. Electric Vehicle Power Consumption

In order to know the energy consumption of a fleet made up entirely of electric vehicles,
the homologated consumption of models representative of the type of transport can be
used as a basis for calculation. From the lowest to the highest electric energy consumption
per kilometre traveled, the following models were selected:
• Motorcycles: Silence S01 [28]. This electric motorcycle is the number 1 seller on the
Spanish market, with sales of 1433 units in 2022. It has a range of 133 km with a
battery of 5.6 kWh of stored energy, giving an energy consumption per 100 km of
4.2 kWh/100 km.
• Passenger cars: Tesla Model 3 [29]. This car is the number 1 in the Spanish electric ve-
hicle market, with sales of 2677 units in 2022. It has an approved energy consumption
of 16.2 kWh/100 km.
• Vans: Peugeot Expert [30]. It is a mid-size van and therefore covers the entire range
of this type of vehicle. In 2022, 722 units were sold on the Spanish market. It has an
energy consumption of 20.5 kWh/100 km, slightly higher than that of a car.
• Trucks: DAF LF Electric [31]. The disadvantage of the battery electric vehicle mar-
ket for trucks is the significant energy consumption required to move such a large
89
Energies 2023, 16, 5006
and heavy vehicle. Therefore, to achieve a homologated range of 280 km, a battery
with 254 kWh of stored energy is required, resulting in an energy consumption of
110.2 kWh/100 km, more than five times higher than that of the reference van.
To disaggregate the van and truck mobility data in Table 2, we have assumed that 55%
are vans and 45% are trucks. This proportion is about the same proportion as their share in
the PB fleet (see Section 3.2).
3.5. Electric Power Generation

In 2022, a total of 276,316 GWh of electricity was produced in Spain, corresponding
to an average daily production of 757 GWh [32]. Up to 44% of this energy came from
renewable sources, i.e., around 330 GWh per day, mainly from wind, photovoltaics, and
hydropower (Table 3).
Table 3. Electricity production by technology in Spain by 2022 [32].
Energy Generated
Generation Type Source Category %
(GWh)
Combined cycle Pollutant 60,652 24.66
Wind Renewable 59,805 22.14
Nuclear No emissions * 55,984 20.26
Photovoltaic Renewable 27,283 10.08
Hydropower Renewable 17,860 6.46
Co-generation Pollutant 17,732 6.43
Coal Pollutant 7687 2.81
Other renewable
Renewable 4646 1.69
sources
Solar thermal Renewable 4123 1.49
Turbine pumping Renewable 3776 1.37
Diesel engines Pollutant 2548 0.92
Non-renewable
Pollutant 1761 0.69
wastes
Steam turbine Pollutant 1207 0.44
Renewable wastes Renewable 739 0.32
Gas turbine Pollutant 657 0.24
Hydro-wind Renewable 23 0.01
TOTAL 276,316 100.00
* but radioactive wastes
3.6. Pollutant Emission Factors for ICEVs

To estimate the emissions of ICEVs, we used the European Environment Agency
emission factors for road transport according to the Tier 2 methodology, which take into
account the technology or legislation (age) of the vehicle [33]. We only considered the
emissions of nitrogen oxides (NOx ) and particulate matter (PM). Table 4 shows the emission
factors expressed in grams of pollutant emitted per kilometer driven, used in this study
for both gasoline and diesel vehicles. The passenger car factors corresponded to mid-size
passenger cars. The truck factors corresponded to trucks between 7.5 and 16 tons. And the
motorcycle factors corresponded to four-stroke motorcycles between 250 and 750 cm3 .
In the absence of factors, the same values are assumed for gasoline trucks as for vans.
Similarly, the PM factors for diesel motorcycles are assumed to be the same as for cars.
90
Energies 2023, 16, 5006
Table 4. NOx and PM emission factors (in g/km) for ICEVs (European gasoline and diesel vehicles)
taken from [33]. Values in italics are estimates used in this study.
Passenger Cars Vans Trucks Motorcycles

Technology/Regulation a a a a
NOx PM b NOx PM b NOx PM b NOx PM b,c
Gasoline
ECE (until 1977) 2.53 0.0022 3.09 0.0023 3.09 0.0023 0.233 0.2
ECE (1978–1980) 2.40 0.0022 3.09 0.0023 3.09 0.0023 0.233 0.2
ECE (1981–1985) 2.51 0.0022 3.09 0.0023 3.09 0.0023 0.233 0.2
ECE (1985–1992) 2.66 0.0022 3.09 0.0023 3.09 0.0023 0.233 0.2
Euro 1 (1992–1996) 0.485 0.0022 0.563 0.0023 0.563 0.0023 0.233 0.2
Euro 2 (1996–2000) 0.255 0.0022 0.23 0.0023 0.23 0.0023 0.477 0.08
Euro 3 (2000–2005) 0.097 0.0011 0.129 0.0011 0.129 0.0011 0.317 0.04
Euro 4 (2005–2010) 0.061 0.0011 0.064 0.0011 0.064 0.0011 0.194 0.04
Euro 5 (2010–2016) 0.061 0.0014 0.064 0.0014 0.064 0.0014 0.194 0.01
Euro 6 until 2016 0.061 0.0014 0.064 0.0012 0.064 0.0012 0.194 0.01
Euro 6 2017–2019 0.061 0.0016 0.064 0.0012 0.064 0.0012 0.194 0.01
Euro 6 2020+ 0.061 0.0016 0.064 0.0012 0.064 0.0012 0.194 0.01
Diesel
Conventional (until 1992) 0.546 0.2209 0.87 0.356 8.92 0.3344 0.546 0.2209
Euro 1 (1992–1996) 0.690 0.0842 0.69 0.117 5.31 0.201 0.690 0.0842
Euro 2 (1996–2000) 0.716 0.0548 0.716 0.117 5.5 0.104 0.716 0.0548
Euro 3 (2000–2005) 0.773 0.0391 0.77 0.0783 4.3 0.0881 0.773 0.0391
Euro 4 (2005–2010) 0.58 0.0314 0.58 0.0409 2.65 0.0161 0.58 0.0314
Euro 5 (2010–2016) 0.55 0.0021 0.55 0.001 1.51 0.0161 0.55 0.0021
Euro 6 until 2016 0.45 0.0015 0.45 0.0009 0.291 0.0008 0.45 0.0015
Euro 6 2017–2019 0.35 0.0015 0.35 0.0009 0.291 0.0008 0.35 0.0015
Euro 6 2020+ 0.17 0.0015 0.17 0.0009 0.291 0.0008 0.17 0.0015
ECE = Economic Commission for Europe; a NOx = nitrogen oxides, expressed as NO2 ; b PM = particulate matter,
expressed as PM2.5 ; c values for quadbikes.
4. Results
The total energy needed to cover the mobility in PB with a fleet composed exclusively
of BEVs can be estimated by multiplying the homologated power consumption of each
vehicle type selected in Section 3.4 by the total distance traveled by each vehicle in PB. The
results are shown in Table 5.
Table 5. Daily energy consumption estimated for a fleet of BEVs in Barcelona province.
Power Consumption Total Daily

Private Transport Distance (km)
(kWh/100 km) Consumption (kWh)
Cars 47,798,589 16.2 7,743,371
Motorcycles 3,855,347 4.2 162,310
Vans 1,795,268 20.5 368,030
Trucks 1,468,856 110.2 1,619,267
TOTAL 54,918,060 9,892,978
In summary, the daily electrical energy required to switch from fossil-fuel-based

ground transportation to clean electric transportation would be about 10 GWh. This
amount would require a 3% increase in clean electricity generation. This increase seems
more than affordable for Spain as a whole, with the exception of hydropower.
Traditionally, hydropower has been very important in electricity generation. The main
problem with this energy source is that due to the drought situation in 2022, the total
production of hydroelectricity in Spain has decreased by 39.7% compared to 2021. The
foreseeable increase in droughts and the abandonment of the construction of hydroelectric
dams in Spain means that an increase in hydroelectric capacity has not been considered.
Therefore, it seems reasonable to consider an increase in electricity generation capacity
with only wind and PV solar energies.
91
Energies 2023, 16, 5006
Unlike nuclear power and combined cycle power plants, wind and solar energies
cannot be controlled by existing human technology, but depend on the weather. This fact
means that the installation of new electricity generation mechanisms must take into account:
(1) the hourly schedules for charging BEVs, and (2) the hourly production of electricity for
each technology.
The main characteristic of the BEV is the slowness of recharging the vehicle compared
to refueling an ICEV. Since the BEV needs to be parked for a relatively long time to recharge,
we propose relating the period of electricity generation needed to recharge the BEV fleet
inversely proportionally to the mobility schedule in PB (Figure 3). That is, we assume
that the fewer trips are made in PB, the more likely it is that the BEV will be recharging
batteries in the standard recharge mode. With this assumption, the distribution of electric
power throughout the day would be as shown in Figure 4. Therefore, most of the additional
electricity generation to charge the batteries of the BEVs that would replace the current
fleet would have to be generated at night.
Figure 4. Estimated availability of electric power to recharge a BEV in PB.

A proposal to increase wind and PV solar power generation capacity is
presented below.
4.1. PV Solar Energy

Solar energy reaches its daily production peak in Barcelona between 10 a.m. and
5 p.m. [34]. In winter, approximately 90% of the total energy is produced in this hourly
window. However, this is also the season with the lowest total production. In summer,
about 70% is produced in this time window [34]. Therefore, it is estimated that 75% of the
total energy production on an annual average is in this time window.
From Figure 4, it can be deduced that there would be an average of 29.7% charge
availability in this period. To achieve this charging scenario, it is necessary to find a
compromise between the recommendation of charging the BEV with non-fast charging and
the availability of using PV solar energy. Therefore, in order to maximize the possibilities of
producing 100% of the electrical energy from clean sources, it seems reasonable to consider
that 30% of this energy can come from PV solar power.
The installed PV capacity in Spain in 2022 was 19,348 MW [35]. Therefore, its load
factor L can be calculated as follows:
E
L= (1)
P·t
where E is the total annual energy generated by photovoltaic source, i.e., 27,283 GWh (see
Table 3); P is the installed PV capacity, and t is the number of hours per year available for
92
Energies 2023, 16, 5006
PV generation, i.e., 2557 h. Applying Equation (1) with these values gives a load factor
of 55.1%.
Then, the additional PV power, PA , to be installed for only BEV charging in PB during
the period between 10 a.m. and 5 p.m., i.e., 7 h, can be estimated as follows:
f · EC
PA = (2)
L·tc
where f is the fraction of energy from PV solar power, i.e., 0.3; EC is the total daily con-
sumption of energy, i.e., 9893 MWh (see Table 5); and tc is the number of charging hours,
i.e., 7 h. Applying Equation (2) with these values gives PA = 770 MW.
As a reference, there are currently 196 PV park projects in Catalonia, with the construc-
tion of PV modules planned to occupy 4977 hectares (ha), with a total PV solar power of
2888 MW [36]. This corresponds to an average power of 0.58 MW/ha. Applying the same
ratio to the additional power of 770 MW, at least 1327 ha of land would need to be covered
with solar panels.
As a reference for estimating the investment required to install this amount of electric-
ity, we have extrapolated the budget of the Ancar II PV solar farm [37]. This farm, located
in the province of Teruel, has a nominal capacity of 41.58 MW, generated by 116,032 PV
generation modules covering an area of 106.53 ha. The budget for the Ancar II PV solar
park is EUR 21.2 M, of which the cost of the solar panels represents 62% [37]. Assuming
the same cost per hectare, an investment of around EUR 265 M would be required to install
1327 ha.
If we extrapolate these calculations to the whole of Spain, assuming that the mobility
of private transport in Spain as a whole follows a behavior similar to that of PB, as shown
in Figure 3, it would be necessary to occupy an area of approximately 14,000 ha with an
investment of around EUR 2800 M.
4.2. Wind Energy

The remaining 70% of the electrical energy to be produced would come from wind.
The installed capacity of wind energy in Spain in 2022 was 29,417 MW [35]. When
we applied Equation (1) to the data for wind power (i.e., E = 59,805 GWh, P = 29,417 MW,
t = 8766 h), the load factor of the entire wind power generation grid was only 23.2%.
As shown in Figure 4, there is virtually 100% hourly availability for BEV charging
between 0 and 6 a.m. Therefore, the additional wind power to be installed for only BEV
charging in PB can be estimated by applying Equation (2) with f = 0.7, EC = 9893 MWh,
L = 0.232, and tc = 6 h. The result is PA = 4976 MW, or about 5 GW.
There are currently 45 wind farm projects in Catalonia (region in the northeast of Spain
where PB is located), with 207 wind turbines in the pipeline, representing a total capacity
of 1184 MW [36]. This corresponds to an average capacity of 5.72 MW per wind turbine.
Applying the same ratio to the additional 5 GW of wind power, at least 870 turbines would
need to be installed.
To estimate the budget of this proposal, we can extrapolate the actual budgets of
existing wind farms. For example, the budget of the Cabigordo wind farm in the province
of Teruel (Spain) is EUR 32.6 M, of which the main budget item is the purchase and
installation of the wind turbines, which amounts to EUR 22 M [38]. This wind farm consists
of nine wind turbines with a total capacity of 50 MW. Considering a similar cost for the
wind turbines, the total cost of just installing the 870 wind turbines, without considering all
the other costs of the project, would be approximately EUR 2127 M.
At the country level, assuming the same mobility pattern as in PB, it is estimated that
approximately 9240 turbines would need to be installed at a cost of around EUR 22,600 M.
4.3. Reduction of Pollutant Emissions

Table 6 shows the estimated NOx and PM emissions in kg/day that would be avoided
if the ICEVs currently circulating in PB were replaced by BEVs. For this purpose, it was
93
Energies 2023, 16, 5006
assumed that the mobility envisaged in Figure 3 is composed of ICEVs of the four categories
analyzed (cars, motorcycles, vans, and trucks) and grouped in five age periods (up to 1993,
1994–2008, 2009–2013, 2014–2018, and 2019–2022) in the proportions that exist in the current
Spanish fleet (see Table 1). The emission factors in Table 4 have been applied to each
category by averaging the factors for the specified periods. These factors are multiplied by
the daily distance traveled by each category of ICEV, as shown in Table 5.
Table 6. Estimated pollutant emissions from ICEVs in PB (kg/day).
Gasoline Diesel
Vehicle Age Period
NOx PM NOx PM
<5 years 2019–2022 375 10 1045 9
5–9 years 2014–2018 210 5 1546 6
10–14
Passenger cars 2009–2013 133 3 1232 37
years
15–19
1994–2008 789 6 2424 184
years
>19 years up to 1993 15,283 15 4321 1067
<5 years 2019–2022 153 2 134 1
5–9 years 2014–2018 114 1 265 1
10–14
Motorcycles 2009–2013 120 4 350 10
years
15–19
1994–2008 231 33 538 41
years
>19 years up to 1993 283 186 638 158
<5 years 2019–2022 2 0 6 0
5–9 years 2014–2018 1 0 4 0
10–14
2009–2013 1 0 5 0
Vans years
15–19
1994–2008 6 0 17 2
years
>19 years up to 1993 553 0 162 49
<5 years 2019–2022 0 0 2 0
5–9 years 2014–2018 0 0 2 0
10–14
2009–2013 7 0 5 0
Trucks years
15–19
1994–2008 3 0 48 1
years
>19 years up to 1993 62 0 164 6
Total 18,325 265 12,910 1572
According to these calculations, more than 31 tons of NOx and almost two tons of PM
would be avoided daily.
Assuming that the mobility pattern throughout Spain is similar to that of PB, it can be
estimated that the daily emission reduction would be about 314 tons of NOx and 17 tons of
PM in Spain.
5. Discussion
The methodology used could be applied to other regions or countries where the vehicle
fleet is mainly composed of ICEVs and for which disaggregated mobility information
is available.
To understand the scale of the challenge of installing the required wind power capacity,
we can compare it with the current situation of existing wind farms. The total number
of wind turbines currently operating in Catalonia is 846 [36]. These turbines are mainly
located in four areas where the wind tends to blow with greater intensity and frequency. As
a result, project proposals in Catalonia are concentrated in these areas [39]. However, this
may foreshadow a land use problem, as the best sites are limited. For this reason, there is a
94
Energies 2023, 16, 5006
first wind farm project, called Tramuntana Park, directly in the sea, on the coast of the Gulf
of Roses, which foresees an initial installation of 35 wind turbines [40]. This would make it
possible to take advantage of the high wind speeds in this area, although it is expected to
have a negative impact on the ecosystem.
In terms of the estimated demand for solar energy, the 1327 ha of land required is
equivalent to 1858 soccer fields, or 13.1% of the surface area of the city of Barcelona. The
requirements to meet this demand appear to be lower than those for wind energy because,
on the one hand, the surface area to be installed is relatively small, since only 30% of the
electrical energy would be covered by PV solar energy, and, on the other hand, there are
not as many restrictions on the location of the solar panels.
In the case of wind energy, the occupied area would be much smaller. For example,
the total area occupied by nine wind turbines and foundations at the Cabigordo wind farm
site is 65,413 m2 [38]. We can then estimate that 870 turbines would occupy an area of about
63 ha. The impact on the land would therefore be small compared to that of a PV solar
installation with a similarly rated output. However, the environmental impact generated
by wind turbines is not harmless, as it mainly affects birds [41].
The cost per unit of installed power is lower for PV solar farms (EUR 0.34 M/MW)
than for wind farms (EUR 0.43 M/MW). The main issue in achieving an effective use of
clean energy for a BEV fleet is the need to ensure the production of clean energy at night.
For this reason, it is essential to prioritize electrical energy produced by wind power, as it
is not possible to produce electrical energy from PV solar sources during the night.
In addition to this investment in electricity generation, there is the cost of the infrastruc-
ture required to distribute and install the charging points. This cost is very variable because
it depends on many factors, such as the price of the chosen charging point, the charging
power, the distance to the meter or electrical panel, and the location of the charging point,
as well as the auxiliary work that needs to be carried out to supply the charging point.
However, as a guideline, the cost of installing one charging point can range from one to
several thousand euros [42].
In terms of environmental benefits, the significant reduction in pollutant emissions
that would be achieved during vehicle operation is noteworthy. The estimated values are
relatively higher than those published in other studies, such as the one carried out for
the urban areas of Berlin and Stuttgart [17]. This is probably due to the high age of the
Spanish vehicle fleet, which is one of the oldest in Europe, with an average age of more
than 13.5 years, compared to the European average of 11.5 years [43]. Likewise, the Spanish
fleet has one of the lowest percentages of BEVs (3.8% [44]) in Europe (average 14.2% [45]).
In conclusion, if the assumptions made in the calculations are accepted, the resulting
estimates allow us to state that the impact of the mobility transformation would be positive
in terms of reducing pollutant emissions during the service life of BEVs. However, the
environmental impact of the rest of the life cycle of the vehicle (extraction of raw materials,
production, and recycling) would have to be taken into account.
The estimated investment required in Spain to generate the additional electricity from
wind and PV solar power is enormous (more than 25.4 billion euros), representing, for
example, 5.5% of the total national budget in 2022. On the other hand, the installation of
the estimated quantities of wind turbines and solar panels, although important, would not
be able to meet the electricity needs alone, unless they are accompanied by other auxiliary
sources to meet demand when the weather is unfavorable.
Estimated energy requirements could be reduced if batteries with higher energy
densities than those currently available are developed in the future. Another possibility
for the future, perhaps in the longer term, is to advance the development of hydrogen
production/use technology from renewable sources to make the use of FCEVs competitive.
Author Contributions: Conceptualization, J.R. and L.V.C.; methodology, J.R. and L.V.C.; validation,
L.V.C.; formal analysis, J.R. and L.V.C.; investigation, J.R. and L.V.C.; data curation, J.R.; writing—
original draft preparation, L.V.C.; writing—review and editing, J.R. and L.V.C.; visualization, J.R. and
95
Energies 2023, 16, 5006
L.V.C.; supervision, L.V.C.; project administration, L.V.C. All authors have read and agreed to the
published version of the manuscript.
References
1. United Nations Climate Change. The Paris Agreement. 2023. Available online: https://unfccc.int/process-and-meetings/the-
paris-agreement (accessed on 19 April 2023).
2. News European Parliament. EU Ban on the Sale of New Petrol and Diesel Cars from 2035 Explained. 2023. Available on-
line: https://www.europarl.europa.eu/news/en/headlines/economy/20221019STO44572/eu-ban-on-sale-of-new-petrol-and-
diesel-cars-from-2035-explained (accessed on 23 April 2023).
3. Eurocities. Low Emission Zones: Challenges and Solutions. 2021. Available online: https://eurocities.eu/latest/low-emission-
zones-challenges-and-solutions/ (accessed on 29 April 2023).
4. Krajinska, A. Electric Vehicles Are Far Better than Combustion Engine Cars When it Comes to Air Pollution. Here’s Why. 2021.
Available online: https://www.transportenvironment.org/discover/electric-vehicles-are-far-better-than-combustion-engine-
cars-when-it-comes-to-air-pollution-heres-why/ (accessed on 29 April 2023).
5. EUR-Lex. Commission Regulation (EU) 2017/1151 Supplementing Regulation (EC) No 715/2007 of the European Parliament and
of the Council on Type-Approval of Motor Vehicles with Respect to Emissions from Light Passenger and Commercial Vehicles
(Euro 5 and Euro 6). Available online: http://publications.europa.eu/resource/cellar/7d1c640d-62d8-11e7-b2f2-01aa75ed71a1.0
006.02/DOC_1 (accessed on 29 April 2023).
6. Motorway. Euro 6 Compliance and Emission Standards—The 2023 Guide. 2023. Available online: https://motorway.co.uk/sell-
my-car/guides/euro-6-emission-standards-compliance (accessed on 29 April 2023).
7. BUYACAR. What Are Euro 6 Emissions Standards? Why You Need to Know. 2022. Available online: https://www.buyacar.co.
uk/cars/euro-6-emissions-standard-cars/522/what-are-euro-6-emissions-standards-why-you-need-to-know (accessed on 29
April 2023).
8. Brzezanski, M.; Sala, R. The influence of AdBlue dosage on the process of selective catalytic reduction of nitrogen oxides.
ResearchGate 2013, 154, 1032–1037.
9. Toyota. 2023 MIRAI. 2023. Available online: https://www.toyota.com/mirai/ (accessed on 29 April 2023).
10. Hyundai. NEXO. 2023. Available online: https://www.hyundai.com/worldwide/en/eco/nexo/because-of-you (accessed on 29
April 2023).
11. Albatayneh, A.; Juaidi, A.; Jaradat, M.; Manzano-Agugliaro, F. Future of Electric and Hydrogen Cars and Trucks: An Overview.
12. McKinsey & Company. EVolution. Electric Vehicles in Europe: Gearing Up for A New Phase? 2014. Available online: https:
//www.mckinsey.com/featured-insights/europe/electric-vehicles-in-europe-gearing-up-for-a-new-phase (accessed on 29
April 2023).
13. Schneider Electric. Electric Vehicle and EV Charging Fundamentals. 2021. Available online: https://www.electrical-installation.
org/enwiki/Electric_Vehicle_and_EV_charging_fundamentals (accessed on 29 April 2023).
14. go-e. The 4 EV Charging Modes You Need to Know about. 2022. Available online: https://go-e.com/en/magazine/charging-
modes-electric-car (accessed on 29 April 2023).
15. Wallbox. EV Charging Connector Types: What You Need to Know. 2023. Available online: https://wallbox.com/en_catalog/
faqs-plug-types (accessed on 29 April 2023).
16. Lewison, T. Full fuel cycle emissions reductions through the replacement of ICEVs with BEVs. In Climate Change 2-Canadian
Technology Development Conference; Canadian Nuclear Society: Toronto, ON, Canada, 2001.
17. Fournier, G.; Boos, A.; Wörner, R.; Jaroudi, I.; Morozova, I.; Nemoto, E. Substituting individual mobility by mobility on demand
using autonomous vehicles—A sustainable assessment simulation of Berlin and Stuttgart. Int. J. Automot. Technol. Manag. 2020,
20, 369–407. [CrossRef]
18. Li, M.; Zhang, X.; Li, G. A comparative assessment of battery and fuel cell electric vehicles using a well-to-wheel analysis. Energy
2016, 94, 693–704. [CrossRef]
19. Choma, E.; Ugaya, C. Environmental impact assessment of increasing electric vehicles in the Brazilian fleet. J. Clean. Prod. 2017,
152, 497–507. [CrossRef]
20. Held, M.; Schücking, M. Utilization effects on battery electric vehicle life-cycle assessment: A case-driven analysis of two
commercial mobility applications. Transp. Res. Part D Transp. Environ. 2019, 75, 87–105. [CrossRef]
96
Energies 2023, 16, 5006
21. Autoritat del Transport Metropolità. Enquesta de Mobilitat en Dia Feiner (EMEF). 2023. Available online: https://www.atm.cat/
comunicacio/publicacions/emef (accessed on 11 May 2023). (In Catalan).
22. Padron.com.es. Población de Barcelona (Provincia). 2023. Available online: https://padron.com.es/barcelona/ (accessed on 14
May 2023). (In Spanish).
23. Repsol. Repsol Opens the Most Powerful Electric Vehicle Charging Station in Europe. 2019. Available online: https:
//www.repsol.com/en/press-room/press-releases/2019/repsol-opens-the-most-powerful-electric-vehicle-charging-station-
in-europe.cshtml (accessed on 29 April 2023).
24. ANFAC. Mapa de Infraestructuras de Recarga de Acceso Público en España Para 2021–2030. 2021. Available online: https:
//anfac.com/publicaciones/mapa-de-infraestructuras-de-recarga-de-acceso-publico-en-espana-a-2021-2030/ (accessed on 29
April 2023).
25. Dirección General de Tráfico. Parque de Vehículos—Tablas Estadísticas 2022. 2023. Available online: https://www.dgt.
es/menusecundario/dgt-en-cifras/dgt-en-cifras-resultados/dgt-en-cifras-detalle/?id=00866 (accessed on 29 April 2023).
(In Spanish).
26. Dirección General de Tráfico. Parque de Vehículos—Tabla Estadística Auxiliar 2022. 2023. Available online: https://www.
dgt.es/menusecundario/dgt-en-cifras/dgt-en-cifras-resultados/dgt-en-cifras-detalle/?id=00867 (accessed on 2 May 2023).
(In Spanish).
27. Institut d’Estudis Regionals i Metropolitans de Barcelona. Enquesta de Mobilitat en Dia Feiner 2021 (EMEF 2021). 2022. Available
online: https://omc.cat/documents/662112/1025112/EMEF+2021_Informe+SIMMB_juny2022+%281%29.pdf/17ff87aa-7005
-193e-ad0e-e08330798c53?t=1666599585086 (accessed on 12 May 2023). (In Catalan).
28. Silence. Silence S01. 2023. Available online: https://www.silence.eco/lang/en/s01/ (accessed on 11 May 2023).
29. Tesla. Model 3. 2023. Available online: https://www.tesla.com/en_eu/model3 (accessed on 11 May 2023).
30. Peugeot. Peugeot e-Expert. 2023. Available online: https://www.peugeot.es/gama/expert-furgon.html (accessed on 11
May 2023).
31. DAF. DAF LF Electric. 2023. Available online: https://www.daf.es/es-es/descubra-los-camiones-daf/alternativas-de-
combustibles-y-cadenas-cinematicas/vehiculos-electricos-de-bateria/daf-lf-electric (accessed on 11 May 2023). (In Spanish).
32. Red Eléctrica de España. Generación Total. 2023. Available online: https://www.sistemaelectrico-ree.es/informe-del-
sistema-electrico/generacion/generacion-de-energia-electrica/generacion-total-de-energia-electrica (accessed on 15 May 2023).
(In Spanish).
33. European Environment Agency. EMEP/EEA air pollutant emission inventory guidebook 2019. In European Monitoring and
Evaluation Programme (EMEP); European Environment Agency (EEA): Copenhagen, Denmark, 2019.
34. Red Eléctrica de España. Península—Seguimiento de la Demanda de Energía Eléctrica. 2023. Available online: https://demanda.
ree.es/visiona/peninsula/demandaqh/total/2023-03-07 (accessed on 16 May 2023). (In Spanish).
35. Red Eléctrica de España. Potencia Instalada. 2023. Available online: https://www.sistemaelectrico-ree.es/informe-del-sistema-
electrico/generacion/potencia-instalada (accessed on 16 May 2023). (In Spanish).
36. Generalitat de Catalunya. Visor Sobre el Estado de Tramitación Ambiental de los Proyectos de Energías Renovables en Catalunya.
2023. Available online: https://mediambient.gencat.cat/es/detalls/Articles/visor (accessed on 16 May 2023). (In Spanish).
37. Gobierno de Aragón. Planta Fotovoltaica Ancar II en los Término Municipales de Alfambra y Orrios (Teruel). 2023. Available
online: https://www.aragon.es/-/planta-fotovoltaica-ancar-ii (accessed on 16 May 2023).
38. Gobierno de Aragón. Parque Eólico Cabigordo en los Términos Municipales de Corbalán, Cedrillas y El Pobo (Teruel). 2023. Avail-
able online: https://www.aragon.es/-/parque-eolico-cabigordo?p_l_back_url=%2Fbuscador%3Fq%3Deolico%2Bcabigordo%
26category-td%3D466682%26category%3D465154%26type%3Dcom.liferay.journal.model.JournalArticle (accessed on 16 May
2023). (In Spanish).
39. Generalitat de Catalunya. Hipermapa. Sol·Licituds de Parcs Eòlics. 2023. Available online: https://sig.gencat.cat/visors/
hipermapa_antic.html#param=param&color=vermell&background=topo_ICC_grisos&BBOX=134915.697674,4485000,6590
84.302326,4752000&layers=ENERGIA_PARCSEOLICS_MULTIPUNT,ENERGIA_PARCSEOLICSSOLARS_LIN,ENERGIA_
PARCSSOLARS_MULTIPOLI (accessed on 16 May 2023).
40. Parc Tramuntana. Parc Tramuntana. Una Oportunidad Azul Para la Sostenibilidad Del Empordà i Girona. 2023. Available online:
https://parctramuntana.com/es/inicio/ (accessed on 16 May 2023). (In Spanish).
41. Miao, R.; Ghosh, P.N.; Khanna, M.; Wang, W.; Rong, J. Effect of wind turbines on bird abundance: A national scale analysis based
on fixed effects models. Energy Policy 2019, 132, 357–366. [CrossRef]
42. Chargeguru. Cuánto Cuesta la Instalación de un Punto de Recarga Para Vehículo Eléctrico? 2023. Available online: https:
//chargeguru.com/es/2021/09/17/cual-es-el-precio-de-la-instalacion-de-un-punto-de-recarga-para-vehiculo-electrico/ (ac-
cessed on 15 May 2023). (In Spanish).
43. ANFAC. La Antigüedad del Parque de Automóviles en España Sigue al Alza y Alcanza Los 13,5 Años de Edad Media. 2022.
Available online: https://anfac.com/actualidad/la-antiguedad-del-parque-de-automoviles-en-espana-sigue-al-alza-y-alcanza-
los-135-anos-de-edad-media/ (accessed on 15 May 2023). (In Spanish).
97
Energies 2023, 16, 5006
44. ANFAC. Vehículo Electrificado. Informe Anual 2022. 2023. Available online: https://anfac.com/wp-content/uploads/2023/04/
2022_Informe_electrificados-Anual.pdf (accessed on 15 May 2023). (In Spanish).
45. Rodríguez, M. Electric Vehicle Sales in Europe in the First Quarter of 2023. 2023. Available online: https://www.electromaps.
com/en/blog/electric-vehicle-sales-europe-first-quarter-2023 (accessed on 15 May 2023).
98
energies
Article
Degradation Prediction and Cost Optimization of Second-Life
Battery Used for Energy Arbitrage and Peak-Shaving in an
Electric Grid
Rongheng Li, Ali Hassan, Nishad Gupte, Wencong Su and Xuan Zhou *
Department of Electrical and Computer Engineering, University of Michigan-Dearborn, Dearborn, MI 48128, USA;
rongheng@umich.edu (R.L.); alihssn@umich.edu (A.H.); nishadg@umich.edu (N.G.); wencong@umich.edu (W.S.)
* Correspondence: xuanzhou@umich.edu
Abstract: With the development of the electric vehicle industry, the number of batteries that are retired
from vehicles is increasing rapidly, which raises critical environmental and waste issues. Second-life
batteries recycled from automobiles have eighty percent of the capacity, which is a potential solution
for the electricity grid application. To utilize the second-life batteries efficiently, an accurate estimation
of their performance becomes a crucial portion of the optimization of cost-effectiveness. Nonetheless,
few works focus on the modeling of the applications of second-life batteries. In this work, a general
methodology is presented for the performance modeling and degradation prediction of second-life
batteries applied in electric grid systems. The proposed method couples an electrochemical model of
the battery performance, a state of health estimation method, and a revenue maximization algorithm
for the application in the electric grid. The degradation of the battery is predicted under distinct
charging and discharging rates. The results show that the degradation of the batteries can be slowed
down, which is achieved by connecting numbers of batteries together in parallel to provide the
same amount of required power. Many works aim for optimization of the operation of fresh Battery
Energy Storage Systems (BESS). However, few works focus on the second-life battery applications.
In this work, we present a trade-off between the revenue of the second-life battery and the service
life while utilizing the battery for distinct operational strategies, i.e., arbitrage and peak shaving
against Michigan’s DTE electricity utility’s Dynamic Peak Pricing (DPP) and Time of Use (TOU)
tariffs. Results from case studies show that arbitrage against the TOU tariff in summer is the best
Citation: Li, R.; Hassan, A.; Gupte,
choice due to its longer battery service life under the same power requirement. With the number
N.; Su, W.; Zhou, X. Degradation
of retired batteries set to increase over the next 10 years, this will give insight to the retired battery
Prediction and Cost Optimization of
owners/procurers on how to increase the profitability, while making a circular economy of EV
Second-Life Battery Used for Energy
Arbitrage and Peak-Shaving in an
batteries more sustainable.
Electric Grid. Energies 2023, 16, 6200.
https://doi.org/10.3390/en16176200 Keywords: second-life battery; electricity grid application; electrochemical modeling; degradation
prediction; battery operational strategy
Received: 31 July 2023

Revised: 19 August 2023
Accepted: 21 August 2023 1. Introduction
Published: 26 August 2023 With the widespread use of electric vehicles, huge numbers of batteries are employed.
As the battery is charged and discharged continuously, the capacity degrades gradually.
When it reaches a certain level, i.e., 80% of the original state of health, the battery should
be retired from the electric vehicles [1,2]. With the continuous development of the electric
vehicle industry, the number of batteries retired from vehicles is increasing rapidly. The
expense of disassembling and recycling these batteries is relatively high, which raises a
substantial issue on how to handle them appropriately. A recent work that utilizes an
conditions of the Creative Commons indicator, i.e., global warming potential, demonstrates that reusing the existing electric
Attribution (CC BY) license (https:// vehicle battery in a secondary application provides a significant environmental benefit
creativecommons.org/licenses/by/ compared to manufacturing a new battery for the same purpose [3]. By appropriate
4.0/). assessment and treatment, the retired batteries can be utilized for other applications. Such

99
Energies 2023, 16, 6200
reused batteries are referred to as second-life batteries [4–6]. The implementation of second-
life batteries is also beneficial for protecting the environment and saving money [7].
Among the various applications of the second-life battery, it is found to be suitable
for employment as an energy storage system in the electricity grid system [8]. Energy
storage systems have proven to be a game changer for the integration of renewable energy
and the stability of modern power systems [9]. Their use in the grid can be for various
purposes. In [10], G. Fitzgerald et al. mentioned thirteen different uses for BESSs in general.
Main grid utilization of BESSs is for energy arbitrage [11–13], frequency regulation [14–16],
peak shaving [17–19], or power smoothing [20–22]. The second-life batteries have been
deployed by industrial consortiums as discussed in [23–25]. In [26], Zhang et al. proposes
a remaining useful life prediction methodology using a deep learning integrated approach.
In [27], Xiofan et al. discusses a power processing methodology for power converters for
power optimization of second-life batteries. Many research studies [28–30] focus on the
aging of a fresh energy storage system, however, there is a clear research gap in predicting
the degradation of second-life batteries. Therefore, accurate battery performance estimation
becomes essential in the choice of operational grid strategies. In the short-term range,
correct estimation of the battery charging and discharging responses can enhance the
stabilization of the grid system [31]. For the long-term consideration, accurate state of
health prediction of the second-life battery is beneficial for reducing the cost and improving
the grid system efficiency [32].
For distinct working conditions, i.e., operational strategies and user loads, the cost-
effectiveness of the batteries can be optimized with appropriate configurations, i.e., series
and parallel connections [33]. Compared to costly and time-consuming experimental test-
ing, numerical modeling is a better approach to speed up the simulation of the operating
process [34]. The three main categories of modeling methods are the physics-based electro-
chemical models [35], the electrical equivalent circuit models [36,37], and the data-driven
models [38,39]. The advantages and limitations of these approaches are summarized in
Table 1. The physics-based model shows the mechanisms of electrochemical reactions,
which are particularly important in predicting degradation. Among various models, the
Doyle-Fuller-Newman (DFN) model attracts great attention due to its good accuracy and
suitability for diverse working conditions [40]. For engineering practice, the P2D model is
simplified to a more concise form [41,42]. Xu et al. presented an electrochemical-thermal-
capacity model that minimizes capacity fade and reduces the temperature rise to prevent a
thermal runaway [43]. Song et al. developed an electro-chemo-mechanical model, which
couples the mechanical and electrochemical factors [44]. The framework of the P2D model
is implemented in commercial software, such as COMSOL [45], and open-source codes,
i.e., DUALFOIL [46], LIONSIMBA [47], and PyBaMM [48].
Table 1. A summary of the battery performance modeling approaches, including their benefits
and limitations.
Category Benefits Limitations

Reveal the physical
Physics-based electrochemical phenomenon behind the Complex to implement;
models [35,40–43] battery; suitable for wide time-consuming
scenario ranges
Electrical equivalent circuit
Robust; easy to implement Less accurate
models [36,37]
Accuracy depends on the
Moderate accuracy;
quality of training data;
Data-driven models [38,39] knowledge of the underlying
usually requires a large set of
process is not required
charging and discharging data
A number of major degradation mechanisms are proposed to accurately model the bat-
tery aging process. They are solid electrolyte interphase (SEI) layer growth, lithium plating,
100
Energies 2023, 16, 6200
particle cracking, and active material losses. Safri et al. developed a solvent-decomposition
reaction model to simulate the growth of solid electrolyte interphase at the anode [49].
Single et al. revealed the mechanism of SEI formation [50]. Their work demonstrated that
the diffusion of neutral radicals is the cause of long-term SEI growth. Luo et al. revealed the
mechanism of SEI formation [51]. They simulated the battery degradation under various
depths of discharge, state-of-charge swing ranges, and temperatures. O’Kane et al. coupled
four degradation mechanisms in the electrochemical model [52]. They reported that five
distinct pathways can result in end-of-life, which depends on how the cell is charged
and discharged.
This article outlines a technique for predicting the performance and degradation of
second-life batteries utilized in electric grid systems. The approach consists of an elec-
trochemical model of the battery’s performance, a health monitoring method, and an
algorithm to reduce costs for grid applications. The governing equations of the electro-
chemical process and the degradation mechanisms are demonstrated. Different charging
and discharging rates are utilized to predict battery degradation. The study shows that the
connection of batteries in parallel can slow down the degradation, but this option requires a
balance between the cost of the battery and its lifetime. Different charging and discharging
strategies are considered in the research, including DPP and TOU. The strategy with both a
longer battery life and mediate cost is depicted.
The main contributions of this work are summarized as follows:
• A methodology that couples the DFN electrochemical model and a revenue maximiza-
tion algorithm is introduced to model the performance of a second-life battery in the
application of the electric grid.
• Simulations of distinct configurations demonstrate that a trade-off between the revenue
of the battery and the service life should be optimized.
• Through the implementation of distinct battery operational strategies, the study il-
lustrates that the arbitrage against the TOU tariff in summer is the optimal solution
among various combinations due to its longer battery service life while providing the
same amount of power.
2. Methodology
The structure of a battery under the P2D model configuration is displayed in Figure 1.
The cell consists of three regions: the positive electrode, the separator, and the negative
electrode. The active materials are modeled as spherical particles that fill in the positive
and negative electrodes. Two dimensions are considered in the P2D model. The x-axis
is defined in the direction perpendicular to these layers to account for the diffusion and
migration of Li ions in the liquid phase. The origin of the r-axis is at the center of the solid
sphere, which describes the diffusion of Li ions in the active materials.
Figure 1. Structure configuration of the P2D model of the Li-ion battery.
2.1. P2D Models

The electrochemical process described in P2D models is governed by the following
five partial differential equations. They are the equations for: charge conservation in the
homogeneous solid, mass conservation in the homogeneous solid, mass conservation in the
101
Energies 2023, 16, 6200
homogeneous electrolyte, charge conservation in the homogeneous electrolyte, and lithium

movement between the solid and electrolyte phases [40,53,54].
2.1.1. Mass Conservation in Solid

The lithium diffusion inside the solid particles is driven by the gradient of its concen-
tration, which is governed by Fick’s second law as:

∂cs ( x, r, t) 1 ∂ ∂cs ( x, r, t)
= 2 Ds r 2 (1)
∂t r ∂r ∂r
where cs (x,t,r) and Ds are the concentration and the diffusion coefficient of lithium in the
solid phase, respectively. The initial condition is:
cs ( x, r, t)|t = 0 = Cs,0 (2)
The boundary conditions are:

∂cs ( x, r, t)
Ds =0 (3)
∂r r =0

∂cs ( x, r, t) j( x, t)
Ds = (4)
∂r r = Rs F
where cs,0 , Rs , F, and j( x, t) are the initial concentration of lithium, the radius of solid-
state particles, the Faraday constant, and the current density of the surface particle’s
electrochemical reaction rate, respectively.
2.1.2. Mass Conservation in Electrolyte

The movement of lithium ions inside the electrolyte is governed by the diffusion and
migration process, which is described by:
j( x, t)
∂ce ( x, t) ∂ e f f ∂ce ( x, t )
εe = De + as 1 − t0+ (5)
∂t ∂x ∂x F
with ce ( x, t), ε e , and t0+ being the lithium ion concentration, the electrolyte volume fraction,
ef f
and the lithium ion transfer number, respectively. De is the effective diffusion coefficient
in the electrolyte and is expressed as:
ef f brugg
De = εe De (6)
brugg
where ε e is the Bruggeman correction coefficient. De is the electrolyte diffusion coeffi-
cient. The specific surface area of the solid particles is calculated as:
3ε s
as = (7)
Rs
where ε s is the solid phase volume fraction.

The corresponding boundary conditions are:

∂ce ( x, t) ∂ce ( x, t)
= =0 (8)
∂ x x =0 ∂ x x= L

ef f ∂ce ( x, t) e f f ∂ce ( x, t )
De,neg = D (9)
∂ x x= xneg ∂ x x= xneg
e,sep
+
− = c e ( x, t )|
ce ( x, t)| x= xneg + (10)
x = xneg
102
Energies 2023, 16, 6200

ef f ∂ce ( x, t) e f f ∂ce ( x, t )
De,sep = De,pos (11)
∂x x = xsep ∂ x x= xsep+
− = c e ( x, t )|
ce ( x, t)| x= xsep + (12)
x = xsep
where x = 0 represents the boundary between the current collector and the negative elec-
trode; x = xneg is the boundary between the negative electrode and the separator; x = xsep
defines the boundary between the separator and the positive electrode. In addition, the
superscripts “−” and “+” denote the negative and positive portions of the battery.
2.1.3. Charge Conservation in Solid

Charge conservation shows that there is rarely a loss of net charge. Ohm’s law can be
seen in the electrode solid-phase potential distribution, which can be stated as follows:

∂is ( x, t) ∂ ef f ∂
− = σs ∅s ( x, t) = as j( x, t) (13)
∂x ∂x ∂x
where is ( x, t) is the electrical current density in the solid phase, φs ( x, t) is the potential
ef f
present in the solid phase, and σS is the material’s effective electrical conductivity in the
solid state defined as:
ef f
σS = εbu s σs (14)
with σs denoting the material’s conductivity in the solid phase.
The boundary conditions are as follows:

∂ ef f ∂ Iapp (t)
∅s ( x, t) ∅s ( x, t)
ef f
− σs = − σs = (15)
∂x x =0 ∂x x= L Acell
where cell A is the electrode area and Iapp (t) is the charge/discharge current value of the
external circuit when the battery is functioning.
2.1.4. Charge Conservation in Electrolyte

The change is conserved in the liquid phase. The lithium ions are intercalated and
deintercalated inside the solid particles, which is governed by:

∂ie ( x, t) ∂ ef f ∂ ∂ ef f ∂
− = ke ∅e ( x, t) + kD lnce ( x, t) = − as j( x, t) (16)
∂x ∂x ∂x ∂x ∂x
ef f
where ie ( x, t), ce ( x, t), φe ( x, t), and k e are the ionic current density, the lithium concen-
tration, the potential, and the effective ionic conductivity in the electrolyte. Then:
ef f brugg
ke = εe Ke (17)
with Ke being the ionic conductivity of the electrolyte. The effective diffusion conductivity
of the electrolyte is expressed as:
ef f
ef f 2RTk e d ln f ± 0
kD = (1 + )(t+ − 1) (18)
F d lnce
with R, T, and f ± being the gas constant, the temperature, and the molar activity coefficients
of the electrolyte. The boundary conditions are:

∂∅e ( x, t) ∂∅e ( x, t)
= =0 (19)
∂x x =0 ∂x x= L
103
Energies 2023, 16, 6200
Iapp (t)
− = i e ( x, t )|
ie ( x, t)| x= xneg + = (20)
x = xneg Acell
∅e ( x, t)| x= xneg
− = ∅e ( x, t )| +
x = xneg (21)
Iapp (t)
− = i e ( x, t )|
ie ( x, t)| x= xsep + = (22)
x = xsep Acell
∅e ( x, t)| x= xsep
− = ∅e ( x, t )| +
x = xsep (23)
2.1.5. Electrochemical Kinetics Equation

To link the surface over potential η ( x, t) with the electrochemical reaction rate j( x, t),
the Bulter–Volmer kinetic equation is employed.

αa F −αc F
j( x, t) = i0 exp η ( x, t) − exp η ( x, t) (24)
RT RT
with i0 , αa , and αc being the exchange current density, the transfer coefficient of the anode,
and the transfer coefficient of the cathode, respectively. The overpotential is expressed as:
re f
η ( x, t) = ∅s ( x, t) − ∅e ( x, t) − Ej (θ j ( x, t)) − j( x, t) RSEI (25)
re f
with RSEI being the resistance of the SEI film, and Ej being the equilibrium potential of
the electrodes. In addition:
cs,sur f ( x,t)
θ j ( x, t) = , j = neg, pos (26)
cs,max,j
where cs,sur f ( x, t) stands for the lithium concentration of the active material surface; cs, max is
the maximal lithium concentration of electrodes. The exchange current can be calculated as:
αa
i0 = Fk0 cs,max − cs,sur f cs,sur f αc ce αa (27)
with k0 as the electrochemical reaction rate constant.
2.2. State of Health

For the long-term lifespan, four mechanisms can be utilized for the prediction of
the state of health. In this paper, the interstitial-diffusion-limited SEI growth model is
implemented [51]. The four mechanisms are the solid electrolyte interphase layer growth,
lithium plating, particle cracking, and active material losses.

∗
F ∗ DLi,i
∗
∗ F∗ ∗ ∗
jinter = − c Li,i,0 exp − ∗ ∗ (∅s,n − ∅s,e ) (28)
L∗σ Rg T
∗
Here, jinter represents the interfacial current density, DLi,i ∗ is the diffusivity of lithium
ions in the inner SEI, ∅∗s,n is the negative electrode potential, and c∗Li,i,0 is the inner SEI’s
concentration of lithium ion interstitials when ∅∗s,n = ∅∗e,n . Here, ∅∗e,n is the true electrolyte
potential, F ∗ shows Faraday’s constant, R∗g is the universal gas constant, and T ∗ is the
reference temperature. The lithium plating is considered irreversible in this work. For the
lithium plating portion, the irreversible model is selected. The SEI on cracks and loss of
active material losses are not considered in this work. The description and derivation of the
governing equations for the degradation mechanisms are well demonstrated in [48–52,55].
The default settings of PyBaMM are utilized for the rest simulations.
104
Energies 2023, 16, 6200
2.3. Grid Optimization

This section will present an optimization problem for the LG M50 cylindrical cell
used in a pack acting as an energy storage system to provide power to the grid for two
management strategies: energy arbitrage and peak shaving. They are under extensive
investigation on the demand side in the area of power systems. Arbitrage involves ex-
ploiting temporal price variations in the electricity market by purchasing electricity during
periods of lower prices and selling it during periods of higher prices, thereby maximizing
economic gains. On the other hand, peak shaving aims to strategically decrease electricity
consumption during peak demand periods to smooth out load curves and reduce grid
stress. The advancement of sustainable energy systems is aided by these strategies, which
play a crucial role in enhancing grid stability, cost-effectiveness, and the integration of
renewable energy sources.
The mathematical statement of the optimization problem is described in the following
equations.
96
max ∑ Edch (t)C pk (t) − Ech (t)Co f pk (t) (29)
t =1
s.t.
(30)
Cbattery (t) = Cbattery (t − 1) + Ech (t)ηch − Edch (t)ηdch
SOC (t = 1) = SOC (t = 96) (31)
Ech (t) ≤ Ech_max (t) ∗ 0.95 ∗ u(t) ∀u ∈ 0, 1 (32)
Edch (t) ≤ Edch_max (t) ∗ 0.95 ∗ u(t) ∀u ∈ 0, 1 (33)
SOC min ≤ SOC ≤ SOC max (34)
) ∗ u(t) + Ech (t) ∗ (1 − u(t)) = 0

Edch (t
u(t) = 1, t = 44 to 80 (35)
where
u(t) = 0, t = 1 to 43 or 81 to 96
In Equation (30), Edch is the discharge energy in kWh from the battery, Ech is the
charging energy in kWh into the battery, C pk is the cost of energy in USD/kWh for the peak
hours, and Co f pk is the cost of energy in USD/kWh for the off-peak hours, respectively. ηch
and ηdch are the charge and discharge efficiencies taken as 0.95 (95%). The variable u(t)
is a binary variable that makes sure that the battery either charges or discharges at one
15-min time interval. The t represents 15-min intervals in a 24-h period, i.e., (25 × 15 = 96).
The cost function maximizes the profit earned by selling the power (revenue) subtracted
from the money incurred by charging the battery (cost). Constraint (31) calculates the
capacity Cbattery of the battery at each time step t, i.e., Cbattery decreases if the battery is
discharged or increases if it is charged. Equation (31) sets the state of charge SOC at t = 1
at the start of the day equal to the SOC at t = 96 at the end of the day. Equations (32)–(34)
keep the charge energy, discharge energy, and state of charge within the maximum physical
limit of the battery, respectively. Equation (35) is an additional constraint only for the
peak-shaving strategy that forces the battery to discharge during the peak demand hours.
The optimization problem is an MILP formulated using the abstract model in Pyomo 6.6.0
and solved using the CPLEX optimization solver.
The load profile of Southeast Michigan on a typical weekday is given in Figure 2, which
shows the load profile for one day of residential customers in Michigan. The load peaks
at around 12 p.m. and dips around 8 p.m. This is known as peak load for which utilities
must fire up the peak power plants, which is both expensive and inefficient. However, with
105
Energies 2023, 16, 6200
recent advances, energy storage systems can be used to provide the stored power to the
grid to meet the peak load, which is known as peak shaving.
Figure 2. The load profile of Southeast Michigan on a typical weekday.
The optimization problem in this work is solved for the two different residential tariffs
of a DTE utility in Michigan: Time of Use (TOU) and Dynamic Peak Pricing (DPP) rate as
shown in Table 2. The concept of TOU rate structures is to divide the day into different
periods, each with varying electricity prices, which encourages consumers to switch their
electricity consumption to off-peak hours when prices are lower. On the other hand, the DPP
rate dynamically adjusts electricity prices in real time depending on the supply and demand
conditions of the network, favoring a more flexible and reactive consumption behavior.
Table 2. DTE TOU and DPP tariff.
Summer (Jun through Sep) Summer (Jun through Sep)

Tariff Type On-Peak Mid-Peak Off-Peak (₡/kWh (USD)) (₡/kWh (USD))
Hours Hours Hours
On-Peak Off-Peak On-Peak Off-Peak
Monday–
Friday
12 a.m.–3
Monday–
Time of Use p.m., 7
Friday 3 p.m. --- 7.941 4.828 5.560 4.828
(TOU) p.m.–12 a.m.
to 7 p.m.
and all-day
Saturday,
Sunday
Dynamic
Peak Pricing Monday– Monday– 11 p.m. to 7 Same rates all around the year
(DPP) Friday 3 p.m. Friday a.m.
to 7 p.m. 7 a.m. to 3 and all-day On-Peak Mid-Peak Off-Peak
p.m. and Saturday,
7 p.m. to 11 Sunday 12.658 5.486 1.184
p.m.
3. Results
3.1. Model Validation
To validate the proposed calculation method, we have conducted simulations using
the LG M50 cylindrical cell. The detailed values of battery parameters were taken from
the reported works [52,56]. To represent different operating conditions, three distinct
discharging rates were selected, which are 0.05C, 1C, and 2C. The voltage responses
obtained from our model, which was implemented using PyBaMM v23.5, were compared
with experimental data available online and depicted in Figure 3. Due to the significant
106
Energies 2023, 16, 6200
difference in discharging timespans between 0.05C and the other rates, their responses
were plotted separately in Figure 3a. The model exhibited good agreement with the
experimental data for the discharging curve. Minor deviations between simulation results
and experimental data were observed for higher discharging rates, i.e., 1C and 2C in
Figure 3b, yet remained within acceptable tolerances.
Figure 3. The voltage responses of the battery at (a) 0.05C and (b) 1C and 2C.
With the short time performance of the battery validated, the long-term degradation
estimation is also validated. The following cycling protocol is utilized, which is the same
as a testing report. Firstly, the cell is charged at a constant current of 0.33C to 4.2 V. Then,
the cell is discharged at the same constant rate of 2.85 V. The above protocol is repeated for
1000 cycles with the discharged capacity displayed in Figure 4. As depicted in the figure,
the discharge capacity drops to 4.10 A.h, which is around 80% of the initial capacity. It is
well agreed with the reported testing results.
Figure 4. The discharge capacity of the battery for 1000 charging and discharging cycles.
3.2. Optimization for Distinct Strategies

Using the second-life battery cell in a grid would need a scheduled charge and dis-
charge, which is calculated by an optimization problem explained in Section 2. The charge
and discharge energy (kWh) of the battery cell is optimized using an optimization problem
formulation explained in Section 2.3. while following the constraints. Figure 5a,b shows
the charging and discharging energy (kWh) of a battery cell for the DPP tariff for arbitrage
and peak-shaving. The battery charges and discharges aggressively for DPP-Arbitrage
during hours 8 to 16 while for DPP—Peak-Shaving the charge and discharge trend is
concentrated from hours 19 to 22. Likewise, Figure 6a,b show charge and discharge energy
107
Energies 2023, 16, 6200
(kWh) calculated for the TOU tariff for summer arbitrage and peak-shaving. Figure 7a,b
provides similar results for the TOU tariff for winter. The overall charge and discharge
profile is rigorous for TOU summer peak shaving as compared to the rest. In terms of
battery degradation, such a profile will lessen cell life. For longer life, the charging and
discharging of the battery needs to be less rigorous for TOU peak shaving during winters.
The peaks are widely distributed with a uniform discharge during the peak hours.
Figure 5. (a) Charge and discharge energy (kWh) for arbitrage against DPP tariff; (b) charge and
discharge energy (kWh) for peak shaving against DPP tariff.
Figure 6. (a) Charge and discharge energy (kWh) for arbitrage against TOU tariff (summer); (b) charge
and discharge energy (kWh) for peak shaving against TOU tariff (summer).
Figure 7. (a) Charge and discharge energy (kWh) for arbitrage against TOU tariff (winter); (b) charge
and discharge energy (kWh) for peak shaving against TOU tariff (winter).
3.3. Battery Cell Capacity

To further expand on the results above, Figures 8–10 show the battery cell’s capacity
in kWh for DPP and TOU tariffs. Figure 8 shows that the battery capacity profile for DPP
peak shaving is more rigorous as compared to DPP arbitrage. For Figure 9, the battery
capacity profile for TOU peak shaving in summer is more rigorous than TOU arbitrage in
summer. The battery capacity profile for TOU winter peak shaving is more rigorous than
108
Energies 2023, 16, 6200
TOU winter arbitrage. Overall, the battery capacity profile for TOU winter arbitrage and
peak shaving looks the smoothest as compared to all other cases, which is in accordance
with the charge and discharge profile results above.
Figure 8. (a) Battery capacity (kWh) for 1 day (15 min interval) for DPP arbitrage; (b) battery capacity
(kWh) for 1 day (15 min interval) for DPP peak shaving.
Figure 9. (a) Battery capacity (kWh) for 1 day (15 min interval) for TOU (summer) EA; (b) battery
capacity (kWh) for 1 day (15 min interval) for TOU (summer) peak shaving.
Figure 10. (a) Battery capacity (kWh) for 1 day (15 min interval) for TOU (winter) arbitrage; (b) battery
capacity (kWh) for 1 day (15 min interval) for TOU (winter) peak shaving.
3.4. Usage Days for Different Operational Strategies

In this part, the usage days of two different operational strategies are investigated.
Case 1 simulates the configuration in that the required power is supplied using one LG M50
cell, while case 2 is for the same power supplied using two identical cells. For each case,
the battery cell(s) starts from 80% of its original state of health, justifying it as a second-life
battery use. It will be terminated at 64%, which is 80% of the starting point of the second-life
use. Figure 11 shows the cell life (usage days) calculated for two different configurations.
109
Energies 2023, 16, 6200
The first configuration considers only one cell providing (discharge) and receiving (charge)
energy/power to and from the grid, respectively, against DPP and TOU tariffs for EA and
peak shaving. Conversely, the second configuration considers the same for two battery
cells. The bar graph shows that cell life is the shortest in the first configuration where the
peak shaving is provided by one cell for summer TOU. This is because the cell charges and
discharges aggressively for this scenario as compared to the other cases, resulting in heavy
cell degradation. The cell life is the longest for the arbitrage case against the DPP tariff. For
configuration 2, again the cell life is lowest for peak shaving for TOU (summer), while the
cell life is longest for TOU (winter) peak shaving.
Figure 11. Usage days when required power is provided by one vs. two battery cells.
Table 3 shows the revenue (in USD) calculated by optimizing the cost function. The
profit is at a maximum (USD 21.25) for the arbitrage for TOU in summer. However, the
degradation for the same battery cell as shown above was the maximum. This again shows
that excessive charge and discharge may earn revenue but will also degrade the battery
fast. Therefore, the battery needs to generate revenue while at the same time degrading
slowly for long life. Considering the results from Figure 11 and Table 3, it is recommended
to use the second-life battery pack made of used LG M50 cells for arbitrage against the
TOU tariff in summer.
Table 3. DTE TOU and DPP tariff.
Scenario Profit per 24 h (USD)

DPP—Arbitrage 0.7067
DPP—Peak Shaving 1.77
TOU—Arbitrage (Summer) 20.08
TOU—Peak Shaving (Summer) 21.25
TOU—Arbitrage (Winter) 0.32
TOU—Peak Shaving (Winter) 0.32
4. Conclusions
A method for predicting the performance and deterioration of second-life batteries
deployed in electric grid systems is presented in this paper. The approach incorporates
an electrochemical model to evaluate battery performance, a health monitoring technique,
and a cost-reduction algorithm designed for grid applications. The method predicts battery
degradation based on various charging and discharging rates, which reveals that parallel
battery connections can mitigate degradation, despite necessitating a balance between
battery cost and longevity. Various charging and discharging strategies, including en-
ergy arbitrage and peak-shaving against DPP and TOU tariffs of Michigan’s DTE utility,
110
Energies 2023, 16, 6200
are explored. Note that the proposed methodology utilizes the SEI layer growth and
lithium plating mechanisms for battery degradation prediction to demonstrate the idea of
its implementation in the performance estimation of second-life batteries. More complex
mechanisms, i.e., the particle cracking model, loss of active material model, and the cou-
pling of these mechanisms, could be employed to further improve the estimation accuracy.
This methodology offers an appropriate framework for the analysis of the second-life
battery in grid applications. It could be extended further for electric grid systems with
photovoltaic panels and wind turbines. Moreover, combined with detailed cost estimation
approaches for the battery, i.e., replacement and depreciation models, a more comprehen-
sive computational framework could be developed for the revenue optimization of the
electric grid system.
Author Contributions: Writing—original draft, R.L.; Grid Optimization, A.H.; Governing Equations
of Battery, N.G.; Supervision, W.S. and X.Z. All authors have read and agreed to the published version
of the manuscript.
Funding: This research was funded by the Research Initiation & Development Grant (FY23 Winter
Cycle) of the University of Michigan—Dearborn Office of Research.
Data Availability Statement: The data are available from the corresponding author upon reasonable
request.
Acknowledgments: We are grateful to acknowledge the comments and suggestions from the editor
and reviewers for their help in improving the quality of this paper.
References
1. Skeete, J.P.; Wells, P.; Dong, X.; Heidrich, O.; Harper, G. Beyond the EVent Horizon: Battery Waste, Recycling, and Sustainability
in the United Kingdom Electric Vehicle Transition. Energy Res. Soc. Sci. 2020, 69, 101581. [CrossRef]
2. Wang, L.; Wang, X.; Yang, W. Optimal Design of Electric Vehicle Battery Recycling Network—From the Perspective of Electric
Vehicle Manufacturers. Appl. Energy 2020, 275, 115328. [CrossRef]
3. Ioakimidis, C.S.; Murillo-Marrodán, A.; Bagheri, A.; Thomas, D.; Genikomsakis, K.N. Life Cycle Assessment of a Lithium Iron
Phosphate (LFP) Electric Vehicle Battery in Second Life Application Scenarios. Sustainability 2019, 11, 2527. [CrossRef]
4. Hua, Y.; Liu, X.; Zhou, S.; Huang, Y.; Ling, H.; Yang, S. Toward Sustainable Reuse of Retired Lithium-Ion Batteries from Electric
Vehicles. Resour. Conserv. Recycl. 2021, 168, 105249. [CrossRef]
5. Hua, Y.; Zhou, S.; Huang, Y.; Liu, X.; Ling, H.; Zhou, X.; Zhang, C.; Yang, S. Sustainable Value Chain of Retired Lithium-Ion
Batteries for Electric Vehicles. J. Power Sources 2020, 478, 228753. [CrossRef]
6. Chen, W.; Liang, J.; Yang, Z.; Li, G. A Review of Lithium-Ion Battery for Electric Vehicle Applications and Beyond. Energy Procedia
2019, 158, 4363–4368. [CrossRef]
7. Haram, M.H.S.M.; Lee, J.W.; Ramasamy, G.; Ngu, E.E.; Thiagarajah, S.P.; Lee, Y.H. Feasibility of Utilising Second Life EV Batteries:
Applications, Lifespan, Economics, Environmental Impact, Assessment, and Challenges. Alex. Eng. J. 2021, 60, 4517–4536.
[CrossRef]
8. White, C.; Thompson, B.; Swan, L.G. Repurposed Electric Vehicle Battery Performance in Second-Life Electricity Grid Frequency
Regulation Service. J. Energy Storage 2020, 28, 101278. [CrossRef]
9. Sahoo, S.; Timmann, P. Energy Storage Technologies for Modern Power Systems: A Detailed Analysis of Functionalities, Potentials,
and Impacts. IEEE Access 2023, 11, 49689–49729. [CrossRef]
10. Fitzgerald, G.; Mandel, J.; Morris, J.; Touati, H. The Economics of Battery Energy Storage How Multi-Use, Customer-Sited Batteries
Deliver the Most Services and Value to Customers and the Grid the Economics of Battery Energy Storage; Rocky Mountain Institute:
Snowmass, CO, USA, 2015.
11. Aaslid, P.; Korpas, M.; Belsnes, M.M.; Fosso, O.B. Stochastic Optimization of Microgrid Operation With Renewable Generation
and Energy Storages. IEEE Trans. Sustain. Energy 2022, 13, 1481–1491. [CrossRef]
12. Gu, B.; Mao, C.; Liu, B.; Wang, D.; Fan, H.; Zhu, J.; Sang, Z. Optimal Charge/Discharge Scheduling for Batteries in Energy
Router-Based Microgrids of Prosumers via Peer-to-Peer Trading. IEEE Trans. Sustain. Energy 2022, 13, 1315–1328. [CrossRef]
13. Wu, X.; Zhao, J.; Conejo, A.J. Optimal Battery Sizing for Frequency Regulation and Energy Arbitrage. IEEE Trans. Power Deliv.
2022, 37, 2016–2023. [CrossRef]
14. Hassan, A.; Altin, M.; Bingöl, F. Inertia and Droop Controller for a Modern Variable Speed Wind Turbine to Provide Frequency
Control in a Microgrid. Politek. Derg. 2020, 23, 771–777. [CrossRef]
111
Energies 2023, 16, 6200
15. Guan, M. Scheduled Power Control and Autonomous Energy Control of Grid-Connected Energy Storage System (ESS) with
Virtual Synchronous Generator and Primary Frequency Regulation Capabilities. IEEE Trans. Power Syst. 2022, 37, 942–954.
[CrossRef]
16. Hassan, A. Frequency Control in an Isolated Power System with High Penetration of Wind Power. Ph.D. Thesis, Izmir Institute of
Technology, İzmir, Turkey, 2019.
17. Li, X.; Wang, L.; Yan, N.; Ma, R. Cooperative Dispatch of Distributed Energy Storage in Distribution Network with PV Generation
Systems. IEEE Trans. Appl. Supercond. 2021, 31, 1–4. [CrossRef]
18. Argiolas, L.; Stecca, M.; Elizondo, L.R.; Soeiro, T.B.; Bauer, P. Optimal Battery Energy Storage Dispatch in Energy and Frequency
Regulation Markets While Peak Shaving an EV Fast Charging Station. IEEE Open Access J. Power Energy 2022, 9, 374–385.
[CrossRef]
19. Li, X.; Cao, X.; Li, C.; Yang, B.; Cong, M.; Chen, D. A Coordinated Peak Shaving Strategy Using Neural Network for Discretely
Adjustable Energy-Intensive Load and Battery Energy Storage. IEEE Access 2020, 8, 5331–5338. [CrossRef]
20. Rasool, S.; Muttaqi, K.M.; Sutanto, D. A Multi-Filter Based Dynamic Power Sharing Control for a Hybrid Energy Storage System
Integrated to a Wave Energy Converter for Output Power Smoothing. IEEE Trans. Sustain. Energy 2022, 13, 1693–1706. [CrossRef]
21. Fan, F.; Kockar, I.; Xu, H.; Li, J. Scheduling Framework Using Dynamic Optimal Power Flow for Battery Energy Storage Systems.
CSEE J. Power Energy Syst. 2022, 8, 271–280. [CrossRef]
22. Abdalla, A.A.; El Moursi, M.S.; El-Fouly, T.H.M.; Al Hosani, K.H. A Novel Adaptive Power Smoothing Approach for PV Power
Plant with Hybrid Energy Storage System. IEEE Trans. Sustain. Energy 2023, 14, 1457–1473. [CrossRef]
23. Europe’s Largest Energy Storage System Now Live at the Johan Cruijff Arena. Available online: https://global.nissannews.com/
en/releases/europes-largest-energy-storage-system-now-live-at-the-johan-cruijff-arena (accessed on 15 April 2021).
24. Schimpe, M.; Piesch, C.; Hesse, H.C.; Paß, J.; Ritter, S.; Jossen, A. Power Flow Distribution Strategy for Improved Power Electronics
Energy Efficiency in Battery Storage Systems: Development and Implementation in a Utility-Scale System. Energies 2018, 11, 533.
[CrossRef]
25. Second Life Energy Storage: VHH and MAN Testing Use of Second Life of Batteries for EBus Charging Station. Available
online: https://press.mantruckandbus.com/second-life-energy-storage-vhh-and-man-testing-use-of-secondlife-of-batteries-
for-ebus-charging-station (accessed on 30 June 2020).
26. Chen, Z.; Chen, L.; Shen, W.; Xu, K. Remaining Useful Life Prediction of Lithium-Ion Battery via a Sequence Decomposition and
Deep Learning Integrated Approach. IEEE Trans. Veh. Technol. 2022, 71, 1466–1479. [CrossRef]
27. Cui, X.; Ramyar, A.; Contreras, V.; Judge, G.; Siegel, J.; Stefanopoulou, A.; Avestruz, A.T. Lite-Sparse Hierarchical Partial Power
Processing for Parallel Batteries in Heterogeneous Energy Storage Systems. In Proceedings of the 2021 IEEE 22nd Workshop on
Control and Modelling of Power Electronics, COMPEL 2021, Cartegena, Colombia, 2–5 November 2021; Institute of Electrical and
Electronics Engineers Inc.: New York, NY, USA, 2021.
28. Li, S.; Zhao, P.; Gu, C.; Li, J.; Huo, D.; Cheng, S. Aging Mitigation for Battery Energy Storage System in Electric Vehicles. IEEE
Trans. Smart Grid 2023, 14, 2152–2163. [CrossRef]
29. Fan, F.; Xu, Y.; Zhang, R.; Wan, T. Whole-Lifetime Coordinated Service Strategy for Battery Energy Storage System Considering
Multi-Stage Battery Aging Characteristics. J. Mod. Power Syst. Clean Energy 2022, 10, 689–699. [CrossRef]
30. Zhang, L.; Yu, Y.; Li, B.; Qian, X.; Zhang, S.; Wang, X.; Zhang, X.; Chen, M. Improved Cycle Aging Cost Model for Battery Energy
Storage Systems Considering More Accurate Battery Life Degradation. IEEE Access 2022, 10, 297–307. [CrossRef]
31. Janota, L.; Králík, T.; Knápek, J. Second Life Batteries Used in Energy Storage for Frequency Containment Reserve Service. Energies
2020, 13, 6396. [CrossRef]
32. Steckel, T.; Kendall, A.; Ambrose, H. Applying Levelized Cost of Storage Methodology to Utility-Scale Second-Life Lithium-Ion
Battery Energy Storage Systems. Appl. Energy 2021, 300, 117309. [CrossRef]
33. Wang, B.; Ji, C.; Wang, S.; Sun, J.; Pan, S.; Wang, D.; Liang, C. Study of Non-Uniform Temperature and Discharging Distribution
for Lithium-Ion Battery Modules in Series and Parallel Connection. Appl. Therm. Eng. 2020, 168, 114831. [CrossRef]
34. Wang, Y.; Tian, J.; Sun, Z.; Wang, L.; Xu, R.; Li, M.; Chen, Z. A Comprehensive Review of Battery Modeling and State Estimation
Approaches for Advanced Battery Management Systems. Renew. Sustain. Energy Rev. 2020, 131, 110015. [CrossRef]
35. Zhang, Q.; Wang, D.; Yang, B.; Cui, X.; Li, X. Electrochemical Model of Lithium-Ion Battery for Wide Frequency Range Applications.
Electrochim. Acta 2020, 343, 136094. [CrossRef]
36. Nejad, S.; Gladwin, D.T.; Stone, D.A. A Systematic Review of Lumped-Parameter Equivalent Circuit Models for Real-Time
Estimation of Lithium-Ion Battery States. J. Power Sources 2016, 316, 183–196. [CrossRef]
37. Seaman, A.; Dao, T.S.; McPhee, J. A Survey of Mathematics-Based Equivalent-Circuit and Electrochemical Battery Models for
Hybrid and Electric Vehicle Simulation. J. Power Sources 2014, 256, 410–423. [CrossRef]
38. Wang, Y.; Yang, D.; Zhang, X.; Chen, Z. Probability Based Remaining Capacity Estimation Using Data-Driven and Neural
Network Model. J. Power Sources 2016, 315, 199–208. [CrossRef]
39. Wei, J.; Dong, G.; Chen, Z. Remaining Useful Life Prediction and State of Health Diagnosis for Lithium-Ion Batteries Using
Particle Filter and Support Vector Regression. IEEE Trans. Ind. Electron. 2018, 65, 5634–5643. [CrossRef]
40. Doyle, M.; Fuller, T.; Newman, J. Modeling of Galvanostatic Charge and Discharge of the Lithium/Polymer/Insertion Cell. J.
Electrochem. Soc. 1993, 140, 1526–1533. [CrossRef]
112
Energies 2023, 16, 6200
41. Dao, T.S.; Vyasarayani, C.P.; McPhee, J. Simplification and Order Reduction of Lithium-Ion Battery Model Based on Porous-
Electrode Theory. J. Power Sources 2012, 198, 329–337. [CrossRef]
42. Han, X.; Ouyang, M.; Lu, L.; Li, J. Simplification of Physics-Based Electrochemical Model for Lithium Ion Battery on Electric
Vehicle. Part II: Pseudo-Two-Dimensional Model Simplification and State of Charge Estimation. J. Power Sources 2015, 278,
814–825. [CrossRef]
43. Xu, M.; Wang, R.; Zhao, P.; Wang, X. Fast Charging Optimization for Lithium-Ion Batteries Based on Dynamic Programming
Algorithm and Electrochemical-Thermal-Capacity Fade Coupled Model. J. Power Sources 2019, 438, 227015. [CrossRef]
44. Song, X.; Lu, Y.; Wang, F.; Zhao, X.; Chen, H. A Coupled Electro-Chemo-Mechanical Model for All-Solid-State Thin Film Li-Ion
Batteries: The Effects of Bending on Battery Performances. J Power Sources 2020, 452, 227803. [CrossRef]
45. Cai, L.; White, R.E. Mathematical Modeling of a Lithium Ion Battery with Thermal Effects in COMSOL Inc. Multiphysics (MP)
Software. J. Power Sources 2011, 196, 5985–5989. [CrossRef]
46. FORTRAN Programs for the Simulation of Electrochemical Systems. Available online: http://www.cchem.berkeley.edu/jsngrp/
fortran.html (accessed on 15 July 2023).
47. Torchio, M.; Magni, L.; Gopaluni, R.B.; Braatz, R.D.; Raimondo, D.M. LIONSIMBA: A Matlab Framework Based on a Finite
Volume Model Suitable for Li-Ion Battery Design, Simulation, and Control. J. Electrochem. Soc. 2016, 163, A1192–A1205. [CrossRef]
48. Sulzer, V.; Marquis, S.G.; Timms, R.; Robinson, M.; Chapman, S.J. Python Battery Mathematical Modelling (PyBaMM). J. Open Res.
Softw. 2021, 9, 14. [CrossRef]
49. Safari, M.; Morcrette, M.; Teyssot, A.; Delacourt, C. Multimodal Physics-Based Aging Model for Life Prediction of Li-Ion Batteries.
J. Electrochem. Soc. 2009, 156, A145. [CrossRef]
50. Single, F.; Latz, A.; Horstmann, B. Identifying the Mechanism of Continued Growth of the Solid–Electrolyte Interphase. Chem-
SusChem 2018, 11, 1950–1955. [CrossRef] [PubMed]
51. Luo, G.; Zhang, Y.; Tang, A. Capacity Degradation and Aging Mechanisms Evolution of Lithium-Ion Batteries under Different
Operation Conditions. Energies 2023, 16, 4232. [CrossRef]
52. O’Kane, S.E.J.; Ai, W.; Madabattula, G.; Alonso-Alvarez, D.; Timms, R.; Sulzer, V.; Edge, J.S.; Wu, B.; Offer, G.J.; Marinescu, M.
Lithium-Ion Battery Degradation: How to Model It. Phys. Chem. Chem. Phys. 2022, 24, 7909–7922. [CrossRef]
53. Doyle, M.; Newman, J. Comparison of Modeling Predictions with Experimental Data from Plastic Lithium Ion Cells. J. Electrochem.
Soc. 1996, 143, 1890–1903. [CrossRef]
54. Plett, G.L. Battery Management Systems, Volume I: Battery Modeling; Artech House: London, UK, 2015; Volume 1, p. 343.
55. Marquis, S.G. Long-Term Degradation of Lithium-Ion Batteries. Ph.D. Thesis, University of Oxford, Oxford, UK, 2020.
56. Chen, C.-H.; Brosa Planella, F.; O’Regan, K.; Gastol, D.; Widanage, W.D.; Kendrick, E. Development of Experimental Techniques
for Parameterization of Multi-Scale Lithium-Ion Battery Models. J. Electrochem. Soc. 2020, 167, 080534. [CrossRef]
113
energies
Article
HPPC Test Methodology Using LFP Battery Cell Identification
Tests as an Example
Tadeusz Białoń 1,2 , Roman Niestrój 1,2 , Wojciech Skarka 3,4, * and Wojciech Korski 2
1 Department of Electrical Engineering and Computer Science, Faculty of Electrical Engineering,

Silesian University of Technology, 44-100 Gliwice, Poland; tadeusz.bialon@polsl.pl (T.B.);
roman.niestroj@polsl.pl (R.N.)
2 Łukasiewicz Research Network—Institute of Innovative Technologies EMAG, 40-189 Katowice, Poland;
wojciech.korski@emag.lukasiewicz.gov.pl
3 Department of Fundamentals of Machinery Design, Faculty of Mechanical Engineering,
Silesian University of Technology, 44-100 Gliwice, Poland
4 Bumech S.A., 40-389 Katowice, Poland
* Correspondence: wojciech.skarka@polsl.pl
Abstract: The aim of this research was to create an accurate simulation model of a lithium-ion battery
cell, which will be used in the design process of the traction battery of a fully electric load-hull-dump
vehicle. Discharge characteristics tests were used to estimate the actual cell capacity, and hybrid pulse
power characterization (HPPC) tests were used to identify the Thevenin equivalent circuit parameters.
A detailed description is provided of the methods used to develop the HPPC test results. Particular
emphasis was placed on the applied filtration and optimization techniques as well as the assessment
of the quality and the applicability of the acquired measurement data. As a result, a simulation model
of the battery cell was created. The article gives the full set of parameter values needed to build a
fully functional simulation model. Finally, a charge-depleting cycle test was performed to verify the
created simulation model.
Keywords: lithium-ion iron phosphate (LFP) battery; hybrid pulse power characterization (HPPC);
Thevenin equivalent circuit
Citation: Białoń, T.; Niestrój, R.;

Skarka, W.; Korski, W. HPPC Test 1. Introduction
Methodology Using LFP Battery Cell
The research described here aims to create an accurate simulation model of a battery
Identification Tests as an Example.
cell, which will be used in the design process of the traction battery of a fully electric
Energies 2023, 16, 6239. https://
doi.org/10.3390/en16176239
load-hull-dump (LHD) vehicle. The model-based design (MBD) method [1,2] was used
to create numerical models of vehicle subassemblies in order to test them by means of
Academic Editor: Simone Barcellona simulation. The MBD method is frequently used to design vehicles and mobile robots [3], as
Received: 30 June 2023
well as to design manned and unmanned aerial vehicles [4,5]. The practical importance of
Revised: 20 August 2023 model-based techniques in energy storage analysis and design is also underlined in [6–10].
Accepted: 24 August 2023 The most basic element of a traction battery is a single cell. Its equivalent circuit,
Published: 28 August 2023 describing the static and dynamic properties, is the starting point for creating a simulation
model. In practice, many different forms of battery cell equivalent circuits are used [11–21],
taking into account various physical and chemical phenomena. The appropriate choice
of model depends on its intended use and the method by which its parameters will
Copyright: © 2023 by the authors. be identified.
Licensee MDPI, Basel, Switzerland. Battery cell models are often used in the form of an equivalent circuit [11–13,16,22–25],
This article is an open access article and many authors emphasize the advantages of this method. The first advantage is the
simplicity of the model, most often composed of resistors and capacitors, and a voltage
source to represent the battery’s OCV [15–17,26]. The battery’s electrical properties are
described by their characteristics dependent on the state of charge (SOC) of the cell. Another
advantage is the flexibility. The complexity of the equivalent circuit (i.e., number of RC
4.0/).

114
Energies 2023, 16, 6239
pairs in the Thevenin model) can be adjusted according to the desired model accuracy and
applied parameter identification method. To represent battery cell properties over a wide
frequency range, an additional inductance may be introduced [27,28]. In this work it is
omitted, because the inductance value cannot be identified with the HPPC tests. The third
important advantage is the composability. Introducing additional elements to the model,
phenomena such as self-discharge can be taken into consideration [18]. The equivalent
circuit element characteristics can also be extended with the thermal and aging model,
creating a multi-physical model [29].
The equivalent circuit model approach also shows high fidelity in simulating battery
performance characteristics [30–33]. Model fidelity, which measures how closely a model
or simulation mimics the state and behavior of a real-world item, is crucial in MBD.
For the purposes of battery design using the MBD method, models describing the
dependence of the open-circuit voltage (OCV) characteristics on SOC [11,18,34,35] and the
dynamic properties of the cell with one or two time constants are most often used. The
Thevenin circuit [17,20,25,36–38] is such model and was used in the research described here.
In general, it is difficult to estimate battery parameters quickly and accurately from
input–output cycling data [23,24,39]; therefore, special identification tests must be used.
The pulse charge or discharge test [14,40–42] or hybrid pulse power characterization (HPPC)
tests [43–47], combined with charge and discharge characteristics [44,48,49], are the most
commonly used. These tests (performed once) can reflect the properties of the cell for the
current state of health (SOH), so they cannot identify the effects of cell aging [21] or changes
in its parameters during long-term operation. They are also unable to identify self-discharge
effects [50]. These effects, however, have little impact on the basic operational properties of
the battery and are usually neglected in the design process with the MBD method.
Identification of resistor–capacitor (RC) parameters of the Thevenin equivalent cir-
cuit depends on the HPPC impulse voltage approximation quality with an exponential
function [40,48,49,51] or multi-exponential function [35,47,48]. However, usually two expo-
nential terms are used. Approximation may be performed by optimization. Deterministic
optimization methods may be used [21,44,52]. However, in this case, the optimization result
depends on the starting point of the algorithm, which is not always easy to choose. How-
ever, this problem does not occur in population-based metaheuristic algorithms [53]. Such
algorithms, i.e., genetic algorithms [17,54,55], particle swarm optimization (PSO) [21,52,56],
and others [57], are also used for HPPC results processing. PSO was also applied in the
research described here.
This article details the step-by-step process of preparing HPPC tests and processing
their results. Typical technical problems, including those resulting from the physical
properties of lithium-ion iron phosphate (LFP) cells, are discussed and methods of solving
them are proposed. LFP battery cells have a lower energy density than the most popular
electromobility applications of nickel manganese cobalt (NMC) cells [58,59], but they ensure
greater safety of use due to much lower susceptibility to thermal runaway [59,60]. LFP
battery cells also perform more favorably in terms of product sustainability [61,62].
The article describes the issues that are a continuation of the research described in the
article [63].
The novelties are as follows:
• An optimization-based battery cell time constant identification algorithm is imple-
mented in software written by the authors.
• An HPPC-based method for OCV vs. SOC characteristic determination is established.
• Other contributions of the article are as follows:
• This paper gives the values of all parameters necessary to build a fully parameterized
mathematical model of the cell.
• The paper explains the HPPC test development methodology step by step. In the
literature, usually only the results of HPPC are given, but the process of obtaining
them is not described. This paper fills that gap.
115
Energies 2023, 16, 6239
• The paper discusses potential flaws in the HPPC test results. Not every HPPC pulse
recorded during measurements is suitable for further analysis and must be omitted.
In the literature, this problem is hardly commented on. This paper fills that gap.
• The paper applies edge detection techniques in the analysis of the HPPC test results.
• The paper remarks on battery cell true capacity experimental estimation.
2. Materials and Methods

The general research methodology is schematically presented in Figure 1.
Figure 1. Research methodology overview.
The research was divided into three main stages: laboratory tests, identification of
cell parameters based on test results, and creation of a simulation model and simulations
(Figure 1). In the laboratory phase, tests were performed that were the basis for model
identification and its subsequent verification. Identification of the parameters of the cell
model was carried out using software written by the authors, using innovative optimization
algorithms based on particle swarm optimization (PSO) and the Levenberg–Marquardt
method. Known signal processing techniques, such as edge detection and filtering of
measurement data, were also used in an original way. As a result, a simulation model was
created in the MATLAB/Simulink environment, using the Simscape Electrical library.
The tests were carried out for the LFP (LiFePO4 ) battery cell with the rated parameters
given in Table 1. The following laboratory tests were carried out: discharge characteristics
to estimate the actual cell capacity, HPPC tests to identify the equivalent circuit parameters,
and a charge-depleting cycle (CDC) test [64] to verify the identified mathematical model.
Table 1. ThunderSky Winston LFP040AHA cell nominal parameters.
Parameter Value
Capacity Qn 40 Ah
Energy density 82.5 Wh/kg
Voltage (min./nominal/max.) 2.5/3.3/4.0
Current (typical/max. discharge) 20 A (0.5C 1 )/400 A (10C 1 )
1 Battery cell C-rating, based on nominal capacity: 1C = 40 A.
The tests were carried out in the laboratory setup shown in Figure 2. The voltage at
the cell terminals and at the shunt was recorded using a National Instruments NI 6251 M
A series data acquisition device was equipped with a 16-bit analog-to-digital converter.
116
Energies 2023, 16, 6239
The sampling rate was 135 Hz during HPPC tests and 100 Hz during the CDC test. The
main element of the setup was a programmable power supply with a load function ITECH
IT6522C, additionally equipped with a dedicated power dissipater module IT-E502. This
set enables both power supply and active load up to 3000 W and 120 A DC. The power
supply may operate in constant current (CC) and constant voltage (CV) modes. The battery
cell was operated in CC mode, in accordance with the given reference current test profile,
only if the cell voltage value was within the tolerable limits given in Table 1. When the
cell voltage reached the minimum or maximum value, the power supply was switched
into CV mode, in which the current was limited to keep the voltage within specified limits.
Measurements were carried out at an ambient temperature of about 22 ◦ C, with deviations
up to 1 ◦ C.
Figure 2. Laboratory setup overview.
The simulation model created during the research directly reflects the structure of the
laboratory setup. Not only a battery cell model based on Thevenin’s equivalent circuit
(Section 3.1) was created, but also an active power supply model was created, including the
implementation of CC/CV mechanisms. This makes it possible to compare the simulation
results with the laboratory CDC test results. In both cases, i.e., in the real power supply
and its simulation model, the same profile of the reference current was implemented.
3. Results
The aim of the research was to create an accurate simulation model of an LFP battery
cell. The basis of the model was the equivalent circuit described in Section 3.1, the parame-
ters of which were determined on the basis of HPPC test results as described in Section 3.3.
The cell capacity set in the simulation model was determined by the methods described in
Section 3.2. The method of verifying the identified model is described in Section 3.4.
3.1. Battery Cell Equivalent Circuit

A mathematical model of the battery cell in the form of a Thevenin equivalent cir-
cuit [21,37,38,46,65–67] was used (Figure 3). The circuit contains two RC pairs, thus simu-
lating two time constants of the dynamic model [65,66,68,69]:
τ 1 = R1 C1 , τ 2 = R2 C2 , (1)
Figure 3. Thevenin equivalent circuit of the battery cell.
117
Energies 2023, 16, 6239
Thevenin models are usually used with one to five RC pairs and their corresponding
time constants. In the case of LFP cells, the first (smallest) time constant has values of a
few seconds, whereas the second has values of tens of seconds. The third time constant is
measured in tens of minutes, etc. HPPC tests can only identify the first two RC pairs (see
Section 3.3), so the others are omitted [22].
All the resistances, capacities as well as the OCV (UOC in Figure 3) depend on the SOC
of the cell [31], which is estimated on the basis of the cell current [22,37,51,57,67,68,70]:
t
1
SOC = SOC0 − Idτ, (2)
Q
0
where SOC0 is the initial SOC of the cell, and Q is the cell capacity. Note that the actual cell
capacity depends on many factors, such as temperature and SOH of the cell, and is usually
different from the rated one, Qn . Here, it was estimated based on the measurement results
as described in Section 3.2.
Determination of the OCV vs. SOC characteristic is described in Section 3.3.2. The deter-
mination of the dependence of RC parameters on SOC is described in Sections 3.3.4 and 3.3.5.
3.2. Capacity and State of Charge Estimation

The actual capacity of the cell is usually different from the nominal one and is crucial
from the point of view of correct parameterization of the created mathematical model. The
correctness of determining the SOC depends on this, and more precisely, it is necessary to
determine the value of Q in Formula (2), corresponding to SOC = 1.
By definition, the charge drawn from the battery is equal to the integral of the current
over time [71]:
t
Q= Idτ. (3)
0
In geometric interpretation, Q is the area under the current waveform. However,
recording of the battery discharge current can be made at different values of I, and different
assumptions as to the operating conditions. Two different methods were used here. The first
was a measurement based on the discharge characteristics [44,48,49]. The second, proposed
by the authors, was the use of current waveforms recorded during HPPC tests [63].
Discharge tests were performed at four different current values (0.5 C, 1 C, 2 C, and
3 C) by recording the current and voltage. The tests were started with the cell charged
to the maximum voltage (Table 1) and after several hours of relaxation. The current was
recorded while the laboratory setup was operating in the CC mode, and the charge taken
under these conditions was called QCC (Figure 4).
Figure 4. Cell capacity estimation based on discharge characteristic.
When the cell voltage reached the minimum value, and the system went into CV
mode, the recording continued until the current completely dropped to 0. The charge
determined under these conditions was called QCV . The results obtained during the tests
are summarized in Table 2, and the recorded transients are shown in Figure 5.
118
Energies 2023, 16, 6239
Table 2. Cell capacity measurement results based on discharge characteristics.
Relative Discharge Total Discharge Discharge in CC Mode Discharge in CV Mode

Current Q [Ah] QCC [Ah] QCV [Ah]
0.5C 47.71 46.30 1.407
1C 47.71 45.78 1.934
2C 47.70 45.21 2.495
3C 47.62 45.41 2.212
Figure 5. Discharge characteristics for cell capacity estimation.
Table 2 shows that the total charge taken from the cell during the discharge tests
was slightly dependent on the current. However, the proportion between the QCC and
QCV values changes. The higher the discharge current, the lower the QCC and the higher
the QCV .
Then, the total charge taken from the cell during the HPPC tests (described in
Section 3.3) was determined by integrating the currents recorded in each of the 18 tests and
summing the results. The value of Q = 50.71 Ah was obtained (see Table 3 in Section 3.3).
Table 3. HPPC tests and impulses summary.
HPPC Impulse No., Type and Relative Current Value

ΔQ ΔQ/Qn Q
Test 1 2 3 4 5 6 7 8 [Ah] [%] [Ah]
No. 0.5 C 0.5 C 1C 1C 2C 2C 3C 3C
1 (−) (+) (−) (+) (−) (+) (−) (+) 2.41 6.03 2.41
2 (−) (+) (−) (+) (−) (+) (−) (+) 2.21 5.53 4.63
3 (−) (+) (−) (+) (−) (+) (−) (+) 4.21 10.53 8.84
4 (−) (+) (−) (+) (−) (+) (−) (+) 4.21 10.53 13.05
5 (−) (+) (−) (+) (−) (+) (−) (+) 4.22 10.54 17.27
6 (−) (+) (−) (+) (−) (+) (−) (+) 4.23 10.57 21.50
7 (−) (+) (−) (+) (−) (+) (−) (+) 4.22 10.56 25.72
8 (+) (−) (+) (−) (+) (−) (+) (−) 4.21 10.54 29.93
9 (+) (−) (+) (−) (+) (−) (+) (−) 4.22 10.55 34.16
10 (+) (−) (+) (−) (+) (−) (+) (−) 2.16 5.40 36.31
11 (+) (−) (+) (−) (+) (−) (+) (−) 2.18 5.45 38.49
12 (+) (−) (+) (−) (+) (−) (+) (−) 2.20 5.51 40.70
13 (+) (−) (+) (−) (+) (−) (+) (−) 2.21 5.52 42.90
14 (+) (−) (+) (−) (+) (−) (+) (−) 2.22 5.54 45.12
15 (+) (−) (+) (−) (+) (−) (+) (−) 2.20 5.49 47.32
16 (+) (−) (+) (−) (+) (−) (+) (−) 2.22 5.55 49.54
17 (+) (−) (+) (−) (+) (−) (+) (−) 0.92 2.31 50.46
18 (+) (−) (+) (−) (+) (−) (+) (−) 0.25 0.62 50.71
Impulses: •—healthy, •—trimmed, •—distorted, (+)—charging, (−)—discharging.
It should be noted that all measured charge values (in discharge and HPPC tests)
were greater than the nominal cell capacitance Qn , but the differences between them were
119
Energies 2023, 16, 6239
significant. Therefore, the question as to which of them should be treated as the final
one (the total capacity of the cell) should be asked, which will be used in the created
mathematical model. In order to find the answer, a number of simulations were carried out
for all the values obtained and the results of the selected values are presented in Section 3.4.
The best result was obtained for the value of Q = 45.7 Ah, calculated as the average of
the QCC values for all four current values (average of the values from the third column of
Table 2).
3.3. HPPC Tests

The basic idea of an HPPC test is to analyze the cell voltage response to a rectangular
current pulse. This response is a multi-exponential waveform, the time constants of which
should be determined in the identification process [17,44,46,48,69]:

− t − t
U = UOC − IR0 − IR1 1 − e τ1 − IR2 1 − e τ2 . (4)
In order for the determination of the time constant to be possible and precise, the
recorded voltage response should last several times longer than the expected length of
the time constant [43]. So, the longer the recorded transient, the better. On the other
hand, a single HPPC pulse should be as short as possible so as not to change the SOC
of the battery, which results directly from (2). The HPPC pulse length used in practice is
therefore a compromise between these two requirements. In the case of nickel manganese
cobalt (NMC) cells, the first two time constants are relatively short [63] and do not exceed
a dozen or so seconds, so pulses from 9 s to 18 s are sufficient. Usually 10 s pulses are
used [16,43–45]. In the case of the considered LFP cell, the time constants are longer, so the
duration time of the HPPC pulses was extended to 60 s.
In practice, HPPC profiles containing different numbers of pulses are used, but it
is always an even number [16,44,55]. This is because the pulses always occur in pairs (a
charging pulse with a discharging pulse), so that the series of pulses does not change the
SOC of the cell. The number of pairs of pulses may be different, sometimes only one is
used [69]. When there is more than one pair, then individual pairs differ in current values.
Here, four pairs of pulses were used, successively with current values of 0.5 C, 1 C, 2 C,
and 3 C (Figure 6). The greater the current value, the greater the voltage change in response
to the impulse, so the easier it is to record (see Section 3.3.1). On the other hand, the greater
the current value, the greater the SOC change during the pulse duration, which may cause
the problems described in Section 3.3.3. The order of the pulses in the pair also matters.
For high SOC values, the discharge pulse was used first. Starting with a charging pulse
would risk increasing the cell voltage during the pulse duration, which for a high SOC
value (close to 1) could cause the measurement system to switch from CC to CV mode and
cut the pulse (see Section 3.3.3). For small SOC values (close to 0), for the same reason,
the order was reversed with the charging pulse used first. The sequence change is seen in
Table 3 after test number 7.
Another important consideration is the relaxation time between pulses. In principle, it
should be much longer than the expected values of the time constants of the cell, so that
before the next pulse occurs, the cell voltage has time to stabilize after the preceding pulse.
However, due to the very large time constants of the examined LFP cell, it was difficult to
meet this assumption. In the conducted tests, a relaxation time between pulses of 20 min
was used (Figure 6).
The last element of the HPPC profile shown in Figure 6 is the discharge of the cell
before the next HPPC test. The values of the cell equivalent circuit parameters change most
rapidly for very small and very large SOC values, but for intermediate values (SOC ≈ 0.5),
they are almost constant. To capture the shape of the characteristics, a discharge of 0.05 Qn
was used for large SOC values, and then the interval was increased to 0.1 Qn to return to
0.05 Qn for small SOC values.
120
Energies 2023, 16, 6239
Figure 6. Exemplary HPPC test result.
Successive HPPC tests were carried out until the total charge taken from the battery
during the test dropped significantly below the set discharge value (0.05 Qn ), which resulted
from the fact that the discharge pulses in the profile were cut off by the CC/CV mechanism
due to reaching the minimum voltage. All the performed tests are summarized in Table 3,
where ΔQ is the charge taken from the cell during the whole HPPC test (including final
discharge by 0.05/0.1 Qn ). For the last two tests (17 and 18), this value drops significantly,
which means that the cell is already discharged. In Table 3, Q is the total value of the charge
taken from the cell at the end of the given test, taking into account the charge taken in the
preceding tests.
3.3.1. Filtering and Slope Detection

The recorded HPPC test transients shown in Figure 6 contain 1,500,000 samples over
time. For this reason, precise localization of the beginnings and ends of HPPC pulses is a
challenge. In addition, the recorded waveforms, in particular the voltages, contain a lot of
noise that hinders further analysis. The high noise content results from the unfavorable
proportion of the analyzed voltage changes to the measuring range of the data acquisition
device. The voltage changes caused by an HPPC pulse range from ten to several tens of mV,
whereas the measured voltage values reach up to 4 V. Therefore, a measuring transducer
with a range of 10 V was used. Consequently, the analyzed changes constitute only a few
percent of the measurement range, which, even with good quality converters and 16-bit
sampling, results in a relatively large amount of noise.
Both problems, noise removal in the voltage transient and detection of the beginnings
and ends of pulses based on the current transient, can be solved by using appropriate
data-filtering techniques.
To detect the edges marking the beginnings and ends of the pulses, a method based
on the analysis of the transient of the difference of two exponential averages was used:

Ii i=1
ri ( α ) = (5)
αIi + (1 − α)ri−1 i > 1,
Δi = ri (αfast ) − ri (αslow ). (6)
In (5) and (6), Ii is the i-th sample of the current waveform, ri is the i-th sample of the
exponential moving average, α is the weight coefficient, and Δi is the i-th sample of the
waveform difference. The principle of operation of the method is shown in Figure 7. The
difference Δ between two waveforms averaged with different weight values α (αslow = 0.02,
121
Energies 2023, 16, 6239
αfast = 0.1) contains peaks at moments when there is a rapid change in the trend of the
source waveform.
Figure 7. Edge detection method on the example of trimmed HPPC impulse.
An edge is detected when the value of Δ exceeds the set threshold, which was 0.5 here.
It should be noted that the values of the weight coefficient α are selected according to the
sampling frequency of the source waveform, and the threshold value should be selected
according to the noise content and disturbances in the source waveform.
A running average of order N = 5 was used to filter out the noise from the
voltage waveform:
i+ N
1
2N + 1 k=∑
Ufiltered i = Uk . (7)
i− N
In (7), i is the sample number of the measured voltage U that corresponds to the
Ufiltered i filtered voltage sample. The filtration consists of calculating the average for N
samples preceding and following the sample with the number i.
This simple method gave good results due to the high sampling frequency of the
recorded voltage waveform and the random character of the filtered noise. Order N = 5
was sufficient, and its low value introduced negligible distortion of the voltage waveform,
having no significant impact on the subsequent identification of time constants. The
filtration results for an exemplary HPPC pulse recorded at the smallest of the applied
currents of 0.5 C (i.e., in conditions where the relative noise content is the highest) are
shown in Figure 8.
3.3.2. OCV vs. SOC Characteristic

The OCV characteristic, represented by UOC (i.e., the voltage source in the Thevenin
equivalent circuit), is identified by measurement. The averaged charging and discharging
characteristics may be used here [44,48,49]. However, this method has some disadvantages.
The measured cell voltage contains not only the OCV but also the voltage drop at the
impedance, which also depends on the SOC. Moreover, the measured charge and discharge
122
Energies 2023, 16, 6239
capacities differ due to power losses. This makes it difficult to correlate them before
the averaging.
Figure 8. Exemplary HPPC impulse before and after data filtering. Near t = 10 s, a false peak
generated by the control system of the active power supply is visible.
To avoid these problems, the authors proposed a method for determining the OCV
characteristics based on the results of HPPC tests, consisting of averaging (over a 10 s time
period) the voltage recorded in the no-current state before each pulse. The SOC value
corresponding to the voltage obtained this way is calculated in reference to the total charge
of all HPPC tests, that is, the Q value from the last row of Table 3.
A measurement-based OCV characteristic is too irregular to be directly applied in
the cell mathematical model and must be approximated [11,70,72–76]. Choice of the
appropriate approximating function is a further problem. Several types of functions were
tested, but a log-linear exponential (LLE) function [11,75] gave the best result [63]. The LEE
function has the following form:
UOC (SOC ) = a + b ln(SOC + c) + d SOC + ee(SOC− f ) . (8)
Its coefficients a to f were obtained by optimization with the particle swarm method
(PSO) described in [63]. The resulting function plot and its coefficient values are given in
Figure 9.
Figure 9. OCV characteristic approximated with LEE function.
123
Energies 2023, 16, 6239
3.3.3. Impulse Evaluation and Selection

Once identified and filtered (Section 3.3.1), the HPPC impulses need to be selected
for further time constant identification. The problem is that not all recorded impulses are
suitable for further analysis. They may contain defects resulting from the measurement
method (CC/CV mechanism) or from the physical properties of the cell.
Figure 10 schematically shows the shape and interpretation of a healthy (suitable for
further analysis) pulse [11,14] and two cases of faulty pulses. Examples of recorded healthy
impulses are shown in Figures 6 and 8.
Figure 10. Shape of a single (charging) HPPC impulse and its potential flaws.
Pulse trimming (Figure 10) is easy to detect by analyzing the sequence of previously
detected edges and the distances between them. A healthy pulse has two edges, falling and
rising (which one is the first depends on whether the impulse is charging or discharging),
separated in time by the assumed pulse length. Deviation from this pattern suggests that
the pulse has been trimmed (Figure 7). Trimming occurs when the cell voltage reaches the
limit during the duration of the pulse, and therefore, the measurement system switches
from CC to CV mode.
In the performed tests, trimming always occurred together with the second defect,
distortion by the OCV characteristic (Figure 10). Pulse distortion occurs for small SOC
values, close to 0, and large ones, close to 1. This is because in these areas the OCV
characteristic is the steepest (Figure 9). Therefore, even a slight change in SOC during the
pulse duration causes a significant change in the voltage UOC of the cell, which translates
into the shape of the recorded waveform U (Figure 10). The shape of the waveform ceases
to depend only on the time constants τ 1 and τ 2 , which is a necessary assumption to make
the identification of these constants possible. In the extreme case, the recorded waveform
bends in a direction opposite (Figure 11) so that it results from (4), assuming that the time
constants τ 1 and τ 2 are positive.
124
Energies 2023, 16, 6239
Figure 11. Exemplary HPPC impulse distorted by OCV characteristic.
In Table 3, summarizing the HPPC tests, trimmed pulses are marked in red and dis-
torted pulses in yellow. Only healthy pulses marked in green were used in further analysis.
3.3.4. Impulse Waveform Approximation

The filtered and selected HPPC pulses were approximated by Function (4) to identify
the time constants τ 1 and τ 2 and the resistances R0 , R1 , and R2 . Then, on the basis of
Formula (1), capacities C1 and C2 were calculated. The approximation was carried out
using the PSO optimization method. At this stage of the research, a configuration of the
PSO algorithm was found that guaranteed high repeatability of the obtained results. The
fully informed particle swarm cognition method and the 8th order ring lattice swarm
topology were used. The cognition factor was 4.1, the swarm consisted of 64 particles, and
the number of iterations of the algorithm was set to 180.
The optimization method used and the experiments performed with it were described
in a separate article [63].
3.3.5. R and C vs. SOC Characteristics Approximation

The results of the approximation described in Section 3.3.4 are the values of the
cell equivalent circuit parameters and the corresponding SOC values. These values are
presented in the form of points on the graphs in Figure 12. These points are arranged in
more or less regular bands, which should be approximated with continuous functions in
order to create a mathematical model of the cell. A polynomial approximation was used,
and several experiments were performed with polynomials of various orders. The best
results were obtained for 3rd order polynomials:
f (SOC ) = a + b SOC + c SOC2 + d SOC3 . (9)
The approximation was carried out using the Levenberg–Marquardt method. The re-
sistance and capacitance characteristics were approximated and the polynomial coefficients
obtained are summarized in Table 4. In Figure 12, in the graphs of R and C values, the blue
lines are the waveforms of Function (9) with the parameters from Table 4. The blue lines in
the graphs τ 1 and τ 2 are the product of the approximating functions, respectively R1 and
C1 for τ 1 , R2 and C2 for τ 2 , according to (1).
3.4. Model Verification

The approximated OCV (Section 3.3.2), R0 , R1 , R2 , C1 , and C2 (Section 3.3.5) character-
istics fully describe the Thevenin equivalent circuit shown in Figure 3. This circuit, together
125
Energies 2023, 16, 6239
with functions describing its parameters, was implemented in the MATLAB/Simulink

environment by creating a simulation model of the cell, in a similar way as in [31]. The
last parameter describing the model is the charge value Q corresponding to SOC = 1. Due
to the problems with determining the actual capacity of the cell described in Section 3.2,
this value was found by performing a series of simulations of the cell operating in model
conditions and comparing their results with the transients recorded in the laboratory.
Figure 12. Thevenin equivalent circuit R and C parameter characteristics approximated with 3rd
order polynomial.
As model operating conditions, a current load profile based on the CDC [64], used
to test batteries of hybrid vehicles, was used. The application of cycle-based tests is a
typical strategy for battery cell mathematical model verification. Other popular cycle-
based tests are the dynamic stress test (DST) [46,51,65,68], ARTEMIS [57,77], and oth-
ers [17,28,54,66,69,75,78,79]. The applied CDC cycle consists of a set of discharge pulses
of different value (acceleration and driving at a constant speed) as well as charging ones
126
Energies 2023, 16, 6239
(regenerative braking). A single cycle was about 7 min long and discharged the cell by 3.3%
of Qn , so it was repeated over 30 times until the cell was fully discharged.
Table 4. R and C parameters 3rd order polynomial approximation coefficients values.
a b c d
R0 3.551 × 10−3 −6.172 × 10−3 8.993 × 10−3 −4.267 × 10−3
R1 9.601 × 10−4 −1.154 × 10−3 1.611 × 10−3 −5.716 × 10−4
R2 6.169 × 10−3 −2.678 × 10−2 4.690 × 10−2 −2.485 × 10−2
C1 5549 −1.359 × 104 5.058 × 104 −3.397 × 104
C2 1.712 × 104 8.510 × 104 −2.850 × 104 −4.243 × 104
Figure 13 shows the selected simulation results compared with the waveform recorded
in the laboratory. The values of capacitance Q used in the simulation model are summarized
in Table 5. The simulation accuracy has been evaluated with root-mean-square (RMS)
error [49,65,68,69,74] given by the formula:

1 K
K k∑
erms = (Umeasurement (tk ) − Usimulation (tk ))2 . (10)
=1
Table 5. Cell capacities used in simulations and resulting voltage error value.
Voltage Cell Capacity

Comment
RMS Error Q [Ah]
0.0432 45.7 Average for discharge characteristics, CC mode only
0.0487 47.7 Average for discharge characteristics, CC + CV
0.120 50.7 HPPC tests total discharge
0.167 40.0 Qn —nominal cell capacity
The data in Table 5 are ordered from the lowest erms value (best result) to the highest.
In Figure 13, the voltage relative error transients are shown, calculated as follows:
Umeasurement − Usimulation
δU = 100%. (11)
Umeasurement
The δU error statistics for transients presented in Figure 13 are presented in Table 6.
The order of the data in Table 6 is the same as in Table 5 and Figure 13.
Table 6. Simulation results—voltage relative error statistics.
Cell Capacity Average Error Average Error for t from 5 Peak Error Peak Error for t from 5 min to
Q [Ah] |δU| [%] min to 180 min |δU| [%] |δU| [%] 180 min |δU| [%]
45.7 0.977 0.751 14.9 9.62
47.7 1.07 0.805 14.6 9.83
50.7 2.44 0.873 22.9 10.1
40 2.73 0.579 20.4 9.04
127
Energies 2023, 16, 6239
Figure 13. Comparison of simulation and measurement CDC test results. Simulations performed for
various values of cell capacity.
128
Energies 2023, 16, 6239
4. Discussion
When identifying the parameters of the mathematical model of the cell, a major
problem was determining its actual capacity, which comes from a comparison of the
transients obtained for various Q values shown in Figure 13. The values obtained by the
different methods (Section 3.2) varied considerably. They also differed from the nominal
capacity Qn . It should be noted that according to the discharge characteristics provided in
the cell data sheet by the cell manufacturer, the cell capacity at normal temperature (i.e., the
temperature at which the tests described herein were performed) varied with the discharge
current from about 1.04 Qn (3 C) to 1.15 Qn (0.5 C).
These values correspond to QCC values in Table 2. Particularly significant here was
the value of the discharge current. This is why the capacity determined from the HPPC
tests was the largest. This was because, during these tests, the charge was taken from the
cell in small increments separated by long relaxation times. Thus, the cell had a lot of time
to regenerate and rebuild the voltage lowered by the discharge.
For these reasons, we decided to treat the result of the CDC test (Section 3.4) as an
indication, because the working conditions during this test were close to the real working
conditions of the battery in the vehicle. Cell capacitance identified using the method giving
a result consistent with the CDC test will, therefore, have the highest value in use.
The applied edge detection algorithm was an effective method of extracting individual
pulses from the entire recorded HPPC test transient (Section 3.3.1). It was also helpful in
identifying trimmed pulses (Section 3.3.3). Nevertheless, the detection of slopes sometimes
encountered problems resulting from the properties of the equipment used in the labo-
ratory setup. For example, in Figure 7, the enlarged fragment of the Δ waveform shows
disturbances in the form of short peaks. These disturbances often occurred just after the
power supply switched from CC to CV mode. They probably resulted from the way the
operation of the control system was implemented in the applied active power supply. In
Figure 7, this disturbance caused the detection of an additional, non-existent edge. The
result of the operation of the power supply control system is also visible in the current
waveform in Figure 8, this being the cause of the “false impulse”. This pulse, despite the
high peak value, was very short, so it had no significant effect on the SOC of the cell. These
types of pulses appeared in the no-current state when the change took place in the set
value of the cut-off voltage (voltage at which the power supply switched from CC to CV
mode). Such a change was performed before each change in the direction of the current
flow: before charging, the value was set to 4 V, and before discharging, the value was set to
2.5 V. These pulses also sometimes resulted in the detection of a non-existent edge, which
had to be taken into account in the analysis.
After identifying the time constants, the applied HPPC pulses, extended to 60 s, were
too short to correctly identify the second time constant of Thevenin’s model. The graphs
in Figure 12 show that the points corresponding to the identified R0 values are arranged
in a narrow, regular band, which proves good quality of identification. In the case of the
time constant τ 1 , the obtained band is much wider and the dispersion of values is greater,
but some regularity is still visible. In the case of the time constant τ 2 , the dispersion of
the results is very large, and their arrangement on the graph does not show any regularity.
Note that the values of the time constant τ 2 in Figure 12 changed in the interval from 30 s to
120 s, i.e., by 400%. Probably, in individual cases, values greater than 120 s would have been
obtained, if not for the fact that such a value was set as a limitation of the search space in
the applied PSO algorithm. It should be noted that, as stated in Section 3.3, to ensure good
quality identification of the exponential waveform time constants, the length of its recorded
fragment should be several times greater than the length of its time constants. However,
with the applied HPPC pulse length equal to 60 s, more than half of the identified τ 2 values
were greater, even up to two times. Increasing the duration of the HPPC pulses would
be undesirable, because it would cause changes too large in the SOC during the pulse
duration. In the case of the tested LFP type cell, resignation from determining two time
constants in favor of only one should be considered, as well as shortening the duration
129
Energies 2023, 16, 6239
of the HPPC pulse. Let us also pay attention to the obtained resistance and capacitance
values, given in Figure 12. The resistances are of the order of mΩ, which results in high
short-circuit currents of lithium-ion cells. Capacitances are of the order of kF. Similar values
were obtained, for example, in [32].
Reducing the HPPC pulse duration would reduce problems with distortion of their
voltage response by the OCV characteristics (Section 3.3.3). It should be noted that the
distortion effect in the form of a voltage waveform bent in the opposite direction shown in
Figures 10 and 11 is an extreme case. When the distortion was small, the distorted impulse
did not differ in shape from the healthy one, but the time constants identified on its basis
had overestimated values. It is possible that this effect (at least partially) is responsible for
the lack of regularity of the results presented in the τ 2 graph in Figure 12. This problem,
however, requires confirmation and further analysis.
Among the functions known in the literature, the LEE function was selected to approx-
imate the OCV characteristic. It has three SOC-dependent terms that are the most suitable
for the specific shape of the LFP cell OCV characteristic. The logarithmic term describes
the shape of the characteristic for SOC close to 0, linear describes the slope of the middle
part of the characteristic, and exponential describes its shape for SOC close to 1. As the
comparison of the measurement and simulation results showed, the OCV characteristic
had the greatest impact on the accuracy of the simulation model. The enlarged fragment
of the graph in Figure 13 shows that the simulated and measured voltage waveforms had
a very similar shape, but there was a slowly varying offset between them. The Thevenin
equivalent circuit (Figure 3) shows that the cell impedance, composed of the R and C
elements, was responsible for the shape of the waveform, this being the response to current
changes. The offset, on the other hand, is the result of differences in UOC (OCV) voltages.
Table 6 summarizes the voltage RMS error statistics, corresponding to the waveforms
in Figure 13. Error values averaged over time and peak values are presented. The data are
presented for entire transients and for a limited time range, from 5 min to 180 min. In this
range, the cell operates on the almost linear part of the OCV characteristic, i.e., in the most
typical conditions from a practical point of view.
The data show that for the optimal cell capacity (Q = 45.7 Ah) the average voltage
errors were less than 1%, which proves the very good fidelity of the obtained simulation
model. Error peaks under typical operating conditions (5–180 min) are at an acceptable level
of about 10%. The peak values correspond to the dynamic states (with rapid changes in the
load current), and their values are influenced primarily by the quality of identifying the
parameters of the RC pairs related to the time constants. Note, that for the reasons described
in Section 3.3, only two time constants have been identified, which affects the precision of
the model in dynamic states. It should also be noted that in the literature, models with only
one time constant [24,25,30,31,39] are sufficiently considered to be accurate.
In conclusion, despite the previously described problems, the identification of time
constants and R and C elements had a satisfactory effect, and the accuracy of the obtained
simulation model can be improved by better methods of identification and approximation
of the OCV characteristics.
5. Conclusions
The research showed the following:
• Among the various cell capacity values obtained as measurements, the best perfor-
mance of the mathematical model was obtained for the averaged charge taken from
the cell during discharge in the CC mode for different current values. Therefore, this
method is recommended for determining the actual capacity of the cell.
• The OCV characteristics of the LFP cell are best approximated by the LEE function.
• Identification of the second time constant of the LFP cell is difficult, because of its large
value, greater than a typical HPPC impulse duration.
• Suggestions for further research:
130
Energies 2023, 16, 6239
• It would be advisable to develop methods for automatic quality evaluation of HPPC

impulses, based on the criteria given in Section 3.3.3, which would enable full automa-
tion of the HPPC test results processing.
• A method should be developed to detect the occurrence of distortion of HPPC pulses
in cases where the distortion is small and does not significantly change the shape of the
voltage waveform yet, but already overestimates the obtained values of time constants.
• Simulation model accuracy may be improved by better OCV characteristic approximation.
Author Contributions: Conceptualization, T.B. and R.N.; methodology, T.B. and R.N.; software,
T.B.; validation, T.B., R.N., W.S. and W.K.; formal analysis, T.B.; investigation, T.B., R.N. and W.K.;
resources, T.B., R.N., W.S. and W.K.; data curation, T.B.; writing—original draft preparation, T.B.;
writing—review and editing, T.B., R.N., W.S. and W.K.; visualization, T.B.; supervision, W.S. and
W.K.; project administration, W.K.; funding acquisition, W.S. and W.K. All authors have read and
agreed to the published version of the manuscript.
Funding: This project was co-financed by the European Regional Development Fund in accordance
with the contract POIR.01.01.01-00-1427/20-00.
Acknowledgments: The authors would like to acknowledge the help of Katarzyna Lota and Kamil
Fraczek
˛ (Łukasiewicz Research Network—Institute of Non-Ferrous Metals) in supplying cells for
the research.
References
1. Diampovesa, S.; Hubert, A.; Yvars, P.A. Designing physical systems through a model-based synthesis approach. Example of a
Li-ion battery for electrical vehicles. Comput. Ind. 2021, 129, 103440. [CrossRef]
2. Skarka, W. Model-Based Design and Optimization of Electric Vehicles. In Proceedings of the 25th ISPE International Conference
on Transdisciplinary Engineering, Modena, Italy, 3–6 July 2018; Volume 7, pp. 566–575.
3. Niestrój, R.; Rogala, T.; Skarka, W. An Energy Consumption Model for Designing an AGV Energy Storage System with a PEMFC
Stack. Energies 2020, 13, 3435. [CrossRef]
4. Mateja, K.; Skarka, W.; Peciak, M.; Niestrój, R.; Gude, M. Energy Autonomy Simulation Model of Solar Powered UAV. Energies
2023, 16, 479. [CrossRef]
5. Peciak, M.; Skarka, W.; Mateja, K.; Gude, M. Impact Analysis of Solar Cells on Vertical Take-Off and Landing (VTOL) Fixed-Wing
UAV. Aerospace 2023, 10, 247. [CrossRef]
6. Giannelos, S.; Borozan, S.; Aunedi, M.; Zhang, X.; Ameli, H.; Pudjianto, D.; Konstantelos, I.; Strbac, G. Modelling Smart Grid
Technologies in Optimisation Problems for Electricity Grids. Energies 2023, 16, 5088. [CrossRef]
7. Giannelos, S.; Djapic, P.; Pudjianto, D.; Strbac, G. Quantification of the Energy Storage Contribution to Security of Supply through
the F-Factor Methodology. Energies 2020, 13, 826. [CrossRef]
8. Raventós, O.; Bartels, J. Evaluation of Temporal Complexity Reduction Techniques Applied to Storage Expansion Planning in
Power System Models. Energies 2020, 13, 988. [CrossRef]
9. Tang, Z.; Song, A.; Wang, S.; Cheng, J.; Tao, C. Numerical Analysis of Heat Transfer Mechanism of Thermal Runaway Propagation
for Cylindrical Lithium-ion Cells in Battery Module. Energies 2020, 13, 1010. [CrossRef]
10. Li, N.; Zhang, H.; Zhang, X.; Ma, X.; Guo, S. How to Select the Optimal Electrochemical Energy Storage Planning Program? A
Hybrid MCDM Method. Energies 2020, 13, 931. [CrossRef]
11. Davis, K.; Hayes, J.G. Comparison of Lithium-Ion Battery Pack Models Based on Test Data from Idaho and Argonne National
Laboratories. In Proceedings of the IEEE Energy Conversion Congress and Exposition (ECCE), Detroit, MI, USA, 11–15 October
2020; pp. 5626–5632. [CrossRef]
12. Rahmoun, A.; Biechl, H. Modelling of li-ion batteries using equivalent circuit diagrams. Electr. Rev. 2012, 2, 152–156.
13. Tremblay, O.; Dessaint, L.; Dekkiche, A. A Generic Battery Model for the Dynamic Simulation of Hybrid Electric Vehicles. In
Proceedings of the IEEE Vehicle Power and Propulsion Conference, Arlington, TX, USA, 9–12 September 2007; pp. 274–279.
[CrossRef]
14. Cipin, R.; Toman, M.; Prochazka, P.; Pazdera, I. Identification of Li-ion Battery Model Parameters. In Proceedings of the
International Conference on Electrical Drives & Power Electronics (EDPE), The High Tatras, Slovakia, 24–26 September 2019;
15. Chen, S.X.; Tseng, K.J.; Choi, S.S. Modeling of Lithium-Ion Battery for Energy Storage System Simulation. In Proceedings of the
Asia-Pacific Power and Energy Engineering Conference, Wuhan, China, 28–30 March 2009; pp. 1–4. [CrossRef]
131
Energies 2023, 16, 6239
16. Huang, K.; Wang, Y.; Feng, J. Research on equivalent circuit Model of Lithium-ion battery for electric vehicles. In Proceedings of
the 3rd World Conference on Mechanical Engineering and Intelligent Manufacturing (WCMEIM), Shanghai, China, 4–6 December
2020; pp. 492–496. [CrossRef]
17. He, H.; Xiong, R.; Fan, J. Evaluation of Lithium-Ion Battery Equivalent Circuit Models for State of Charge Estimation by an
Experimental Approach. Energies 2011, 4, 582–598. [CrossRef]
18. Sibi Krishnan, K.; Pathiyil, P.; Sunitha, R. Generic Battery model covering self-discharge and internal resistance variation. In
Proceedings of the IEEE 6th International Conference on Power Systems (ICPS), New Delhi, India, 4–6 March 2016; pp. 1–5.
[CrossRef]
19. Wu, W.; Qin, L.; Wu, G. State of power estimation of power lithium-ion battery based on an equivalent circuit model. J. Energy
20. Khattak, A.A.; Khan, A.N.; Safdar, M.; Basit, A.; Zaffar, N.A. A Hybrid Electric Circuit Battery Model Capturing Dynamic Battery
Characteristics. In Proceedings of the IEEE Kansas Power and Energy Conference (KPEC), Manhattan, KS, USA, 13–14 July 2020;
21. Mueller, K.; Schwiederik, E.; Tittel, D. Analysis of parameter identification methods for electrical Li-Ion battery modelling. In
Proceedings of the World Electric Vehicle Symposium and Exhibition (EVS27), Barcelona, Spain, 17–20 November 2013; pp. 1–9.
[CrossRef]
22. Meng, J.; Yue, M.; Diallo, D. Nonlinear extension of battery constrained predictive charging control with transmission of Jacobian
matrix. Int. J. Electr. Power Energy Syst. 2023, 146, 108762. [CrossRef]
23. Komal, S.; Kamyar, M.; Zunaib, A. Online reduced complexity parameter estimation technique for equivalent circuit model of
lithium-ion battery. Electr. Power Syst. Res. 2020, 185, 106356. [CrossRef]
24. Maletić, F.; Deur, J. Analysis of ECM-based Li-Ion Battery State and Parameter Estimation Accuracy in the Presence of OCV and
Polarization Dynamics Modeling Errors. In Proceedings of the IEEE 29th International Symposium on Industrial Electronics
(ISIE), Delft, The Netherlands, 17–19 June 2020; pp. 1318–1324. [CrossRef]
25. Simin, P.; Gang, S.; Yunfeng, C.; Xu, C. Control of different-rating battery energy storage system interface to a microgrid. Przeglad
˛
Elektrotechniczny (Electr. Rev.) 2011, 87, 256–262.
26. Meng., J.; Boukhnifer, M.; Diallo, D. Lihtium-ion battery monitoring and observability analysis with extended equivalent circuit
model. In Proceedings of the IEEE Mediterranean Conference on Control and Automation (MED), Saint-Raphaël, France, 15–18
September 2020; pp. 764–769.
27. Westerhoff, U.; Kurbach, K.; Lienesch, F.; Kurrat, M. Analysis of lithium-ion battery models based on electrochemical impedance
spectroscopy. Energy Technol. 2016, 4, 1620–1630. [CrossRef]
28. Stroe, D.I.; Swierczynski, M.; Stroe, A.I.; Knudsen Kær, S. Generalized Characterization Methodology for Performance Modelling
of Lithium-Ion Batteries. Batteries 2016, 2, 37. [CrossRef]
29. Cordoba-Arenas, A.; Onori, S.; Rizzoni, G.A. Control-oriented lithium-ion battery pack model for plug-in hybrid electric vehicle
cycle-life studies and systems design with consideration of health management. J. Power Sources 2015, 279, 791–808. [CrossRef]
30. Liaw, B.Y.; Nagasubramanian, G.; Jungst, R.G.; Doughty, D.H. Modeling of lithium ion cells—A simple equivalent-circuit model
approach. Solid State Ion. 2004, 175, 835–839. [CrossRef]
31. Huria, T.; Ceraolo, M.; Gazzarri, J.; Jackey, R. High fidelity electrical model with thermal dependence for characterization and
simulation of high power lithium battery cells. In Proceedings of the IEEE International Electric Vehicle Conference, Greenville,
SC, USA, 4–8 March 2012; pp. 1–8. [CrossRef]
32. Sockeel, N.; Shahverdi, M.; Mazzola, M.; Meadows, W. High-Fidelity Battery Model for Model Predictive Control Implemented
into a Plug-In Hybrid Electric Vehicle. Batteries 2017, 3, 13. [CrossRef]
33. Li, K.; Soong, B.H.; Tseng, K.J. A high-fidelity hybrid lithium-ion battery model for SOE and runtime prediction. In Proceedings
of the IEEE Applied Power Electronics Conference and Exposition (APEC), Tampa, FL, USA, 26–30 March 2017; pp. 2374–2381.
[CrossRef]
34. Hemi, H.; M’Sirdi, N.K.; Naamane, A.; Ikken, B. Open Circuit Voltage of a Lithium ion Battery Model Adjusted by Data Fitting.
In Proceedings of the 6th International Renewable and Sustainable Energy Conference (IRSEC), Rabat, Morocco, 5–8 December
2018; pp. 1–5.
35. Zhang, Q.; Shang, Y.; Li, Y.; Cui, N.; Duan, B.; Zhang, C. A novel fractional variable-order equivalent circuit model and parameter
identification of electric vehicle Li-ion batteries. ISA Trans. 2020, 97, 448–457. [CrossRef]
36. Baczyńska, A.; Niewiadomski, W.; Gonçalves, A.; Almeida, P.; Luís, R. Li-NMC Batteries Model Evaluation with Experimental
Data for Electric Vehicle Application. Batteries 2018, 4, 11. [CrossRef]
37. Somakettarin, N.; Funaki, T. Study on Factors for Accurate Open Circuit Voltage Characterizations in Mn-Type Li-Ion Batteries.
Batteries 2017, 3, 8. [CrossRef]
38. Gao, Y.; Ji, W.; Zhao, X. SOC Estimation of E-Cell Combining BP Neural Network and EKF Algorithm. Processes 2022, 10, 1721.
[CrossRef]
39. Rothenberger, M.J.; Docimo, D.J.; Ghanaatpishe, M.; Fathy, H.K. Genetic optimization and experimental validation of a test
cycle that maximizes parameter identifiability for a Li-ion equivalent-circuit battery model. J. Energy Storage 2015, 4, 156–166.
[CrossRef]
132
Energies 2023, 16, 6239
40. Nemes, R.; Ciornei, S.; Ruba, M.; Hedesiu, H.; Martis, C. Modeling and simulation of first-order Li-Ion battery cell with
experimental validation. In Proceedings of the 8th International Conference on Modern Power Systems (MPS), Cluj-Napoca,
Romania, 21–23 May 2019; pp. 1–6. [CrossRef]
41. Nemes, R.O.; Ciornei, S.M.; Ruba, M.; Martis, C. Parameters identification using experimental measurements for equivalent circuit
Lithium-Ion cell models. In Proceedings of the 11th International Symposium on Advanced Topics in Electrical Engineering
(ATEE), Bucharest, Romania, 28–30 March 2019; pp. 1–6. [CrossRef]
42. Hua, X.; Zhang, C.; Offer, G. Finding a better fit for lithium ion batteries: A simple, novel, load dependent, modified equivalent
circuit model and parameterization method. J. Power Sources 2021, 484, 229117. [CrossRef]
43. Li, Z.; Shi, X.; Shi, M.; Wei, C.; Di, F.; Sun, H. Investigation on the Impact of the HPPC Profile on the Battery ECM Parameters’
Offline Identification. In Proceedings of the Asia Energy and Electrical Engineering Symposium (AEEES), Chengdu, China, 28–31
May 2020; pp. 753–757. [CrossRef]
44. Haghjoo, Y.; Khaburi, D.A. Modeling, simulation, and parameters identification of a lithium-ion battery used in electric vehicles.
In Proceedings of the 9th Iranian Conference on Renewable Energy & Distributed Generation (ICREDG), Mashhad, Iran, 23–24
February 2022; pp. 1–7. [CrossRef]
45. Tran, M.K.; Mathew, M.; Janhunen, S.; Panchal, S.; Raahemifar, K.; Fraser, R.; Fowler, M. A comprehensive equivalent circuit
model for lithium-ion batteries, incorporating the effects of state of health, state of charge, and temperature on model parameters.
46. Deng, S.D.; Liu, S.Y.; Wang, L.; Xia, L.L.; Chen, L. An improved second-order electrical equivalent modeling method for the
online high power Li-ion battery state of charge estimation. In Proceedings of the IEEE 12th Energy Conversion Congress &
Exposition—Asia (ECCE-Asia), Singapore, 24–27 May 2021; pp. 1725–1729. [CrossRef]
47. Parthasarathy, C.; Laaksonen, H.; Halagi, P. Characterisation and Modelling Lithium Titanate Oxide Battery Cell by Equivalent
Circuit Modelling Technique. In Proceedings of the IEEE PES Innovative Smart Grid Technologies—Asia (ISGT Asia), Brisbane,
Australia, 5–8 December 2021; pp. 1–5. [CrossRef]
48. Navas, S.J.; Cabello González, G.M.; Pino, F.J.; Guerra, J.J. Modelling Li-ion batteries using equivalent circuits for renewable
energy applications. Energy Rep. 2023, 9, 4456–4465. [CrossRef]
49. Wang, J.; Jia, Y.; Yang, N.; Lu, Y.; Shi, M.; Ren, X.; Lu, D. Precise equivalent circuit model for Li-ion battery by experimental
improvement and parameter optimization. J. Energy Storage 2022, 52, 104980. [CrossRef]
50. Sörés, M.A.; Hartmann, B. Overview of possible methods of determining self-discharge. In Proceedings of the IEEE Interna-
tional Conference on Environment and Electrical Engineering and IEEE Industrial and Commercial Power Systems Europe
(EEEIC/I&CPS Europe), Madrid, Spain, 9–12 June 2020; pp. 1–5. [CrossRef]
51. Tang, A.; Gong, P.; Li, J.; Zhang, K.; Zhou, Y.; Zhang, Z. A State-of-Charge Estimation Method Based on Multi-Algorithm Fusion.
World Electr. Veh. J. 2022, 13, 70. [CrossRef]
52. Jarrraya, I.; Degaa, L.; Rizoug, N.; Chabchoub, M.H.; Trabelsi, H. Comparison study between hybrid Nelder-Mead particle swarm
optimization and open circuit voltage—Recursive least square for the battery parameters estimation. J. Energy Storage 2022,
50, 104424. [CrossRef]
53. Castanho, D.; Guerreiro, M.; Silva, L.; Eckert, J.; Antonini Alves, T.; Tadano, Y.d.S.; Stevan, S.L., Jr.; Siqueira, H.V.; Corrêa, F.C.
Method for SoC Estimation in Lithium-Ion Batteries Based on Multiple Linear Regression and Particle Swarm Optimization.
54. Pizarro-Carmona, V.; Castano-Solís, S.; Cortés-Carmona, M.; Fraile-Ardanuy, J.; Jimenez-Bermejo, G. GA-based approach to
optimize an equivalent electric circuit model of a Li-ion battery-pack. Expert Syst. Appl. 2021, 172, 114647. [CrossRef]
55. Huang, Y.; Li, Y.; Jiang, L.; Qiao, X.; Cao, Y.; Yu, J. Research on Fitting Strategy in HPPC Test for Li-ion battery. In Proceedings of
the IEEE Sustainable Power and Energy Conference (iSPEC), Beijing, China, 21–23 November 2019; pp. 1776–1780. [CrossRef]
56. Wang, C.; Xu, M.; Zhang, Q.; Feng, J.; Jiang, R.; Wei, Y.; Liu, Y. Parameters identification of Thevenin model for lithium-ion
batteries using self-adaptive Particle Swarm Optimization Differential Evolution algorithm to estimate state of charge. J. Energy
57. Hamida, M.A.; El-Sehiemy, R.A.; Ginidi, A.R.; Elattar, E.; Shaheen, A.M. Parameter identification and state of charge estimation of
Li-Ion batteries used in electric vehicles using artificial hummingbird optimizer. J. Energy Storage 2022, 51, 104535. [CrossRef]
58. Szewczyk, P.; Łebkowski, A. Comparative Studies on Batteries for the Electrochemical Energy Storage in the Delivery Vehicle.
59. Łebkowski, Ł. Temperature, Overcharge and Short-Circuit Studies of Batteries used in Electric Vehicles. Prz. Elektrotechniczny
2017, 93, 67–73. [CrossRef]
60. Ohneseit, S.; Finster, P.; Floras, C.; Lubenau, N.; Uhlmann, N.; Seifert, H.J.; Ziebert, C. Thermal and Mechanical Safety Assessment
of Type 21700 Lithium-Ion Batteries with NMC, NCA and LFP Cathodes–Investigation of Cell Abuse by Means of Accelerating
Rate Calorimetry (ARC). Batteries 2023, 9, 237. [CrossRef]
61. Kiemel, S.; Glöser-Chahoud, S.; Waltersmann, L.; Schutzbach, M.; Sauer, A.; Miehe, R. Assessing the Application-Specific
Substitutability of Lithium-Ion Battery Cathode Chemistries Based on Material Criticality, Performance, and Price. Resources 2021,
10, 87. [CrossRef]
62. Forte, F.; Pietrantonio, M.; Pucciarmati, S.; Puzone, M.; Fontana, D. Lithium Iron Phosphate Batteries Recycling: An Assessment
of Current Status. Crit. Rev. Environ. Sci. Technol. 2021, 51, 2232–2259.
133
Energies 2023, 16, 6239
63. Białoń, T.; Niestrój, R.; Korski, W. PSO-Based Identification of the Li-Ion Battery Cell Parameters. Energies 2023, 16, 3995.
[CrossRef]
64. Belt, J.R. Battery Test Manual for Plug-In Hybrid Electric Vehicles, 2nd ed.; U.S. Department of Energy Vehicle Technologies Program:
Idaho Falls, ID, USA, 2010. [CrossRef]
65. Yang, Z.; Wang, X. An improved parameter identification method considering multi-timescale characteristics of lithium-ion
batteries. J. Energy Storage 2023, 59, 106462. [CrossRef]
66. Karimi, D.; Behi, H.; Van Mierlo, J.; Berecibar, M. Equivalent Circuit Model for High-Power Lithium-Ion Batteries under High
Current Rates, Wide Temperature Range, and Various State of Charges. Batteries 2023, 9, 101. [CrossRef]
67. Guenther, C.; Barillas, J.K.; Stumpp, S.; Danzer, M.A. A dynamic battery model for simulation of battery-to-grid applications. In
Proceedings of the 3rd IEEE PES Innovative Smart Grid Technologies Europe (ISGT Europe), Berlin, Germany, 14–17 October
2012; pp. 1–7. [CrossRef]
68. Shi, J.; Guo, H.; Chen, D. Parameter identification method for lithium-ion batteries based on recursive least square with sliding
window difference forgetting factor. J. Energy Storage 2021, 44, 103485. [CrossRef]
69. Tran, M.-K.; DaCosta, A.; Mevawalla, A.; Panchal, S.; Fowler, M. Comparative Study of Equivalent Circuit Models Performance in
Four Common Lithium-Ion Batteries: LFP, NMC, LMO, NCA. Batteries 2021, 7, 51. [CrossRef]
70. Feng, D.; Huang, J.; Jin, P.; Chen, H.; Wang, A.; Zheng, M. Parameter Identification and Dynamic Simulation of Lithium-Ion
Power Battery Based on DP Model. In Proceedings of the 14th IEEE Conference on Industrial Electronics and Applications
(ICIEA), Xi’an, China, 19–21 June 2019; pp. 1275–1279. [CrossRef]
71. Einhorn, M.; Conte, V.F.; Kral, C.; Fleig, J.; Permann, R. Parameterization of an electrical battery model for dynamic system
simulation in electric vehicles. In Proceedings of the IEEE Vehicle Power and Propulsion Conference, Lille, France, 1–3 September
2010; pp. 1–7. [CrossRef]
72. Yu, Q.; Wan, C.; Li, J.; E, L.; Zhang, X.; Huang, Y.; Liu, T. An Open Circuit Voltage Model Fusion Method for State of Charge
Estimation of Lithium-Ion Batteries. Energies 2021, 14, 1797. [CrossRef]
73. Gao, L.; Liu, S.; Dougal, R.A. Dynamic lithium-ion battery model for system simulation. IEEE Trans. Compon. Packag. Technol.
2002, 25, 495–505. [CrossRef]
74. Wen, F.; Duan, B.; Zhang, C.; Zhu, R.; Shang, Y.; Zhang, J. High-Accuracy Parameter Identification Method for Equivalent-Circuit
Models of Lithium-Ion Batteries Based on the Stochastic Theory Response Reconstruction. Electronics 2019, 8, 834. [CrossRef]
75. Baccouche, I.; Jemmali, S.; Manai, B.; Omar, N.; Amara, N.E.B. Improved OCV Model of a Li-Ion NMC Battery for Online SOC
Estimation Using the Extended Kalman Filter. Energies 2017, 10, 764. [CrossRef]
76. Pillai, P.; Sundaresan, S.; Kumar, P.; Pattipati, K.R.; Balasingam, B. Open-Circuit Voltage Models for Battery Management Systems:
A Review. Energies 2022, 15, 6803. [CrossRef]
77. Shaheen, A.M.; Hamida, M.A.; El-Sehiemy, R.A.; Elattar, E.E. Optimal parameter identification of linear and non-linear models
for Li-Ion Battery Cells. Energy Rep. 2021, 7, 7170–7185. [CrossRef]
78. Plett, G.L. High-performance battery-pack power estimation using a dynamic cell model. IEEE Trans. Veh. Technol. 2004, 53,
1586–1593. [CrossRef]
79. Marušić, D.; Vašak, M. Efficient Method of Identifying a Li-Ion Battery Model for an Electric Vehicle. In Proceedings of the
IEEE 20th International Power Electronics and Motion Control Conference (PEMC), Brasov, Romania, 25–28 September 2022;
134
energies
Article
A Model-Aware Comprehensive Tool for Battery Energy Storage
System Sizing
Matteo Spiller 1 , Giuliano Rancilio 1 , Filippo Bovera 1 , Giacomo Gorni 2 , Stefano Mandelli 3 , Federico Bresciani 2
and Marco Merlo 1, *
1 Politecnico di Milano—Department of Energy, Via Lambruschini 4a, 20156 Milano, Italy;

matteo.spiller@polimi.it (M.S.); giuliano.rancilio@polimi.it (G.R.); filippo.bovera@polimi.it (F.B.)
2 Eni S.p.A., Renewable, New Energies and Material Science Research Center, Via Fauser 4, 28100 Novara, Italy;
giacomo.gorni@eni.com (G.G.); federico.bresciani@eni.com (F.B.)
3 Plenitude, Via Giuseppe Ripamonti 85, 20141 Milano, Italy; stefano.mandelli@plenitude.com
* Correspondence: marco.merlo@polimi.it
Abstract: This paper presents a parametric procedure to size a hybrid system consisting of renewable
generation (wind turbines and photovoltaic panels) and Battery Energy Storage Systems (BESS). To
cope with the increasing installation of grid-scale BESS, an innovative, fast and flexible procedure
for evaluating an efficient size for this asset has been developed. The tool exploits a high-fidelity
empirical model to assess stand-alone BESS or hybrid power plants under different service stacking
configurations. The economic performance has been evaluated considering the revenue stacking
that occurs when participating in up to four distinct energy markets and the degradation of the
BESS performances due to both cycle- and calendar-aging. The parametric nature of the tool enables
the investigation of a wide range of system parameters, including novel BESS control logic, market
prices, and energy production. The presented outcomes detail the techno-economic performances of a
hybrid system over a 20-year scenario, proposing a sensitivity analysis of both technical and economic
parameters. The case study results highlight the necessity of steering BESS investment towards the
coupling of RES and accurate planning of the service stacking. Indeed, the implementation of a
storage system in an energy district improves the internal rate of return of the project by up to 10% in
the best-case scenario. Moreover, accurate service stacking has shown a boost in revenues by up to
44% with the same degradation.
Citation: Spiller, M.; Rancilio, G.;
Bovera, F.; Gorni, G.; Mandelli, S.;
Keywords: battery energy storage system; renewables; market service stacking
Bresciani, F.; Merlo, M. A
Model-Aware Comprehensive Tool
for Battery Energy Storage System
Sizing. Energies 2023, 16, 6546.
https://doi.org/10.3390/en16186546 1. Introduction
Academic Editor: Quanqing Yu

The energy sector is responsible for a large share of anthropogenic carbon emissions
that lead to climate change and global warming. The 2015 United Nations Climate Change
Received: 3 July 2023 Conference in Paris set the milestone of limiting the average temperature increase to below
Revised: 16 August 2023 1.5 ◦ C [1]. A total of 160 countries around the world have agreed to combat global heating
Accepted: 25 August 2023 through the installation of renewable energy sources (RES) for a more sustainable energy
Published: 12 September 2023 scenario. Consequently, the overall energy production by RES has increased from 20 to 28%
since 2010, and the target is for production to reach at least 43% by 2030 [2]. Focusing on
the European Union (EU), the EU Green Deal and the “Fit-for-55” package set the goal is of
reducing greenhouse gas emissions by 55% with respect to 1990 by 2030 [3].
RES have a low carbon footprint and are, therefore, among the main candidates
for energy sources that could be used to reach these decarbonization targets. In any
case, their integration into the energy system is not straightforward. These sources are
conditions of the Creative Commons characterized by the high intermittency and non-programmability of the energy output,
Attribution (CC BY) license (https:// which complicates the balance between supply and demand in the power system [4]. Grid-
creativecommons.org/licenses/by/ connected battery energy storage systems (BESS) represent a viable resource to cope with
4.0/). those issues and guarantee the balance, stability, and adequacy of a decarbonizing power

135
Energies 2023, 16, 6546
system. Compared to other storage systems, BESS are close to market maturity, with prices
dropping by 87% from 2010 to 2019 [5]. The installed grid-scale battery storage capacity
will expand 44-fold between 2021 and 2030 to 680 GW according to IEA [6]. In addition,
lithium-ion batteries outperform other storage technologies in terms of energy density,
power density, and round-trip efficiency. Moreover, their operational reliability can last
up to 20 years with suitable management strategies, such as temperature regulation and
capacity augmentation [7]. The success of this technology is attributed to the flexibility
and scalability of these assets, coupled with their ability to behave as bulk energy systems.
These characteristics enable the provision of a wide set of services to system operators
and active users. Such services could reduce RES uncertainty and make BESS attractive
investments.
Despite the possible enhancement that BESS could provide to electric power system
operation, this technology still has shortcomings. Although battery prices are decreasing
yearly, the overall cost per kWh is still high, representing the most common challenge.
To cope with high capital costs, it is crucial to capture multiple cash flows to increase the
financial viability of the project. The dynamic stacking of BESS services ensures higher
profits for the asset [8]. Therefore, it is of paramount importance to evaluate an efficient
system operation that creates multiple streams of revenues to increase the economic benefits.
To properly assess and optimize the cash flow, the owner of the system must optimally
size the BESS, accounting for several economic aspects such as cell technology, installation
cost, and maintenance cost. These expenditures are mainly affected by lifetime, battery
capacity, and overall performance, which interact in a non-linear manner. The proper
evaluation of these aspects allows for more accurate modeling and, consequently, a more
precise economic analysis.
High-fidelity BESS modelization is mandatory to ensure accurate economic evaluation.
This paper proposes a model-aware BESS-sizing procedure that accurately represents the
performance of BESS in different energy markets during their lifetime, accounting for the
main non-linearities. In general, the stacking of the services is mostly addressed by constant
BESS models that do not consider the non-linearities of this technology and the presence of
auxiliaries. Furthermore, service stacking with grid-scale storage is mainly investigated in
a stand-alone configuration, without evaluating the possible services that the system can
provide to an RES power plant. Lastly, services such as the capacity market have not been
assessed to date. Based on the state-of-art described in the next section, the novelties of this
work are as follows:
• A sizing procedure is developed that investigates a 20-year BESS investment with a
high-fidelity empirical model developed in [9] and updated with equations capable
of emulating the capacity degradation of the system. The latter structure has been
exploited to create a fast and flexible tool that is able to evaluate the most cost-effective
storage investment, ensuring an efficient trade-off between computational effort and
accuracy.
• Innovative algorithms are developed that are capable of stacking multiple services
with a sequential approach. The implemented solutions produce results seamlessly,
with two distinct configurations: stand-alone and hybrid-renewable power plants.
The work is structured as follows. Section 2 reviews the modelization and the algo-
rithms exploited for sizing BESS in the literature. Section 3 describes the proposed empirical
model, the methodology of the sizing procedure, and the novel algorithm proposed for
stacking the energy markets. Section 4 introduces the study cases. Section 5 discusses the
main results. Lastly, Section 6 summarizes the activities and lists future works.
2. Literature Review
The BESS sizing procedure consists of identifying the most cost-effective configuration
for the stakeholders. The application is complex and non-linear. This section aims to
describe two different fundamental aspects of the procedure: the modeling, and solution
methods [10].
136
Energies 2023, 16, 6546
2.1. Modeling
Modeling focuses on the mathematical representation of the key components of BESS.
A wide number of approaches have been developed, with different levels of complexity
and computational effort. According to the degree of physical insight, battery models can
be divided into three different levels: electrochemical model, circuit-oriented (or electrical)
model, and black box models (empirical or stochastic) [11]. The selection of a model is
bounded to the application area. Namely, a specific model ensures a different degree of
accuracy and computational effort, and depending on the details required by the imple-
mentation, an efficient balance between the two characteristics improves the quality of
the analysis. For instance, electrochemical models are the most accurate approach to bat-
tery representation. They describe the chemical reactions that take place in the electrodes
and the electrolytes using a set of non-linear differential equations [12]. The expressions
detail the effect of the electrochemical reactions, such as the diffusion, migration, kinetic
phenomena, and lithium concentration, to truly represent the state variables of the sys-
tem [13]. A description of the batteries at a microscopic scale accurately represents the
key behaviors of the nonlinear system [14]. Although there is no doubt regarding the
accuracy of electrochemical models, they are used in low-speed applications, such as online
capacity estimation [15,16], predictive maintenance [17] or the validation of complemen-
tary models [18]. Circuit-oriented models are electrical equivalent models that are able
to represent the state variables of the batteries [11]. Electrical models consist of electrical
circuits made by capacitors and resistances, whose proper connection emulates the behavior
of the batteries. The basic electrical model, known as Rint, has a big capacitor that can
describe the open-circuit voltage of the cell, and a series resistance that can simulate the
battery’s internal resistance [19]. Moreover, RC circuits can be connected in series to the
Rint model to increase the order of the system and emulate relaxation and polarization
effects [20]. Despite the empirical nature of those elements, these modelizations are widely
adopted thanks to their computational efficiency in capturing the dynamic response of the
system [21]. The ability to estimate the state of batteries in real-time makes these models
suitable for applications such as EV state estimation [20], and grid stability [22,23]. Further-
more, electrical models ensure a good performance when evaluating the terminal voltage
and SOC of the batteries. Articulated modelization, such as a three-order model, allows
for these state variables to be described with errors lower than 1% [21,24]. However, if the
accuracy of the model increases, the same occurs regarding the computational effort [25].
In sizing applications, computational efficiency is a priority since it requires the evaluation
of steady-state performances over a long time-window (e.g., 15–20 years). Therefore, the
complexity of the electrical model could not produce results that are valuable for the pro-
cedure in a reasonable time. Empirical models ensure the optimal computational effort
for this kind of analysis. The modelization is characterized by mathematically constructed
models that utilize the observed data and measurements to represent the behavior and
performance of BESS systems. These models are developed by an analysis of real-world
operational data and the characteristics gathered from datasheets or experiments, allowing
for them to capture the key relationships and patterns between various parameters [26,27].
By leveraging statistical techniques, regression analyses, or other mathematical approaches,
empirical models provide a parametric representation of how different factors, such as
battery lifetime, efficiency, and capacity, interact and impact the overall performance of
BESS systems [9].
A fair share of BESS sizing procedures are based on scalar linear empirical systems that
resemble the BESS performances. In [28], a constant battery and inverter efficiency model
has been exploited to size a PV residential system by evaluating the economic criteria of
the annuity method. Study [29] investigates the optimal sizing of BESS through a life-cycle
cost model. This model inspects the different phases of the system, also accounting for
the final decommissioning, maintenance and recycling, and disposal as a reduction in
the economics of the system. Ref. [30] proposes a constant-efficiency empirical model to
size an energy storage system, accounting for hosting capacity and reductions in wind
137
Energies 2023, 16, 6546
curtailment. Although these simplified empirical models represent the easiest approach
to configuring and producing results for BESS, they are generally the least accurate [31].
Nevertheless, high-fidelity empirical models that represent varying efficiencies and power
capabilities depending on SoC and C-rate improve the accuracy of the analysis [32]. These
approaches rely on experimental campaigns to create a computationally efficient BESS
model that contains an error. In [33], a regression technique has been exploited to build
a non-linear BESS with an average SOC root mean square error of 3%. In [34], a detailed
non-linear power losses model has been implemented to more precisely capture the low-
efficiency working region of the storage system. In [9], a grid-scale BESS has been modeled
with lookup tables to represent the non-linear efficiency and the auxiliary consumption
of the system, ensuring an average SOC error of 0.168%. Despite the high fidelity of the
modelization, the latter applications are limited to the operation perspective, and no one
has used the accuracy of the model to size the BESS.
2.2. Solution Method

The solution method consists of the approach used to investigate the optimal size of
the BESS, i.e., the mathematical procedure adopted to set the schedule and the dispatching
of different services. The goal of the algorithm selected in the sizing procedure is to identify
the best power setpoint to efficiently maximize the benefits generated by the BESS. The
algorithm needs to emulate the provision of diverse and multiple services by the BESS.
Namely, at each time step, given the actual SOC, the storage system is charged or discharged
according to a logic that improves the welfare of the BESS owner. Most existing studies
are based on mathematical programming due to their ability to identify a global optimum
for the objective function under analysis. Stochastic Dynamic Programming (DP), and
stochastic Mixed-Integer Linear Programming (MILP) or techniques derived from these
two approaches are the most-adopted solutions to the sizing problem [35]. For instance,
studies [36,37] propose a stochastic MILP and DP, respectively, to optimize the sizing
of a grid-scale storage system with constant efficiency. However, these methodologies
show important limitations in terms of the modeling and objectivity of the solution [38].
The optimal sizing problem is a non-convex and non-linear combinatorial optimization
problem [39]. Therefore, MILP-sizing procedures need to rely on techniques such as
relaxation, piecewise linear approximation, or the implementation of a constant efficiency
to generate a problem that can be treated by the standard solvers [32,40,41]. However, DP is
afflicted by dimensionality problems that need to be solved with the so-called approximated
dynamic programming [42,43]. Although these techniques solve the main issues related to
computational effort, their implementation comes at the expense of accuracy. Moreover,
sizing problems are computationally intensive, and many works consider a timeframe that
does not evaluate the whole investment lifespan. Lastly, to properly manage technical
and economic targets, multi-objective functions are typically required [44]. These lead to
a lack of objectivity in the solutions, which needs to be properly evaluated with Pareto
analysis [45]. It follows that mathematical programming cannot provide diversified system
operation. However, without the support provided by these techniques, it is necessary
to develop algorithms that are capable of efficiently scheduling the assets. Furthermore,
considering the actual cost of BESS, the profits from multiple applications are fundamental
to generating a positive investment. Therefore, algorithms capable of effectively stacking
the services are necessary to improve the economics of a storage system. Implementing
those algorithms is not straightforward, since BESS are limited in power and energy. The
desired output of the service stacking problem is a strategy for optimal capacity allocation
during a given period, accounting for market prices and system dynamics. Three distinct
types of service stacking are defined in the literature: sequential, parallel, and dynamic [46].
In parallel stacking, a constant allocation of storage capacity is given to the services whilst
the sequential multi-use provides these services in turn. The dynamic multi-use aims to
increase profit as it combines the advantages of the two predecessors. It follows that service
stacking in sizing applications is a complex topic and aspects such as the type of stacking
138
Energies 2023, 16, 6546
or the services’ optimum sizing represent important challenges for the stakeholder [47].
Furthermore, power plant configuration has a relevant role in the BESS-sizing procedure
due to the different services that this technology can provide. For instance, a BESS coupled
with an energy district could perform tasks such as load peak reductions and smooth power
injections. Vice versa, in a stand-alone setup, the main activities are exclusively grid- and
market-oriented. BESS-sizing approaches have been categorized into four main categories
to deal with different configurations: microgrids, distributed renewable energy systems,
standalone hybrid renewable energy systems, and renewable energy power plants [48].
Despite the wide interest in BESS-sizing in the literature, the proposed approaches typically
focus on a single configuration. Conversely, industrial and utility-scale BESS stakeholders
are interested in flexible tools that are able to evaluate heterogeneous configurations of
power plants, storage systems, markets, and services.
Table 1 lists the studies that have addressed the stacking of services for a grid-scale
BESS.
Table 1. List of studies that have inspected the stacking of services for a grid-scale BESS.
Performance
Reference Services Configuration Scope
Assumptions
Efficiency as a function of SOC
[49] Fast Frequency—Balancing market Stand-alone Operation
and power
[50] Arbitrage—Frequency regulation Constant efficiency Stand-alone Operation
[51] Arbitrage—Frequency regulation Constant efficiency Stand-alone Operation
[52] Arbitrage—Frequency regulation- Constant efficiency Stand-alone Sizing
[53] Power shifting—aFRR Constant efficiency Wind farm coupling Operation
[54] Ancillary services market Constant efficiency Stand-alone Operation
Arbitrage—distribution investment
[55] Constant efficiency Stand-alone Operation
deferral—frequency regulation
Non-linear power losses
[34] Arbitrage—Frequency regulation Stand-alone Operation
depending on C-rate and SOC
Frequency regulation—power
[56] Constant efficiency Microgrid Sizing
shifting
Arbitrage—Frequency
Efficiency as a function of SOC Stand-alone and
This work regulation—mFRR— Sizing
and power PV–wind coupled
capacity market
As stated in the Introduction, this work improves on the state-of-the-art by developing

a high-fidelity model of BESS and novel stacking algorithms, considering both stand-alone
and RES-coupled operations.
3. Methodology
This work proposes a novel flexible tool evaluating the size of a BESS and computing
its cost-effectiveness while providing multiple services. The exploitation of the tool has been
eased thanks to a graphic user interface (GUI) developed in MATLAB TM . The structure
allows for the easy evaluation of candidate groups of BESS and identifies the configuration
that ensures the best economic return. The GUI compares different specific BESS sizes to
accurately assess the trade-off between the size and the cost. The procedure is designed to
be compatible with the generic EU market (e.g., it implements standard balancing products,
and considers the day-ahead market according to the EU framework), meaning that it can
be easily adapted to different requirements. However, the focus of the proposed algorithm
in this paper is specifically related to the Italian market (e.g., the ancillary services market
price scenarios and award rates, a well as the rules and prizes of the capacity remuneration
mechanism, suit the Italian case).
This procedure may investigate both stand-alone and BESS hybrid-renewable en-
ergy power plants with wind and photovoltaic services connected to the national grid.
Furthermore, the power managed at the point of delivery (POD) with the grid can be
139
Energies 2023, 16, 6546
limited to a specified value to address study cases where there is an existing contractual
connection power, and the curtailment of the overproduction is foreseen. The proxy model
adopted in the procedure is a high-fidelity (cf. realistic) empirical battery model based
on an experimental campaign at the Joint Research Centre (JRC) of Ispra (VA-Italy) on
nickel–manganese cobalt (NMC) BESS [9]. The main features are the efficiency of the overall
system, including transformer and power-conversion systems, and the auxiliary consump-
tion. The nonlinear round-trip efficiency is expressed as a function of power and SOC
through look-up tables. Instead, the auxiliaries’ consumption relies on the power flown in
the BESS and the ambient temperature. Although the model emulates the performances of
NMC technology, its structures easily allow for the fitting of the data of different electro-
chemical technologies (e.g., lithium iron phosphate or even non-lithium-based batteries).
Figure 1 depicts the block diagram of the high-fidelity empirical model adopted in this
work using SI MULI NK TM . Each block has a specific purpose that ensures the emulation
of a large-scale BESS. The overall efficiency block takes the AC power, and the SOC as input
and converts them into DC power using the previously mentioned look-up table. After
this, the DC input is processed by the capability curve that limits the C-rate depending on
the actual SOC. Lastly, the actual C-rate is used to update the SOC of the system. Inside the
latter block, the energy content of the system is evaluated at each time sample.
Figure 1. Block diagram of the sizing tool developed in this work.
To ensure a realistic representation of BESS performances, an aging model from the

literature was inserted into this framework. It is well-known that capacity degradation
is affected by various chemical reactions, which can typically be classified into two major
phenomena: calendar and cycle aging [57]. To account for these processes, two equations
that describe the cycle and the calendar aging of the system have been inserted into the SOC
update block to reduce the nominal capacity throughout the simulation. Both expressions
were obtained from the literature. The cycle aging is related to the use of the BESS and its
C-rate. The complete expression implemented in the model was obtained in [58].
Cycledegrad.[%] = −[SOC(t) − SOC(t − 1)] × 3.57 × 10−5 × e0.465×C-rate (1)
Equation (1) describes the system degradation due to cycling by evaluating the depth
of discharge between each time sample [SOC(t) − SOC(t − 1)] and the specific C-rate at
which the BESS is cycled. The proposed tool investigates the BESS operation with a 15-min
granularity, over a very long time window of up to several years; consequently, the equation
140
Energies 2023, 16, 6546
computes the percentage of capacity reduction generated in 15 min by a specific constant

C-rate.
The second degradation mechanism has been modeled as a constant-capacity degra-
dation in the square root of time function. This trend has been widely adopted to describe
the chemical reaction inside the system that occurs due to calendar aging. The expression
adopted in this work is based on [59], assuming an average SOC of 50% for the system.
− 42577 √
Calendardegrad. = −99.43 × 103 × e RT
× t (2)
Equation (2) reports the capacity calendar degradation depending on the age of
the storage system, considering the gas constant R and the temperature T at which the
system is kept by the auxiliaries. It has been assumed that the two effects superimpose
to continuously update the SOH of the system. Therefore, at each time sample the BESS’
nominal energy is reduced depending on the C-rate and the time that has elapsed.
The left side of Figure 1 reports the possible power plant setup evaluated by the tool.
The inputs for the analysis are time series that can be ascribed into three different categories:
energy, market, and ambient inputs. An additional block considering the optional presence
of RES plants was also included to investigate hybrid configurations. Energy input returns
the MW production of a given RES plant, if present. Market input contains fundamental
information about the structure of each service being addressed. Lastly, temperature inputs
are necessary for the auxiliary consumption of the BESS model. These parameters can be
easily changed thanks to the GUI, which allows for the selection of the Excel file that is
necessary for the analysis.
The evaluation of the nominal energy of the BESS depending on a given set of inputs
is evaluated by the tool with iterations of different sizes through the definition of two
distinct sets of parameters: nominal power and energy-to-power ratio (EPR). Moreover, if
the storage system is coupled with RES, the tool can evaluate variable power plant sizes.
The cost-effectiveness of BESS relies on the profits that the asset can generate by
participating in different services. Therefore, different algorithms were developed to
emulate the participation of the BESS in various energy markets. Each algorithm elaborates
the inputs and returns a power setpoint that simulates the dispatching of the BESS power
flows or the possibility of charging using the assets in the energy district.
The algorithms proposed in this paper are analytical adaptive algorithms. Every
day, the input data are fed to the algorithms that compute the operation of the BESS as
a combination of charge and discharge signals. The simulation calculates the cash flow
for twenty years for each size. The outcomes of each analysis are exploited to compute
the performance indicators fundamentals to determine the optimal size of the system.
The tool evaluates the internal rate of return (IRR) and the CAPEX that is covered to
quantify the cost-effectiveness of each investment. The IRR was computed with the financial
toolbox implemented in MATLAB TM . Instead, the CAPEX covering 20 years was obtained
following Equation (3), using the constant interest rate r.
100 20
Revenues(t) − OPEX(t) − Marketpenalties
CAPEXcovered [%] = ×∑ (3)
CAPEX t=1 (1 + r)t
CAPEX and OPEX are properly addressed in the analysis. Equation (4) describes the
capital cost as a product of the nominal energy and an exponential function of the duration.
This expression was obtained through an interpolation of the values reported in [60].

CAPEX[€] = Enom 220 × duration−0.9795 + 287.1 (4)
This condition is justified by the fact that equipment cost is strictly correlated with the
power being handled and not only with the nominal energy. The hypothesis adopted for
this work is that yearly OPEX corresponds to 2.5% of the CAPEX [61].
141
Energies 2023, 16, 6546
Lastly, due to the novelties of the application, a validation procedure based on previous
works has not been added. Nevertheless, the high-fidelity empirical model represents
an updated version of a previously developed structure. In [62], the adopted model has
been compared with state-of-the-art models: the accuracy of performance representation
increases when using a variable BESS efficiency and considering auxiliary system demand.
The algorithms implemented in the tools aim to emulate participation in different
services. The next paragraphs briefly describe the logic adopted for their implementation.
Four services typically present in EU markets have been modeled: energy arbitrage, capac-
ity market participation, the provision of manual Frequency Restoration Reserve (mFRR),
and the provision of fast frequency regulation within the Fast Reserve (FR) project. Each
algorithm generates a signal that simulates the charge and discharge requirement for the
BESS that participates in the specific service.
3.1. Energy Arbitrage

The term energy arbitrage refers to the possibility of storage systems exploiting day-
ahead market spread to generate profits. To achieve this control, the algorithm identifies the
minimum and maximum prices of the market each day to set up the proper control logic.
The algorithm was limited to a cycle per day to cope with the uncertainty of the prices that
are always present in a multi-year analysis. To grant an economic profit, the procedure has
a feedback control that checks whether the cycling cost of the system cancels the profits
per cycle. Namely, this cost can be described as the total lifetime cost of the investment
in electricity storage, divided by the cumulative delivered electricity. This parameter is
customizable from the GUI and impacts the arbitrage provision. Furthermore, in case of
BESS coupled with RES, if the RES power plant produces a power higher than the POD
limit (i.e., the contractual connection power), the algorithm directs the exceeding generation
toward the BESS if it is not fully charged. The latter charging process is considered free of
charge since that energy will be curtailed if not stored. To properly address this aspect, an
SOC control is implemented inside the algorithm to check the boundary condition of the
storage system. Lastly, it has been hypothesized that the BESS acts as a price-taker in the
day-ahead market. Algorithm 1 describes the structure of the energy arbitrage algorithm in
detail.
3.2. Capacity Market

The electricity capacity market is coupled with the energy market. The scope of this
service is to ensure power plant owners have sufficient capacity to meet the system demands
and reliability. Participation in the market is defined through a tender procedure. The
winners receive the capacity payment if they correctly provide power during the scarcity
hours defined by the authority. In Italy, the assets that won the tender are entitled to the
capacity payment if they participate in the day-ahead market in 1000 mandatory hours.
The capacity payment differs from each power plant and is based on the probability of
derating a given technology. Namely, the Italian authority associated a parameter with each
type of power plant, related to their reliability in producing the nominal power [63]. This
value reduces the capacity payment depending on the characteristics of each technology.
For instance, BESS derating is shown in Table 2 and is proportional to the duration of
the storage system, which means that a larger duration ensures better reliability in the
provision of the capacity. Thus, the remunerated power of a BESS is its qualified power
times (1—derating).
142
Energies 2023, 16, 6546
Algorithm 1: Energy Arbitrage

prod
Input: Pi , Pnom , Enom , PODlimit , PDAM
i , SOCinitial
0
Output: Pcha , Pdis
CHARGE PHASE
for I in 96 (quarter of hours in a day) do
prod
if Pi > PODlimit
prod
if Pi − PODlimit < Pnom
prod
Pcha
i = Pi − PODlimit
Else
Pcha
i = Pnom
Pcha
Echa = Echa + i
4
If sum(E cha < Enom
i
j = find min PMGPi
for i in 96 (quarter of hours in a day) do
if i == j
prod
if Pi < Pnom
prod
Pcha
i = Pi
Else
Pcha
i = Pnom
Pcha
E = Echa + i4
cha
DISCHARGE PHASE
Edis = Echa
j = find max PMGP i
while Edis > 0
if i == j
prod
if Pnom
i + Pi > PODlimit
prod
i = PODlimit − Pi
Pdis
Else
i =P
Pdis nom
Pdis
Edis = Edis − 4i
SOC CONTROL
Pdis Pcha
SOCi = SOCi−1 − 4i + i4
If SOCi > 100
SOCi = 100
−SOCi−1
Pcha
i = 1004*E nom
Else if SOCi < 0

SOCi = 0
Pcha
i = SOC i−1
4*Enom
PROFITS PER CYCLE CONTROL
prod
If Pi > PODlimit
prod
P = min Pnom , Pi
chargefree
− PODlimit
charge
Pchargepaid = Pi − Pchargefree

discharge DAMi
profit(i) = 0.9 × Pi − Pchargepaid × 4

If sum profit − ProfitsCycle /Enom ≤ 0
discharge
P1:96 =0
charge
P1:96 =0
143
Energies 2023, 16, 6546
Table 2. Storage derating depending on the duration of the capacity payment in the capacity market
defined by the Italian authority.
Duration [h] Derating [%]

1 76
2 66
4 33
6 19
8 10
Furthermore, the non-programmable RES, due to their reliance on natural phenomena,

face difficulties in providing a constant power setpoint. Hereby, the Italian authority has
defined a derating factor for the capacity payment that these power plants can receive.
Table 3 details the derating factor for the capacity payment of wind and PV power plants.
Table 3. Power plant derating factor defined by the Italian authority for the capacity payment.
Technology Derating [%]

PV 84
Wind 88
In this work, when the capacity algorithm is enabled, it is assumed that the hybrid
power plant has won the tender for, and participates in, the capacity market. The capacity
payment can be extended to account for a hybrid system with multiple energy sources fol-
lowing Equation (5), where the overall remunerated capacity CP is equal to the summation
peak
of the product between the peak power Ppowerplant and the derating Deratingpowerplant of
each technology multiplied by the capacity payment CapacityMWpayment specified by the user.
Algorithm 2 describes the logic used to emulate the capacity market inside the model.
N
peak
CP[€] = ∑ Ppowerplant × 1 − Deratingpowerplant × CapacityMW
payment (5)
powerplant=1
Algorithm 2: Capacity Market Algorithm

prod input charge
Input: Pi , Pnom , PODlimit , CDP, CMSi , CMSi
Output: Pcha , Pdis
input
If sum(MDCi )>0
CHARGE PHASE
charge
if CMSi == 1
prod prod
if Pi > 0 and Pi < Pnom
prod
Pcha
i = Pi
prod prod
Else if Pi > 0 and Pi > Pnom
Pi = Pnom
cha
DISCHARGE PHASE
input
if CMSi == 1
prod
if Pi < CDP
prod
i = CDP − Pi
Pdis
Else
Energy Arbitrage Algorithm (Algorithm 1)
144
Energies 2023, 16, 6546
The derating parameters adopted in this work are those proposed by the Italian
authority and reported in Tables 2 and 3. Nevertheless, the tool allows for those values to
be adapted to a different regulatory scenario.
Lastly, the inadequacy of an asset to properly participate in the capacity market is
computed ex-post. This calculation is mandatory to verify if the system correctly provides
the requested capacity and is entitled to the capacity payment. The criteria established by
the Italian authority foresee a neglection of the payment when the power plant does not
provide more than 80% of the capacity for at least three months.
3.3. Manual Frequency Restoration Reserve (mFRR)

mFRR is the manual activation of frequency reserve that has the purpose of restoring
the power balance of the electric power system. In Italy, this service is traded in the ancillary
services market with a pay-as-bid approach. In [64], it has been pointed out that distrusted
energy sources such as hybrid RES plants have a very low share of acceptance in the market.
Despite the small liquidity of the ancillary services market, the participation of the asset in
these services could be profitable. Indeed, as an inherent structure of the process, the mFRR
prices tend to be less competitive than those in the day-ahead market and consequently
generate more profit. To properly model the opportunity generated by participation in
the mFRR, an acceptance criterion has been developed in [65]. The procedure, based on
historical results on mFRR, develops around two important pieces of information. Firstly,
among all the bids that were submitted, only 13% were accepted. Secondly, a correlation
between bid acceptance and the submission price is present. In light of these statements,
the tool was programmed to create a binary input vector that expresses the acceptance of
the bids in the mFRR. The acceptance array changes depending on the average price that
the user foresees being submitted in the market. Lastly, the accepted bids are remunerated
at a value that is equal to the constant price times a gain that is proportional to the hour
at which the bid is accepted. This solution was implemented to grasp the hourly price
fluctuations in the bids in the ancillary services market.
Algorithm 3 describes the logic behind the control of the system participating in the
mFRR. In particular, the approach evaluates whether the mFRR bids are more profitable
than those in the day-ahead market to ensure higher revenues.
Algorithm 3: mFRR Algorithm

prod
Input: Pi , Pnom , Enom , PODlimit , PMGP
i , PmFRR
i
Output: Pcha , Pdis
DISCHARGE PHASE
for I in 96 (quarter of hours in a day) do
if sum(P mFRR
i >0

if j = find max PmFRR
i
if PMSD
j > PMGP
j
prod
if Pj > PODlimit
Pdis
j =0
prod
Else if PODlimit − Pj > Pnom
Pdis
j =P nom
Else
prod
Pdis
j = PODlimit − Pj
Pdis
Edis = Edis − j
4
3.4. Fast Frequency Reserve

Fast frequency reserve is a service enabled by the Italian authority with the resolution
200/2020/R/eel [66]. It consists of the rapid provision of power from BESS to counteract
the more usual frequency swings in the electric power system. The service is exclusive to
145
Energies 2023, 16, 6546
BESS selected through a tender procedure. In detail, the pilot project launched in 2021 in
Italy mandated the provision of a fast frequency reserve for 1000 specified hours for the
storage systems entitled to participate in this market. In those periods, the assets must
ensure the provision of the power submitted during the tender to the service and grant it in
all the hours specified by the authority. The correct interaction of the BESS with the grid is
remunerated with a capacity payment from the TSO that is proportional to the power made
available to the service. Fast frequency reserve requirements limit both the power and
energy of the storage system. Algorithm 4 describes the fast frequency reserve algorithm as
a limitation of the available power and energy during the hour mandated by the authority
for the service.
Algorithm 4: Fast Frequency Reserve

prod
Input: Pi , Pnom , Enom , PODlimit , PMGP
i , PmFRR
i , Pqualified , PFRU , FRUsignal
Output: Pcha , Pdis
SOC Control
for i in 96 (quarter of hours
in a day) do
qualifified
SOCmax
i = 100 − P4×Enom × 100 × FRUsignal
qualifified
SOCmin
i = P4×Enom × 100 × FRUsignal
If Pcha
i > Pnom − PFRU and FRUsignal = 1
Pcha
i = Pnom − PFRU
Else if Pdis
i > Pnom − PFRU
i =P
Pdis − PFRU
nom
If SOCi > SOCmax i

SOCi = SOCmax
i
SOCmax −SOCi−1
i =
Pdis i
4×Enom
Else if SOCi < SOCi min
SOCi = SOCmin i
Pcha
i = SOC4i−×1 − SOCmin
Enom
i
dis cha
Pi Pi
SOCi = SOCi−1 − 4 + 4
4. Study Cases
The sizing procedure was divided into two different study cases, labeled as stand-
alone and RES-coupled. The purpose of this classification is to highlight the different
benefits and shortcomings of the two configurations and the tool’s ability to operate in
different configurations. The stand-alone case analyzes the performance of the sole BESS
interacting with the grid. The second one foresees a hybrid RES power plant coupled
with a BESS that is capable of exchanging energy with the grid and the RES. In each study
case, four different sets of simulations were carried out with different combinations of the
algorithms, as described in Section 3. Each set of simulations aims to properly allocate
different combinations of services to investigate the impact that a multi-use BESS has on
its economics. The first simulation exclusively addresses the provision of energy arbitrage
by the storage system. The second set couples the first service that was investigated with
the capacity market through sequential stacking. In detail, the capacity algorithm has the
dispatch priority due to the necessity of satisfying the energy provision mandated by the
service regulations. The third simulation package couples the mFRR with the arbitrage and
capacity algorithms. The ancillary service structure interacts with the arbitrage to identify
the most profitable trading strategy for each day. It follows that the most remunerative
dispatch is selected depending on the day-ahead market (DAM) prices and the acceptance
in the mFRR. Lastly, the fast frequency reserve is evaluated in the fourth set of simulations
stacked with the other services. The provision of this application is superimposed with
the previous markets that were modeled. As a consequence, the fast reserve and a further
service can be provided in parallel, and the latter is derated by the power and capacity
required for the frequency regulation. The list of simulations described has allowed for
146
Energies 2023, 16, 6546
an investigation of the impact that service stacking has on the revenues of the BESS. The
tool iterates the power and EPR of the system and identifies the optimal size of different
configurations using the CAPEX being covered and IRR. Figure 2 summarizes the sets
selected for each study case. Furthermore, the figure specifies the stacking classification of
the investigated service. The number of services included in the stacking configuration is
incremental. The first one investigates sole arbitrage. Configuration 2 adds the capacity
market to the arbitrage. The third one involves the mFRR in the previous services. Finally,
configuration 4 investigates the stacking of all the services.
Figure 2. Summary of the study case and the service stacking proposed for this study.
It is worth stating that the parameters reported in Figure 2 are completely customizable
through the GUI to cover different setups. The tool evaluates the economic performance of
the BESS, allowing for the charge from the grid or the power plants.
Moreover, a POD limit was established to limit the power injection into the grid.
Table 4 details the parameters adopted for the study cases. The capacity payment duration
complies with the current Italian framework [67]. Furthermore, the fast reserve being
implemented reflects the actual one in place in Italy [68]. Both capacity payments refer to
the outcomes of the auction for the Sardinia market zone. To correctly compare the latter
service with BESS of various sizes, the power enslaved to the fast reserve is defined as a
percentage of the nominal power of the system.
Table 4. Parameters adopted in the study case presented in this work.
Parameter Stand-Alone RES-Coupled

Power [MW] [10, 20, 30, 40, 50] [10, 20, 30, 40, 50]
Duration [h] [3, 4, 5, 6] [3, 4, 5, 6]
Wind peak power [MWp] none [20, 30, 40]
PV peak power [MWp] none [20, 30, 40]
POD limit [MW] none [40]
Temperature Sardinia Italy, 2021 Sardinia Italy, 2021
Prices [years] 2019–2022 2019–2022
Capacity market duration [years] 15 15
Fast reserve duration [years] 3 3
Capacity market payment [€ × year/MW] 51,012 51,012
Fast Reserve payment [€ × year/MW] 64,890 64,890
Fast Reserve Power [%] 20% Pnom 20% Pnom
Interest rate [%] 6 6
Profits per cycle [€/MWh] 25 25
147
Energies 2023, 16, 6546
The input vectors describe a twenty-year time series with a time sample of 15 min.
The main input vectors are reported in Figure 3. The wind and PV profiles were generated
by the open-access software AtlanteEolico-RSE [69] and PVGIS, respectively, for Southern
Italy. The lack of seasonality of the wind was modeled by randomly sampling the weeks of
the three years wind profile used as input. The PV power profile was reduced every year
by a degradation equal to 1%/year. The historical temperatures were downloaded from
the ARPA website for Southern Italy and linearly increased up to 2 ◦ C to account for global
warming [70]. Lastly, the day-ahead market prices of the year 2019 were exploited for the
study case [71]. The profile was chosen because it represents the last year of business as
usual before the pandemic and the gas shortages. Furthermore, due to the high prices
recorded in 2021 and 2022, the 2019 price profile can be adopted as a conservative solution
for the investment cash flow. A one-year simulation exclusively involving the arbitrage with
the 2019-2020-2021 prices profile was developed, resulting in different realized revenues, as
depicted in Table 5.
Figure 3. Input arrays for the BESS sizing tool.
Table 5. Revenues for one year of arbitrage made by a 20 MW/60 MWh BESS with three different
price scenarios.
Price Scenario Revenues [k€] CAPEX Covered in the First Year [%]
DAM 2019 388.77 1.82
DAM 2021 762.22 3.58
DAM 2022 2125.54 9.98
The scope of this brief estimation has been the identification of the price profile that
grants a lower economic return for the investment and can be used as input for the main
analysis. The considered trend was extended to the period under evaluation and coupled
with two different yearly gains. The scope of the gains is to create two distinct price
scenarios to cope with the possible evolution of the prices in the future and provide a wider
economic analysis of the investment. The yearly gains, defined as low- and high-price
scenarios, were extrapolated from the results reported in [72]. The profits per cycle were
kept low to understand the main differences that arise in the BESS cycling due to this
parameter.
148
Energies 2023, 16, 6546
5. Results
This section presents the results obtained from the different algorithms implemented
in the sizing tool. The results focus strictly on the economic indexes that are fundamental
for evaluating the profitability of an investment. For the sake of completeness, both the
covered CAPEX and IRR are reported. Furthermore, technical pieces of information are
provided through the full equivalent cycle (FEC) and the capacity fades of the storage
system under analysis.
Firstly, it is worth stating that the tool takes from 4.3 to 12.1 min (Intel i7 1260 CPU—
16 GB) to simulate 20 years of service of a sole BESS with the lightest and heaviest algorithm;
such a limited computational effort validates the approach’s viability in tecno-economical
studies. In the specified period, the procedure evaluates the power flow, degradation, and
revenues of the BESS to provide valuable information on the investment.
The first reported outcomes concern the identification of a price profile that can
act as a conservative solution. Table 5 details the revenues generated in one year by
a 20 MW/60 MWh BESS, which performs energy arbitrage with three different price
scenarios: the day-ahead market prices of 2019, 2021, and 2022.
It is possible to note that the larger volatility of the prices during the gas shortages
returns better economics for the storage system. In particular, almost 10% of the 21.3 M€ of
BESS CAPEX is covered in a single year in the most profitable price scenario. Nevertheless,
these profits were generated by the abnormal market conditions of 2022 that, thanks to the
more widespread adoption of RES, will not recur in the coming years. It follows that DAM
prices for 2019 were exploited as a base to develop high and low price scenarios for the
next 20 years.
The algorithms have different impacts on the economics of the system. The aim of
the following paragraphs is to detail the benefits of each approach and their limitations in
both stand-alone and RES coupled configurations to understand the outcomes of the tool
in more depth.
5.1. Study Case 1: Stand-Alone BESS

5.1.1. Stacking Configuration 1: Energy Arbitrage
The arbitrage algorithm investigated in stacking configuration 1 is by far the less
remunerative scheme for a stand-alone system. Firstly, the profit is strictly related to the
spot market outcomes and, generally, a certain degree of uncertainty characterizes these
variables. Furthermore, the operation of the BESS is connected to the market price spread.
The high-price scenario has a larger spread than the low-price profile; consequently, it
represents the most remunerative profile. The IRR difference between the same BESS
with the two price trends corresponds to 4.5% on average. This value corresponds to a
difference in revenues between the two scenario prices equal to 12.0% of the CAPEX for the
whole investment lifespan. Furthermore, it is important to correctly estimate the profits
per cycle to seize the correct trade-off between the operation of the system and the covered
CAPEX. A reduction in the capital cost will directly impact the cycling cost of the system
and increase the possibility of profits for arbitrage. The degradation is strictly related to
these parameters, since a larger cycling will correspond to a higher cycle aging. The final
capacity after 20 years for the arbitrage service is, on average, equal to 81.3%, considering
a periodic capacity augmentation. Revenues are linearly impacted by the degradation.
The phenomenon leads to a reduction in the profit at the end of the investment equal to
12.2%. A possible solution to increase the revenues from the arbitrage algorithm consists of
increasing the cycling limits of the BESS set to one cycle per day. This condition may seem
quite conservative for a system with a small EPR. Indeed, it is possible to increase profits if
the market price spread allows for multiple cycles. However, profits larger than the profits
per cycle must be always granted to perform the arbitrage. This condition is difficult to
ensure with a higher number of cycles because the spread between the charge and discharge
must be higher than the cycling cost. Furthermore, for a large EPR (>5 h), the benefits of
multiple cycles are null due to the inherent structure of the market outcomes, which, in
149
Energies 2023, 16, 6546
general, has two peaks lasting 2–3 h each [73]. It follows that BESS with durations over
5 h are less-suited to the daily energy arbitrage. As a consequence, the best configuration
for the sole arbitrage is a 3 h BESS that correctly grasps the variation in the day-ahead
market. Lastly, the revenues for BESS with the same EPR are linear with power because
the efficiency of the service is exclusively related to the duration of the storage system.
Therefore, BESS with the same duration, which mostly provide the arbitrage at the nominal
power, have a similar IRR.
5.1.2. Stacking Configuration 2: Energy Arbitrage + Capacity Market

The capacity market represents an important source of income for the storage system
as evaluated in configuration 2. The fixed revenues ensured by the provision of the services
increase the economic return of the asset. However, in the best scenario, the profits cover
only 55.8% of the CAPEX. Nevertheless, a stand-alone system must participate in this
service to cover the missing money due to the excessive capital cost. Moreover, the service
does not increase the degradation compared to the sole arbitrage strategy: indeed, in Italy,
capacity remuneration is due if the obligation of bidding the remunerated capacity on the
market is respected, as BESS would do in the case of energy arbitrage. However, capacity
fade is slightly lower since the algorithms limit the BESS cycling to ensure the provision
of the energy required by the regulation. Furthermore, due to the structure of the Italian
capacity market, a longer duration of BESS benefits more from the capacity market. Indeed,
as reported in Table 2 a lower derating factor is mandated by the authority of storage
systems with a longer duration. This condition coupled with the economy of scale granted
by a larger system EPR, which identifies BESS with a 5 h duration as the best candidate for
the provision of arbitrage and capacity services.
5.1.3. Stacking Configuration 3: Energy Arbitrage + Capacity Market + mFRR

The addition of the mFRR to the services provided in stacking configuration 3 grants
slightly better revenues thanks to the higher profits generated by the bids in the ancillary
services market. On average, the increase in profits attests to around 1–2% of the CAPEX
for the whole investment lifespan. In this case, the limits are mainly driven by the low
award rate of assets on the Italian ancillary services market. However, with the increased
penetration of RES, the liquidity of these assets inside the ancillary services market could
increase. Therefore, future updates based on new market analysis may lead to an important
increase in the cash flows. In the upper part of Figure 4, it is possible to observe how the
BESS is dispatched by Algorithm 3, which selects the most profitable solution between
mFRR and DAM. Furthermore, an SOC control, as depicted in the bottom part of the image,
ensures that the BESS can efficiently participate in multiple services.
Figure 4. Algorithm 3 output for a stand-alone 20 MW/60 MWh BESS.
150
Energies 2023, 16, 6546
5.1.4. Stacking Configuration 4: Energy Arbitrage + Capacity Market + mFRR + FR

Lastly, the fast frequency reserve algorithm has been coupled with the other three in
the stacking configuration 4. It has been noted that the impact of these initiatives on the
economics of the system is inversely proportional to the nominal energy of the asset. Indeed,
the regulation foresees a capacity payment based on the percentage of the nominal power
enslaved to the service. However, the CAPEX of the asset increases with the duration
of the BESS. Therefore, the service leads to a higher improvement in the economics of
systems with a low duration. The covered CAPEX, thanks to three years of fast reserve
service, goes from 2.0% of the 6 h BESS to 3.5% of the 3 h system. It is worth stating that the
revenues generated by the capacity market maintain the 5 h duration system as the best
investment for the stand-alone configuration. Nevertheless, the hypothesis adopted for
the fast reserve is conservative. Indeed, after the three years mandated by the authority,
no further service is foreseen for fast reserve provision inside the tool. However, the
fast reserve provided by BESS may be standardized in the future and remunerated with
regulations that are not yet specified. Table 6 lists the best results for the stand-alone study
case for each stacking configuration. It is possible to observe that, despite being negative in
all stacking configurations under our assumptions, the best IRR is granted by configuration
4, which is the solution that couples the higher number of services. Nevertheless, it is
worth noting that the adopted price scenarios are conservative and return the minimum
profit that BESS can generate. Furthermore, better economics are not strictly related to the
asset cycling if capacity payments are granted to the storage system. Indeed, the same
number of cycles allows for the revenues of the storage system to be tripled thanks to its
participation in the capacity market and fast reserve provision. The general outcome of the
stand-alone configuration is that, for a one-cycle-per-day system, the 20-year investment
is not remunerative, no matter the adopted revenue stacking strategy: this unprofitable
result is mainly due to the conservative assumption of using the 2019 DAM price scenario.
However, the coupling of various services boosts the CAPEX covered by the profits from
49% to 60%. Therefore, the asset operator must carefully evaluate which services the system
can provide to increase profit and ensure capacity payment. This is fundamental, especially
for those services where participation is granted through an auction mechanism. Indeed,
thanks to the capacity payment, the best investment is represented by 10 MW/50 MWh
BESS. In general, systems with a duration of 5 h have better economics because they
represent the optimal trade-off between capacity payment and CAPEX.
Table 6. Main results for the stand-alone study case (study case 1).
Stacking Best BESS CAPEX Energy

IRR [%] FEC
Configuration Sizing Covered [%] Fade [%]
1 10 MW/30 MWh −1.72 49 3183.3 18.68
2 10 MW/50 MWh −1.03 55 3104.7 18.05
3 10 MW/50 MWh −1.00 57 3141.1 18.40
4 10 MW/50 MWh −0.95 60 3141.1 18.51
5.2. Study Case 2: RES-Coupled BESS

5.2.1. Stacking Configuration 1: Energy Arbitrage
The second study case exploits the presence of RES to increase profits. Indeed, the
energy curtailment mandated by the POD limit is equivalent to free energy that will be lost
without the storage system. This parameter, together with the PV and wind size, show a
large sensitivity regarding the free charging energy for the BESS and the energy lost by the
power plant. Although the overproduction of RES provides advantages to the BESS, it is
fundamental to highlight that the configurations with the lowest peak power production
barely improve the economics of the BESS. In the worst case, namely, a covered CAPEX of
50 MW/300 MWh BESS this increases by only 3% following the addition of 20 MWp of
both PV and wind compared to the stand-alone case. A further observation related to the
151
Energies 2023, 16, 6546
coupling of the asset to the RES is the higher number of cycles. The availability of cheap
energy for charging the system due to the presence of RES increases the use of the BESS and
its degradation, which reaches over 30% for the system with a smaller duration. However,
increased use of the storage system corresponds to better economics; as augmentation of
the asset is usually included in the OPEX, it is always better to cycle the system and boost
the profits.
The presence of a power plant allows for the BESS to be charged for free and save
energy that, without the storage system, will be curtailed. This condition leads to benefits
proportional to the energy being stored. In the best scenarios, namely high prices and high-
power plant peak power, the investment of a BESS generates value with the sole arbitrage.
The revenues from the sole arbitrage are almost doubled in the best-case scenario. The most
cost-effective BESS size with the provision of the sole arbitrage is 10 MW/30 MWh. This
system represents an optimal trade-off between the CAPEX and the ability to collect the
byproducts caused by RES plant overproduction. Higher powers reduce the curtailment
but not linearly. Therefore, the expenditure increase is higher than the revenue generated
for the increment in energy savings. Instead, the EPR increases only the expenditures
without further energy savings, leading to a higher cost and similar profit.
5.2.2. Stacking Configuration 2: Energy Arbitrage + Capacity Market

Stacking configuration 2, which couples the capacity market with the arbitrage, im-
proves the economic results. Following Equation (5), coupling with RES led to a higher-
capacity payment compared to the stand-alone case. This condition drastically improves
the economics of the system. However, this increase in capacity may lead to a mandatory
power provision value that is larger than the nominal power of the BESS. Therefore, if the
unpredictable production of RES goes to zero, in some cases, the BESS does not have the
necessary characteristics to fulfill the constraints of the capacity market. It follows that,
for a small BESS and large RES-size capacity, payment is not received due to the power
limits of the storage system. However, larger BESS benefit from this condition, with an
improvement of up to 30% in the CAPEX covered compared to the sole arbitrage provision.
This aspect is fundamental to justify the presence of a storage system in a hybrid energy
district. Indeed, for small-scale BESS, the capacity payment is not granted and the same
will occur if no BESS is present in the energy district. Therefore, for an RES owner that is
willing to participate in the capacity market, it is mandatory to purchase a BESS to fulfill
the requirements of the Italian authorities. Lastly, as in the previous study, participation in
this new service has a very low impact on cycling and degradation. Figure 5 depicts the
covered CAPEX for the RES-coupled study case with Algorithm 2. The heatmap produced
by the GUI can identify the most cost-effective BESS for the scenario under analysis, namely
the 20 MW/60 MWh system.
5.2.3. Stacking Configuration 3 and 4: Energy Arbitrage + Capacity Market + mFRR + FR

Lastly, mFRR and fast reserve provision has a similar influence on the RES-coupled
study case to the stand-alone. Indeed, the two services are strictly related to the operation
of the BESS rather than the presence of RES power plants.
The best results for the RES-coupled study case are listed in Table 7. The outcomes of
the second study case are more impacted by service stacking than the first analysis. Indeed,
the capacity prize is larger due to the necessity of coupling RES with a BESS to provide
a constant power setpoint. The service stacking improves the covered CAPEX by up to
30% compared to the sole arbitrage provision. Furthermore, as stated in [74], BESS have an
important economic advantage in supporting RES systems compared to the stand-alone
case. The coupling with power sources reduces the expenditures on the energy sale on the
day-ahead market and stabilizes the unpredictability of RES granting the capacity payment.
152
Energies 2023, 16, 6546
Figure 5. CAPEX covering 20 years with different sizes of BESS for the RES-coupled study case with
40 MWp wind and PV.
Table 7. Main results for RES-coupled study case 2 with 40 MWp for both wind and PV.
Stacking Best CAPEX Capacity

IRR [%] FEC
Configuration Configuration Covered [%] Fade [%]
1 10 MW/30 MWh 5.97 99 6410.3 32.45
2 20 MW/60 MWh 10.46 138 6233.3 30.23
3 20 MW/60 MWh 10.55 139 6180.5 30.17
4 20 MW/60 MWh 10.84 144 6124.2 29.87
6. Conclusions
This article proposes a model-aware analysis to resolve the BESS sizing issue con-
sidering different applications that implement service stacking. The economic analysis
evaluated the investment throughout a 20-year scenario, exploiting a high-fidelity empirical
BESS model equipped with equations that are able to describe both (i) the performance of
BESS in different operating and ambient conditions and (ii) the capacity degradation of
the asset. Novel algorithms were developed to properly emulate the BESS control strategy
participating in the capacity and energy markets. Flexibility and reduced computational
effort make the tool valid for the rapid assessment of different system configurations (e.g.,
stand-alone or RES-integrated), the sizing (e.g., the varying power and duration of the
BESS), and the services (i.e., a control strategy block can implement complex algorithms to
emulate service provision under different regulatory frameworks).
In this work, two systems were simulated, featuring a stand-alone and RES-coupled
BESS, which served as case studies. The provided services are implemented to fit Italian
regulation and market conditions. The outcomes highlighted the wide difference between
the stand-alone and RES-coupled BESS’ return of the investment. A stand-alone system
that is cycled once per day could not cover more than 60% of the CAPEX in 20 years.
Although the inclusion of capacity payments led to an improvement in terms of economics,
the current Italian regulations do not provide sufficient revenue to return on the investment,
as per the assumptions made in the study. However, in the future, reductions in the system
153
Energies 2023, 16, 6546
inertia will require more BESS use for frequency regulation. As a consequence, a new
market or an extension of the existing mechanisms (e.g., fast reserve, limited to 1000 h per
year as of 2023) could enter the pitch and provide new remuneration schemes for BESS.
Conversely, the configuration presenting a BESS coupled with RES shows a positive IRR
thanks to the optimal synergy between the storage and the intermittent power production.
A comparison between the two study cases highlighted the advantage that BESS has in
supporting the RES system.
This can be used to obtain the best result from the energy sale on the day-ahead market
and allows for a better integration of the RES + BESS system in the grid. This is testified
by the decreased connection power required for the hybrid RES + BESS plant and the
exploitation of a larger capacity payment.
Optimal sizing is fundamental to obtaining a positive economic outcome, as high-
lighted by the results. Moreover, the system configuration play a key role in the economics
of the asset. Indeed, the same services and the same power rating with RES can provide a
2.3 times higher return on the investment compared to the stand-alone case.
Additionally, the outcomes highlighted the necessity of service stacking to fully cover
the investment costs. In particular, it has been pointed out that the sole energy service
(i.e., arbitrage) cannot cover the capital costs of the assets, increasing the interest for
BESS participation in capacity remuneration mechanisms or long-term contracts to ensure
flexibility. The most efficient service stacking led to an increase in the covered CAPEX
that was equal to 11% in the stand-alone study case and an astonishing 45% in the hybrid
configuration. This improvement was achieved without increasing the number of cycles
performed by the BESS.
The limitations of the study include the fact that the iterative procedure is not an
optimization; thus, the optimal solution can be hidden between two tested configurations.
Given the possibility of increasing the granularity of the simulations thanks to the low
computational effort required, this is not considered a major issue. Additionally, the study
was performed using a BESS model from the literature, featuring a Li-NMC battery. It is
well-known that, as of 2023, LFP chemistry has overcome Li-NMC, especially in utility-scale
applications. The generalization of the procedure to include an LFP battery model is of
great interest.
Future works could focus on improving the algorithms proposed in this article. These
include the possibility of cycling the BESS more than once per day (to better test the trade-
off between BESS life and NPV), the inclusion of an aging-aware algorithm to identify an
efficient trade-off between profit and degradation, and the testing of innovative ancillary
services that can fulfil the new needs of power systems. Furthermore, the possibility of
easily evaluating different BESS technologies will drive the comparison and identification
of efficient services for a specific cell chemistry.
Author Contributions: Conceptualization, G.G., M.M., S.M. and G.R.; methodology, F.B. (Filippo
Bovera),G.R., M.M. and M.S.; software M.S.; Validation F.B. (Federico Bresciani) and M.S., writing—
original draft, M.S.; writing—review and editing G.R., F.B. (Federico Bresciani) and M.M.; visualiza-
tion M.S.; supervision, M.M., G.G. and S.M.; project administration M.M. and G.G. All authors have
read and agreed to the published version of the manuscript.
Funding: This work has been funded by Eni in the context of the Joint Research Center between Eni
and Politecnico di Milano (Prot. Nr. 4310481954—Project B3M).
Data Availability Statement: Data will be made available on request.
154
Energies 2023, 16, 6546
Nomenclature
Parameters Pqualified : Power qualified to provide FRU

prod
Pi : Power produced by RES at instant i PFRU : Power enslaved to FRU provision
Pnom : BESS nominal power FRUsignal : FRU participation signal
Enom : BESS nominal energy Decision Variables
PODlimit : Point of Delivery maximum power Pcha : BESS charging power
PDAM
i : Day-ahead market price at instant i Pdis : BESS discharging power
SOCinitial
0 : SOC at the beginning of the day Echa : Energy associated with the daily charge
CDP: Awarded capacity Edis : Energy associated with the daily discharge
input
CMSi : Capacity Market Signal Pchargefree : Power without charge from the hybrid energy district
charge
CMSi : Capacity Market charge Signal Pchargepaid : Power bought from the DAM
PmFRR
i : mFRR price profit: Profits daily generated by the BESS
References
1. Horowitz, C.A. Paris Agreement. Materials 2016, 55, 740–755. [CrossRef]
2. Robert, C. Pietzcker, Sebastian Osorio, Renato Rodrigues, Tightening EU ETS targets in line with the European Green Deal:
Impacts on the decarbonization of the EU power sector. Appl. Energy 2021, 293, 116914. [CrossRef]
3. European Commision. The European Green Deal. Bruxelles. June 2021. Available online: https://sustainabledevelopment.un.
org/post2015/transformingourworld (accessed on 18 April 2023).
4. Antonelli, M.; Desideri, U.; Franco, A. Effects of large scale penetration of renewables: The Italian case in the years 2008–2015. In
Renewable and Sustainable Energy Reviews; Elsevier Ltd.: Amsterdam, The Netherlands, 2018; Volume 81, pp. 3090–3100. [CrossRef]
5. Bloomberg; EATON; Statkraft. Flexibility Solutions for High-Renewable Energy Systems. 2018. Available online:
https://data.bloomberglp.com/professional/sites/24/2018/11/UK-Flexibility-Solutions-for-High-Renewable-Energy-
Systems-2018-BNEF-Eaton-Statkraft.pdf (accessed on 5 June 2023).
6. International Energy Agency. World Energy Outlook 2022. 2022. Available online: www.iea.org/t&c/ (accessed on 20 May 2023).
7. Dunn, B.; Kamath, H.; Tarascon, J.-M. Electrical Energy Storage for the Grid: A Battery of Choices System Power Ratings, Module
Size. Available online: www.sciencemag.org (accessed on 20 March 2023).
8. Englberger, S.; Jossen, A.; Hesse, H. Unlocking the Potential of Battery Storage with the Dynamic Stacking of Multiple Applications.
Cell Rep. Phys. Sci. 2020, 1, 100238. [CrossRef]
9. Rancilio, G.; Lucas, A.; Kotsakis, E.; Fulli, G.; Merlo, M.; Delfanti, M.; Masera, M. Modeling a large-scale battery energy storage
system for power grid application analysis. Energies 2019, 12, 3312. [CrossRef]
10. Vykhodtsev, A.V.; Jang, D.; Wang, Q.; Rosehart, W.; Zareipour, H. A review of modelling approaches to characterize lithium-ion
battery energy storage systems in techno-economic analyses of power systems. Renew. Sustain. Energy Rev. 2022, 166, 112584.
[CrossRef]
11. Tamilselvi, S.; Gunasundari, S.; Karuppiah, N.; Razak RK, A.; Madhusudan, S.; Nagarajan, V.M.; Sathish, T.; Shamim, M.Z.M.;
Saleel, C.A.; Afzal, A. A review on battery modelling techniques. Sustainability 2021, 13, 42. [CrossRef]
12. Moura, S.J.; Chaturvedi, N.A. Adaptive PDE Observer for Battery SOC/SOH Estimation via an Electrochemical Model. Am. Soc.
Mech. Eng. 2012, 45295, 101–110.
13. Doyle, M.; Newman, J. The use of mathematical modeling in the design of lithium polymer battery system. Electrochim. Acta 1995,
40, 2191–2196. [CrossRef]
14. Xu, L.; Lin, X.; Xie, Y.; Hu, X. Enabling high-fidelity electrochemical P2D modeling of lithium-ion batteries via fast and non-
destructive parameter identification. Energy Storage Mater. 2022, 45, 952–968. [CrossRef]
15. Allam, A.; Onori, S. Online Capacity Estimation for Lithium-Ion Battery Cells via an Electrochemical Model-Based Adaptive
Interconnected Observer. IEEE Trans. Control Syst. Technol. 2020, 29, 1636–1651. [CrossRef]
16. He, W.; Pecht, M.; Flynn, D.; Dinmohammadi, F. A physics-based electrochemical model for lithium-ion battery state-of-charge
estimation solved by an optimised projection-based method and moving-window filtering. Energies 2018, 11, 2120. [CrossRef]
17. Barzacchi, L.; Lagnoni, M.; Di Rienzo, R.; Bertei, A.; Baronti, F. Enabling early detection of lithium-ion battery degradation by
linking electrochemical properties to equivalent circuit model parameters. J. Energy Storage 2022, 50, 104213. [CrossRef]
18. Rahman, M.A.; Anwar, S.; Izadian, A. Electrochemical model parameter identification of a lithium-ion battery using particle
swarm optimization method. J. Power Sources 2016, 307, 86–97. [CrossRef]
19. Chan, H.L.; Sutanto, D. A new battery model for use with battery energy storage systems and electric vehicles power systems.
In Proceedings of the 2000 IEEE Power Engineering Society Winter Meeting. Conference Proceedings (Cat. No.00CH37077),
Singapore, 23–27 January 2000; Institute of Electrical and Electronics Engineers Inc.: New York, NY, USA, 2000; pp. 470–475.
[CrossRef]
20. He, H.; Xiong, R.; Guo, H.; Li, S. Comparison study on the battery models used for the energy management of batteries in electric
vehicles. In Energy Conversion and Management; Elsevier Ltd.: Amsterdam, The Netherlands, 2012; pp. 113–121. [CrossRef]
155
Energies 2023, 16, 6546
21. Campagna, N.; Castiglia, V.; Miceli, R.; Mastromauro, R.A.; Spataro, C.; Trapanese, M.; Viola, F. Battery models for battery
powered applications: A comparative study. Energies 2020, 13, 85. [CrossRef]
22. Berger, M.; Kocar, I.; Farantatos, E.; Haddadi, A. Modeling of Li-ion battery energy storage systems (BESSs) for grid fault analysis.
Electr. Power Syst. Res. 2021, 196, 107160. [CrossRef]
23. Pichetjamroen, A.; Chindamanee, P.; Teerakawanich, N.; Somakettarin, N. A Study on Performances of Flexible Power Control
with Empirical Lithium-ion Battery Modeling in PV Power Systems. In Proceedings of the 19th International Conference on
Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology, ECTI-CON 2022, Prachuap
Khiri Khan, Thailand, 24–27 May 2022; Institute of Electrical and Electronics Engineers Inc.: New York, NY, USA, 2022. [CrossRef]
24. Perinov; Garniwa, I.; Hudaya, C.; Sudiarto, B. Comparison of Battery Models for Battery Energy Storage System Development. J.
Phys. Conf. Ser. 2021, 1858, 012046. [CrossRef]
25. Hussein, A.A.-H.; Batarseh, I. An overview of generic battery models. In 2011 IEEE Power and Energy Society General Meeting;
IEEE: New York, NY, USA, 2011; pp. 1–6.
26. Nebuloni, R.; Meraldi, L.; Bovo, C.; Ilea, V.; Berizzi, A.; Sinha, S.; Tamirisakandala, R.B.; Raboni, P. A hierarchical two-level MILP
optimization model for the management of grid-connected BESS considering accurate physical model. Appl. Energy 2023, 334,
120697. [CrossRef]
27. Shamarova, N.; Suslov, K.; Ilyushin, P.; Shushpanov, I. Review of Battery Energy Storage Systems Modeling in Microgrids with
Renewables Considering Battery Degradation. Energies 2022, 15, 6967. [CrossRef]
28. Weniger, J.; Tjaden, T.; Quaschning, V. Sizing of residential PV battery systems. In Energy Procedia; Elsevier Ltd.: Amsterdam, The
Netherlands, 2014; pp. 78–87. [CrossRef]
29. Marchi, B.; Pasetti, M.; Zanoni, S. Life Cycle Cost Analysis for BESS Optimal Sizing. In Energy Procedia; Elsevier Ltd.: Amsterdam,
The Netherlands, 2017; pp. 127–134. [CrossRef]
30. Etherden, N.; Bollen, M.H.J. Dimensioning of energy storage for increased integration of wind power. IEEE Trans. Sustain. Energy
2013, 4, 546–553. [CrossRef]
31. Ravishankar, R.; Vrudhula, S.; Rakhmatov, D. Battery Modeling for Energy-Aware System Design. Computer 2003, 36, 77–87.
[CrossRef]
32. Wu, D.; Ma, X. Modeling and Optimization Methods for Controlling and Sizing Grid-Connected Energy Storage: A Review. Curr.
Sustain./Renew. Energy Rep. 2021, 8, 123–130. [CrossRef]
33. Wu, D.; Balducci, P.; Crrawford, A.; Mongrid, K.; Ma, X. Building Battery Energy Storage System Performance Data into economic
assesment. In IEEE Power & Energy Society General Meeting (PESGM); IEEE: New York, NY, USA, 2020; pp. 1–5.
34. Nguyen, T.A.; Copp, D.A.; Byrne, R.H.; Chalamala, B.R. Market Evaluation of Energy Storage Systems Incorporating Technology-
Specific Nonlinear Models. IEEE Trans. Power Syst. 2019, 34, 3706–3715. [CrossRef]
35. Hannan, M.A.; Faisal, M.; Ker, P.J.; Begum, R.A.; Dong, Z.Y.; Zhang, C. Review of optimal methods and algorithms for sizing
energy storage systems to achieve decarbonization in microgrid applications. In Renewable and Sustainable Energy Reviews; Elsevier
Ltd.: Amsterdam, The Netherlands, 2020; Volume 131. [CrossRef]
36. Fernández-Blanco, R.; Dvorkin, Y.; Xu, B.; Wang, Y.; Kirschen, D.S. Optimal Energy Storage Siting and Sizing: A WECC Case
Study. IEEE Trans. Sustain. Energy 2017, 8, 733–743. [CrossRef]
37. Shuai, H.; Fang, J.; Ai, X.; Tang, Y.; Wen, J.; He, H. Stochastic optimization of economic dispatch for microgrid based on
approximate dynamic programming. IEEE Trans. Smart Grid 2019, 10, 2440–2452. [CrossRef]
38. Pozo, D. Linear battery models for power systems analysis. Electric. Power Syst. Res. 2022, 212, 108565. [CrossRef]
39. Akram, U.; Khalid, M.; Shafiq, S. Optimal sizing of a wind/solar/battery hybrid grid-connected microgrid system. IET Renew.
Power Gener. 2018, 12, 72–80. [CrossRef]
40. Javadi, M.; Gong, Y.; Chung, C.Y. Frequency Stability Constrained BESS Sizing Model for Microgrids. IEEE Trans. Power Syst.
2023, 1–13. [CrossRef]
41. Duman, A.C.; Erden, H.S.; Gönül, Ö.; Güler, Ö. Optimal sizing of PV-BESS units for home energy management system-equipped
households considering day-ahead load scheduling for demand response and self-consumption. Energy Build 2022, 267, 112164.
[CrossRef]
42. Powell, W. Approximate Dynamic Programming: Solving the Curses of Dimensionality Informs Computing Society Tutorial.
2008. Available online: http://www.castlelab.princeton.edu (accessed on 30 May 2023).
43. Wang, X.; Li, F.; Zhang, Q.; Shi, Q.; Wang, J. Profit-Oriented BESS Siting and Sizing in Deregulated Distribution Systems. IEEE
Trans. Smart Grid 2023, 14, 1528–1540. [CrossRef]
44. Hannan, M.A.; Wali, S.B.; Ker, P.J.; Abd Rahman, M.S.; Mansor, M.; Ramachandaramurthy, V.K.; Muttaqi, K.M.; Mahlia, T.M.I.;
Dong, Z.Y. Battery energy-storage system: A review of technologies, optimization objectives, constraints, approaches, and
outstanding issues. J. Energy Storage 2021, 42, 103023. [CrossRef]
45. Lampinen, J. Multiobjective Nonlinear Pareto-Optimization Multiobjective Nonlinear Pareto-Optimization A Pre-Investigation
Report LAPPEENRANTA 2000. Available online: https://www.researchgate.net/publication/239560754 (accessed on 12 January
2023).
46. Truong, C.N.; Schimpe, M.; Bürger, U.; Hesse, H.C.; Jossen, A. Multi-use of stationary battery storage systems with blockchain
based markets. In Energy Procedia; Elsevier Ltd.: Amsterdam, The Netherlands, 2018; pp. 3–16. [CrossRef]
156
Energies 2023, 16, 6546
47. Hjalmarsson, J.; Thomas, K.; Boström, C. Service stacking using energy storage systems for grid applications—A review. J. Energy
48. Yang, Y.; Bremner, S.; Menictas, C.; Kay, M. Battery energy storage system size determination in renewable energy systems: A
review. In Renewable and Sustainable Energy Reviews; Elsevier Ltd.: Amsterdam, The Netherlands, 2018; Volume 91, pp. 109–125.
[CrossRef]
49. Rancilio, G.; Bovera, F.; Merlo, M. Revenue Stacking for BESS: Fast Frequency Regulation and Balancing Market Participation in
Italy. Int. Trans. Electr. Energy Syst. 2022, 2022, 1894003. [CrossRef]
50. Brogan, P.V.; Best, R.; Morrow, J.; Duncan, R.; Kubik, M. Stacking battery energy storage revenues with enhanced service provision.
IET Smart Grid 2020, 3, 520–529. [CrossRef]
51. Alharbi, A.M.; Alsaidan, I.; Gao, W. Optimal Scheduling of Battery Energy Storage System Performing Stacked Services. In
Proceedings of the IEEE Green Technologies Conference, Houston, TX, USA, 30 March–1 April 2022; IEEE: New York, NY, USA,
2022; pp. 110–115. [CrossRef]
52. Zhang, Y.; Anvari-Moghaddam, A.; Peyghami, S.; Dragicevic, T.; Li, Y.; Blaabjerg, F. Optimal sizing of behind-the-meter BESS for
providing stackable services. In Proceedings of the 2022 IEEE 13th International Symposium on Power Electronics for Distributed
Generation Systems (PEDG 2022), Kiel, Germany, 26–29 June 2022; IEEE: New York, NY, USA, 2022. [CrossRef]
53. Watson, D.; Hastie, C.; Gaudette, B.; Rodgers, M. Demonstrating Stacked Services of a Battery in a Wind R&D Park. IEEE Trans.
Power Syst. 2018, 33, 1411–1419. [CrossRef]
54. Hameed, Z.; Træholt, C.; Hashemi, S. Investigating the participation of battery energy storage systems in the Nordic ancillary
services markets from a business perspective. J. Energy Storage 2023, 58, 106464. [CrossRef]
55. Tsagkou, A.S.; Kerasidis, E.D.; Doukas, D.I.; Labridis, D.P.; Marinopoulos, A.G.; Tengnér, T. Stacking Grid Services with Energy
Storage Techno-Economic Analysis. In Proceedings of the IEEE Manchester PowerTech, Manchester, UK, 18–22 June 2017; pp. 1–6.
56. Liu, Y.; Du, W.; Xiao, L.; Wang, H.; Bu, S.; Cao, J. Sizing a Hybrid Energy Storage System for Maintaining Power Balance of an
Isolated System with High Penetration of Wind Generation. IEEE Trans. Power Syst. 2016, 31, 3267–3275. [CrossRef]
57. Pelletier, S.; Jabali, O.; Laporte, G.; Veneroni, M. Battery degradation and behaviour for electric vehicles: Review and numerical
analyses of several models. Transp. Res. Part B Methodol. 2017, 103, 158–187. [CrossRef]
58. Moncecchi, M.; Brivio, C.; Mandelli, S.; Merlo, M. Battery energy storage systems in microgrids: Modeling and design criteria.
59. Ecker, M.; Nieto, N.; Käbitz, S.; Schmalstieg, J.; Blanke, H.; Warnecke, A.; Sauer, D.U. Calendar and cycle life study of
Li(NiMnCo)O2 -based 18650 lithium-ion batteries. J. Power Sources 2014, 248, 839–851. [CrossRef]
60. Ramasamy, V.; Zuboy, J.; O’Shaughnessy, E.; Feldman, D.; Desai, J.; Woodhouse, M.; Basore, P.; Margolis, R.U.S. Solar Photovoltaic
System and Energy Storage Cost Benchmarks, With Minimum Sustainable Price Analysis: Q1 2022. 2022. Available online:
https://www.osti.gov/biblio/1891204 (accessed on 15 June 2023).
61. Cole, W.; Frazier, A.W.; Augustine, C. Cost Projections for Utility-Scale Battery Storage: 2021 Update. 2021. Available online:
https://www.nrel.gov/docs/fy21osti/79236.pdf (accessed on 15 June 2023).
62. Rancilio, G.; Merlo, M.; Lucas, A.; Kotsakis, E.; Delfanti, M. BESS modelling: Investigating the role of auxiliary system
consumption in efficiency derating. In Proceedings of the International Symposium on Power Electronics, Electrical Drives,
Automation and Motion, Sorrento, Italy, 24–26 June 2020; pp. 1–6.
63. Terna Mercato della Capacità—Disposizioni Tecniche di Funzionamento n.2. 2021. Available online: https://www.terna.it/it/
sistema-elettrico/pubblicazioni/news-operatori/dettaglio/CM-Disposizioni-tecniche-di-funzionamento-DTF (accessed on 10
March 2023).
64. Gulotta, F.; Rossi, A.; Bovera, F.; Falabretti, D.; Galliani, A.; Merlo, M.; Rancilio, G. Opening of the Italian Ancillary Service
Market to Distributed Energy Resources: Preliminary Results of UVAM project. In Proceedings of the HONET 2020—IEEE 17th
International Conference on Smart Communities: Improving Quality of Life Using ICT, IoT and AI, Charlotte, NC, USA, 14–16
December 2020; IEEE: New York, NY, USA, 2020; pp. 199–203. [CrossRef]
65. Bovera, F.; Blaco, A.; Rancilio, G.; Delfanti, M. Assessing the Accuracy of Different Machine Learning Classification Algorithms
in Forecasting Results of Italian Ancillary Services Market. In Proceedings of the 2019 16th International Conference on the
European Energy Market (EEM), Ljubljana, Slovenia, 18–20 September 2019.
66. Arera. Deliberazione 3 Giugno 2020 200/2020/R/EEL Approvazione del Regolamento, Predisposto da Terna S.P.A., Ai Sensi
della Deliberazione Dell’autorità 300/2017/R/EEL, Relativo al Progetto Pilota per L’erogazione del Servizio di Regolazione
Ultra-Rapida di Frequenza. 2020. Available online: https://www.arera.it/it/docs/20/200-20.htm (accessed on 25 February
2023).
67. Terna Mercato Della Capacità Rendiconto Degli Esiti—Asta Madre 2024. 2022. Available online: https://www.terna.it/it/
sistema-elettrico/pubblicazioni/news-operatori/dettaglio/rendiconto-asta-capacity-market-2024 (accessed on 14 March 2023).
68. Terna Progetto Pilota Fast Reserve—Esiti Asta. 2020. Available online: https://www.terna.it/it/sistema-elettrico/pubblicazioni/
news-operatori/dettaglio/esiti-asta-Fast-reserve (accessed on 28 January 2023).
69. RSE Atlante Eolico RSE. Available online: https://atlanteeolico.rse-web.it/ (accessed on 7 June 2023).
70. ARPA Puglia Temperature Puglia. Available online: https://www.arpa.puglia.it/pagina2839_meteo.html (accessed on 7 June
2023).
157
Energies 2023, 16, 6546
71. Gestore Mercati Energetici (GME) Esiti Mercato Elettrico. Available online: https://www.mercatoelettrico.org/it/ (accessed on 7
June 2023).
72. Afman, M.; Hers, S.; Scholten, T. Energy and Electricity Price Scenarios 2020-2023-2030. 2017. Available online: chrome-extension:
//efaidnbmnnnibpcajpcglclefindmkaj/https://cedelft.eu/wp-content/uploads/sites/2/2021/04/CE_Delft_3H58_Energy_
and_electricity_price_scenarios_DEF.pdf (accessed on 7 June 2023).
73. Mercier, T.; Olivier, M.; De Jaeger, E. The value of electricity storage arbitrage on day-ahead markets across Europe. Energy Econ.
2023, 123, 106721. [CrossRef]
74. Lazard LCOE Lazard. 2023. Available online: https://www.lazard.com/research-insights/2023-levelized-cost-of-energyplus/
(accessed on 7 June 2023).
158
energies
Article
Data Augmentation and Feature Selection for the Prediction of
the State of Charge of Lithium-Ion Batteries Using Artificial
Neural Networks
Sebastian Pohlmann 1, *, Ali Mashayekh 2 , Manuel Kuder 3 , Antje Neve 1 and Thomas Weyh 2
1 Institute of Distributed Intelligent Systems, University of the Bundeswehr, Werner-Heisenberg-Weg 39,

85577 Neubiberg, Germany; antje.neve@unibw.de
2 Institute of Electrical Energy Systems, University of the Bundeswehr, Werner-Heisenberg-Weg 39,
85577 Neubiberg, Germany; ali.mashayekh@unibw.de (A.M.); thomas.weyh@unibw.de (T.W.)
3 Bavertis GmbH, Marienwerderstraße 6, 81929 Munich, Germany; manuel.kuder@unibw.de
* Correspondence: sebastian.pohlmann@unibw.de
Abstract: Lithium-ion batteries are a key technology for the electrification of the transport sector and
the corresponding move to renewable energy. It is vital to determine the condition of lithium-ion
batteries at all times to optimize their operation. Because of the various loading conditions these
batteries are subjected to and the complex structure of the electrochemical systems, it is not possible
to directly measure their condition, including their state of charge. Instead, battery models are used
to emulate their behavior. Data-driven models have become of increasing interest because they
demonstrate high levels of accuracy with less development time; however, they are highly dependent
on their database. To overcome this problem, in this paper, the use of a data augmentation method
to improve the training of artificial neural networks is analyzed. A linear regression model, as well
as a multilayer perceptron and a convolutional neural network, are trained with different amounts
of artificial data to estimate the state of charge of a battery cell. All models are tested on real data
to examine the applicability of the models in a real application. The lowest test error is obtained
for the convolutional neural network, with a mean absolute error of 0.27%. The results highlight
Citation: Pohlmann, S.; Mashayekh, the potential of data-driven models and the potential to improve the training of these models using
A.; Kuder, M.; Neve, A.; Weyh, T. artificial data.
Data Augmentation and Feature
Selection for the Prediction of the Keywords: lithium-ion batteries; state of charge; machine learning; artificial neural networks; data
State of Charge of Lithium-Ion augmentation
Batteries Using Artificial Neural
Networks. Energies 2023, 16, 6750.
https://doi.org/10.3390/en16186750
Academic Editor: Simone 1. Introduction

Barcellona As the transportation sector is responsible for a large share of greenhouse gas emissions,
Received: 28 July 2023
it is crucial for the automotive and mobility industry to turn towards renewable energy [1].
Revised: 11 August 2023
Lithium-ion batteries (LIBs) have taken a predominant role as electrochemical energy
Accepted: 15 August 2023 storage solutions in many applications, ranging from portable consumer electronics to
Published: 21 September 2023 integration in power grids and battery electric vehicles (BEVs) [2]. Next to applications in
stationary energy storage systems with particularly high efficiency needs in third-world
countries [3], LIBs are best-suited for empowering BEVs due to their high energy densities
and their long lifespans [4]. Nevertheless, the operation of LIBs has to be optimized
Copyright: © 2023 by the authors. to exceed the performance of internal combustion engine vehicles. The demand for
Licensee MDPI, Basel, Switzerland. BEVs is growing quickly, and the materials they require are rare. The optimization of
This article is an open access article BEVs is essential for their worldwide success, especially in developing countries, such as
distributed under the terms and Latin American countries, which have large populations and uncertain future markets [5].
The whole lifecycle of the BEV, including production, operation, recycling, and reuse, has
to be considered. Additionally, the price pressure seen regarding BEVs intensifies the need
for further enhancements during operation [6]. The battery management system (BMS) of
4.0/).

159
Energies 2023, 16, 6750
a BEV is responsible for determining the condition of the vehicle’s battery. This system
monitors and controls the battery cells [7]. During operation, the condition of a battery is
influenced by various intertwined parameters and the ambient conditions [8]. Additionally,
there are several mechanisms for the degradation of an LIB, all of which directly affect
the performance and the state of the battery. These mechanisms are caused by chemical
processes, mechanical damage, temperature, and different loading conditions [9].
A key challenge for the application of LIBs is to accurately predict their state of charge
(SoC), which is necessary to ensure their safety and facilitate their efficient charge and
discharge cycles [10]. Other than physical estimation methods, such as coulomb counting
or other electrochemical models [11–13], there are primarily two different approaches:
model-based approaches and data-driven approaches. The main representatives of the
model-based methods are from the Kalman filter family, including the extended Kalman
filter, dual extended Kalman filter, and unscented Kalman filter models [14–16].
The data-driven methods utilize machine learning (ML) or other statistical algorithms to
estimate the condition of a battery. Because these algorithms approximate the electrochemical
processes inside a battery cell with high levels of accuracy while having decreased levels of
complexity, they have gained considerable interest [17]. The main reason for the success
of a data-driven method is the data the method is based on, which should be reliable and
capture the behavior of the cell [18]. Poor data impede the state estimation for batteries,
as the parameters are highly dependent on the loading as well as the ambient conditions
and are further internally correlated [19]. The methods used to estimate the SoC of a battery
include support vector machines (SVM) [20], regression algorithms [21], and artificial
neural networks (ANNs) [22]. Different types of ANNs can be distinguished from one
another. The conventional type are multilayer perceptrons (MLP), which are feedforward
neural networks [23]. The other applied types are recurrent neural networks (RNNs) [24,25]
and convolutional neural networks (CNNs), where the convolution is typically performed
along the time axis [26,27]. The aim of the data-driven models is to approximate a function
between the measurable parameters of a battery and the nonmeasurable conditions, such
as the SoC. As ML models are highly dependent on their input data, it is crucial to have
a sufficiently large dataset to replicate the behavior of a battery. For LIBs in particular,
which have several working conditions and respond differently to changing ambient
conditions, creating an appropriate dataset is a key challenge. Furthermore, battery tests
are time- and cost-consuming [28]. A possible solution to overcome these problems is the
usage of artificially augmented data. In the last few years, it has been shown that data
augmentation techniques can lead to improved results for ML models, thus making it
possible to successfully tackle the problem of limited datasets [29–31].
In this work, the SoC of an LIB is estimated using different ML algorithms. To decrease
the effort required for time-consuming battery tests to a minimum, a real-world dataset is
enriched using artificially augmented data. The goal is to approximate a function for the
SoC with the current, voltage, and temperature as input variables. After preprocessing,
the data are used to train and test the ML models. The results are compared to a reference
model, which is a linear regression model. Two types of ANNs, an MLP and a CNN, are
trained and tested to evaluate the impact of the data augmentation technique. The MLP
is chosen because of its lower complexity and simpler structure in comparison with other
ANNs. Therefore, less computing power is needed to train the model. An advantage
of CNNs is their additional data processing step along the time axis, which allows the
accuracy of the model to benefit from determining more complex correlations, which are
not identifiable using only the rare input data.
2. Materials and Methods

2.1. Data Origin and Data Augmentation
The parameters of a battery are highly intercorrelated. During operation, it is important
to accurately monitor the condition of the cells to ensure efficient usage and adequate
loading. In the process of developing an ML model, considerable attention must be paid
160
Energies 2023, 16, 6750
to the selection of features, as the results are highly dependent on the input data. The
data processing steps include data collection, data cleaning, and data transformation [32].
Real-world data were obtained by loading an LIB in a battery test system. The experiments
were conducted using the Molicel 21700 P42a battery cell, which has a capacity of 4.2 Ah
with an end-of-charge voltage of 4.2 V ± 0.05 and a cut-off voltage of 2.5 V. The battery was
cycled in a temperature chamber at a constant temperature of 23 °C. The test system was
an OctoStat5000 from Ivium Technologies. The cell was discharged in ten percent intervals
from a SoC of 100% to a SoC of 0%. Every 0.5 s, the voltage, current, and temperature of
the cell were measured. A discharge cycle of the analyzed battery cell is shown in Figure 1.
Figure 1. Discharge cycle of the analyzed battery cell with voltage plotted over the SoC. The cell is
discharged in 10% intervals and rested after each interval.
Based on the real-world data, a data augmentation technique is applied to enrich the
data basis for training the ML models. The input data for the algorithms, also called features,
are current, temperature, and voltage. The output of the models, referred to as target value,
is the SoC. To artificially create the new data, a whole discharge cycle is used and the current
as well as the SoC values are kept constant. Two different regression models are trained
for voltage and temperature, using a ridge regression. The loss function L, consisting of
the squared error between observed y and predicted values y∗ , is supplemented by an
L2-penalty for the weight parameters w to decrease the risk of overfitting and to process
the highly correlated data. This is summarized in (1).
n p
1
L(w) =
2n ∑ (y∗ (xij ) − yi )2 + λ ∑ w2j (1)
i =1 j =1
Based on the previous ten time steps, which is equivalent to five seconds cumulated
with a sample rate of 0.5 s, the next value for voltage, respectively, temperature, is estimated.
Both models are regression algorithms and, therefore, can be demonstrated in a regression
formula. The regression equation for the voltage estimation is shown in (2), and for the
temperature in (3). The target value is the current value of timestep t for voltage and
temperature, respectively.
Vt = 0.003 + 0.078Vt−10 + 0.079Vt−9 + 0.080Vt−8 + 0.085Vt−7 + 0.090Vt−6

(2)
+0.094Vt−5 + 0.103Vt−4 + 0.117Vt−3 + 0.122Vt−2 + 0.154Vt−1
Tt = 0.080 − 0.008Tt−10 + 0.006Tt−9 − 0.002Tt−8 − 0.029Tt−7 + 0.059Tt−6

(3)
−0.022Tt−5 − 0.023Tt−4 + 0.040Tt−3 + 0.055Tt−2 + 0.924Tt−1
To keep the computation time and the processing power low, only every twentieth
measurement point is calculated using the regression model. The data points between
these reference points are interpolated. At first, the estimation models for voltage and
temperature are tested against real measurements to validate their accuracy. Even though
161
Energies 2023, 16, 6750
these models are highly accurate, a small error benefits the data augmentation as further
randomness is included in the process. Following that, the real input data are slightly
modified. Therefore, new reference points are estimated, and a new discharge curve is
created. Again, the data between the reference points are interpolated. Demonstrated
in Figure 2, the accuracy and two artificially created discharge curves are shown as an
example for the voltage.
Figure 2. Results of the data augmentation method. On the left, the accuracy of the voltage estimation
model for real data is presented. Real measurements are compared to the results of the voltage
estimation model. On the right, results of the estimation model with slightly modified input values
resulting in two artificially created discharge curves are shown.
Before the data are used in the ML models, they are preprocessed. At first, the data are
normalized by using a standard scaler. By means of the mean values x and the standard
deviation σ, the data are transformed, resulting in a data distribution with zero mean and
unit variance [33]. This is performed for each battery parameter.
In a crucial step for the high performance of ML models [34]), the dimension of the
data is reduced after the normalization, which can be referred to as the actual feature
selection. In context of the curse of dimensionality and to reduce the risk of overfitting,
a principal component analysis (PCA) is conducted. In sum, the curse of dimensionality
describes the problem of an exponentially growing search space with an increasing amount
of features [35]. A PCA is a method to reduce the dimension of a problem by analyzing
the variance of the data [36], which is transformed into a new coordinate space with a
lower dimension. This is especially useful while working with highly intercorrelated
data [37]. The principal components are orthonormal axes, which cover a certain value
of variance of the initial data. These components are given by the dominant eigenvectors
of the sample covariance matrix. They can be identified by the largest corresponding
eigenvalues [38]. The PCA is connected to the singular value decomposition (SVD), which
is a matrix factorization that can also be used to reduce the dimension. A SVD is typically
computationally more efficient, and because of the relation between the singular and the
eigenvalues, the SVD in (4) can be used to calculate the principal components. M is the
feature matrix, V is the orthonormal basis of the eigenvectors of M M, ∑ is a diagonal
matrix with the singular values θi , which are the square roots of the eigenvalues λ, and U
is the orthonormal matrix, which can be calculated using (5) [39].
M = U∑V (4)
1
ui = Mvi (5)
θi
After preprocessing the data, the features are used to train different ML models.
The training data consist of 80% of the initial data plus the augmented data, and the test
data make up the remaining 20%. These models are introduced in the following.
162
Energies 2023, 16, 6750
2.2. Machine Learning Models

2.2.1. Linear Regression
As a reference model, and to compare the results to a less complex model, a linear
regression is implemented. Based on weight parameters wi , which are determined during
the training, a linear function between the input features voltage, current, and temperature,
as well as the target values, is approximated. The estimated values y∗ are calculated
using (6) [40].
n
y ∗ = w0 + ∑ wi x i (6)
i =1
By minimizing the squared error between the real and predicted values, the values for
the weight parameters are determined. The corresponding loss function L is shown in (7).
n
1
L(w) =
2n ∑ (y∗ (xij ) − yi )2 (7)
i =1
The reference model is an ordinary least square approach to determine the SoC based
on measurement values of voltage, current, and temperature.
2.2.2. Artificial Neural Networks

Two types of ANNs are used to predict the SoC. The first one is an MLP, which
is a simple feedforward neural network. All neurons from a layer are connected to
the neurons from the next layer. The information is passed on in one direction to the
output layer, where the target value is calculated. The inputs in a neuron are summed up
and are then further processed in an activation function, where the output is calculated.
The chain of mathematical functions is used to approximate the target values. To determine
the appropriate parameters of the model, a four-fold-cross-validation was conducted by
dividing the training data into four equally sized groups. One group acts as the validation
set and the remaining groups as training data [41]. There are approaches in the literature
using an MLP to estimate the SoC of a battery cell [42]. The structure of applied feedforward
neural networks in the area of SoC estimation is mainly kept simple, with only a few hidden
layers and a low number of neurons [43–45]. The structure of the proposed MLP consists
of three hidden layers with ten neurons in both the first and the second hidden layer and
five neurons in the third hidden layer. As activation function, the rectified linear unit
(ReLU) function is used, which has another advantage of efficient model training. The final
learning rate is 0.1 and the Adam optimizer is used to improve the training speed and
ensure the accuracy of the estimation results [46].
The second applied model is a CNN. While the main application of CNNs is image
processing, it is gaining more and more interest for other areas as well [47,48]. The first
approaches to estimate the condition of a battery cell can be found in the literature. Most
of the CNN-based models are used in combination with other ANNs [49,50]. The main
difference for an MLP is the convolution by means of a kernel function. When analyzing
time series data, the convolution is conducted along the time axis [51]. Typically, the kernel
filter is followed by a pooling layer, where several points can be pooled in a single data
point [52]. Instead of an MLP, where the neurons are fully connected, CNNs reduce
and reassemble the feature matrix to learn new and complex patterns of the input data.
The structure and the general approach of a CNN is shown in Figure 3 [53].
163
Energies 2023, 16, 6750
Figure 3. Structure of a CNN with a feature map followed by a convolution, a pooling layer, and a
fully connected layer. The filter is moved across the features.
The applied CNN consists of two convolution layers with a pooling layer followed
by another two convolutions and a pooling layer. During the convolution, 32 randomly
initialized filters are used in the first part and 16 filters are used in the second part. When
applying the filters, the window size is five timesteps along the temporal axis and the
dimension is not affected during the convolution. The final layer is a fully connected layer,
where the value for the SoC is estimated. As a metric during the training of the model,
the coefficient of determination R2 is used, which is shown in (8). The R2 score is calculated
by means of the real values yi , the predicted values yi∗ , and the mean value y.
∑in=1 (yi − yi∗ )2

R2 = 1 − (8)
∑in=1 (yi − y)2
To validate the ML models, two metrics with the mean absolute error (MAE) and the
root mean square error (RMSE) are used. Both use the difference between the real target
values y and the estimated values y∗ and are shown in (9) and (10).
n
1
MAE(yi , yi∗ ) = ∑ |yi − yi∗ | (9)
n i =1

n
1
RMSE(yi , yi∗ ) = ∑ (yi − yi∗ )2 (10)
n i =1
3. Results
Three different ML models are trained, respectively, with and without data augmentation
to evaluate the accuracy of the SoC prediction. The raw data are temperature, voltage,
and current values. The test data are exclusively real data. To have a benchmark and
a comparison for the neural networks, a linear regression model is used as reference.
The input data in all models are the same. To analyze the impact of the data augmentation
technique, the results are calculated without artificial data, with ten times the initial data,
and with 20 times the initial data. All models are retrained five times and the mean values as
well as the standard deviation are presented, as the initialization of the data augmentation
technique and the neural networks is random and therefore slightly different. The results
for the linear regression are shown in Table 1. The MAE and the RMSE are separated for
the training and the test of the models. All mean values and the corresponding standard
deviation are listed for the three different input datasets.
164
Energies 2023, 16, 6750
Table 1. Results of the linear regression with the three different sizes of training datasets. The mean
of five times retraining the model and the corresponding standard deviations are shown.
Linear Training Test

Regression MAE RMSE MAE RMSE
Without
3.874 (±0.021)% 4.941 (±0.012)% 4.089 (±0.205)% 4.999 (±0.095)%
augmented data
With augmented
3.914 (±0.029)% 4.970 (±0.018)% 4.041 (±0.220)% 4.980 (±0.100)%
data (10×)
With augmented
4.004 (±0.027)% 5.066 (±0.036)% 3.977 (±0.050)% 5.044 (±0.052)%
data (20×)
The impact of the data augmentation method on the linear regression is small.
In comparison to other ML algorithms, the dependence of a linear regression on a large
database is slight, and an impact or a significant improvement was not expected. Accordingly,
the results for all three different sizes of training data are similar and no influence of the
data augmentation method can be measured. Nevertheless, it is possible to estimate the
SoC with a simple linear model with error values below 5%. There is no overfitting and,
consequently, the test errors are similar to the training errors. Nevertheless, the main focus
of the linear model is to have reference accuracies for the neural networks.
Before analyzing the accuracies of both models, the convergence during the training
of the models is examined. During the training, the weights are updated after each epoch.
The training phase with the MAE over the epochs for the MLP is shown in Figure 4a and
for the CNN in Figure 4b.
Figure 4. Convergence of the different models. MAE is shown over the epochs for the MLP (a) and
the CNN (b) with and without augmented data.
A comparison of the convergence is drawn between the performance with different

input data for both models. The convergence for the model without augmented data is
indicated in blue, with ten times the initial data in orange, and with 20 times the initial
data in green. As expected, the error is decreasing quickly over the first epochs, and then
converges against a certain value. This behavior can be determined for all input data,
but there are differences in the number of epochs to reach the final value and the final
error itself. For both models, the augmented data are favorable for convergence, as the
error decreases faster than it does without the augmented data. The difference is greater
for the MLP, but also visible for the CNN. A faster convergence has the advantages of
less training effort and a higher robustness, as there is less chance to become stuck in a
local minima. Although the difference in the convergence behavior is apparent, it is in a
smaller range and is, therefore, not significant for the optimization of the model. Further,
regardless of the training dataset, each model converges to a similar value. In comparison,
the CNN drops faster below an MAE of 1%, and the differences between the model with
and without augmented data are slightly greater. This can also be shown when analyzing
165
Energies 2023, 16, 6750
the accuracies of each model. The first examined ANN is the MLP. The preprocessed input
data are passed through the layers of the MLP and the SoC is estimated. The results are
summarized in Table 2. Similar to Table 1, the MAE and RMSE with the corresponding
standard deviations are shown for the training and test of both models with the same three
different input datasets.
Table 2. Results of the MLP with the three different sizes of training datasets. The mean of five times
of retraining the model and the corresponding standard deviations are shown.
MLP Training Test

MAE RMSE MAE RMSE
Without
0.828 (±0.292)% 1.072 (±0.329)% 0.553 (±0.051)% 0.805 (±0.072)%
augmented data
With augmented
0.626 (±0.184)% 0.848 (±0.190)% 0.539 (±0.087)% 0.758 (±0.109)%
data (10×)
With augmented
0.722 (±0.222)% 0.977 (±0.242)% 0.727 (±0.217)% 0.978 (±0.246)%
data (20×)
Firstly, it can be noted that the model shows significantly higher accuracies than the
linear model. Even the test RMSE is below 1% for all three datasets. Despite this notion,
the impact of the data augmentation technique is small. The errors of the dataset with
ten times the initial data are lower, but they are increasing with a higher amount of data.
The standard deviation is also not highly impacted. The general ability to estimate the
SoC can be determined. Additionally, the PCA is working efficiently, as no indications for
overfitting can be detected. Further optimization with data augmentation is not necessary.
Even though it does not deteriorate the performance of the model, the influence on the
error and the standard deviation is low. The dimension reduction method improves the
ratio between dimension and number of features and, thus, the influence of the data
augmentation method on the simpler neural network is low.
In comparison to that, a CNN is used to learn new and complex patterns in the input
data. This is conducted by using a convolution filter along the time axis. The results are
summarized in Table 3.
Table 3. Results of the CNN with the three different sizes of training datasets. The mean of five times
of retraining the model and the corresponding standard deviations are shown.
CNN Training Test

MAE RMSE MAE RMSE
Without
0.975 (±0.459)% 1.173 (±0.531)% 0.505 (±0.201)% 0.723 (±0.249)%
augmented data
With augmented
0.371 (±0.269)% 0.494 (±0.315)% 0.315 (±0.140)% 0.478 (±0.124)%
data (10×)
With augmented
0.261 (±0.071)% 0.392 (±0.102)% 0.270 (±0.068)% 0.437 (±0.101)%
data (20×)
Several aspects are striking when analyzing the results of the CNN. Without the
augmented data, the CNN is slightly worse than the MLP. Additionally, the difference
between training and test errors is higher, which is an indicator that there is potential for
optimization. Further, the standard deviation is higher. In comparison to the MLP, there is
a higher degree of randomness, as the filters for the convolution are randomly initialized.
The model learns complex patterns in the data, but is not able to reproduce them with the
limited database. Still, the standard deviation is not high and acceptable, but the fluctuation
of the CNN is higher. By increasing the amount of training data, the errors are decreasing.
While using ten times the initial data, the training MAE could be reduced by over 50%
from 0.975% to 0.371%. With 20 times the initial data, the error could again be decreased to
0.261%. Further, the standard deviation is also decreasing, which means that the patterns in
166
Energies 2023, 16, 6750
the data can be learned regardless of the filters. The filters vary while retraining the model.
With a sufficiently large dataset, the impact of the filters and the uncertainty of the model
is decreasing. The same behavior can be shown for the test data. The difference between
training and test errors is small and, therefore, there is no overfitting. On the contrary, some
test errors are slightly lower than the training errors using augmented data. The reason for
that is the data augmentation technique, which should reflect a wider range of discharge
behavior. Consequently, there is a higher variety in the training dataset, which could lead
to higher errors; yet, still, the errors are nearly the same. The direct comparison of the errors
of MLP and CNN with the corresponding standard deviations is shown in Figure 5. The
error bars displaying the MAE are demonstrated for the different input data and, further,
the uncertainty for retraining the model is shown. While the impact on the MLP is small,
the optimization potential using augmented data is clearly visible for the CNN.
As stated before, the test data consist of only real data of a discharge cycle. Therefore,
a direct comparison between real data and the prediction of the model can be drawn.
For the MLP, this is shown in Figure 6a and for the CNN in Figure 6b. For a full discharge
cycle, the voltage is shown over the SoC. The real experimental data are indicated in gray.
The results of the final model with augmented data are demonstrated in orange.
Figure 5. Error bars and standard deviation of the five times retrained models without data
augmentation, with ten times the initial data and 20 times the initial data. The MAE is shown
for the MLP and the CNN.
Figure 6. Test results of the SoC estimation model in comparison with the real values for the
MLP (a) and for the CNN (b). The results for the SoC are shown over the last voltage value for
the estimation.
167
Energies 2023, 16, 6750
The estimations and the real data are mainly overlapping. The drawbacks of the MLP
can be seen in the high and low SoC areas and in the transition area from linear to nonlinear,
where there are higher fluctuations in the estimation results. In the mainly linear range
between a SoC of 20% to 80%, the performance of both models is similar. The CNN shows
fewer outliers and, especially, the low SoC area can be accurately approximated. This
behavior can explain the differences between a test MAE of the MLP with 0.727% and the
CNN with 0.270%. As the edge areas are critical for an accurate state estimation, the CNN
shows better results and is more suitable.
Overall, the performance of the CNN is slightly better. The impact of the data
augmentation method is huge in relative values, but small in absolute values, as both neural
networks show high accuracies by estimating the SoC without further data processing.
4. Discussion
When it comes to LIBs, a major challenge for balanced and safe loading cycles
is the accurate determination of the condition of a battery cell. In comparison with
conventional modeling techniques, data-driven models require less development time
and no electrochemical characterization. Nevertheless, they need a reliable and large
enough data basis, which covers the behavior of the cell. Especially for battery cells, for
which tests and experiments are time- and cost-intensive, this is a key problem. Data
augmentation, which is usually used for image processing, is a method to overcome this
challenge by artificially creating new training data. In this case, two time-series forecasting
models for voltage and temperature are developed. The accuracy of both models is high,
which is shown in Section 2.1, but, moreover, the small error is favorable because of the
additional randomness of the results. Therefore, new grid points are estimated and a wider
range of input values is covered. Simultaneously, the current and the SoC values are kept
constant. This technique relies on creating a whole discharge cycle, but it can also be used
to enrich certain SoC areas. Only grid points are estimated because of the computing time.
As a compromise between computing costs and accuracy, only every twentieth point is
estimated. The points in between are interpolated. This interpolation does not impair
the estimation and is hence sufficient for the data augmentation method. At first, a linear
model is developed to predict the SoC. The errors are mainly below 5%. Although it could
be shown that it is possible to determine the SoC with a linear regression, the accuracies
cannot compete with conventional methods. Nevertheless, it is a reference model and a
starting point to evaluate the results of the ANNs. In an MLP, the information is processed
from input to output layer and a function between the features and the target values is
approximated. The data augmentation technique has low impact on the accuracy of the
MLP. A reason for that is the preprocessing method. ML algorithms are prone to overfit,
when the dimension is equal to or greater than the number of features. This is the case
for a feature matrix xij , where j is equal or greater than i. By reducing the dimension
using the PCA, the tendency to overfit could be reduced. This is sufficient for the MLP,
and the impact of the approach to artificially create additional data is decreased. Further,
the testing data cover a limited range of loading conditions. Hence, analyzing the test
errors does not capture the full capabilities of the optimized model. On the contrary,
the CNN can determine more complex patterns, in which a huge database is beneficial.
When examining the results of the CNN, it becomes apparent that the data augmentation
technique increases the accuracy of the CNN. Two other advantages can be observed. First,
the model converges earlier with the artificial data. Even though the improvement is small,
there is less chance to become stuck in local minima, and fewer iterations are needed to
train the model. Second, the standard deviation is decreasing, which results in a more
robust model and an improved ability to reproduce the results.
Overall, the CNN shows better results and more potential for optimization. As the
different loading conditions and their combinations are infinite in a real application,
the ability to capture complex patterns in the input data is a key advantage of the CNN.
The data augmentation method leads to improved results and the impact is expected
168
Energies 2023, 16, 6750
to be higher for testing against several loading conditions. Still, the potential of data
augmentation for the optimization of ANNs is evident. Further, the errors below 0.5%
show that the CNN is able to accurately estimate the SoC. In comparison to conventional
estimation models, such as representatives of the Kalman filter family, the errors could be
slightly reduced [54–56]. Next to that, there is no need for an elaborately electrochemical
characterization of the cell. This shows the huge potential to determine the condition of a
battery cell with data-driven methods.
5. Conclusions
The electrification of the transport sector is inevitable to reach environmental goals.
LIBs as electrochemical energy storage systems are a crucial factor for the success of electric
vehicles. For efficient and optimized driving cycles, it is important to be able to determine
the condition of the battery at all times. As ML algorithms reach high accuracies, are
robust, and need less development and computing time, they are a promising alternative
to conventional battery models. Nevertheless, they need a reliable and huge dataset to
represent reality; however, battery tests are time- and cost-consuming. Therefore, the
applicability of data augmentation was examined in this paper to optimize the ML models.
The data of real-world experiments with the battery cell Molicel 21700 P42a were used.
The training data were enriched by artificially created data by means of linear estimation
models for voltage and temperature, and the impact on the ML models was analyzed.
The additional data improve the performance of the models in terms of convergence,
robustness, and accuracy. Both neural networks succeed the linear model and are able
to estimate the SoC with errors below 1%. The linear model serves as a reference model,
but the final results with error values around 4% to 5% cannot compete with the ANNs.
Comparing both ANNs, the CNN reaches the lower test error with an MAE of 0.27% and
outperforms the MLP with an MAE of 0.539%. Therefore, the CNN is identified as the
most suitable model. Further, the optimization methods have a higher impact on the CNN.
By means of the data augmentation method, it is possible to nearly halve the test error from
0.505% to the lowest error of 0.27%. Thus, the data augmentation method shows itself as
an effective way to optimize the estimation model.
In the future, it is planned to further develop the data augmentation technique.
The current method consists of linear estimations for grid points. The values between the
grid points are interpolated. It should be examined if the estimations can be improved using
other approaches while keeping the computing time and efforts low. Further, the proposed
algorithm should be tested against a higher variety of loading conditions to evaluate
possible fields of application and to compete with traditional estimation approaches.
Author Contributions: Conceptualization, S.P.; methodology, S.P.; software, S.P.; validation, S.P. and
A.N.; formal analysis, S.P. and A.N.; investigation, S.P.; resources, S.P. and A.M.; data curation, A.M.;
writing—original draft preparation, S.P.; writing—review and editing, S.P., A.M., M.K., A.N., and
T.W.; visualization, S.P.; supervision, M.K., A.N., and T.W.; project administration, M.K., A.N., and
T.W.; funding acquisition, M.K., A.N., and T.W. All authors have read and agreed to the published
version of the manuscript.
Funding: This research [project MORE] is funded by dtec.bw—Digitalization and Technology
Research Center of the Bundeswehr, which we gratefully acknowledge. dtec.bw is funded by the
European Union—NextGenerationEU. Further, we acknowledge financial support by the University
of the Bundeswehr Munich.
169
Energies 2023, 16, 6750
Abbreviations
The following abbreviations are used in this manuscript:
ANN Artificial neural network

BEV Battery electric vehicle
BMS Battery management system
CNN Convolutional neural network
LIB Lithium-ion battery
MAE Mean absolute error
ML Machine learning
MLP Multilayer perceptron
PCA Principal component analysis
RMSE Root mean square error
RNN Recurrent neural network
SoC State-of-Charge
SVD Singular value decomposition
SVM Support vector machine
References
1. Buberger, J.; Kersten, A.; Kuder, M.; Eckerle, R.; Weyh, T.; Thiringer, T. Total CO2 -equivalent life-cycle emissions from commercially
available passenger cars. Renew. Sustain. Energy Rev. 2022, 159, 112158. [CrossRef]
2. Bernhart, W. Challenges and Opportunities in Lithium-ion Battery Supply. In Future Lithium-Ion Batteries; The Royal Society of
Chemistry: London, UK, 2019; pp. 316–334. [CrossRef]
3. Schulte, J.; Figgener, J.; Woerner, P.; Broering, H.; Sauer, D.U. Forecast-based charging strategy to prolong the lifetime of
lithium-ion batteries in standalone PV battery systems in Sub-Saharan Africa. Sol. Energy 2023, 258, 130–142. [CrossRef]
4. Stock, S.; Pohlmann, S.; Günter, F.J.; Hille, L.; Hagemeister, J.; Reinhart, G. Early Quality Classification and Prediction of Battery
Cycle Life in Production Using Machine Learning. J. Energy Storage 2022, 50, 104144. .: 10.1016/j.est.2022.104144. [CrossRef]
5. Castro, F.D.; Cutaia, L.; Vaccari, M End-of-life automotive lithium-ion batteries (LIBs) in Brazil: Prediction of flows and revenues
by 2030. Resour. Conserv. Recycl. 2021, 169, 105522. [CrossRef]
6. Boxall, N.J.; King, S.; Cheng, K.Y.; Gumulya, Y.; Bruckard, W.; Kaksonen, A.H. Urban mining of lithium-ion batteries in Australia:
Current state and future trends. Miner. Eng. 2018, 128, 45–55. [CrossRef]
7. Wang, Y.; Tian, J.; Sun, Z.; Wang, L.; Xu, R.; Li, M.; Chen, Z. A comprehensive review of battery modeling and state estimation
approaches for advanced battery management systems. Renew. Sustain. Energy Rev. 2020, 131, 110015. [CrossRef]
8. Lee, J.H.; Lee, I.S. Lithium Battery SOH Monitoring and an SOC Estimation Algorithm Based on the SOH Result. Energies 2021,
14, 4506. [CrossRef]
9. Shchurov, N.I.; Dedov, S.I.; Malozyomov, B.V.; Shtang, A.A.; Martyushev, N.V.; Klyuev, R.V.; Andriashin, S.N. Degradation of
Lithium-Ion Batteries in an Electric Transport Complex. Energies 2021, 14, 8072. [CrossRef]
10. Bonfitto, A. A Method for the Combined Estimation of Battery State of Charge and State of Health Based on Artificial Neural
Networks. Energies 2020, 13, 2548. [CrossRef]
11. Ng, K.S.; Moo, C.S.; Chen, Y.P.; Hsieh, Y.C. Enhanced coulomb counting method for estimating state-of-charge and state-of-health
of lithium-ion batteries. Appl. Energy 2009, 86, 1506–1511. [CrossRef]
12. Chaoui, H.; Mandalapu, S. Comparative Study of Online Open Circuit Voltage Estimation Techniques for State of Charge
Estimation of Lithium-Ion Batteries. Batteries 2017, 3, 12. [CrossRef]
13. Marcicki, J.; Canova, M.; Conlisk, A.T.; Rizzoni, G. Design and parametrization analysis of a reduced-order electrochemical
model of graphite/LiFePO4 cells for SOC/SOH estimation. J. Power Sources 2013, 237, 310–324. [CrossRef]
14. Hossain, M.; Haque, M.E.; Arif, M.T. Kalman filtering techniques for the online model parameters and state of charge estimation
of the Li-ion batteries: A comparative analysis. J. Energy Storage 2022, 51, 104174. [CrossRef]
15. Shrivastava, P.; Soon, T.K.; Idris, M.Y.I.B.; Mekhilef, S. Overview of model-based online state-of-charge estimation using Kalman
filter family for lithium-ion batteries. Renew. Sustain. Energy Rev. 2019, 113, 109233. [CrossRef]
16. Luo, Y.; Qi, P.; Kan, Y.; Huang, J.; Huang, H.; Luo, J.; Wang, J.; Wei, Y.; Xiao, R.; Zhao, S. State of charge estimation method
based on the extended Kalman filter algorithm with consideration of time–varying battery parameters. Int. J. Energy Res. 2020,
44, 10538–10550. [CrossRef]
17. Sharma, P.; Bora, B.J. A Review of Modern Machine Learning Techniques in the Prediction of Remaining Useful Life of Lithium-Ion
Batteries. Batteries 2023, 9, 13. [CrossRef]
18. Hannan, M.A.; Lipu, M.S.H.; Hussain, A.; Ker, P.J.; Mahlia, T.M.I.; Mansor, M.; Ayob, A.; Saad, M.H.; Dong, Z.Y. Toward Enhanced
State of Charge Estimation of Lithium-ion Batteries Using Optimized Machine Learning Techniques. Sci. Rep. 2020, 10, 4687.
[CrossRef]
19. Basia, A.; Simeu-Abazi, Z.; Gascard, E.; Zwolinski, P. Review on State of Health estimation methodologies for lithium-ion
batteries in the context of circular economy. CIRP J. Manuf. Sci. Technol. 2021, 32, 517–528. [CrossRef]
170
Energies 2023, 16, 6750
20. Álvarez Antón, J.C.; García Nieto, P.J.; Blanco Viejo, C.; Vilán Vilán, J.A. Support Vector Machines Used to Estimate the Battery
State of Charge. IEEE Trans. Power Electron. 2013, 28, 5919–5926. [CrossRef]
21. Deng, Z.; Hu, X.; Lin, X.; Che, Y.; Xu, L.; Guo, W. Data-driven state of charge estimation for lithium-ion battery packs based on
Gaussian process regression. Energy 2020, 205, 118000. [CrossRef]
22. Cui, Z.; Wang, L.; Li, Q.; Wang, K. A comprehensive review on the state of charge estimation for lithium-ion battery based on
neural network. Int. J. Energy Res. 2022, 46, 5423–5440. [CrossRef]
23. Li, X.; Jiang, H.; Guo, S.; Xu, J.; Li, M.; Liu, X.; Zhang, X.; Bhardwaj, A. SOC Estimation of Lithium-Ion Battery for Electric Vehicle
Based on Deep Multilayer Perceptron. Comput. Intell. Neurosci. 2022, 2022, 3920317. [CrossRef] [PubMed]
24. Li, S.; Ju, C.; Li, J.; Fang, R.; Tao, Z.; Li, B.; Zhang, T. State-of-Charge Estimation of Lithium-Ion Batteries in the Battery Degradation
Process Based on Recurrent Neural Network. Energies 2021, 14, 306. [CrossRef]
25. Jiao, M.; Wang, D.; Qiu, J. A GRU-RNN based momentum optimized algorithm for SOC estimation. J. Power Sources 2020,
459, 228051. [CrossRef]
26. Qian, C.; Xu, B.; Chang, L.; Sun, B.; Feng, Q.; Yang, D.; Wang, Z. Convolutional neural network based capacity estimation using
random segments of the charging curves for lithium-ion batteries. Energy 2021, 227, 120333. [CrossRef]
27. Bian, C.; Yang, S.; Liu, J.; Zio, E. Robust state-of-charge estimation of Li-ion batteries based on multichannel convolutional and
bidirectional recurrent neural networks. Appl. Soft Comput. 2022, 116, 108401. [CrossRef]
28. Hu, W.; Peng, Y.; Wei, Y.; Yang, Y. Application of Electrochemical Impedance Spectroscopy to Degradation and Aging Research of
Lithium-Ion Batteries. J. Phys. Chem. C 2023, 127, 4465–4495. [CrossRef]
29. Naaz, F.; Herle, A.; Channegowda, J.; Raj, A.; Lakshminarayanan, M. A generative adversarial network-based synthetic data
augmentation technique for battery condition evaluation. Int. J. Energy Res. 2021, 45, 19120–19135. . [CrossRef]
30. Qiu, X.; Wang, S.; Chen, K. A conditional generative adversarial network-based synthetic data augmentation technique for
battery state-of-charge estimation. Appl. Soft Comput. 2023, 142, 110281. [CrossRef]
31. Channegowda, J.; Maiya, V.; Joshi, N.; Raj Urs, V.; Lingaraj, C. An attention-based synthetic battery data augmentation technique
to overcome limited dataset challenges. Energy Storage 2022, 4, e354. [CrossRef]
32. Fayyad, U.; Piatetsky-Shapiro, G.; Smyth, P. From Data Mining to Knowledge Discovery in Databases. AI Mag. 1996, 17, 37.
[CrossRef]
33. Zhang, F.; Lai, T.L.; Rajaratnam, B.; Zhang, N.R. Cross-Validation and Regression Analysis in High-Dimensional Sparse Linear Models;
Stanford University: Stanford, CA, USA, 2011.
34. Kondo, M.; Bezemer, C.P.; Kamei, Y.; Hassan, A.E.; Mizuno, O. The impact of feature reduction techniques on defect prediction
models. Empir. Softw. Eng. 2019, 24, 1925–1963. [CrossRef]
35. Debie, E.; Shafi, K. Implications of the curse of dimensionality for supervised learning classifier systems: Theoretical and
empirical analyses. Pattern Anal. Appl. 2019, 22, 519–536. [CrossRef]
36. Gruosso, G.; Storti Gajani, G.; Ruiz, F.; Valladolid, J.D.; Patino, D. A Virtual Sensor for Electric Vehicles’ State of Charge Estimation.
Electronics 2020, 9, 278. [CrossRef]
37. Yanai, H.; Takeuchi, K.; Takane, Y. Projection Matrices, Generalized Inverse Matrices, and Singular Value Decomposition, 1st ed.;
Statistics for Social and Behavioral Sciences; Springer: New York, NY, USA, 2011.
38. Tipping, M.E.; Bishop, C.M. Mixtures of Probabilistic Principal Component Analyzers. Neural Comput. 1999, 11, 443–482.
[CrossRef] [PubMed]
39. Zhou, N.; Zhao, X.; Han, B.; Li, P.; Wang, Z.; Fan, J. A novel quick and robust capacity estimation method for Li-ion battery cell
combining information energy and singular value decomposition. J. Energy Storage 2022, 50, 104263. [CrossRef]
40. Joshi, A.V. Machine Learning and Artificial Intelligence; Springer International Publishing: Berlin/Heidelberg, Germany, 2019.
41. Yu, Y.; Feng, Y. Modified Cross-Validation for Penalized High-Dimensional Linear Regression Models. J. Comput. Graph. Stat.
2013, 23, 1009–1027. [CrossRef]
42. Chemali, E.; Kollmeyer, P.J.; Preindl, M.; Emadi, A. State-of-charge estimation of Li-ion batteries using deep neural networks: A
machine learning approach. J. Power Sources 2018, 400, 242–255. [CrossRef]
43. Feng, F.; Teng, S.; Liu, K.; Xie, J.; Xie, Y.; Liu, B.; Li, K. Co-estimation of lithium-ion battery state of charge and state of temperature
based on a hybrid electrochemical-thermal-neural-network model. J. Power Sources 2020, 455, 227935. [CrossRef]
44. He, W.; Williard, N.; Chen, C.; Pecht, M. State of charge estimation for Li-ion batteries using neural network modeling and
unscented Kalman filter-based error cancellation. Int. J. Electr. Power Energy Syst. 2014, 62, 783–791. [CrossRef]
45. Hannan, M.A.; Lipu, M.S.H.; Hussain, A.; Saad, M.H.; Ayob, A. Neural Network Approach for Estimating State of Charge of
Lithium-Ion Battery Using Backtracking Search Algorithm. IEEE Access 2018, 6, 10069–10079. [CrossRef]
46. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2017, arXiv:1412.6980.
47. Pedrycz, W.; Chen, S.M. Interpretable Artificial Intelligence: A Perspective of Granular Computing, 1st ed.; Studies in Computational
Intelligence; Springer International Publishing: Cham, Switzerland, 2021. [CrossRef]
48. Kiranyaz, S.; Avci, O.; Abdeljaber, O.; Ince, T.; Gabbouj, M.; Inman, D.J. 1D convolutional neural networks and applications: A
survey. Mech. Syst. Signal Process. 2021, 151, 107398. [CrossRef]
49. Hu, C.; Ma, L.; Guo, S.; Guo, G.; Han, Z. Deep learning enabled state-of-charge estimation of LiFePO4 batteries: A systematic
validation on state-of-the-art charging protocols. Energy 2022, 246, 123404. [CrossRef]
171
Energies 2023, 16, 6750
50. Cui, Z.; Kang, L.; Li, L.; Wang, L.; Wang, K. A hybrid neural network model with improved input for state of charge estimation of
lithium-ion battery at low temperatures. Renew. Energy 2022, 198, 1328–1340. [CrossRef]
51. Hannan, M.A.; How, D.N.T.; Lipu, M.S.H.; Ker, P.J.; Dong, Z.Y.; Mansur, M.; Blaabjerg, F. SOC Estimation of Li-ion Batteries With
Learning Rate-Optimized Deep Fully Convolutional Network. IEEE Trans. Power Electron. 2021, 36, 7349–7353. [CrossRef]
52. Rebala, G.; Ravi, A.; Churiwala, S. An Introduction to Machine Learning; Springer International Publishing: San Ramon, CA, USA;
San Jose, CA, USA; Hyderabad, Telangana, India , 2019. [CrossRef]
53. Zhang, D.; Zhong, C.; Xu, P.; Tian, Y. Deep Learning in the State of Charge Estimation for Li-Ion Batteries of Electric Vehicles: A
Review. Machines 2022, 10, 912. [CrossRef]
54. Tian, Y.; Xia, B.; Sun, W.; Xu, Z.; Zheng, W. A modified model based state of charge estimation of power lithium-ion batteries
using unscented Kalman filter. J. Power Sources 2014, 270, 619–626. [CrossRef]
55. Li, X.; Huang, Z.; Tian, J.; Tian, Y. State-of-charge estimation tolerant of battery aging based on a physics-based model and an
adaptive cubature Kalman filter. Energy 2021, 220, 119767. [CrossRef]
56. Wang, L.; Lu, D.; Liu, Q.; Liu, L.; Zhao, X. State of charge estimation for LiFePO4 battery via dual extended kalman filter and
charging voltage curve. Electrochim. Acta 2019, 296, 1009–1017. [CrossRef]
172
energies
Article
3D Heterogeneous Model for Electrodes in Lithium-Ion
Batteries to Study Interfacial Detachment of Active Material
Particles and Carbon-Binder Domain
Mohammadali Mirsalehian 1, *, Bahareh Vossoughi 1 , Jörg Kaiser 2 and Stefan Pischinger 1,2
1 Chair of Thermodynamics of Mobile Energy Conversion Systems (TME), RWTH Aachen University,
Forckenbeckstraße 4, 52074 Aachen, Germany; bahareh.vossoughi@rwth-aachen.de (B.V.);
pischinger_s@tme.rwth-aachen.de (S.P.)
2 FEV Europe GmbH, Neuenhofstraße 181, 52078 Aachen, Germany
* Correspondence: mirsalehian@tme.rwth-aachen.de
Abstract: Mechanics plays a crucial role in the performance and lifespan of lithium-ion battery
(LIB) cells. Thus, it is important to address the interplay between electrochemistry and mechanics
in LIBs, especially when aiming to enhance the energy density of electrodes. Accordingly, this
work introduces a framework for a fully coupled electro-chemo-mechanical heterogeneous 3D
model that allows resolving the inhomogeneities accompanied by electrochemical and mechanical
responses of LIB electrodes during operation. The model is employed to numerically study the
mechanical degradation of a nickel manganese cobalt (NMC) cathode electrode, assembled in a half-
cell, upon cycling. As opposed to previous works, a virtual morphology for a high-energy electrode
with low porosity is developed in this study, which comprises distinct domains of active material
(AM) particles, the carbon-binder domain (CBD), and the pore domain to resemble real commercial
electrodes. It is observed that the mechanical strain mismatch between irregularly and randomly
positioned AM particles and the CBD might lead to local contact detachment. This interfacial gap,
in combination with the diminishing contact strength over cell cycling, continuously deteriorates
the electrode performance upon cycling by impedance rise and capacity drop. In agreement with
Citation: Mirsalehian, M.; Vossoughi, previous experimental reports, the presented simulation results exhibit that the contact loss mostly
B.; Kaiser, J.; Pischinger, S. 3D takes place in the regions closer to the separator. Eventually, the resulting gradual capacity drop and
Heterogeneous Model for Electrodes change in impedance spectrum over cycling, as the consequence of interfacial gap formation, are
in Lithium-Ion Batteries to Study discussed and indicated.
Interfacial Detachment of Active
Material Particles and Carbon-Binder
Keywords: lithium-ion batteries; electrode microstructure; heterogeneous physical model;
Domain. Energies 2023, 16, 7391.
mechanical degradation; electrochemical impedance spectroscopy
https://doi.org/10.3390/en16217391
Academic Editors: Daniel-Ioan Stroe

and Simone Barcellona
1. Introduction
Received: 5 September 2023
Revised: 11 October 2023 Lithium-ion batteries have established themselves as the predominant energy stor-
Accepted: 24 October 2023 age system due to their superior energy and power densities, as well as their extended
Published: 1 November 2023 operational lifespan. Therefore, they have become the prevailing choice for energy stor-
age across a diverse spectrum of applications. These applications encompass electric
vehicles (EVs), consumer electronics, energy storage systems, and various eco-friendly in-
dustries [1]. Nonetheless, given the ever-growing demand within industries for heightened
energy density, enhanced performance, and cost-effective solutions, ongoing optimization
investigations remain a vital necessity in the pursuit of advancing LIB technology [2].
Given the substantial time and resources required for the development and char-
acterization of new lithium-ion cell types in both research laboratories and commercial
cell manufacturing facilities, computational modeling and simulation emerge as efficient
tools, offering fast and in-depth insights at a microscopic level into the intricate processes
4.0/).
occurring within lithium-ion electrodes during charging and discharging. In this context,

173
Energies 2023, 16, 7391
the continuum model, a widely employed physics-based framework [3,4], simplifies the
complex electrode microstructure by treating it as a continuum through the application
of approximation methods. This simplification substantially reduces computational costs.
However, it is essential to acknowledge that the existing insufficiencies inherent in the
continuum model limit its capability in a range of applications [5–7].
The microstructural architecture and spatial arrangement of domains constituting
Lithium-ion battery electrodes exert profound influence over the complicated interplay
between diverse electrochemical and mechanical processes during LIB operation [8]. Con-
sequently, to attain a comprehensive understanding of the effect of electrode structure on
LIB performance, it becomes imperative to construct a virtual representation that emu-
lates the actual, heterogeneous electrode structure. Creating a model from tomography
imaging is an expensive and time-intensive endeavor, entailing the inevitable destruction
of samples, thereby imposing limitations on iterative optimization and model enhance-
ment, especially in the context of prototype electrodes or cells [9–11]. Additionally, the
data derived from imaging techniques might be insufficient for comprehending battery
electrochemistry and mechanics due to various constraints. In contrast, the computational
modeling approach offers a conceptual design in the initial phases of research projects,
facilitating optimization with respect to desired functionalities. In other words, virtual
material testing and experimentation with diverse electrode properties can be conducted
efficiently and cost-effectively [10].
Developing a heterogeneous model provides deep insights into localized non-uniformities
within the electrode structure, leading to uneven utilization of active materials. Such inho-
mogeneity may result in degradation phenomena, like lithium plating on anode material
particles [12,13], particularly those near the separator interface during fast charging of
electric vehicles. Additionally, this model enables investigations into optimized battery
operating parameters and contributes to the advancement of microstructural engineering
strategies [5,14–16] aimed at mitigating aging phenomena within the electrode.
Mechanics play a pivotal role in determining both the performance and longevity
of lithium-ion batteries. With the growing demand for extended cycle life, fast charging,
and increased driving range in EV applications, mechanical degradation is one of the
obstacles that directly regulates the mechanisms of capacity deterioration [17,18]. There-
fore, the mutual impacts of electrochemistry and mechanics in LIB cells need to be ad-
dressed in the pursuit of electrodes with high energy density. In this regard, the study of
electrochemical–mechanical interactions within an electrode structure can be effectively
conducted through the application of heterogeneous models. This encompasses the exami-
nation of the impact of lithiation and delithiation on the evolution of mechanical strain and
stress within active material particles [19,20], the influence of the mechanics on electrochem-
ical response [21–24], and the exploration of potential fracture nucleation and propagation
phenomena within the electrode structure [25–27]. Accordingly, the model enables the de-
velopment of measures to suppress mechanical degradation, ranging from the electrode’s
geometrical features to mechanical properties to operational considerations such as cycling
voltage windows.
In addition to the fracture inside active material as one degradation mechanism,
the mismatch between the mechanical behaviors of AM particles and inactive materials
may cause decohesion at the interfaces between them. The conductive additive, which is
typically carbon black, provides the electronic conduction pathways between the active
particles and the current collector (CC). The contact between the AM particle and the
conductive additive could be lost during cycling, partially hindering the electronic path.
As a result, the impedance of the electrode increases, and the capacity retention is reduced,
leading to electrode performance degradation. In contrast to the particle fracture, the
investigation of mechanical failure at the interface has received comparatively less attention.
This disparity arises from the difficulty in experimentally resolving the intricate interfacial
interactions between the active materials and the conductive agents. Nevertheless, the
scanning electron microscope (SEM) images presented by some researchers confirmed that
174
Energies 2023, 16, 7391
the continuing deformation of AM particles during cycling can result in decohesion at the
interface between the active material and the conductive agents [28]. Figure 1 shows cross-
sectional SEM images of an NMC cathode electrode before and after 10 cycles at 5 C-rate
to visualize the interfacial debonding. While Figure 1a illustrates the existing cohesion
between the particle and surrounding inactive matrix, the incurred interfacial decohesion
in the cycled state is highlighted by the red dashed lines in Figure 1b. In the rapidly
advancing field of solid-state batteries, the occurrence of delamination between particles
and the solid electrolyte (SE) can be more pronounced compared to conventional batteries
employing liquid electrolytes. Figure 2 contains SEM images showing the developing bond
at the interface between NMC particles and solid electrolyte in three states: the pristine
state, after the first cycle, and after 50 cycles. The SEM images distinctly reveal the NMC
particles due to their characteristic, nearly spherical morphology. While the NMC particles
and the solid electrolyte form a densely packed composite electrode in the pristine state,
(Figure 2a,b), the NMC particles are surrounded by gaps after the first charging or after
50 cycles, (Figure 2c–f).
Figure 1. Cross-sectional SEM images of NMC particles and surrounding inactive matrix in (a) pris-
tine state and (b) cycled state after 10 cycles at 5 C-rate of a lithium-ion cell cathode reported by
Xu et al. [28] (shown with publisher’s permission).
The mechanistic comprehension of LIB cell degradation faces a critical challenge due
to the intrinsic heterogeneity in the structural configuration and the composition of the
electrode. While Yang et al. [29] have demonstrated a noteworthy spatial and temporal de-
pendency of damage in a commercial electrode, many prior experimental studies primarily
concentrated on an idealized structure, such as thin films and individual particles. Others
focused on localized regions within the electrode, characterized by limited dimensions,
thereby being unable to provide a statistical representation of the overall chemomechanical
behavior exhibited by the electrodes at a large scale. Hence, the exploration of heteroge-
neous chemomechanics and damage within lithium-ion batteries remains at its early stages
of development, with a multitude of questions yet to be elucidated. The above-mentioned
experimental challenges also exist in numerical investigations. The common continuum
modeling approach and the extended single-particle approach oversimplify and assume
free-standing spherical particles of the same size that are not subject to mechanical or
electrochemical constraints from the adjacent environment. Recently, Baboo et al. [30]
employed a simple single-particle model (SPM) and adjusted the solid diffusivity and
specific active surface area to predict the effects of the formation of solid electrolyte inter-
phase (SEI), side reactions, and AM particle fragmentation on the degradation of a LIB
full cell performance during cycling. Their investigation was conducted for three different
binders with diverse adhesion strengths. However, the simple SPM was unable to replicate
the decreasing capacity during cycling, which was observed in the experiment. It can be
attributed to the existing simplifications in SPM to express the complicated and spatially-
resolved physicochemical interactions taking place inside the LIB cell during operation.
Nonetheless, in commercial batteries, both the cathode and anode exhibit a composite
175
Energies 2023, 16, 7391
nature characterized by substantial heterogeneity on the nano- to microscale, comprising

irregularly shaped and sized active particles within a matrix of polymeric binders and
conductive additives, along with pores containing electrolyte.
Figure 2. SEM images of the cathode composite of NMC particles embedded in the solid electrolyte
(Li3 PS4 ) (a,b) as prepared in pristine state, (c,d) after single charging at 0.1 C-rate and, (e,f) after 50 full
cycles in the discharged state provided, by Koerver et al. [31] (shown with publisher’s permission).
Recently, a few researchers have developed experimental setups and fully coupled
electro-chemo-mechanical heterogeneous models to investigate the interface of AM parti-
cles and inactive matrix in electrodes. Xu et al. [28] employed experimental tomography
to visualize nanoscale interfacial debonding between NMC particles and the conductive
matrix. Their observations revealed that the interfacial detachment during cycling pri-
marily occurs in the electrode region near the separator, rather than in the vicinity of the
current collector. Moreover, Xu et al. built a microstructural-resolved model utilizing tomo-
graphic data to examine the heterogeneous damage within composite cathode particles
situated within the binder matrix. However, the volumetric fraction of AM particles in
the reconstructed virtual morphology by Xu et al. equates εave AM = 40% and is thus not as
dense as the current commercial electrodes with high energy densities. Moreover, the pore
and carbon-binder domains in the re-built morphology in their work are simplified and
represented by a single integrated composite domain. Later, Liu et al. [32] also investigated
the degradation of contacts between the AM particles and the inactive matrix in an NMC
176
Energies 2023, 16, 7391
half-cell electrode using a reconstructed morphology with similar simplified geometrical

features as developed in the work of Xu et al. [28]. Accordingly, Liu et al. adopted a
phenomenological approach to reproduce the mechanical fatigue of cohesion along the AM
particles and the inactive matrix during cycling. It was observed that the steady decay of
the interfacial strength causes a growing interfacial debonding at cycling.
In this work, a 3D microstructural-resolved model is developed to study the impact of
degradation of contact strength between the AM particles and CBD in the electrode in LIB
cells during cycling. In contrast to previous models, the constructed virtual morphology
of the electrode encompasses distinguished domains to represent the individual existing
components inside the real electrode. Such a heterogeneous model enables predicting the
electrochemical and mechanical behavior of the electrode’s constituents more reliably. In
the end, the influence of interfacial decohesion upon cycling is studied and discussed in
terms of capacity retention and electrochemical impedance spectroscopy (EIS) in the time
and frequency domains, respectively.
2. Methods
2.1. Electrode Microstructure
In this work, the heterogeneous model, introduced by the authors in their previous
work [2], is employed to build a fully coupled electro-chemo-mechanical model and study
the interfacial bonding of NMC622 AM particles and carbon-binder domain in a cathode
electrode with conventional liquid electrolyte. The presented model allows to observe
the heterogeneous electrochemistry, stress, and interfacial detachment in a commercial
cathode electrode. Therefore, the spatial and temporal variations on different levels can be
investigated. The offered modeling approach in this work can be further utilized to study
the mechanical degradation in solid-state batteries as well.
As opposed to the developed cathode electrode’s virtual morphology with low AM
domain volumetric fraction by Xu et al. [28] and Liu et al. [32], a denser cathode electrode
AM = 62.83% has been built in this study to represent an electrode with higher
with εave
energy density. In addition, the generated morphology contains separate distinguished
domains of pore and CBD that increase the reliability of simulation results as compared to
the assumed simplification made by the previous research to consider these two domains
as one integrated domain.
Subsequently, a half-cell representative volume element (RVE) is assembled by in-
corporating the developed cathode electrode microstructure and two cubic domains. The
resulting model explicitly encompasses the crucial half-cell components, including the
separator, current collector, and lithium foil, as vividly illustrated in Figure 3. Table 1
includes the considered geometrical properties in the generation of the half-cell RVE.
The simulations are carried out in COMSOL Multiphysics® version 6.0 using an Intel®
Xeon® CPU @ 3.70 GHz (2 processors) with 128 GB of RAM. Moreover, COMSOL LiveLink®
for MATLAB® is used in the current work to transfer model’s mesh data from MATLAB ®
version R2020b to COMSOL Multiphysics® software to create the FEM model.
Table 1. Half-cell RVE microstructure generation specifications.
Property Value
Cathode microstructure dimension 50 × 50× 25 μm3
Average NMC particle radius r p 2 μm and 5 μm
Particle radius standard deviation 0.2 μm
Active material volumetric fraction (εave
AM ) 62.83%
Carbon-binder domain volumetric fraction (εave
CBD ) 10.55%
ave
Porosity (ε e ) 26.6%
Separator thickness 11 μm
Al current collector thickness 6.5 μm
177
Energies 2023, 16, 7391
Figure 3. Developed 3D half-cell RVE model and associated boundary conditions (BCs) to study
AM-CBD degrading contact.
2.2. Electro-Chemo-Mechanical Model

The governing equations to describe diverse coupled electrochemical and mechan-
ical interactions within various domains, along with their corresponding parameters,
are delineated in Table 2, Table 3 and Table 4, respectively. In the equations and as-
sociated parameters, the subscripts e and s denote the electrolyte and the solid-phase
properties, respectively.
Table 2. List of physics implemented in different domains of the half-cell.
Domain/Boundary Equation No.

Electrochemistry
∗ ∗∗
AM, CBD, CC ∇ · (σseff ∇φs ) − jtotal = 0 (1)
∂cs Ds Ω (2)
∂t = ∇.( Ds ∇ cs ) − ( RT ∇ c · ∇ σh + c ∇ · ∇ σh )
AM
∂(ε e∗∗∗ ce ) 1− t0 ∗∗
(3)
Pore, CBD, separator
∂t = ∇ · Deeff ∇ce + F + jtotal
∗∗
∇.(κ ∇φe ) + ∇ · (κ D ∇ ln(ce )) + jtotal
eff eff =0 (4)
CBD, Separator Deeff = (ε e ) p · Debulk (5)
κeeff = (ε e ) p · κebulk (6)
CBD σseff = (ε s ) p· σsbulk (7)
α a Fη αc Fη (8)
AM-electrolyte interface j f = Av · j0 exp RT − exp − RT
αa
j0 = Fkcαs c (cs,max − cs )αa ce,ref ce (9)
Ωσh (10)
η = φs − φe − U − F
∂(φs −φe ) (11)
jc = Av · ∂t · CDL
j total = j + jc
f (12)
Mechanics
All domains ∇·σ = 0 (13)
σ = C! : e " (14)
= 12 ∇u + (∇u) T (15)
= e + Li∗∗∗∗ (16)
AM Li = Ω3 ( cs − cs,0 )I (17)
AM-CBD interface Kint = Kint,0 − k d ncyc − 1 (18)
∗ σeff for CBD domain while σbulk in AM and CC domains. ∗∗ jtotal = 0 in CC and separator domains. ∗∗∗ εe
s s
excluded in equation for pore domain. ∗∗∗∗ Li = 0 for all domains except AM.
2.2.1. Solid Domains

Ohm’s law, Equation (1) in Table 2, is solved in the AM, CBD, and CC domains to
describe electronic migration. Current collector domains are impermeable to lithium ions,
so normal ion flux is set to zero at their boundaries.
178
Energies 2023, 16, 7391
At the interfaces between AM and the electrolyte, the Butler–Volmer (BV) equation is
used to describe the electrochemical reaction kinetics or the volumetric faradaic current
density j f . In the BV equation, the total overpotential is modified by introducing a stress-
biased term to account for the effect of mechanics on the kinetics of the faradic reaction.
Accordingly, the fourth term on the right side of Equation (10) in Table 2 represents the
stress-biased term when Ω and σh denote the partial molar volume of NMC particles
and the hydrostatic stress, respectively. In addition, to account for the formation of the
double layer (DL) at this interface, a homogeneous volumetric capacitive current density
represented as jc is incorporated alongside the faradaic current density. These components
collectively contribute to the determination of the total volumetric current density, denoted
as jtotal , at these interfaces.
To describe solid-state diffusion, the formulation of Fick’s second law is modified to
account for the stress-induced diffusion flux, Equation (2) in Table 2. The additional terms
inside the second bracket on the right side of this equation reflect the stress-driven flux of
the lithium inside the AM particles. Then, at the interface between the active material and
the electrolyte, the Neumann boundary condition is established to consider the interplay
between the faradaic current density and the diffusion processes occurring within the
AM particles.
2.2.2. Pore and Separator Domains

It is assumed that the pore domain located between the solid domains is completely
saturated with the liquid electrolyte. The separator domain is treated as an electrolyte-
permeable media with a porosity of 50%. To account for the separator’s porous nature,
the Bruggeman relation is utilized to predict the effective transport coefficients of the
separator domain.
Also, a rather small porosity of 27.6% is assigned to the CBD to account for the
reported porosity factor by Daemi et al. [33]. Accordingly, the electrolyte’s effective ionic
conductivity and diffusivity in CBD are obtained by the corresponding bulk values scaled
by a factor of 0.276, following Boyce et al. [34].
The concentrated solution theory is employed to mathematically describe electrolyte
transport. A porosity factor ε e is introduced to Equations (3) and (4) in Table 2 within
domains of the separator and the CBD to consider their porous nature, which is not
explicitly included in their re-built modeled morphology.
2.2.3. Lithium Foil Boundary

The lithium foil is modeled as a boundary, and the BV equation is used to compute
the volumetric current density associated with lithium metal deposition and dissolution at
the interface between the lithium foil and the electrolyte. Here, the Butler–Volmer equation
is formulated with zero overpotential in order to not introduce additional impedance. As a
result, the calculated impedances are primarily attributed to contributions originating from
the porous separator and cathode microstructure, saturated with electrolyte, and a current
collector foil attached.
2.2.4. Mechanics
A mechanical analysis is performed to consider the effects of lithiation-induced de-
formation on both mechanical and electrochemical responses of the various electrode
constituent domains. The heterogeneity of the reaction kinetics and the non-uniform me-
chanical behavior of the active and inactive materials lead to the development of stresses
followed by mechanical degradation in the cell. Due to the rather slow kinetics of solid
diffusion within AM particles, the mechanical equilibrium equation is solved in the absence
of body forces, Equation (13) in Table 2, where σ denotes the Cauchy stress tensor. In the
case of the cathodic NMC electrode with limited expansion upon lithiation, Hooke’s law,
Equation (14), is employed as the constitutive relation to describe the linear elastic material
model. Accordingly, C and e refer to the elastic stiffness tensor and the elastic strain,
179
Energies 2023, 16, 7391
respectively. By assuming small strains and rigid body rotations, the Green-Lagrange strain
tensor is reduced by the removal of quadratic terms, resulting in the well-known compati-
bility relation, Equation (15), which correlates the total strain tensor and the symmetric part
of the gradient of the displacement field. The total strain is further decomposed into the
elastic and lithiation-induced parts, as expressed in Equation (16). The lithiation-induced
strain in the AM particles, Li , is attained by Equation (17), in analogy to the thermal strain
in materials, where I is the identity tensor. An initial active material concentration, cs,0 ,
is prescribed whilst the electrode and all constituent domains are assumed to be initially
in an unstressed state. Equation (17) formulates the contribution of lithiation-induced
deformation by the development of hydrostatic stress. As illustrated in Figure 3, while
symmetrical boundary condition is applied on the RVE sides in the x-y and x-z planes,
fully-clamped and free boundary conditions are allocated for the remaining two sides in
the y-z plane at x = lelectrode and x = 0, respectively.
In order to evaluate the strength along the interfaces between the AM particles and the
carbon-binder domain, an interfacial cohesion model is required. Despite the model devel-
oped by Müller et al. [35], which considers a serial spring and a dissipative damper to repro-
duce the mechanical response of the contact between the AM particles and carbon-binder
domain, the induced mechanical hysteresis effect is ignored in this work. Accordingly,
a spring layer with zero thickness across the electrode represents the interfacial contact
between the AM cathode particles and the CBD. The constant value of this spring per unit
area in the pristine state is denoted as Kint,0 . When the spring expands along the interface
under tensile stress, the contact locally detaches, hindering electron transfer paths and
elevated impedance. On the other hand, the bond is maintained for the remaining contacts,
which are locally under compression or no stress. The above-mentioned spatial contact loss
induces a drop in the local solid potential φs along the debonded contact, which in turn has
a local impact on the charge conservation equation in the solid domain. Moreover, due to
the introduced drop in solid potential by detachment, the overall overpotential η, as the
driving force for electrochemical reaction, is retarded, and consequently, the local reaction
kinetics described via the Butler–Volmer equation is impeded, resulting in degradation of
battery cell performance.
Zhu et al. [36] demonstrated that the interfacial strength gradually decreases due to
mechanical fatigue during cycling. Similar to the phenomenological approach presented
by Liu et al. [32], the fatigue deterioration of contact between AM particles and CBD
during cycling is realized in this work by a decreased spring constant along this interface.
Equation (18) expresses the considered degradation of the spring stiffness per cycle, where
k d represents the rate at which the spring stiffness decreases per cycle, and ncyc indicates
the cycle number.
Table 3. Electrochemical parametrization of the half-cell model.
Parameters Value Ref.

NMC622 Particles
AM solid conductivity (σs ) 1.6 × 10−4 S · m−1 [34]
AM solid diffusivity (Ds ) f (SoL) m2 · s−1 [34]
max AM solid concentration (cs,max ) 48,700 mol · m−3 [34]
Initial AM solid concentration (cs,0 ) 500 mol · m−3
−324.2 · SoL8 + 1034.4 · SoL7 − 129.6 · SoL6 −
Equilibrium potential (U) 777.8 · SoL5 − 214.5 · SoL4 − 9.8 · SoL3 + [34]
8.2 · SoL2 − 2.8 · SoL + 4.4 V
Kinetics
Reaction rate constant (k) 2 × 10−11 m · s−1 [34]
Transfer coefficients (α a , αc ) 0.5
Surface double layer capacitance (CDL ) 0.2 F · m−2 [28]
Bruggeman exponent (p) 1.0 [34]
Current collector
Conductivity (σs ) 3.7 × 107 S · m−1 [34]
180
Energies 2023, 16, 7391
Table 3. Cont.
Parameters Value Ref.

Carbon-binder domain
Conductivity (σs ) 375 S · m−1 [34]
Electrolyte
Initial electrolyte concentration (ce,0 ) 1000 mol · m−3 [34]
Conductivity (κe ) f ( c e ) S · m−1 [37]
Diffusivity (De ) f (ce ) m2 · s−1 [37]
Activity (∂ ln f /∂ ln ce ) 0.43 [37]
Transference (t0+ ) 0.37 [37]
Table 4. Mechanical parameterization of the half-cell model.
Parameter Value Ref.

NMC622 particles
Partial molar volume (Ω) 1.8 × 10−6 m3 · mol−1 [34]
Young’s modulus (E) 140 GPa [34]
Current collector
Young’s modulus (E) 70 GPa [34]
Carbon-binder domain
Young’s modulus (E) 0.3 GPa [34]
Electrolyte
Instantaneous shear modulus (G) 0.3 MPa [38]
AM-CBD interface
Spring constant per unit area in pristine state (Kint,0 ) 2 × 1015 N · m−3 [28]
Decreasing rate of spring constant per cycle (k d ) 0.375 × 1015 N · m−3 [28]

To visualize the expansion of NMC particles upon lithiation and the subsequent
temporally and spatially resolved detachment along the contacts between the AM particles
and the CBD, the half-cell’s first discharge is simulated in the pristine state. The half-cell is
initially in the fully charged state and is then galvanostatically discharged at 1 C-rate to 3 V.
The applied current at different C-rates can be calculated as expressed below.
cs,max · Vtotal
particles · F · Crate
Iapp =
3600 s
where Vtotal
particles represents the total volume of the NMC particles in the reconstructed
cathode electrode. While Figure 4a,b illustrate the state of lithiation (SoL) in the NMC parti-
cles across the cathode electrode at the beginning and end of the discharge process, their
corresponding deformations are depicted in Figure 4c,d. At the beginning of discharge, SoL
looks significantly uniform in the particles across the electrode, and as the discharge contin-
ues, non-uniformity becomes more noticeable. As can be seen in Figure 4b, some particles
are still not highly lithiated at the end of the discharge process, indicating partially isolated
particles, as also observed in previous experimental works on real electrodes [39–41]. Such
heterogeneous electrochemical response in the electrode is due to the spatial arrangement
and orientation of the different constituent domains in the electrode, resulting in different
reaction kinetics.
181
Energies 2023, 16, 7391
(a) Beginning of 1st discharge (b) End of 1st discharge
SoL
0 0.5 1
(c) Beginning of 1st discharge (d) End of 1st discharge
Displacement / μm
0 0.5 1
Figure 4. (a,b) State of lithiation (SoL) of NMC particles and lithium flux stream and, (c,d) mechanical
displacement of the electrode at the beginning (time = 50 s) and end (time = 3220 s) of 1st discharge at
1 C-rate.
(a) (b)
Figure 5. Evolution of the contact area undergoing interfacial detachment outlined by red at
SoL = 50 % (a) at 1st cycle and (b) 5th cycle during discharge at 1 C-rate.
182
Energies 2023, 16, 7391
Figure 4d shows the deformed electrode at the end of discharge compared to the almost
undeformed electrode at the beginning of the lithiation process, Figure 4c. The applied
free boundary condition in the model, as shown in Figure 3, results in the expansion and
contraction of the electrode in the thickness direction. Such displacement can form tension
stress perpendicular to the contact boundaries between AM particles and CBD, depending
on the spatial positioning and arrangement of the constituent domains. Accordingly,
Figure 5a exhibits the interfacial area, which undergoes detachment at SoL = 50% upon
the first discharge process. In agreement with what has been observed by means of the
X-ray contrast tomography technique conducted by Xu et al. [28] and X-ray computed
tomography (XCT) by Parks et al. [42], both the fractured AM particles and the detached
contacts take place mostly close to the separator upon cycling. Moreover, the portion of
contacts experiencing debonding directly relies on the intensity of the spatial mismatch
between the mechanical response of the variously positioned and orientated AM particles
and the CBD at each lithiated state of NMC particles.
To study the effect of the degrading contacts between AM particles and CBD on elec-
trochemical performance during cycling, the fully charged half-cell is galvanostatically
cycled at 1 C-rate for five sequential times in a defined cell voltage window of 3–4.3 V.
Figure 5b exhibits the contact area experiencing decohesion at SoL = 50% during discharge
at the fifth cycle. As noted, compared to the pristine state, a larger area undergoes de-
tachment due to the degraded contact strength between AM particles and CBD during
cycling. This outcome is in agreement with what was observed in SEM images conducted
by Xu et al. [28], shown already in Figure 1, where the gap development between active and
inactive materials occurs in the electrode during cycling. Moreover, Figure 6 illustrates the
evolution of the cell characteristics during cycling. While Figure 6a exhibits the cell voltage
and OCV, Figure 6b,c display the evolution of the averaged values for the state of lithiation
and volumetric strain over the whole NMC particles during the cycling. In Figure 6d, the
averaged interfacial gap associated with the detached contacts at the interfaces between
AM particles and CBD is shown. As can be observed, the degrading contact during cycling,
which is defined by Equation (18), results in a growing detachment as cycling proceeds. In
other words, the diminishing spring constant through cycling leads to not only a larger de-
tached area (Figure 5), but also to the formation of a larger decohesion along these contacts.
Furthermore, the interfacial gap in Figure 6d shows an irreversible growth of detachment
during cycling. This can be explained by the degradation of interfacial strength during
cycling. The calculated interfacial gap maxima in Figure 6d varies from almost 10 nm in
the first cycle to 20 nm in the fifth cycle. Liu et al. [32] performed a numerical study and
attained a range of 50–65 nm for an NMC electrode cycled at 1 C-rate for five times. This
discrepancy primarily originates from two different assumptions in the model built by Liu
et al.: a fully clamped boundary condition was assigned to all RVE sides, and the domains
of pore and CBD were simplified and represented by a single integrated composite domain
with considerably high mechanical stiffness. While the former results in higher absolute
values of interfacial gap during cycling, the latter leads to a narrower interfacial gap range
compared to the computed values in this work.
The effect of interfacial detachment on the polarization of the cell, and hence on the
capacity retention, is better illustrated in Figure 7, which shows the delivered discharge
capacities, normalized by the theoretical capacity, during the cycling. The gradual decrease
in discharge capacity during cycling is due to increasing impedance caused by increasing
detachment. However, the more significant decrease in capacity from the first to the second
discharges is mainly due to the higher initial cell voltage and more complete lithium
insertion in the first discharge.
183
Energies 2023, 16, 7391
Cell potential OCV

4.5
(a)
4
Voltage
3.5
3
1
(b)
SoL
0.5
0
10
(c)
¯V / %
0
Interfacial gap / nm
22
(d)
11
0
0 1 2 3 4 5 6 7
Time / h
Figure 6. Evolution of (a) cell voltage and open circuit voltage (OCV), (b) averaged lithiation state
(SoL), (c) averaged volumetric strain ( ¯ V ) and, (d) growing averaged interfacial gap along AM
particles and CBD, as cycling at 1 C-rate.
1
Normalized discharge capacity
0.85
0.7
1 2 3 4 5
Cycle number (ncyc )
Figure 7. Evolution of normalized deliverable discharge capacity over cycling.
Electrochemical Impedance Spectroscopy (EIS)

As noted earlier, the development of such spatially resolved gaps along the contacts
hinders the electron transfer paths and impedes the reaction kinetics. The resulting elevated
overpotential caused by the interfacial debonding after only five cycles and its subsequent
capacity retention, depicted in Figure 7, indicates that the delivered discharge capacity
drops smoothly over the successive cycles. On the other hand, since electrochemical
interactions are biased by various degradation mechanisms, electrochemical impedance
spectroscopy is often used to characterize the deterioration of the battery’s performance.
Therefore, to quantify the effect of interfacial decohesion of AM particles and CBD on
184
Energies 2023, 16, 7391
the performance of the half-cell, the impedance spectra are also numerically computed in
different cycles. The EIS simulation is conducted around SoL = 50% as the equilibrium
state. Then, a harmonically oscillating voltage signal with varying frequency and small
magnitude is introduced to perturb the half-cell. While the input signal magnitude is set to
10 mV, its frequency is varied from 1 mHz to 100 kHz.
Figure 8 illustrates the computed impedance spectra for different studied cases. The
calculated impedance spectrum is composed of two semicircles in the high- and mid-
frequency ranges, followed by a sloping tail in the low frequencies. To quantify the
effects of interfacial detachment on the dynamic response, the EIS simulation was initially
carried out under the two conditions “with interfacial gap” and “without interfacial gap”
(Figure 8a). As can be observed, when interfacial detachment is accounted for, the spectrum
differs from the case with well-bonded contact. This variation primarily includes a shift
of the spectrum to the right and also a larger dimension of the high-frequency semicircle,
both due to the existing debonded area and hindered electronic transport. Moreover, the
mid-frequency semicircle enlarges slightly, which is explained by the impeded reaction
kinetics caused by the introduced local drop in the solid-phase potential at the particle
surfaces detaching from the CBD. On the other hand, Figure 8b exhibits a similar but
significantly more gradual change in the impedance spectra during cycling, which is due
to the mechanical deterioration of the interfacial cohesion. Similar to the observed slight
capacity decrease in Figure 7, the impedance spectrum slightly alters while cycling. It is
certain that the strength of the interface would continue to decrease by further cycling after
the fifth cycle, resulting in greater detachment at the interface and thus higher impedance.
−Imag(Z) / Ω.cm2
−Imag(Z) / Ω.cm2
Real(Z) / Ω·cm2
Figure 8. Nyquist plots of the impedance spectra of the half-cell calculated (a) with and without
interfacial gaps in the pristine state and (b) upon cycling with growing area undergoing interfa-
cial detachment.
185
Energies 2023, 16, 7391
4. Conclusions
Mechanical degradation is an obstacle on the way to high-energy-density electrodes.
Different types of mechanisms lead to mechanical degradation of the LIB cell by contribut-
ing to the capacity fade and impedance increase. The numerical methods provide a low-cost
and fast alternative to investigate mechanical degradation as compared to experimental
tomography techniques. Such methods play an even more important role in the case of
cyclic mechanical fatigue, which requires more resources for experimental testing.
In this context, a 3D half-cell’s representative volume element (RVE) was built using
the virtual morphology of the NMC622 cathode electrode to allow for exploration of the
degrading contact of the AM particles and CBD in the electrodes. The half-cell was cycled
and its characteristics during cycling were visualized. The degrading interfacial strength
over five cycles was quantified in terms of the evolution of the interfacial gap in the time
domain and the increased impedance in the frequency domain by means of numerical
electrochemical impedance spectroscopy. In agreement with previous experimental tomog-
raphy images, it was observed that most of the detachments occur in the regions close to the
separator. Moreover, it was found that by cycling and degradation of contact between the
AM particles and CBD, the impedance spectrum shifts to the right and the high-frequency
semicircle enlarges. Furthermore, the mid-frequency semicircle’s dimension expands as
well. The former changes in the spectrum are due to the more retarded electronic transport
paths, while the latter reflects the introduced drop in the solid-phase potential, hindering
the reaction kinetics.
The framework presented in this work can be further employed to investigate different
interfacial characteristics in LIBs. For instance, it enables determining the minimum
required CBD cohesion strength, the optimized CBD volumetric content, and the optimized
CBD morphology. Additionally, the framework can be used to study the critical role
of delamination of the solid electrolyte (SE) in solid-state batteries. Such investigations
provide a conceptual design in the early stages of a research project that allows optimization
concerning desired functionality. In other words, virtual material testing and experimenting
with various properties of the electrode constituents is possible in short time and with
little cost.
Author Contributions: Conceptualization, M.M. and B.V.; methodology, M.M. and B.V.; software,
M.M. and B.V.; validation, M.M. and J.K.; formal analysis, M.M. and B.V.; investigation, M.M.; re-
sources, S.P.; data curation, M.M.; writing—original draft preparation, M.M. and B.V.;
writing—review and editing, J.K.; visualization, M.M. and B.V.; supervision, J.K.; project administra-
tion, S.P. All authors have read and agreed to the published version of the manuscript.
Abbreviations
The following abbreviations are used in this manuscript:
Latin Letters Unit

Av Specific active surface area [m−1 ]
C Elastic stiffness tensor [Pa]
CDL Specific double layer capacitance [F · m−2 ]
ce Electrolyte concentration [mol · m−3 ]
cs Solid concentration [mol · m−3 ]
Debulk , Deeff Bulk and effective electrolyte diffusivity [m2 · s−1 ]
Ds Solid diffusivity [m2 · s−1 ]
F Faraday constant [C · mol−1 ]
I Identity tensor
Iapp Applied current [A]
186
Energies 2023, 16, 7391
j0 Exchange current density [A · m−2 ]

jf Volumetric faradaic current density [A · m−3 ]
jc Volumetric capacitive current density [A · m−3 ]
k Reaction rate constant [m · s−1 ]
Kint,0 Spring constant per unit area [N · m−3 ]
Kd Diminishing rate of spring constant [N · m−3 ]
Lelectrode Electrode length [m]
p Bruggeman exponent
R Gas constant [J · mol−1 · K−1 ]
rp Average particle radius [m]
T Temperature [K]
U Open circuit voltage [V]
u Deformation field [m]
Greek Letters
α a , αc Symmetry coefficient
ε Volumetric fraction
e Elastic strain tensor tensor
Li Lithium-induced strain tensor
Ω Partial molar volume [m3 · mol−1 ]
φe Electrolyte electrochemical potential [V]
φs Solid electrical potential [V]
σ Cauchy stress tensor [Pa]
σh Hydrostatic stress [Pa]
κebulk , κeeff Bulk and effective electrolyte conductivity [S · m−1 ]
κ eff
D Effective electrolyte diffusional conductivity [A · m2 · mol−1 ]
σsbulk , σseff Bulk and effective solid conductivity [S · m−1 ]
η Overpotential [V]
Abbreviations
AM Active material
BC Boundary condition
BV Butler–Volmer
CBD Carbon-binder domain
CC Current collector
DL Double layer
EIS Electrochemical impedance spectroscopy
EV Electric vehicle
LIB Lithium-ion battery
RVE Representative volume element
SEM Scanning electron microscope
SoL State of lithiation
XCT X-ray computed tomography
Subscripts
s, e Solid, electrolyte
References
1. Ansah, S.; Hyun, H.; Shin, N.; Lee, J.S.; Lim, J.; Cho, H.H. A modeling approach to study the performance of Ni-rich layered
oxide cathode for lithium-ion battery. Comput. Mater. Sci. 2021, 196, 110559. [CrossRef]
2. Mirsalehian, M.; Vossoughi, B.; Kaiser, J.; Pischinger, S. 3D Heterogeneous Model for Electrodes in Lithium-Ion Batteries and Its
Application to a Modified Continuum Model. Batteries 2023, 9, 298. [CrossRef]
3. Doyle, M.; Fuller, T.F.; Newman, J. Modeling of Galvanostatic Charge and Discharge of the Lithium/Polymer/Insertion Cell.
J. Electrochem. Soc. 1993, 140, 1526–1533. [CrossRef]
4. Fuller, T.F.; Doyle, M.; Newman, J. Simulation and Optimization of the Dual Lithium Ion Insertion Cell. J. Electrochem. Soc. 1994,
141, 1–10. [CrossRef]
5. Fang, R.; Ge, H.; Wang, Z.; Li, Z.; Zhang, J. A two-dimensional heterogeneous model of lithium-ion battery and application on
designing electrode with non-uniform porosity. J. Electrochem. Soc. 2020, 167, 130513. [CrossRef]
6. Tjaden, B.; Cooper, S.J.; Brett, D.J.; Kramer, D.; Shearing, P.R. On the origin and application of the Bruggeman correlation for
analysing transport phenomena in electrochemical systems. Curr. Opin. Chem. Eng. 2016, 12, 44–51. [CrossRef]
187
Energies 2023, 16, 7391
7. Garcıa, R.E.; Chiang, Y.M.; Carter, W.C.; Limthongkul, P.; Bishop, C.M. Microstructural modeling and design of rechargeable
lithium-ion batteries. J. Electrochem. Soc. 2004, 152, A255. [CrossRef]
8. Mistry, A.N.; Smith, K.; Mukherjee, P.P. Secondary-phase stochastics in lithium-ion battery electrodes. ACS Appl. Mater. Interfaces
2018, 10, 6317–6326. [CrossRef]
9. Tu, V. Modeling and Finite Element Simulation of the Bifunctional Performance of a Microporous Structural Battery Electrolyte.
Master’s Thesis, Chalmers University of Technology, Gothenburg, Sweden, 2019.
10. Feinauer, J.; Westhoff, D.; Kuchler, K.; Schmidt, V. 3D Microstructure Modeling and Simulation of Materials in Lithium-ion
Battery Cells. In Proceedings of the Simulation Science: First International Workshop, SimScience 2017, Göttingen, Germany,
27–28 April 2017; pp. 128–144. [CrossRef]
11. Nguyen, T.T.; Demortière, A.; Fleutot, B.; Delobel, B.; Delacourt, C.; Cooper, S.J. The electrode tortuosity factor: Why the
conventional tortuosity factor is not well suited for quantifying transport in porous Li-ion battery electrodes and what to use
instead. NPJ Comput. Mater. 2020, 6, 1–12. [CrossRef]
12. Danner, T.; Singh, M.; Hein, S.; Kaiser, J.; Hahn, H.; Latz, A. Thick electrodes for Li-ion batteries: A model based analysis. J. Power
Source 2016, 334, 191–201. [CrossRef]
13. He, X.; Bresser, D.; Passerini, S.; Baakes, F.; Krewer, U.; Lopez, J.; Mallia, C.T.; Shao-Horn, Y.; Cekic-Laskovic, I.; Wiemers-Meyer,
S.; et al. The passivity of lithium electrodes in liquid electrolytes for secondary batteries. Nat. Rev. Mater. 2021, 6, 1036–1052.
[CrossRef]
14. Lu, X.; Bertei, A.; Finegan, D.P.; Tan, C.; Daemi, S.R.; Weaving, J.S.; O’Regan, K.B.; Heenan, T.M.M.; Hinds, G.; Kendrick, E.; et al.
3D microstructure design of lithium-ion battery electrodes assisted by X-ray nano-computed tomography and modelling. Nat.
Commun. 2020, 11, 2079. [CrossRef] [PubMed]
15. An, F.; Zhou, W.; Li, P. A comparison of model prediction from P2D and particle packing with experiment. Electrochim. Acta 2021,
370, 137775. [CrossRef]
16. Chouchane, M.; Rucci, A.; Lombardo, T.; Ngandjong, A.C.; Franco, A.A. Lithium ion battery electrodes predicted from
manufacturing simulations: Assessing the impact of the carbon-binder spatial location on the electrochemical performance.
J. Power Source 2019, 444, 227285. [CrossRef]
17. Zhao, Y.; Stein, P.; Bai, Y.; Al-Siraj, M.; Yang, Y.; Xu, B.X. A review on modeling of electro-chemo-mechanics in lithium-ion
batteries. J. Power Source 2019, 413, 259–283. [CrossRef]
18. Iqbal, N.; Choi, J.; Lee, C.; Khan, A.; Tanveer, M.; Lee, S. A Review on Modeling of Chemo-mechanical Behavior of Particle–Binder
Systems in Lithium-Ion Batteries. Multiscale Sci. Eng. 2022, 4, 79–93. [CrossRef]
19. Renganathan, S.; Sikha, G.; Santhanagopalan, S.; White, R.E. Theoretical analysis of stresses in a lithium ion cell. J. Electrochem.
Soc. 2009, 157, A155. [CrossRef]
20. Cheng, Y.T.; Verbrugge, M.W. Evolution of stress within a spherical insertion electrode particle under potentiostatic and
galvanostatic operation. J. Power Source 2009, 190, 453–460. [CrossRef]
21. Gao, X.; Lu, W.; Xu, J. Modeling framework for multiphysics-multiscale behavior of Si–C composite anode. J. Power Source 2020,
449, 227501. [CrossRef]
22. Liu, B.; Jia, Y.; Li, J.; Jiang, H.; Yin, S.; Xu, J. Multiphysics coupled computational model for commercialized Si/graphite composite
anode. J. Power Source 2020, 450, 227667. [CrossRef]
23. Wang, M.; Xiao, X.; Huang, X. A multiphysics microstructure-resolved model for silicon anode lithium-ion batteries. J. Power
Source 2017, 348, 66–79. [CrossRef]
24. Wu, W.; Xiao, X.; Wang, M.; Huang, X. A microstructural resolved model for the stress analysis of lithium-ion batteries.
J. Electrochem. Soc. 2014, 161, A803. . jes. [CrossRef]
25. Ahmadi, M. A hybrid phase field model for fracture induced by lithium diffusion in electrode particles of Li-ion batteries. Comput.
Mater. Sci. 2020, 184, 109879. [CrossRef]
26. Lee, S.; Yang, J.; Lu, W. Debonding at the interface between active particles and PVDF binder in Li-ion batteries. Extrem. Mech.
Lett. 2016, 6, 37–44. [CrossRef]
27. Grantab, R.; Shenoy, V.B. Pressure-gradient dependent diffusion and crack propagation in lithiated silicon nanowires. J. Elec-
trochem. Soc. 2012, 159, A584. [CrossRef]
28. Xu, R.; Yang, Y.; Yin, F.; Liu, P.; Cloetens, P.; Liu, Y.; Lin, F.; Zhao, K. Heterogeneous damage in Li-ion batteries: Experimental
analysis and theoretical modeling. J. Mech. Phys. Solids 2019, 129, 160–183. [CrossRef]
29. Yang, Y.; Xu, R.; Zhang, K.; Lee, S.J.; Mu, L.; Liu, P.; Waters, C.K.; Spence, S.; Xu, Z.; Wei, C.; et al. Quantification of heterogeneous
degradation in Li-ion batteries. Adv. Energy Mater. 2019, 9, 1900674. [CrossRef]
30. Baboo, J.P.; Yatoo, M.A.; Dent, M.; Hojaji Najafabadi, E.; Lekakou, C.; Slade, R.; Hinder, S.J.; Watts, J.F. Exploring different binders
for a LiFePO4 battery, battery testing, modeling and simulations. Energies 2022, 15, 2332. [CrossRef]
31. Koerver, R.; Aygün, I.; Leichtweiß, T.; Dietrich, C.; Zhang, W.; Binder, J.O.; Hartmann, P.; Zeier, W.G.; Janek, J. Capacity fade in
solid-state batteries: Interphase formation and chemomechanical processes in nickel-rich layered oxide cathodes and lithium
thiophosphate solid electrolytes. Chem. Mater. 2017, 29, 5574–5582. [CrossRef]
32. Liu, P.; Xu, R.; Liu, Y.; Lin, F.; Zhao, K. Computational modeling of heterogeneity of stress, charge, and cyclic damage in composite
electrodes of Li-ion batteries. J. Electrochem. Soc. 2020, 167, 40527. [CrossRef]
188
Energies 2023, 16, 7391
33. Daemi, S.R.; Tan, C.; Volkenandt, T.; Cooper, S.J.; Palacios-Padros, A.; Cookson, J.; Brett, D.J.; Shearing, P.R. Visualizing the carbon
binder phase of battery electrodes in three dimensions. ACS Appl. Energy Mater. 2018, 1, 3702–3710. [CrossRef]
34. Boyce, A.M.; Martínez-Pañeda, E.; Wade, A.; Zhang, Y.S.; Bailey, J.J.; Heenan, T.M.; Brett, D.J.; Shearing, P.R. Cracking predictions
of lithium-ion battery electrodes by X-ray computed tomography and modelling. J. Power Source 2022, 526, 231119. [CrossRef]
35. Müller, S.; Pietsch, P.; Brandt, B.E.; Baade, P.; De Andrade, V.; De Carlo, F.; Wood, V. Quantification and modeling of mechanical
degradation in lithium-ion batteries based on nanoscale imaging. Nat. Commun. 2018, 9, 2340. [CrossRef] [PubMed]
36. Zhu, J.; Zeng, K.; Lu, L. Cycling effects on surface morphology, nanomechanical and interfacial reliability of LiMn2 O4 cathode in
thin film lithium ion batteries. Electrochim. Acta 2012, 68, 52–59. [CrossRef]
37. Valøen, L.O.; Reimers, J.N. Transport properties of LiPF6-based Li-ion battery electrolytes. J. Electrochem. Soc. 2005, 152, A882.
[CrossRef]
38. Xiao, X.; Wu, W.; Huang, X. A multi-scale approach for the stress analysis of polymeric separators in a lithium-ion battery.
J. Power Source 2010, 195, 7649–7660. [CrossRef]
39. Appiah, W.A.; Park, J.; Byun, S.; Cho, I.; Mozer, A.; Ryou, M.H.; Lee, Y.M. A coupled chemo-mechanical model to study the
effects of adhesive strength on the electrochemical performance of silicon electrodes for advanced lithium ion batteries. J. Power
Source 2018, 407, 153–161. [CrossRef]
40. Li, P.; Zhao, Y.; Shen, Y.; Bo, S.H. Fracture behavior in battery materials. J. Phys. Energy 2020, 2, 022002. [CrossRef]
41. Iqbal, N.; Lee, S. Mechanical failure analysis of graphite anode particles with PVDF binders in Li-ion batteries. J. Electrochem. Soc.
2018, 165, A1961. [CrossRef]
42. Parks, H.C.; Boyce, A.M.; Wade, A.; Heenan, T.M.; Tan, C.; Martínez-Pañeda, E.; Shearing, P.R.; Brett, D.J.; Jervis, R. Direct 4D
Observations of Electrochemically Induced Intergranular Cracking in NMC811 Particles. ChemRxiv 2023. [CrossRef]
189
energies
Article
Load Capacity of Nickel–Metal Hydride Battery and
Proton-Exchange-Membrane Fuel Cells in the
Fuel-Cell-Hybrid-Electric-Vehicle Powertrain
Ireneusz Pielecha *, Filip Szwajca and Kinga Skobiej
Institute of Combustion Engines and Powertrains, Faculty of Civil and Transport Engineering,
Poznan University of Technology, 60-965 Poznan, Poland; filip.szwajca@put.poznan.pl (F.S.);
kinga.skobiej@put.poznan.pl (K.S.)
* Correspondence: ireneusz.pielecha@put.poznan.pl
Abstract: This article investigates the impact of loading on the hybrid powertrain of the FCAT-
30 model, equipped with a proton-exchange-membrane fuel cell (PEMFC) and a nickel–metal hydride
(NiMH) battery. This study involves analyzing structural component performance based on voltage
and current measurements of the fuel cell, battery, and powertrain. Tests conducted under different
load conditions reveal significant differences in battery current and fuel-cell voltage, highlighting the
crucial role of the battery in the powertrain. External loading induces cyclic operation of the fuel cell,
generating peak power. The energy balance analysis demonstrates that, under no-load conditions,
the vehicle consumes 37.3% of its energy from the fuel cell, with a total energy consumption of 3597 J.
Under load, the energy from the battery is significantly utilized, resulting in a constant fuel-cell
share of approximately 19%, regardless of the vehicle’s load. This study concludes that the battery
predominantly drives the powertrain, with the fuel cell acting as a secondary energy source. These
findings provide valuable insights into the power distribution and energy balance in the hybrid
powertrain. Using a load driving profile reduced the fuel-cell-stack energy contribution by 6.85%
relative to driving without an external load.
Keywords: battery; fuel-cell vehicle; load capacity; hybrid powertrain; electric propulsion system;
Citation: Pielecha, I.; Szwajca, F.; energy balance assessment; external load; chassis dynamometer
Skobiej, K. Load Capacity of
Nickel–Metal Hydride Battery and
Proton-Exchange-Membrane Fuel
Cells in the Fuel-Cell-Hybrid-Electric-
1. Introduction
Vehicle Powertrain. Energies 2023, 16,
7657. https://doi.org/10.3390/
With constant technological advances and growing environmental awareness, the
en16227657
transportation sector is constantly looking for innovative solutions that meet the growing
needs of mobility and minimize the negative effects on the environment. The share of
vehicles using electric propulsion is steadily increasing in the global market [1–3]. Fuel-cell
Received: 19 October 2023 hybrid electric vehicles (FCHEVs) are a promising alternative, combining the advantages
Revised: 15 November 2023 of electric vehicles with a proton-exchange-membrane (PEM) fuel cell [2].
Accepted: 16 November 2023 Reducing dependence on fossil fuels, such as oil and natural gas, has gained world-
Published: 19 November 2023 wide acceptance. As a result, the development and intensive use of renewable energy has
become an inherent trend. However, most renewable energy sources are characterized by
intermittent access, which requires the development of efficient energy storage and gener-
ation systems [4,5]. Today, the use of renewable energy to produce hydrogen is gaining
popularity and recognition, further supporting the development of fuel-cell technology.
Hydrogen production is expected to triple by 2050, driven by its falling cost [6]. In
2021, a total of 51,437 FCHEVs were registered worldwide [7]. There were 729 hydrogen
refueling stations in operation at the time [6]. The results of the analysis by Samsun et al. [8]
clearly indicate a very favorable trend in the development of fuel-cell vehicles and hydrogen
refueling stations in 2021. The authors in their article [9] point out that for any vehicle
4.0/).

190
Energies 2023, 16, 7657
with a range greater than 160 km (100 miles), fuel cells are superior to batteries in terms of
weight, energy efficiency, and life-cycle costs.
A comparative analysis of internal-combustion-engine vehicles (ICEVs) and
FCHEVs [10] shows that emissions, maintenance, operating costs, and efficiency are much
more favorable for FCHEVs.
Fuel cells, unlike many traditional energy sources, do not emit pollutants during
the power generation process. The electrochemical process in fuel cells is based on a
chemical reaction between hydrogen and oxygen, generating clean electricity, with the only
byproduct being water. This makes fuel cells one of the greenest solutions for electricity
generation [11–14]. The battery in the vehicle performs the function of managing the
dynamic response of the vehicle under varying load conditions [10].
The fuel cells used in the first prototype vehicles (in 2002) achieved a volumetric
power factor of 1.0 kW/dm3 with a mass power factor of 0.75 kW/kg [15]. In the FCHV
model (in 2008), these ratios were 1.45 kW/dm3 and 0.9 kW/kg, respectively. The first-
generation Toyota Mirai, equipped with nickel–metal hydride (NiMH) batteries, had values
of 3.1 kW/dm3 and 2 kW/kg, while the new generation of the Mirai (Li-ion battery) vehicle
achieves 5.4 kW/dm3 (4.4 kW/kg excluding end plates) and 5.4 kW/kg, respectively [16,17].
Honda used 103 kW fuel cells in the Clarity model, for which the volumetric and mass
power factors were 3.1 kW/dm3 and 2.0 kW/kg, respectively [18,19]. The system uses a
Li-ion battery with a capacity of 25.5 kWh [20]. The Hyundai Nexo uses a 95 kW cell and
a 40 kWh battery [21,22]. The BMW iX5 Hydrogen is equipped with a 125 kW cell and a
battery with a very small capacity of 2 kWh [23,24].
The development of Li-ion batteries can be seen in terms of changes in specific energy:
from 90 Wh/kg in 1990 to more than 250 Wh/kg today [25–27]. Their cost has now been
reduced from $1000 to about $250 per kilowatt-hour (kW/h) [25,28].
A very important aspect is the study of PEM fuel cells during different loads from the
perspective of optimizing their performance under different operating conditions, which
directly affects the efficiency of the entire FCHEV system [29].
Optimal operating conditions for a fuel cell are crucial for its efficient operation. In
this context, maintaining an optimal temperature plays an important role. Fuel cells,
especially those based on PEM and solid-oxide-fuel-cell (SOFC) technology, show their best
performance in a tightly controlled temperature range. For PEM fuel cells, maintaining
a temperature of 60–80 ◦ C is crucial, while SOFCs require higher temperatures, typically
above 500 ◦ C [30–33].
In addition, it is important to maintain appropriate humidity levels. Fuel cells, es-
pecially PEMs, are sensitive to ambient humidity, and ensuring proper humidity levels
helps maintain proper proton conductivity, which affects the efficiency of electrochemical
processes inside the cell [34,35]. The moisture content of the input gases supplied to the
anode and cathode has been shown to have a significant impact on fuel-cell performance.
Yan et al. [36] showed that lowering the humidity at the cathode negatively affects the
steady-state and dynamic performance of the fuel cell. In the context of humidity control,
in addition to the effect of humidity on the proton conductivity of membranes, it is worth
noting that large-scale commercialization of proton-exchange fuel cells requires achieving
higher power and current densities. Nonetheless, at high operating current densities,
liquid water accumulation can lead to flooding problems and impede gas diffusion, which
accelerates cell performance degradation [37,38]. Therefore, it is important to improve
water management capabilities to achieve better cell efficiency.
Controlling the cleanliness of the fuel supplied to the cell also has an important role,
as impurities, such as sulfur and particulate matter, can negatively affect the performance
of the electrodes, resulting in reduced cell performance [39]. Ensuring the quality and
purification of the fuel is key to maintaining optimal operating conditions [40]. Nonetheless,
cells show great flexibility due to their ability to use a variety of fuels, such as hydrogen,
methanol, natural gas, and biogas [41,42]. Despite the significant manufacturing cost of
191
Energies 2023, 16, 7657
fuel cells, the benefits of this advanced solution are significant enough to make investment
in them economically viable [43,44].
This research aims to determine how the external load on a chassis dynamometer af-
fects the control strategy of a hybrid powertrain equipped with a fuel cell (low-temperature)
and a nickel–metal hydride battery. By analyzing similar values of the energy stored in the
battery and fuel cell, changes in the shares of energy coming from the two drive sources
are identified. Due to the availability of the test object, this article examines a vehicle using
nickel-based batteries. The presented research evaluates the impact of external loading
on the drivetrain of the FCAT-30 model, a hydrogen hybrid vehicle equipped with a PEM
fuel cell and NiMH battery. This study aims to understand the performance of structural
components under different load conditions and extends the analysis to include the energy
balance. The vehicle was subjected to tests with and without external loading on a chassis
dynamometer, allowing for the measurement of voltage and current of the battery, fuel cell,
and drivetrain.
2. Research Methodology
2.1. Study Object
This research was conducted using a model vehicle equipped with a hybrid drive
system: a PEM-type fuel cell along with a NiMH battery (Figure 1). The remotely controlled
4 × 4 vehicle uses a low-temperature PEM-fuel-cell stack (Table 1) with two hydrogen
storage tanks of 10 dm3 each, with a maximum hydrogen pressure of 30 bar. Mounted
measurement systems allow real-time recording of typical vehicle movement parameters
along with data acquisition of battery and fuel-cell voltage and current.
Figure 1. FCAT30 hydrogen model of hybrid powertrain with NiMH battery (photo owner).
Table 1. Vehicle model technical parameters.
Parameter Unit Value

Fuel cell
Fuel-cell type – PEM
Number of cells – 14
Power W 30
Hydrogen pressure MPa 0.045–0.055
Cell-stack mass g 280
H2 flow at maximum Ne dm3 /min 0.42
System efficiency % 40 (at max power)
Battery
Type – NiMH
Max output voltage V 7.2
Electric capacity mAh 4200
192
Energies 2023, 16, 7657
Table 1. Cont.
Parameter Unit Value

Hydrogen storage
Tank capacity Ndm3 2 × 10
Purity % ≥99.995
Form of storage – AB5—metal hydrides
Tank pressure MPa 3.0
Tank dimensions mm × mm φ22 × 88
A static analysis of the energy flow (Figure 2) shows comparable values for the energy
stored in the battery and hydrogen tanks, with a difference of about 20%, in favor of the
NiMH battery. The battery used directly interacts with the energy management system,
unlike the hydrogen energy, which is converted to electricity using a PEM fuel cell.
Figure 2. Diagram of energy flow in a hybrid drive system.
The detailed scheme of the system is shown in Figure 3. The vehicle was propelled by
the Mabuchi RS-540SH-7520 electric motor, operating within a voltage range of 4.8–7.2 V.
The motor attains its peak efficiency of 67% under the following operational parameters:
P = 63.2 W, n = 19740 rpm, I = 13 A, Mo = 30.6 mNm. The metering system incorporates a
fuel cell, battery, and energy flow controller from CREA Technologie. Owing to proprietary
“know-how”, the complete system control details remain undisclosed. Arduino circuits
were employed to manage the energy flow. The hybrid system is dependent on the chemical-
energy battery, and its functionality is possible without the fuel-cell system. As depicted in
Figure 3, there is the capability to recharge the electrochemical battery. Measurement data
can be directly showcased in LabView or, alternatively, stored on an SD card for subsequent
offline analysis on a computer.
This implies the possibility of a much higher load on the battery compared with the
fuel cell. It follows that the share of the battery in variable load situations should be greater
than that of the fuel cell. The energy flow diagram indicates a parallel drive system, which
directly affects the voltage and current values in the results.
193
Energies 2023, 16, 7657
Figure 3. A view of the components of the hybrid propulsion system, including the data acquisition
and visualization system.
2.2. Chassis Dynamometer

Tests on the operation of the battery and fuel cell under varying load conditions
were carried out on a chassis dynamometer (Figure 4). The dynamometer model used in
the experiment has two axles, each equipped with two rollers. Load generation for the
drivetrain is carried out through a friction brake, which acts directly on one of the rollers
of the vehicle’s rear axle. To ensure an even load on all wheels, the axles are connected
by a toothed belt. The load generated is transmitted as the force with which the friction
lining is pressed against the roller in direct contact with the vehicle’s tire. In the context
of the analysis, the load on the system is expressed in the basic SI unit of force (N). The
movement of the friction lining is controlled by a servo motor, and the force is measured by
an additional sensor.
Figure 4. View of the chassis dynamometer with its most important components.
Electrical parameters and some mechanical parameters are recorded using the mea-
surement system. The measurement frequency was about 50 Hz during the acquisition of
194
Energies 2023, 16, 7657
all 10 measurement data (Figure 5). The program allows recording electrical parameters
while visualizing the values of excess voltage, battery current, and driving speed (set using
the Setpoint function—Figure 5). The software was developed in the LabView environment
using the Arduino platform. Data recording takes place on an SD card mounted in the
vehicle, but the application functions wirelessly, using Bluetooth data transmission.
Figure 5. View of the chassis dynamometer software screen.
3. Scope of the Study

Research on the FCAT-30 vehicle’s powertrain was carried out in two variants: (a) with-
out external loading and (b) with external loading. The load was introduced by using a
chassis dynamometer. In the first variant, the rollers were not additionally braked; the load
resulted solely from the moment of inertia of the system driven by the dynamometer.
Tests on the hybrid model vehicle were carried out according to the profiles shown
in Figure 6. These profiles are shown as a function of percentage of elapsed time due to
the different durations of the test. The no-load speed profile includes several jumps in
speed on the chassis dynamometer, initially with an increasing and then decreasing trend
(see Figure 6a). The profile with external load included only an incremental increase in
driving speed (see Figure 6b) with a simultaneous cyclic increase in load. In both cases,
the driving speed was limited to 6 m/s, and the maximum external load occurring was
6.2 N. The test durations without load and with load were 100 and 140 s, respectively. The
goal was not to directly compare the two tests but only to evaluate the behavior of the
drivetrain (energy flow) under these conditions. For this reason, differences in distance or
test duration were ignored.
195
Energies 2023, 16, 7657
ϴ ϴ
s΀ŵͬƐ΁ ǁŝƚŚŽƵƚ ǁŝƚŚ
ϳ s΀ŵͬƐ΁
&΀E΁ ĨŽƌĐĞ ϳ ĨŽƌĐĞ
&΀E΁
ϲ ϲ
s΀ŵͬƐ΁͕&΀E΁
ϱ ϱ
s΀ŵͬƐ΁͕&΀E΁
ϰ ϰ
ϯ ϯ
Ϯ Ϯ
ϭ ϭ
Ϭ Ϭ
Ϭ ϮϬ ϰϬ ϲϬ ϴϬ ϭϬϬ Ϭ ϮϬ ϰϬ ϲϬ ϴϬ ϭϬϬ
ƚ΀й΁ ƚ΀й΁
(a) (b)
Figure 6. Vehicle speed profiles on the chassis dynamometer: (a) without load and (b) with load.
4. Hybrid Propulsion Research

The speed profiles presented earlier (Figure 6) clearly show the differences between
external load and no load. Eight parameters were measured during the tests, and their
course as a function of test duration is shown in Figure 7. A direct comparison of the results
for the two cases analyzed (with load and without load) reveals significant differences in
the values of battery current (I_BATT) and cell voltage (U_FC).
ϮϬ ϭϬ ϮϬ ϭϲ
s΀ŵͬƐ΁ &;EͿ /ͺ&΀΁ hͺ&΀s΁
s΀ŵͬƐ΁ &;EͿ /ͺ&΀΁ hͺ&΀s΁
/ͺdd΀΁ /ͺKhd΀΁ hͺKhd΀s΁ hͺdd΀s΁ hͺdd΀s΁ /ͺdd΀΁ /ͺKhd΀΁ hͺKhd΀s΁
ϭϱ ϵ
ϭϱ ǁŝƚŚĨŽƌĐĞ ϭϮ
ǁŝƚŚŽƵƚĨŽƌĐĞ
s͕&͕&͕dd͕Khd
s͕&͕&͕dd͕Khd
ϭϬ ϴ
hͺdd
hͺdd
ϭϬ ϴ
ϱ ϳ
ϱ ϰ
Ϭ ϲ
Ͳϱ ϱ Ϭ Ϭ
Ϭ ϮϬ ϰϬ ϲϬ ϴϬ ϭϬϬ Ϭ ϮϬ ϰϬ ϲϬ ϴϬ ϭϬϬ ϭϮϬ ϭϰϬ
ƚ΀Ɛ΁ ƚ΀Ɛ΁
(a) (b)
Figure 7. Test results of measured quantities during hybrid drive tests: (a) without load and
(b) with load.
Tests of the drivetrain with no load show small changes in the battery voltage
(U_BATT) in the range of 7.0 to 7.6 V, with simultaneous changes in the battery current
(I_BATT) in the range of up to 5 A (see Figure 7a). Under load, the voltage change is similar,
but the current change ranges from zero to almost 15 A at the highest brake load (see
Figure 7b). These changes are proportional to the strength of the brake load.
At no brake load, a fuel cell operating at about 8–9 V generates about 1–1.5 A. When
loaded, the voltage values of U_FC are much higher, exceeding 13 V, and the current is
about 2 A at the minimum voltage value of U_FC.
In both cases, the output voltage is close to the voltage U_BATT, and the current I_OUT
is related to the sum of both intensities: the battery and the cell. In the case of additional
load with force F, cyclic operation of the fuel cell is observed, manifested in the absence of
current (i.e., no-load operation). In this case, the resultant current intensity is close to that
of the battery.
196
Energies 2023, 16, 7657
According to the analysis shown in Figure 7, noticeably higher current values are
obtained from the battery than from the fuel cell. This suggests that the fuel cell is not
the main source of energy for vehicle propulsion. Such operating conditions differ from
standard powertrains, such as in the Toyota Mirai [15] and Hyundai Nexo models. Accord-
ing to the research described in [45,46], the fuel-cell energy contribution is significantly
higher, which is a direct result of the high-powered cell compared with the low energy of
the battery.
5. Power Share of Propulsion System Components

An analysis of the power transferred to the drivetrain, broken down into the power
of the battery, the cell, and the output power of the system in no-load and external-load
modes, is shown in Figure 8. The power was calculated as the product of the voltage and
current on each component. Under no-load conditions (Figure 8a), the powertrain obtains
a maximum of 40 watts from the battery and about 12 watts from the fuel cell. Loading the
system (Figure 8b) increases the power of the battery to 90 W and the cell to a maximum of
15 W (continuous power). Finally, without load, the system obtains about 50 W of power,
while during load, it obtains about 90 W.
ϴϬ ϴ ϭϭϬ ϭϭ
&΀E΁ Eͺ&΀t΁ Eͺ΀t΁ EͺKhd΀t΁ s΀ŵͬƐ΁ Eͺ&΀t΁ Eͺ΀t΁ EͺKhd΀t΁ s΀ŵͬƐ΁ &΀E΁
ϳϬ ϳ
&΀E΁͕Eͺ͕Eͺ&͕EͺKhd΀t΁
ǁŝƚŚŽƵƚĨŽƌĐĞ ϵϬ ϵ
ϲϬ ϲ
ǁŝƚŚĨŽƌĐĞ
Eͺ͕Eͺ&͕EͺKhd΀t΁
ϱϬ ϱ ϳϬ ϳ
&΀E΁͕s΀ŵͬƐ΁
s΀ŵͬƐ΁
ϰϬ ϰ
ϱϬ ϱ
ϯϬ ϯ
ϮϬ Ϯ ϯϬ ϯ
ϭϬ ϭ
ϭϬ ϭ
Ϭ Ϭ
ͲϭϬ Ͳϭ ͲϭϬ Ͳϭ
ƚ΀й΁ ƚ΀й΁
(a) (b)
Figure 8. Magnitudes of cell and battery power during speed profile analysis: (a) without load and
(b) with load.
Under no-load conditions, the power is the resultant power of the battery and the cell,
with the battery contributing significantly. During load, the battery’s share also dominates.
The fuel-cell share in both cases remains at about 15 W.
The fuel-cell value of 15 watts represents 50% of its claimed power. Under no-load
conditions, the fuel-cell share is 12:26 W, which is less than 50%. At high vehicle speeds,
the share drops to 12 W out of 50 W, or 24%. So, the fuel-cell share is not particularly high.
With an external load, these values max out at a ratio of 15 W to 90 W, which is 16% (at
maximum load, t = 85%). At low load (t = 5%), the values are 15 W to 28 W, respectively,
which is 53%. The described method of controlling the hybrid powertrain makes it possible
to protect the fuel cell from power peaks that can contribute to the degradation.
The conclusion is that the drivetrain mainly uses battery power at significant vehicle
speed, with the fuel cell acting as a secondary drive source. Nevertheless, the system
controller always uses both energy sources: the battery and the fuel cell.
6. Analysis of Fuel-Cell Operation

In the current chapter, operating conditions were analyzed for the fuel cell only,
ignoring the other components of the drive train (Figure 9). No load (Figure 9a) results in
small changes in the fuel-cell voltage. During the no-load operation of the cell, there is a
noticeable drop in current between the peaks that occur, probably related to the drop in
ionic conductivity. The decrease in conductivity occurs with the accumulation of water
and nitrogen in the anode channels, which are removed when the anode is flushed with
197
Energies 2023, 16, 7657
hydrogen. The flushing phenomenon is accompanied by the aforementioned voltage and

current peaks, which occur cyclically, and loss of hydrogen [47]. The same phenomenon
occurs when the vehicle’s drivetrain is externally loaded (Figure 9b). This process has also
been described by other researchers [47,48].
ϭϬ ϱ ϭϲ ϴ
hͺ&΀s΁ /ͺ&΀΁ hͺ&΀s΁ /ͺ&΀΁
ϭϰ ϳ
ϵ ϰ ϭϮ ϲ
ϭϬ ϱ
hͺ&΀s΁
hͺ&΀s΁
/ͺ&΀΁
/ͺ&΀΁
ϴ ϯ ϴ ϰ
ϲ ǁŝƚŚĨŽƌĐĞ ϯ
ǁŝƚŚŽƵƚĨŽƌĐĞ
ϳ Ϯ ϰ Ϯ
Ϯ ϭ
ϲ ϭ Ϭ Ϭ
ƚ΀й΁ ƚ΀й΁
(a) (b)
Figure 9. Changes in voltage and current of the cell under operating conditions: (a) without load and
(b) with load.
An externally loaded fuel cell exhibits sequential operation, which is manifested by

the existence of a specific operating frequency. This can be compared to a PWM signal,
where most of the time is active time. This means limiting the voltage while maintaining
a high value of the fuel-cell current (Figure 8b). Visible current peaks appeared only at a
low F-force load during the initial phase of the test. The sudden increase in voltage at no
load on the cell (I_FC) is consistent with the polarization characteristics of the PEM fuel
cell. This is related to the losses occurring in the cell during operation at different load
states [49].
The current–voltage characteristics of the fuel-cell stack differ in the case of no load
and with external load (Figure 10). Operation of the cell without a load (Figure 10a) causes
small changes in voltage and current. The points with voltage U_FC below 8.4 V refer to
the purging of anode channels and activation of the fuel cell, while the rest correspond to
typical operating conditions. The operating points of an externally loaded cell are shown in
Figure 10b. In this case, the points above 8.4 V are characteristic of the fuel-cell activation
and termination states. The applied load increases the load on the fuel cell, increasing the
current. The apparent voltage step changes are due to the dynamics of the load and the
adjustment of the fuel-cell operation.
Static loading of the cell would result in typical cell operating conditions and result in
a current–voltage characteristic in which each load point corresponds to a single voltage
value. However, the dynamic nature of the fuel-cell operation means that such static
characteristics are “filled in” by dynamic changes in the operating conditions. The lack of
current (Figure 10b) indicates typical no-load operation under dynamic conditions since
the change in recorded voltage ranges from 9 V to about 13.5 V, which exceeds the rated
voltage of the cell used.
198
Energies 2023, 16, 7657
(a) (b)
Figure 10. Voltage–current characteristics of the fuel cell: (a) without load and (b) with load.
7. Battery Operation Indicators

A partial analysis of battery operation is included in Chapter 4. The voltage–current
characteristics of battery operation are shown in Figure 11. Operation without an external
load on the drivetrain is characterized by slightly higher voltage values and significantly
lower current values (Figure 11a) than with a load (Figure 11b). The no-load characteristic is
only a part of the typical current–voltage characteristic of a battery. As a result of applying
a variable external load, the drivetrain makes significant use of the energy provided by the
battery, resulting in a significant increase in the operating field to a current I_BATT of just
under 15 A. There are battery charging states in both tests, but they represent a negligible
part of the characteristics. It should be noted that the model system tested is not equipped
with braking energy recovery.
(a) (b)
Figure 11. Voltage–current characteristics of the battery: (a) without load and (b) with load.
8. Indicators of Hybrid Powertrain Operation

The operating parameters of the hybrid powertrain are shown in Figure 12. Analysis
of the no-load powertrain (Figure 12a) shows that the battery contributes almost 75%
(maximum). However, its share is always greater than 50%. The battery share is much
higher at high speed than at low speed. During the early and final stages of the drive at no
speed, the battery was recharged, as indicated by the negative value of the battery share.
199
Energies 2023, 16, 7657
ϭϬϬ ϭϮϬ
^ŚĂƌĞŽĨ&͕dd΀й΁͕sΎϭϬ΀ŵͬƐ΁
^ŚĂƌĞŽĨ&͕dd΀й΁͕sΎϭϬ΀ŵͬƐ΁
ϳϱ
ǁŝƚŚĨŽƌĐĞ Eͺ΀й΁ Eͺ&΀й΁ &΀E΁
ϭϬϬ
ϱϬ
ϴϬ
Ϯϱ
Ϭ ϲϬ
ͲϮϱ ǁŝƚŚŽƵƚĨŽƌĐĞ
ϰϬ
ͲϱϬ ĂƚƚĞƌǇĐŚĂƌŐŝŶŐ
ϮϬ
Ͳϳϱ s΀ŵͬƐ΁ Eͺ΀й΁ Eͺ&΀й΁
ͲϭϬϬ Ϭ
ƚ΀й΁ ƚ΀й΁
(a) (b)
Figure 12. Power shares of hybrid drive system components under operating conditions: (a) without
load and (b) with load.
Battery charging conditions were not recorded under vehicle load conditions
(Figure 12b). When the fuel cell is deactivated and the anode channels are purging, the
battery share is about 100%. As the external load increases, the share of the cell decreases,
and the share of battery power increases. As the load and speed increase (in both the initial
and final phases of the test), the fuel-cell share drops below 50%. The higher the load,
the higher the share of the battery. At high external load, the fuel-cell share decreases to
around 30%.
An analysis of the total energy consumed is shown in Figure 13. Under no-load
conditions (Figure 13a), 2254 J of energy was consumed from the battery, while only 1343 J
was consumed from the fuel cell. This shows that the average proportion of energy from
the fuel cell is 37%. Despite the lack of braking of the dynamometer rollers, i.e., no external
load, the vehicle consumed a total of 3597 J.
ϲϬϬϬ ϲϬ ϲϬϬϬ ϲ
^ͺdd΀:΁ ǁŝƚŚŽƵƚĨŽƌĐĞ ^ͺdd΀:΁ ǁŝƚŚĨŽƌĐĞ
ϱϬϬϬ ^ͺ&΀:΁ ϱϬ ϱϬϬϬ ^ͺ&΀:΁ ϱ
^ͺdd͕^ͺ&΀:΁͕s
^ͺdd͕^ͺ&΀:΁͕s
s΀ŵͬƐ΁ s΀ŵͬƐ΁
ϰϬϬϬ ϰϬ ϰϬϬϬ ϰ
ϯϬϬϬ ϯϬ ϯϬϬϬ ϯ
ϮϬϬϬ ϮϬ ϮϬϬϬ Ϯ
ϭϬϬϬ ϭϬ ϭϬϬϬ ϭ
Ϭ Ϭ Ϭ Ϭ
ƚ΀й΁ ƚ΀й΁
(a) (b)
Figure 13. Changes in energy of the battery and fuel cell under operating conditions: (a) without
load and (b) with load.
Under the load conditions of the vehicle’s drivetrain (Figure 13b), the consumed
energies from the battery and the cell are 5277 J and 1251 J, respectively. This means that
the fuel-cell share is practically constant and does not depend on the vehicle’s load. During
the load analysis, the fuel-cell share decreased significantly and is only 19%.
The study concludes that the battery predominantly drives the powertrain, with the
fuel cell acting as a secondary energy source.
200
Energies 2023, 16, 7657
9. Conclusions
1. The experimental study presented here concerns the evaluation of the effect of the
load on the drivetrain of the FCAT-30 model hybrid vehicle, equipped with a PEM
fuel cell, on the performance of selected structural components. The voltage and
current of the battery (BATT), fuel cell (FC), and drivetrain (OUT) were selected as
the directly measured parameters analyzed. Based on the measured parameters, the
performance of the components was evaluated, and the analysis was extended to
include the energy balance.
2. It was pointed out that there is a variation in power distribution with respect to the
applied load. When operating the system without an external load, the cell generates
an approximately constant power of about 12 W during the test, which is between
20% and 50% of the power transferred to the drive, depending on the speed of the
vehicle. The use of an external dynamic load results in cyclic operation of the cell with
a peak power of 15 W, where the fuel-cell-stack contribution ranges from 0% (off state)
to 38%.
3. Regardless of the test conditions, there is a process of flushing the anode channels,
manifested by momentary jumps in cell voltage and current. For dynamic load
conditions of current decay (deactivation of cell operation), clear jumps in cell-stack
voltage from 8.5 V to 13.7 V were recorded. The no-load fuel-cell-stack operation area
was indicated as a voltage below 8.4 V for no-load operation and above 8.4 V for the
dynamic-external-load test. The results of the analysis of the energy flow within the
NiMH battery indicate a small share of charging from the fuel cell with current in the
0–2 A range. The external-load test significantly increases the power demand, which
puts a significant strain on the battery, which is the main energy source.
4. Analysis of the energy balance shows that there are no situations where the drivetrain
uses only the fuel cell, except at the beginning and end of the test, where the wheel
speed is 0 m/s. For the no-load test, the drivetrain consumed 3597 J, of which 37.3%
was energy from the fuel cell. For the external-load test, the vehicle consumed 6528 J,
of which the energy of the cell accounted for 19.2%. It was noted that there was
no significant effect in the way the drivetrain was loaded on the amount of energy
produced by the fuel cell. In the overall balance, the difference between the test with
and without load was 6.85%.
Author Contributions: Conceptualization, I.P., F.S. and K.S.; methodology, I.P., F.S. and K.S.; software,
I.P. and F.S.; validation, I.P., F.S. and K.S.; formal analysis, F.S. and K.S.; investigation, I.P., F.S. and
K.S.; resources, I.P., F.S. and K.S.; data curation, I.P. and F.S.; writing—original draft preparation,
I.P., F.S. and K.S.; writing—review and editing, I.P., F.S. and K.S.; visualization, I.P., F.S. and K.S.;
supervision, I.P., F.S. and K.S.; project administration, I.P., F.S. and K.S.; funding acquisition, I.P., F.S.
and K.S. All authors have read and agreed to the published version of the manuscript.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
References
1. Szumska, E.; Jurecki, R. Technological developments in vehicles with electric drive. Combust. Engines 2023, 194, 38–47. [CrossRef]
2. Luo, Y.; Wu, Y.; Li, B.; Qu, J.; Feng, S.; Chu, P.K. Optimization and cutting-edge design of fuel-cell hybrid electric vehicles. Int. J.
Energy Res. 2021, 45, 18392–18423. [CrossRef]
3. Liu, Z.; Wu, Y.; Feng, J. Competition between battery switching and charging in electric vehicle: Considering anticipated regret.
Environ. Dev. Sustain. 2023, 1–22. [CrossRef]
201
Energies 2023, 16, 7657
4. Sun, C.; Negro, E.; Vezzù, K.; Pagot, G.; Cavinato, G.; Nale, A.; Herve Bang, Y.; Di Noto, V. Hybrid inorganic-organic proton-
conducting membranes based on SPEEK doped with WO3 nanoparticles for application in vanadium redox flow batteries.
Electrochim. Acta 2019, 309, 311–325. [CrossRef]
5. Hussein, M.M.; Mohamed, T.H.; Mahmoud, M.M.; Aljohania, M.; Mosaad, M.I.; Hassan, A.M. Regulation of multi-area power
system load frequency in presence of V2G scheme. PLoS ONE 2023, 18, e0291463. [CrossRef] [PubMed]
6. Matla, J.; Kaźmierczak, A.; Haller, P.; Trocki, M. Hydrogen as a fuel for spark ignition combustion engines–state of knowledge
and concept. Combust. Engines 2023. [CrossRef]
7. Global Hydrogen Review 2021, IEA. Available online: https://www.iea.org/reports/global-hydrogen-review-2021 (accessed on
17 September 2023).
8. Samsun, R.; Rex, M.; Antoni, L.; Stolten, D. Deployment of fuel cell vehicles and hydrogen refueling station infrastructure: A
global overview and perspectives. Energies 2022, 15, 4975. [CrossRef]
9. Thomas, C. Fuel cell and battery electric vehicles compared. Int. J. Hydrogen Energy 2009, 34, 6005–6020. [CrossRef]
10. Li, Z.P.; Liu, B.H.; Arai, K.; Suda, S. A Fuel cell development for using borohydrides as the fuel. J. Electrochem. Soc. 2003, 150,
A868. [CrossRef]
11. Lewis, A.C. Optimising air quality co-benefits in a hydrogen economy: A case for hydrogen-specific standards for NOx emissions.
Environ. Sci. Atmos. 2021, 1, 201–207. [CrossRef]
12. Dincer, I. Environmental and sustainability aspects of hydrogen and fuel cell systems. Int. J. Energy Res. 2006, 31, 29–55. [CrossRef]
13. Stambouli, A.B. Fuel cells: The expectations for an environmental-friendly and sustainable source of energy. Renew. Sustain.
Energy Rev. 2011, 15, 4507–4520. [CrossRef]
14. Van Mierlo, J.; Maggetto, G.; Lataire, P. Which energy source for road transport in the future? A comparison of battery, hybrid
and fuel cell vehicles. Energy Convers. Manag. 2006, 47, 2748–2760. [CrossRef]
15. Konno, N.; Mizuno, S.; Nakaji, H.; Ishikawa, Y. Development of compact and high-performance fuel cell stack. SAE Int. J. Alt.
Power 2015, 4, 123–129. [CrossRef]
16. Yoshizumi, T.; Kubo, H.; Okumura, M. Development of high-performance FC stack for the new MIRAI. In SAE Technical; SAE
International: Warrendale, PA, USA, 2021. [CrossRef]
17. Nonobe, Y. Development of the fuel cell vehicle Mirai. IEEJ Trans. Electr. Electron. Eng. 2017, 12, 5–9. [CrossRef]
18. Kimura, K.; Kawasaki, T.; Ohmura, T.; Atsumi, Y.; Shimizu, K. Development of New Fuel Cell Vehicle Clarity Fuel Cell. Honda
R&D Technical Review. April 2016. Available online: http://www.uninsim.com/data/ueditor/php/upload/file/20191118/15740
59716840217.pdf (accessed on 17 September 2023).
19. Matsunaga, M.; Fukushima, T.; Ojima, K. Powertrain system of Honda FCX Clarity fuel cell vehicle. World Electr. Veh. J. 2009, 3,
820–829. [CrossRef]
20. Robledo, C.B.; Oldenbroek, V.; Abbruzzese, F.; van Wijk, A.J.M. Integrating a hydrogen fuel cell electric vehicle with vehicle-to-grid
technology, photovoltaic power and a residential building. Appl. Energy 2018, 215, 615–629. [CrossRef]
21. 2023 Hyundai NEXO Fuel Cell|Specs. Available online: https://www.hyundaiusa.com/us/en/vehicles/nexo (accessed on
30 August 2023).
22. Gómez, J.A.; Santos, D.M.F. The status of on-board hydrogen storage in fuel cell electric vehicles. Designs 2023, 7, 97. [CrossRef]
23. Rubio, O. BMW i Hydrogen NEXT está en Fase de Pruebas en Carreteras de Europa. Available online: https://greenracingnews.
com/es/bmw-i-hydrogen-next-esta-en-fase-de-pruebas-en-carreteras-de-europa/ (accessed on 30 August 2023).
24. 2023 BMW IX5 Hydrogen First Drive Review: More than Just a Science Project. Available online: https://www.cnet.com/
roadshow/news/2023-bmw-ix5-hydrogen-first-drive-review/ (accessed on 30 April 2023).
25. Hocking, M.; Kan, J.; Young, P.; Terry, C.; Begleiter, D. Welcome to the Lithium-Ion Age. Deutsche Bank Markets Research. 2016.
Available online: http://www.belmontresources.com/LithiumReport.pdf (accessed on 30 August 2023).
26. Placke, T.; Kloepsch, R.; Dühnen, S.; Winter, M. Lithium ion, lithium metal, and alternative rechargeable battery technologies: The
odyssey for high energy density. J. Solid State Electrochem. 2017, 21, 1939–1964. [CrossRef]
27. Zhang, J.; Wang, C.; Zhang, A. Experimental study on temperature and performance of an open-cathode PEMFC stack under
thermal radiation environment. Appl. Energy 2022, 311, 118646. [CrossRef]
28. Curry, C. Lithium-Ion Battery Costs: Squeezed Margins and New Business Models. Bloomberg New Energy Finance. Avail-
able online: https://about.bnef.com/blog/lithium-ion-battery-costs-squeezed-margins-newbusiness-models/ (accessed on
14 July 2023).
29. Wang, T.; Li, Q.; Wang, X.; Qiu, Y.; Liu, M.; Meng, X.; Li, J.; Chen, W. An optimized energy management strategy for fuel cell
hybrid power system based on maximum efficiency range identification. J. Power Sources 2020, 445, 227333. [CrossRef]
30. Arsalis, A. A comprehensive review of fuel cell-based micro-combined-heat-and-power systems. Renew. Sustain. Energy Rev.
2019, 105, 391–414. [CrossRef]
31. Giddey, S.; Badwal, S.; Kulkarni, A.; Munnings, C. A comprehensive review of direct carbon fuel cell technology. Prog. Energy
Combust. 2012, 38, 360–399. [CrossRef]
32. Olabi, A.; Wilberforce, T.; Sayed, E.T.; Elsaid, K.; Abdelkareem, M.A. Prospects of fuel cell combined heat and power systems.
33. Song, Z.; Pan, Y.; Chen, H.; Zhang, T. Effects of temperature on the performance of fuel cell hybrid electric vehicles: A review.
Appl. Energy 2021, 302, 117572. [CrossRef]
202
Energies 2023, 16, 7657
34. Ogungbemi, E.; Wilberforce, T.; Ijaodola, O.; Thompson, J.; Olabi, A.G. Review of operating condition, design parameters and
material properties for proton exchange membrane fuel cells. Int. J. Energy Res. 2020, 45, 1227–1245. [CrossRef]
35. Ji, M.; Wei, Z. A review of water management in polymer electrolyte membrane fuel cells. Energies 2009, 2, 1057–1106. [CrossRef]
36. Yan, Q.; Toghiani, H.; Causey, H. Steady state and dynamic performance of proton exchange membrane fuel cells (PEMFCs)
under various operating conditions and load changes. J. Power Sources 2006, 161, 492–502. [CrossRef]
37. Yan, S.; Yang, M.; Sun, C.; Xu, S. Liquid water characteristics in the compressed gradient porosity gas diffusion layer of proton
exchange membrane fuel cells using the Lattice Boltzmann Method. Energies 2023, 16, 6010. [CrossRef]
38. Yu, H.; Chen, D.; Ni, X.; Qing, P.; Yan, C.; Wei, W.; Ma, J.; Ji, X.; Chen, Y.; Chen, L. Reversible adsorption with oriented
arrangement of a zwitterionic additive stabilizes electrodes for ultralong-life Zn-ion batteries. Energy Environ. Sci. 2023, 16,
2684–2695. [CrossRef]
39. Cheng, X.; Shi, Z.; Glass, N.; Zhang, L.; Zhang, J.; Song, D.; Liu, Z.S.; Wang, H.; Shen, J. A review of PEM hydrogen fuel cell
contamination: Impacts, mechanisms, and mitigation. J. Power Sources 2007, 165, 739–756. [CrossRef]
40. Rosli, R.; Sulong, A.; Daud, W.; Zulkifley, M.; Husaini, T.; Rosli, M.; Majlan, E.; Haque, M. A review of high-temperature proton
exchange membrane fuel cell (HT-PEMFC) system. Int. J. Hydrogen Energy 2017, 42, 9293–9314. [CrossRef]
41. Cropper, M.A.; Geiger, S.; Jollie, D.M. Fuel cells: A survey of current developments. J. Power Sources 2004, 131, 57–61. [CrossRef]
42. Binti Awang Mat, Z.; Madya Kar, Y.B.; Hasmady Bin Abu Hassan, S.; Azrina Binti Talik, N. Proton exchange membrane (PEM)
and solid oxide (SOFC) fuel cell based vehicles—A review. In Proceedings of the 2017 2nd IEEE International Conference on
Intelligent Transportation Engineering (ICITE), Singapore, 1–3 September 2017. [CrossRef]
43. Bar-On, I.; Kirchain, R.; Roth, R. Technical cost analysis for PEM fuel cells. J. Power Sources 2002, 109, 71–75. [CrossRef]
44. Thompson, S.T.; James, B.D.; Huya-Kouadio, J.M.; Houchins, C.; DeSantis, D.A.; Ahluwalia, R.; Wilson, A.R.; Kleen, G.;
Papageorgopoulos, D. Direct hydrogen fuel cell electric vehicle cost analysis: System and high-volume manufacturing description,
validation, and outlook. J. Power Sources 2018, 399, 304–313. [CrossRef]
45. Pielecha, I.; Szałek, A.; Tchorek, G. Two generations of hydrogen powertrain—An analysis of the operational indicators in real
driving conditions (RDC). Energies 2022, 15, 4734. [CrossRef]
46. Pielecha, I.; Szwajca, F. Cooperation of a PEM fuel cell and a NiMH battery at various states of its charge in a FCHEV drive.
Eksploat. Niezawodn. 2021, 23, 468–475. [CrossRef]
47. Liu, Z.; Chen, J.; Liu, H.; Yan, C.; Hou, Y.; He, Q.; Zhang, J.; Hissel, D. Anode purge management for hydrogen utilization and
stack durability improvement of PEM fuel cell systems. Appl. Energy 2020, 275, 115110. [CrossRef]
48. Eom, T.-H.; Kang, J.-W.; Kim, J.; Shin, M.-H.; Lee, J.-H.; Won, C.-Y. Improved voltage drop compensation method for hybrid fuel
cell battery system. Electronics 2018, 7, 331. [CrossRef]
49. Karami, N.; Outbib, R.; Moubayed, N. Fuel flow control of a PEM fuel cell with MPPT. In Proceedings of the 2012 IEEE
International Symposium on Intelligent Control, Dubrovnik, Croatia, 3–5 October 2012; pp. 289–294. [CrossRef]
203
energies
Article
State of Health Estimation Procedure for Lithium-Ion Batteries
Using Partial Discharge Data and Support Vector Regression
Emil Petkovski, Iacopo Marri, Loredana Cristaldi * and Marco Faifer
emil.petkovski@polimi.it (E.P.); iacopomarri@gmail.com (I.M.); marco.faifer@polimi.it (M.F.)
* Correspondence: loredana.cristaldi@polimi.it
Abstract: Battery aging is a complex phenomenon, and precise state of health (SoH) monitoring
is essential for effective battery management. This paper presents a data-driven method for SoH
estimation based on support vector regression (SVR), utilizing features built from both full and partial
discharge capacity curves, as well as battery temperature data. It provides an in-depth discussion
of the novel features constructed from different voltage intervals. Moreover, three combinations of
features were analyzed, demonstrating how their efficacy changes across different voltage ranges.
Successful results were obtained using the full discharge capacity curves, built from the full interval
of 2 to 3.4 V and achieving a mean R2 value of 0.962 for the test set, thus showcasing the adequacy
of the selected SVR strategy. Finally, the features constructed from the full voltage range were
compared with ones built from 10 small voltage ranges. Similar success was observed, evidenced by
a mean R2 value ranging between 0.939 and 0.973 across different voltage ranges. This indicates the
practical applicability of the developed models in real-world scenarios. The tuning and evaluation of
the proposed models were carried out using a substantial dataset created by Toyota, consisting of
124 lithium iron phosphate batteries.
Keywords: lithium-ion battery; battery degradation; prognostics; machine learning; SoH
Citation: Petkovski, E.; Marri, I.;

1. Introduction
Cristaldi, L.; Faifer, M. State of Health
Estimation Procedure for Lithium-Ion
Lithium-ion batteries have become the preferred choice due to their high energy
Batteries Using Partial Discharge
density, long cycle life, and low self-discharge rate [1]. This makes them especially favorable
Data and Support Vector Regression. for electric vehicles, where minimizing the weight of the battery pack while maintaining
Energies 2024, 17, 206. https:// the desired performance and range is crucial. Furthermore, the increasing presence of
doi.org/10.3390/en17010206 photovoltaic and wind systems in the grid has led to the deployment of massive stationary
battery storage systems to address the intermittency of these renewable sources and to
Academic Editor: Ramon
provide essential support services to the grid [2]. Battery cells are very small units that are
Costa-Castelló
a part of battery packs for the majority of storage applications. The pack necessitates the
Received: 31 October 2023 presence of a Battery Management System (BMS) to ensure safe and efficient operation. It
Revised: 29 November 2023 monitors and controls the charging and discharging process and implements battery cell
Accepted: 28 December 2023 balancing. To accomplish this task, the BMS must accurately estimate key battery metrics
Published: 30 December 2023 such as State of Charge (SoC), State of Health (SoH), and remaining useful life (RUL).
Battery SoH is a measure of the battery’s ability to hold its charge. It is commonly
defined as the ratio of the maximum available capacity of a used battery compared with
its original or brand-new state. On the other hand, battery RUL refers to the estimated
amount of time or number of charge–discharge cycles that a battery can operate for before
experiencing failure or unacceptable performance. Battery degradation is a highly variable
process, dependent on ambient conditions, cell chemistry, use patterns, and the BMS.
conditions of the Creative Commons Consequently, a wide variety of model-based [3–6] and data-driven battery aging methods
Attribution (CC BY) license (https:// used for SoH and End of Life prediction of batteries exist in literature.
creativecommons.org/licenses/by/ Model-based methods rely on physical principles and mathematical equations to
4.0/). describe the behavior of batteries. These methods are characterized by very high accuracy;

204
Energies 2024, 17, 206
however, their development can be very resource-intensive and time-consuming, requiring

both significant computational power and expert knowledge.
With the abundance of data and advancements in the machine learning field, data-
driven methods have become particularly popular. By analyzing large datasets, data-driven
methods can uncover patterns and relationships that may not be easily identifiable using
traditional analytical techniques. Numerous features, also called health indicators (HIs),
have been used to realize different ML strategies. They are very commonly extracted from
voltage, current, and temperature curves during the charging and discharging processes;
for example, the authors of [7] extracted 11 features in the voltage range of 3.8 to 4.1 V.
In [8], eight features were extracted from the discharge voltage, current, temperature,
and the elapsed time of the discharge process, while in [9], features were built from the
same quantities during the constant current (CC) and constant voltage (CV) charging
process. In [10], unique features were constructed from the dependence of battery voltage
and discharge capacity (Q (V), and the integral of battery temperature, through statistical
analysis. Furthermore, incremental capacity analysis (ICA) is another commonly used
method for feature extraction, representing the ratio between an increment of capacity and
a fixed-voltage increment (dQ/dV) for a full or partial cycle. The resulting differential
curves have a peak that becomes less pronounced as the battery ages. Many works have
generated various valuable features from these curves [11–13]. The only drawback of these
curves is that they can be easily affected by noise.
Numerous machine learning methods have been proposed in literature regarding State
of Health (SoH) estimation. These can range from classical machine learning techniques
such as modified linear regression, Gaussian processes regression (GPR), support vector
regression (SVR), and random forest to different neural network (NN)-based approaches.
GPR can provide mean capacity estimates as well as probabilistic bounds using various
features [14]. In [15], the authors use GPR to estimate the SoH and RUL of lithium-ion
batteries while also considering uncertainty in two different attempts: first, by using a
known parametric battery degradation model to exploit prior knowledge, and secondly,
by developing a model based on experimental capacity measurements of battery cells.
Ref. [16] presents an approach where probability prediction together with a GPR model is
used to estimate the SoH, which is then used with three other indirect health indicators to
forecast the RUL of the batteries.
Refs. [17,18] apply regression to model battery aging patterns; additionally, the RUL
prediction capabilities of two fitting functions are analyzed, which are the third-degree
polynomial and a custom hybrid function. An SVR strategy is presented in [19] based
on curves of battery voltage as a function of charging capacity (V-Q). In [20,21], an SVR
strategy based on partial voltage charging curves is proposed, while in [22], a comparison
between linear regression, SVR, and random forest methods is provided.
Finally, many works have focused on different methods based on neural networks
because of their flexibility, adaptability, and ability to learn complex patterns. Refs. [23,24]
employ back-propagation neural networks (BPNN) using various health indicators, result-
ing in accurate SoH estimation. Ref. [25] proposes a solution based on the convolutional
neural network (CNN), incorporating the concepts of transfer learning and network prun-
ing. In [26], an Echo State Network is used together with a single exponential function
to predict the SoH evolution curve of the tested batteries in different cycles. From the
generated curves, the RUL is also inferred. Long Short-Term Memory (LSTM) networks
are a type of neural network architecture that are particularly effective at processing long
sequences of data [27–31]. In [28], the proposed method is based on LSTM NN and signal
processing methods for SoH and RUL prediction of lithium-ion batteries, while [29] created
an SoH estimation method based on LSTM and transfer learning.
This work presents a data-driven method for SoH estimation based on SVR and
three features obtained from full and partial discharge capacity curves as well as battery
temperature data. SVR was selected instead of the recently popular NN-based strategies
because it offers a good trade-off between low computational burden, applicability, and
205
Energies 2024, 17, 206
accuracy of results, while also providing higher model interpretability. The performance
of all models was evaluated using a substantial dataset, consisting of 124 battery cells. A
discussion is provided on the unique extraction procedure of the features, their combination,
as well as the parameter tuning process. Results demonstrate that the SVR models can
successfully use features built from partial discharge capacity curves for accurate SoH
estimation, highlighting the method’s applicability in real-life scenarios.
2. Dataset
Toyota, in cooperation with Stanford University and MIT, created a dataset of
124 commercial lithium–iron phosphate battery cells [32], with 1.1 Ah nominal capac-
ity and 3.3 V nominal voltage, cycled under fast charging conditions until their end of life
(EoL) is reached. The cycling was carried out in a fixed chamber temperature of 30 ◦ C,
applying two steps of constant current charging, according to a policy defined by the format
“C1(Q1) − C2”, where C1 and C2 are the charging current values, and Q1 is the SoC level
at which the current is switched. A total of 72 different policies with different values for
charging current and switching step are used across the dataset, while the discharge is
performed at a constant current of 4.4 A. While the chamber temperature is controlled, the
cell temperature can vary by up to 10 ◦ C within a cycle and between cells due to the vastly
different charging policies and internal impedance values.
The number of cycles to failure for the batteries spans from approximately 150 to
2300, and the average end of life number is 806. Figure 1 shows the batteries ranked by
cycle life, while Figure 2 shows the various SoH evolution profiles. The heterogeneous
charging profiles generate a lot of variety within the dataset, allowing for deep analysis
and insight into the behavior of the batteries and degradation patterns. For every battery
cycle, the dataset includes a range of measured quantities including voltage, charge and
discharge capacity, temperature, and internal resistance. In [14], the dataset authors created
an in-depth analysis of the data and proposed many data-driven features, using which they
developed a linear regression model to predict battery RUL.
Figure 1. Dataset ordered by battery lifespan expressed in number of cycles.
206
Energies 2024, 17, 206
Figure 2. SoH over number of cycles for a set of representative batteries of the dataset.
3. Proposed Approach
This paper proposes a solution to estimate the SoH, which is based on a feature
extraction procedure, where features that are considered strong battery health indicators
are extracted from the discharge phase of every cycle. In the second step, those features are
used as predictors to fit an SVR model, which is well suited to battery degradation modeling
since it is strongly resistant to overfitting and can handle non-linear data. The SVR model
is tuned with a cross-validation procedure and then used to perform SoH estimation.
3.1. Support Vector Regression

Support vector machine (SVM) is a very powerful algorithm for classification and
regression tasks, which has been widely and successfully used in various fields such
as image recognition, text classification, and prognostic and diagnostic processes. SVM
works by finding the hyperplane that separates the data into classes and maximizes the
minimum distance between data points and the hyperplane itself (maximizes the margin).
SVM is effective with highly non-linear data thanks to the kernel trick: it can be used to
transform the input (non-linear) data into higher-dimensional feature spaces, where linear
classification can be performed.
SVR is a version of SVM specifically tailored for regression tasks. Similarly to SVM,
it works by finding the hyperplane (line) that best fits the data and minimizes the error
by using a tunable value ε as the minimum considered error. Any point lying inside
the ε-tube is not considered in the cost function (1). A high value of ε leads to a more
tolerant model, prone to underfitting, while a low value of ε leads to a more accurate but
likely overfitted model. ξ n , ξ n * are the slack variables used to account for the positive and
negative error of the points outside the ε-tube (2). The sum of ξ n and ξ n * is minimized in
the loss function, weighted by a hyperparameter C, called a box constraint, which helps
balance the complexity of the model. Finally, w and w are the weights arrays, normal and
transposed, respectively; Yn is the target value; Xn is the transposed descriptor array; and
b is the bias.
N
1
min w w + C ∑ (ξ n + ξ n∗ ) (1)
2 n =1

Yn − ( Xn w + b) ≤ ε + ξ n ∀n
(2)
( Xn w + b) − Yn ≤ ε + ξ n∗ ∀n
3.2. Feature Selection Procedure

In [10], the authors propose and study a set of features computed over the Toyota-MIT
dataset, which they use as predictors to fit a linear regression model and provide an early
207
Energies 2024, 17, 206
RUL prediction at cycle 100 of the batteries. To do so, they extract one value of each feature
per battery, using the measurements made between cycles 10 and 100. The extraction
approach is revisited in this work and adapted to compute three of the features at every
cycle. The values of the features are then used to feed the SVR model to obtain a continuous
estimation of the SoH at every cycle, using the feature data of that cycle. The first two
features are related to the discharge capacity curves of the batteries as a function of voltage
(Q (V)), i.e., their total moved charge (Q) during each discharge phase.
Figure 3 shows these discharge capacity curves at different cycles of battery life. The
area under each curve can be seen as the total energy output by the battery during that
cycle, and it gradually drops as the battery ages. The discharge capacity curve at cycle 10 is
taken as a reference to compute the values of the features. The reference discharge curve
is then subtracted by the discharge curve of every cycle after it to obtain a measure of the
drop in performance between cycle 10 and the actual cycle (k).
Figure 3. Discharging capacity as a function of voltage for all cycles of a sample battery.
The full voltage range is equal for all the cycles (and batteries) and constitutes a solid
base for comparisons. However, before it could be utilized, interpolation of the different
curves over a common set of voltage values had to be performed so that their subtraction
could be made point by point. In Figure 4, the difference curves are shown for cycle
200 of all batteries, which are the result of the procedure explained above. Blue lines that
are almost flat show no significant drop in performance from cycle 10. Yellow lines, on
the other hand, indicate a major degradation. The first feature (Ftr1) is computed as the
common logarithm of the variance of the difference curves, while the second feature (Ftr2)
is computed as the common logarithm of the minimum of the difference curves. Both the
features are computed for every battery and every cycle except cycles from 1 to 10, for
which the value is considered equal to the 11th. The result, for both features, is a feature
space on which the data are almost linearly distributed over the battery life span, and it
is plotted for cycle 200 of every battery in Figure 5a for Ftr1 and Figure 5b for Ftr2. This
property is maintained over all the aging cycles. While the results for both features seem
identical, drastic differences will come to the surface when partial discharge curves are
discussed in Section 4.2.
208
Energies 2024, 17, 206
Figure 4. Difference in discharge capacity curves between reference cycle 10 and cycle 200, calculated
for all the batteries.
(a) (b)
Figure 5. (a) Common logarithm of the variance in the difference between capacity curves at cycle
200 and cycle 10 for all batteries (Ftr1(200)). (b) Common logarithm of the minimum of the difference
between capacity curves at cycle 200 and cycle 10 for all batteries (Ftr2(200)).
Finally, the third feature (Ftr3) is computed as the sum of temperature: for every cycle
k of each battery, the average temperature of each cycle (T cycle ) is summed from the first up
to the k-th cycle.
In Equation (3), the three features are defined and computed for each cycle of every
battery. More specifically, Q10 (V) and Qk (V) represent the discharge capacities at reference
cycle 10 and cycle k, respectively, as a function of voltage. ΔQk (V) is the difference of
those two curves, while ΔQk (V) represents the average value of that difference for cycle k.
209
Energies 2024, 17, 206
Parameter p refers to the number of points in the curve. Finally, Ti represents a singular
temperature measurement.
⎧ p
⎪ 2
⎪
⎪ Ftr1(k) = log(| p−1 ∑ ΔQk (V) − ΔQk (V) |)
1
⎪
⎪
⎪ i =1
⎪
⎪
⎪
⎪ Ftr2(k ) = log(|min(ΔQk (V ))|)
⎪
⎨ k
Ftr3(k ) = ∑ T cycle (3)
⎪
⎪ cycle=1
⎪
⎪
⎪
⎪ ΔQk (V) = Qk (V) − Q10 (V)
⎪
⎪
⎪
⎪ p
⎪
⎩ ΔQk (V) = 1p ∑ ΔQk,i (V)
i =1

4.1. Experimental Procedure and Full Discharge Window Approach
The voltage operating range of the batteries in the used dataset is approximately
from 2 to 3.4 V. Initially, the features were extracted from the whole interval according to
the procedure defined in Section 3.2. From the 124 available batteries, 15 were randomly
selected to form the test set, while all of the cycles of the remaining 109 batteries formed
the training and validation set. The SVR models were trained using solely the data of the
109 batteries. Initial parameter tuning was performed, followed by two sets of five-fold
cross-validation (CV) campaigns to determine the best feature sets and then optimize the
hyperparameters of the final model. In all stages, the predicted and measured values of
SoH were compared after every cycle of every battery of the training and validation set.
To achieve a fully standardized and comparable results framework, the following
procedure was developed and strictly followed:
The accuracy of the various proposed models was evaluated using the coefficient of
determination (R2 ). It is often defined as the amount of variability in the data explained or
accounted for by the model, and it is represented by Equation (4).
2
∑in=1 (y i − f ( xi ))
R2 = 1 − 2
(4)
∑in=1 (y i − y)
where
n—number of observations;
yi —target value of SoH;
f (xi )—predicted value of SoH;
y —mean of the target values.
Initial tuning of the SVR model hyperparameters (box constraint, epsilon, and kernel
scale) was performed using MATLAB 2021b’s built-in tool for hyperparameter optimiza-
tion with the following properties: Bayesian optimizer, 60 optimization iterations, and
loss metric RMSE. Due to the limited computational power, the number of optimization
iterations had to be limited as well. This step is required to find an acceptable starting point
from which to proceed with subsequent refinements.
Afterward, one of three sets of features is selected, where each feature set is a different
combination of the three constructed features. Five-fold cross-validation (CV) is applied
to the three potential feature sets, and the one with the highest mean CV R2 is chosen.
The considered feature sets along with their mean CV R2 scores for the full discharge
window approach are shown in Table 1. The highest R2 value was achieved for feature set
B, consisting of the first and third features.
Once the set of features is fixed, the model is further tuned by performing a five-fold
CV for a range of values for each single hyperparameter, where the range depends on the
results of the initial tuning of the parameters. This procedure is hierarchical, meaning that
it starts from the most impactful hyperparameter (epsilon), then moves to the second most
impactful (box constraint), and so on. The CV R2 of the final optimized model is 0.976.
210
Energies 2024, 17, 206
The hyperparameters of the final model considering the full window approach are a box
constraint (BC) equal to 0.0055, epsilon equal to 0.0021, and a gaussian kernel and kernel
scale (KS) of 1.
Table 1. Feature sets and achieved R2 values.
Feature Set Features Mean CV R2

A Ftr1, Ftr2, Ftr3 0.968
B Ftr1, Ftr3 0.970
C Ftr2, Ftr3 0.961
Finally, the test set is fed to the tuned model. In every cycle, the model provides a
predicted value of the SoH based on the input data of the same cycle. The R2 value is
calculated for every test battery individually, considering the measured and predicted
value of the SoH after each cycle. The best fit is achieved for battery 11 (T11) of the test
set, shown in Figure 6 with R2 = 0.999; the worst fit for battery 4 (T4), shown in Figure 7
with R2 = 0.794; and the mode is battery 5 (T5), shown in Figure 8 with an R2 = 0.981.
The average R2 value for all batteries of the test set is also very high and equal to 0.962,
indicating that, for the most part, the model fits the SoH evolution of the test set very well.
Figure 6. Best SoH estimation results for test battery 11.
4.2. Partial Discharge Window Approach

The full window approach is subject to limitations in real use cases. The typical
use of battery-powered devices prevents the battery cycle from adequately covering the
entire voltage range since batteries are rarely fully charged and discharged. Therefore,
further analysis was carried out, adopting the same experimental procedure described in
Section 4.1. to study how performance is affected by shrinking and moving the adopted
voltage window over the full voltage span.
Ten voltage intervals were selected and a five-fold CV procedure was performed for
every voltage interval to choose the best feature set. The results in Table 2 demonstrate how
the efficacy of the three feature sets changes for the different voltage intervals based on their
R2 values. Moreover, the full voltage interval is also present in the table for comparison. For
the high voltage intervals of 3.15 to 3.4 V and especially 3.25 to 3.4 V, which are highlighted
in red, the R2 value is low for all feature sets. The explanation for this low accuracy can be
found in Figure 3, which shows that between such high voltage limits the moved charge is
211
Energies 2024, 17, 206
low and does not profoundly change as the number of cycles increase, thereby resulting
in suboptimal features. Table 2 also shows that feature set B performs poorly for voltage
intervals with an upper limit of less than 3.1 V. One of the features of this set is related
to the variance in the difference between the curves of cycle k and the reference cycle 10.
While this set was the best performer in the case of a full voltage window, it can be seen in
Figure 4 that these difference curves have a constant value for voltages lower than 3.1 and,
thus, the variance is no longer a good indicator. On the other hand, this is not the case for
the minimum of the difference curve; thus, feature sets C and A continue to perform well.
For each voltage interval, the feature set that results in the highest R2 value is selected. The
chosen feature set along with the final tuned value of the hyperparameter for each model
are shown in Table 3.
Figure 7. Worst SoH estimation results for test battery 4.
Figure 8. Mode SoH estimation results for test battery 5.
212
Energies 2024, 17, 206
Table 2. Mean cross-validation R2 values for feature set selection.
Voltage Range (V) Feature Set A Feature Set B Feature Set C

Full window (2–3.4) 0.9686 0.9705 0.9620
3–3.4 0.8852 0.9572 0.9538
3.15–3.4 0.8992 0.8306 0.8777
3.25–3.4 0.8517 0.7894 0.8195
3–3.2 0.9624 0.9645 0.9536
3–3.1 0.9429 0.3904 0.9531
3–3.05 0.9608 0.3883 0.9646
2.8–3 0.9691 0.4715 0.9705
2.9–3 0.9657 0.3932 0.9705
2.4–2.6 0.9774 0.3944 0.9699
2.2–2.4 0.9772 0.3269 0.9741
Table 3. Final model hyperparameters for every voltage interval.
Voltage Range (V) BC KS Epsilon Kernel Feature Set

Full window (2–3.4) 0.0055 1.0 0.0021 Gaussian B
3–3.4 0.0055 1.0 0.0005 Gaussian B
3.15–3.4 1.100 2.0 0.0001 Gaussian A
3.25–3.4 0.022 1.0 0.0021 Gaussian A
3–3.2 0.0055 1.0 0.0001 Gaussian B
3–3.1 0.0055 1.0 0.0010 Gaussian C
3–3.05 0.0055 1.0 0.0021 Gaussian C
2.8–3 0.0055 1.0 0.00005 Gaussian C
2.9–3 0.0055 1.0 0.0010 Gaussian C
2.4–2.6 0.1100 2.0 0.0010 Gaussian A
2.2–2.4 0.1100 2.0 0.0005 Gaussian A
Table 4 shows a five-fold CV procedure applied on the final models. The results show
very repeatable R2 values for all folds and all voltage intervals except for 3.15–3.4 V and
3.25–3.4 V. For all other intervals, the minimum R2 value of any fold is higher than 0.91,
and the mean R2 value across all folds ranges between 0.957 and 0.982. These results
demonstrate that for the training and validation data, the models are of appropriate
complexity and have low bias and variance. This is further corroborated when the models
are applied to the remaining test battery data, as shown in Table 5.
Table 4. Five-fold CV R2 values of the final model for every voltage interval.
Voltage Range (V) Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Mean

Full window (2–3.4) 0.9578 0.9682 0.9891 0.9836 0.9816 0.9761
3–3.4 0.9337 0.9554 0.9799 0.9651 0.9701 0.9609
3.15–3.4 0.8672 0.9109 0.8994 0.9135 0.9301 0.9042
3.25–3.4 0.8063 0.8206 0.9146 0.8717 0.8495 0.8525
3–3.2 0.9479 0.9589 0.9846 0.9745 0.9707 0.9673
3–3.1 0.9453 0.9113 0.9784 0.9749 0.9735 0.9567
3–3.05 0.9509 0.9614 0.9816 0.9753 0.9742 0.9687
2.8–3 0.9585 0.9652 0.9857 0.9804 0.9802 0.9740
2.9–3 0.9581 0.9653 0.9855 0.9800 0.9797 0.9737
2.4–2.6 0.9784 0.9722 0.9842 0.9893 0.9866 0.9821
2.2–2.4 0.9760 0.9718 0.9893 0.9850 0.9865 0.9818
213
Energies 2024, 17, 206
Table 5. R2 value of all 15 test batteries for all selected voltage intervals.
Full
3–3.4 3.25–3.4 3.15–3.4 3–3.2 3–3.1 3–3.05 2.8–3 2.9–3 2.4–2.6 2.2–2.4
Window
T1 0.9834 0.9434 0.618 0.5622 0.9514 0.9548 0.9644 0.9744 0.9752 0.9874 0.9839
T2 0.9832 0.9371 0.9826 0.8765 0.9527 0.9512 0.9530 0.9665 0.9660 0.9913 0.9970
T3 0.9945 0.9901 0.9915 0.9734 0.9908 0.9914 0.9906 0.9914 0.9915 0.9954 0.9919
T4 0.7938 0.7817 0.7788 0.7452 0.7598 0.7689 0.7855 0.8038 0.8043 0.8103 0.8133
T5 0.9813 0.9869 0.9851 0.8784 0.9835 0.9887 0.9869 0.9853 0.9856 0.9782 0.9704
T6 0.9771 0.9773 0.9678 0.7470 0.8824 0.9737 0.9717 0.9709 0.9718 0.9905 0.9948
T7 0.9480 0.8937 0.7668 0.8838 0.9007 0.9216 0.9305 0.9465 0.9471 0.9924 0.9910
T8 0.9976 0.9771 0.6677 0.9613 0.9978 0.9783 0.9771 0.9966 0.9952 0.9987 0.9996
T9 0.9945 0.9825 0.9500 0.9521 0.9900 0.9896 0.9896 0.9924 0.9924 0.9970 0.9977
T10 0.9963 0.9796 0.8633 0.8468 0.9775 0.9925 0.9937 0.9967 0.9967 0.9881 0.9995
T11 0.9992 0.9985 0.8278 0.9964 0.9992 0.9990 0.9988 0.9993 0.9992 0.9981 0.9993
T12 0.9709 0.9533 0.8923 0.961 0.9361 0.9427 0.9455 0.9554 0.9537 0.9962 0.9891
T13 0.8234 0.8126 0.5329 0.6732 0.7952 0.8133 0.8196 0.8133 0.8089 0.8735 0.8706
T14 0.9925 0.9763 0.9929 0.9503 0.9782 0.9796 0.9805 0.9886 0.9885 0.9984 0.9987
T15 0.9944 0.9809 0.8917 0.9505 0.9856 0.9835 0.9856 0.9872 0.9880 0.9958 0.9954
Mean 0.9620 0.9447 0.8473 0.8639 0.9387 0.9486 0.9515 0.9579 0.9576 0.9727 0.9728
The columns of Table 5 are the different voltage intervals used to build the features of
the final models, while the rows are each test battery. As was the case for the full voltage
window, the R2 value shows the accuracy with which the predicted value of SoH—as a
function of the number of cycles—compares with the measured one, for every battery
individually. Unsurprisingly, for the voltage ranges of 3.15–3.4 V and 3.25–3.4 V, the
accuracy of the models is low for most batteries. On the other hand, for the rest of the
intervals, the R2 values are high and similar to the full window, and to the R2 values
obtained during the five-fold CV campaign using the training and validation data. More
specifically, the mean R2 for all batteries is in the range of 0.939 and 0.973. Therefore, the
first conclusion is that the models are accurate and not just overfitting to the training data.
The second, but just as important, conclusion is that partial voltage ranges can be used to
build the features of the model, thus giving real-life applicability to the constructed models.
Finally, it can be noted that for batteries 13 (T13) and 4 (T4), highlighted in red in
Table 5, the model accuracy is not very high, regardless of the voltage range used. This
is because they follow a very different SoH evolution trend. To highlight this difference,
for every battery, the SoH as a function-normalized number of cycles was defined by
dividing the value of number of cycles by the value of total number of cycles. Therefore,
the normalized number of cycles value for all cells is equal to 0 at the beginning and 1 when
the battery reaches the end of life. Afterwards, the batteries of the training and validation
set are grouped by their SoH value at every percent of the normalized number of cycles.
The blue curve in Figure 9 is the 50th percentile for the SoH as a function of the normalized
number of cycles, while the dotted blue lines are the 25th and 75th percentile. The same
procedure is repeated for the test set, omitting T4 and T13. The yellow line is the 50th
percentile of the SoH evolution for the test set, while the yellow dotted lines are the 25th
and 75th percentile. Based on the small width of the percentile curves, as well as the mostly
overlapping blue and yellow curves, it can be concluded that the batteries of the train
and test set follow a similar SoH evolution trend, regardless of their vastly different cycle
life. On the contrary, batteries 4 and 13, represented by the two red curves in Figure 9,
follow a vastly different pattern of degradation, having a more sudden and almost linear
decline in SoH rather than a slow descent. This is most likely due to some physical internal
battery issue or extreme usage conditions, making these two cells outliers of the dataset
and resulting in lower accuracy of SoH estimation by the models.
214
Energies 2024, 17, 206
Figure 9. Distribution of the SoH evolution curves for the training data (blue), test data (yellow), and
mis-predicted batteries 4 and 13 (red).
5. Conclusions
Accurate battery SoH estimation is crucial to achieve optimal performance, ensure
safety, and minimize cost and environmental impact. This paper presents an SoH estima-
tion method based on SVR, which offers a favorable compromise between applicability,
computational efficiency, and accuracy of results. The dataset used in this work consists of
124 batteries. A total 109 batteries formed the training and validation set and the remaining
15 batteries formed the test set, which was used only in the final stage, to evaluate the
performance of the developed models. Three features were selected, including two derived
from discharge capacity curves and the third one based on battery temperature.
Moreover, three combinations of features were analyzed to determine how their
effectiveness changes across different voltage ranges. When full discharge capacity curves
are considered, all three feature combinations are successful, with the best results obtained
when considering the logarithm of the variance of the difference curves and the temperature
integral features, reaching a mean CV R2 value of 0.976 for the training set and 0.962 for the
test set. Additionally, the same features were built from 10 small voltage intervals, and the
results show the considerations that must be made when following this approach—namely,
models using features constructed from the upper voltage limits of 3.15–3.4 V or 3.25–3.4 V
demonstrate poor accuracy because at such high voltage limits the moved charge is low
and does not profoundly change as the number of cycles increases. Furthermore, feature
set B performs poorly for voltage intervals with an upper limit of less than 3.1 V because
the difference curves have a constant value for voltages lower than 3.1 V; thus, the feature
related to variance is no longer a good indicator. This is not the case for the minimum
of the difference curve feature; thus, feature sets A and C continue to perform well. The
models that incorporate these limitations are very successful, reaching a CV R2 value in
the range of 0.957 and 0.982, and a mean R2 value for the test set in the range of 0.939
and 0.973, depending on the selected voltage interval. These values are on par with the
values achieved considering the full voltage window. Therefore, it can be concluded that
partial voltage ranges can be used to build the features of the model, thus giving real-life
applicability to the constructed models.
215
Energies 2024, 17, 206
Author Contributions: Conceptualization, E.P., I.M. and L.C.; methodology, E.P. and I.M.; software,
I.M. and E.P.; validation, I.M. and M.F.; formal analysis, E.P.; investigation, I.M.; resources, E.P.; data
curation, L.C. and E.P.; writing—original draft preparation, E.P. and I.M; writing—review and editing,
E.P. and I.M.; visualization, E.P. and I.M.; supervision, L.C., M.F. and E.P.; project administration,
L.C. and M.F.; funding acquisition, L.C. and M.F. All authors have read and agreed to the published
version of the manuscript.
Funding: The research received no external funding.
Data Availability Statement: Data are contained within the article.
Conflicts of Interest: The authors declare no conflicts of interest.
References
1. Pelletier, S.; Jabali, O.; Laporte, G.; Veneroni, M. Battery Degradation and Behaviour for Electric Vehicles: Review and Numerical
Analyses of Several Models. Transp. Res. Part B Methodol. 2017, 103, 158–187. [CrossRef]
2. Cristaldi, L.; Faifer, M.; Laurano, C.; Ottoboni, R.; Petkovski, E.; Toscani, S. Power Generation Control Algorithm for the
Participation of Photovoltaic Panels in Network Stability. IEEE Trans. Instrum. Meas. 2023, 72, 9000809. [CrossRef]
3. Li, X.; Fan, G.; Rizzoni, G.; Canova, M.; Zhu, C.; Wei, G. A Simplified Multi-Particle Model for Lithium Ion Batteries Via a
Predictor-Corrector Strategy and Quasi-Linearization. Energy 2016, 116, 154–169. [CrossRef]
4. Petit, M.; Prada, E.; Sauvant-Moynot, V. Development of an Empirical Aging Model for Li-Ion Batteries and Application to Assess
the Impact of Vehicle-to-Grid Strategies on Battery Lifetime. Appl. Energy 2016, 172, 398–407. [CrossRef]
5. Barcellona, S.; Colnago, S.; Dotelli, G.; Latorrata, S.; Piegari, L. Aging Effect on the Variation of Li-Ion Battery Resistance as
Function of Temperature and State of Charge. J. Energy Storage 2022, 50, 104658. [CrossRef]
6. Xu, B.; Oudalov, A.; Ulbig, A.; Andersson, G.; Kirschen, D.S. Modeling of Lithium-Ion Battery Degradation for Cell Life
Assessment. IEEE Trans. Smart Grid 2018, 9, 1131–1140. [CrossRef]
7. Li, X.; Yuan, C.; Li, X.; Wang, Z. State of Health Estimation for Li-Ion Battery Using Incremental Capacity Analysis and Gaussian
Process Regression. Energy 2020, 190, 116467. [CrossRef]
8. Cui, Z.; Wang, C.; Gao, X.; Tian, S. State of Health Estimation for Lithium-Ion Battery Based on the Coupling-Loop Nonlinear
Autoregressive with Exogenous Inputs Neural Network. Electrochim. Acta 2021, 393, 139047. [CrossRef]
9. Cao, M.; Zhang, T.; Wang, J.; Liu, Y. A Deep Belief Network Approach to Remaining Capacity Estimation for Lithium-Ion Batteries
Based on Charging Process Features. J. Energy Storage 2022, 48, 103825. [CrossRef]
10. Severson, K.A.; Attia, P.M.; Jin, N.; Perkins, N.; Jiang, B.; Yang, Z.; Chen, M.H.; Aykol, M.; Herring, P.K.; Fraggedakis, D.; et al.
Data-Driven Prediction of Battery Cycle Life before Capacity Degradation. Nat. Energy 2019, 4, 383–391. [CrossRef]
11. Ansean, D.; Garcia, V.M.; Gonzalez, M.; Blanco-Viejo, C.; Viera, J.C.; Pulido, Y.F.; Sanchez, L. Lithium-Ion Battery Degradation
Indicators Via Incremental Capacity Analysis. IEEE Trans. Ind. Appl. 2019, 55, 2992–3002. [CrossRef]
12. He, J.; Wei, Z.; Bian, X.; Yan, F. State-of-Health Estimation of Lithium-Ion Batteries Using Incremental Capacity Analysis Based on
Voltage–Capacity Model. IEEE Trans. Transp. Electrif. 2020, 6, 417–426. [CrossRef]
13. Zhou, R.; Zhu, R.; Huang, C.-G.; Peng, W. State of Health Estimation for Fast-Charging Lithium-Ion Battery Based on Incremental
Capacity Analysis. J. Energy Storage 2022, 51, 104560. [CrossRef]
14. Richardson, R.R.; Birkl, C.R.; Osborne, M.A.; Howey, D.A. Gaussian Process Regression for In Situ Capacity Estimation of
Lithium-Ion Batteries. IEEE Trans. Ind. Inform. 2019, 15, 127–138. [CrossRef]
15. Richardson, R.R.; Osborne, M.A.; Howey, D.A. Gaussian Process Regression for Forecasting Battery State of Health. J. Power
Sources 2017, 357, 209–219. [CrossRef]
16. Jia, J.; Liang, J.; Shi, Y.; Wen, J.; Pang, X.; Zeng, J. SOH and RUL Prediction of Lithium-Ion Batteries Based on Gaussian Process
Regression with Indirect Health Indicators. Energies 2020, 13, 375. [CrossRef]
17. Barcellona, S.; Cristaldi, L.; Faifer, M.; Petkovski, E.; Piegari, L.; Toscani, S. State of Health Prediction of Lithium-Ion Batteries.
In Proceedings of the 2021 IEEE International Workshop on Metrology for Industry 4.0 & IoT (MetroInd4.0 & IoT), Rome, Italy,
7–9 June 2021; pp. 12–17.
18. Lashgari, F.; Petkovski, E.; Cristaldi, L. State of Health Analysis for Lithium-Ion Batteries Considering Temperature Effect. In
Proceedings of the 2022 IEEE International Workshop on Metrology for Extended Reality, Artificial Intelligence and Neural
Engineering (MetroXRAINE 2022-Proceedings), Rome, Italy, 26–28 October; pp. 40–45. [CrossRef]
19. Weng, C.; Sun, J.; Peng, H. Model Parametrization and Adaptation Based on the Invariance of Support Vectors With Applications
to Battery State-of-Health Monitoring. IEEE Trans. Veh. Technol. 2015, 64, 3908–3917. [CrossRef]
20. Feng, X.; Weng, C.; He, X.; Han, X.; Lu, L.; Ren, D.; Ouyang, M. Online State-of-Health Estimation for Li-Ion Battery Using Partial
Charging Segment Based on Support Vector Machine. IEEE Trans. Veh. Technol. 2019, 68, 8583–8592. [CrossRef]
21. Marri, I.; Petkovski, E.; Cristaldi, L.; Faifer, M. Lithium-Ion Batteries Soh Estimation, Based on Support-Vector Regression and a
Feature-Based Approach. In Proceedings of the 18th IMEKO TC10 Conference on Measurement for Diagnostic, Optimisation and
Control to Support Sustainability and Resilience 2022, Warsaw, Poland, 26–27 September 2022; pp. 109–113.
216
Energies 2024, 17, 206
22. Marri, I.; Petkovski, E.; Cristaldi, L.; Faifer, M. Comparing Machine Learning Strategies for SoH Estimation of Lithium-Ion
Batteries Using a Feature-Based Approach. Energies 2023, 16, 4423. [CrossRef]
23. Wen, J.; Chen, X.; Li, X.; Li, Y. SOH Prediction of Lithium Battery Based on IC Curve Feature and BP Neural Network. Energy
2022, 261, 125234. [CrossRef]
24. Tian, Y.; Dong, Q.; Tian, J.; Li, X.; Kukkapalli, V.K.; Kim, S.; Thomas, S.A. Capacity Estimation of Lithium-Ion Batteries Based on
Multiple Small Voltage Sections and BP Neural Networks. Energies 2023, 16, 674. [CrossRef]
25. Li, Y.; Li, K.; Liu, X.; Wang, Y.; Zhang, L. Lithium-Ion Battery Capacity Estimation—A Pruned Convolutional Neural Network
Approach Assisted with Transfer Learning. Appl. Energy 2021, 285, 116410. [CrossRef]
26. Catelani, M.; Ciani, L.; Fantacci, R.; Patrizi, G.; Picano, B. Remaining Useful Life Estimation for Prognostics of Lithium-Ion
Batteries Based on Recurrent Neural Network. IEEE Trans. Instrum. Meas. 2021, 70, 3524611. [CrossRef]
27. Marri, I.; Petkovski, E.; Cristaldi, L.; Faifer, M. Battery Remaining Useful Life Prediction Supported by Long Short-Term Memory
Neural Network. In Proceedings of the IEEE International Instrumentation and Measurement Technology Conference (I2MTC),
Kuala Lumpur, Malaysia, 22–25 May 2023; pp. 1–6. [CrossRef]
28. Qu, J.; Liu, F.; Ma, Y.; Fan, J. A Neural-Network-Based Method for RUL Prediction and SOH Monitoring of Lithium-Ion Battery.
IEEE Access 2019, 7, 87178–87191. [CrossRef]
29. Tan, Y.; Zhao, G. Transfer Learning with Long Short-Term Memory Network for State-of-Health Prediction of Lithium-Ion
Batteries. IEEE Trans. Ind. Electron. 2020, 67, 8723–8731. [CrossRef]
30. Wang, S.; Takyi-Aninakwa, P.; Jin, S.; Yu, C.; Fernandez, C.; Stroe, D.-I. An Improved Feedforward-Long Short-Term Memory
Modeling Method for the Whole-Life-Cycle State of Charge Prediction of Lithium-Ion Batteries Considering Current-Voltage-
Temperature Variation. Energy 2022, 254, 124224. [CrossRef]
31. Wang, S.; Wu, F.; Takyi-Aninakwa, P.; Fernandez, C.; Stroe, D.-I.; Huang, Q. Improved Singular Filtering-Gaussian Process
Regression-Long Short-Term Memory Model for Whole-Life-Cycle Remaining Capacity Estimation of Lithium-Ion Batteries
Adaptive to Fast Aging and Multi-Current Variations. Energy 2023, 284, 128677. [CrossRef]
32. Toyota Research Institute. Experimental Data Platform. 2021. Available online: https://data.matr.io/1/ (accessed on
1 September 2023).
217
energies
Article
Battery State of Health Estimation Using the Sliding Interacting
Multiple Model Strategy
Richard Bustos 1 , Stephen Andrew Gadsden 2, *, Mohammad Biglarbegian 3 , Mohammad AlShabi 4
and Shohel Mahmud 1
1 College of Engineering and Physical Sciences, University of Guelph, Guelph, ON N1G 2W1, Canada
2 Department of Mechanical Engineering, McMaster University, Hamilton, ON L8S 4L8, Canada
3 Department of Mechanical and Aerospace Engineering, Carleton University, Ottawa, ON K1S 5B6, Canada
4 Department of Mechanical and Nuclear Engineering, University of Sharjah,
Sharjah 27272, United Arab Emirates
* Correspondence: gadsden@mcmaster.ca
Abstract: Due to their nonlinear behavior and the harsh environments to which batteries are subjected,
they require a robust battery monitoring system (BMS) that accurately estimates their state of charge
(SOC) and state of health (SOH) to ensure each battery’s safe operation. In this study, the interacting
multiple model (IMM) algorithm is implemented in conjunction with an estimation strategy to
accurately estimate the SOH and SOC of batteries under cycling conditions. The IMM allows for an
adaptive mechanism to account for the decaying battery capacity while the battery is in use. The
proposed strategy utilizes the sliding innovation filter (SIF) to estimate the SOC while the IMM serves
as a process to update the parameter values of the battery model as the battery ages. The performance
of the proposed strategy was tested using the well-known B005 battery dataset available at NASA’s
Prognostic Data Repository. This strategy partitions the experimental dataset to build a database of
different SOH models of the battery, allowing the IMM to select the most accurate representation of
the battery’s current conditions while in operation, thus determining the current SOH of the battery.
Future work in the area of battery retirement is also considered.
Keywords: lithium batteries; Kalman filters; sliding innovation filter; interacting multiple model;
Citation: Bustos, R.; Gadsden, S.A.;
state of health; state of charge; battery monitoring system; B005 battery dataset
Biglarbegian, M.; AlShabi, M.;
Mahmud, S. Battery State of Health
Estimation Using the Sliding
Interacting Multiple Model Strategy.
Energies 2024, 17, 536. https://
1. Introduction
doi.org/10.3390/en17020536 Owing to their high specific energy and high operational voltage, lithium-ion batteries
(LiB) have received great attention and are used in many applications [1]. Unfortunately,
LiB have a limited operational area mainly bound by two important parameters: voltage
and temperature. As such, careful monitoring of a battery’s working temperature and
Received: 16 October 2023 voltage is necessary for its optimal and safe operation [2]. If the battery’s voltage exceeds
Revised: 1 January 2024 its limit, the battery may develop dendrites over time, which increases the battery’s internal
Accepted: 4 January 2024 resistance, resulting in a lower output voltage. Moreover, if the working temperature is
Published: 22 January 2024 substantially increased, the battery may release toxic gases or burst into flames [3].
Another challenge presented by LiB technology is the accurate estimation of its avail-
able power or state of charge (SOC). SOC describes the amount of charge available in the
battery at any given time during usage. SOC is often represented as a percentage value
of available power vs. maximum power, or the available capacity vs. maximum capacity
of the battery [3]. The main problem in determining the SOC is the absence of instrumen-
tation that can accurately measure SOC during the battery’s operation. This results in an
conditions of the Creative Commons estimation problem where the SOC must be estimated using indirect measurements such
Attribution (CC BY) license (https:// as the battery’s terminal voltage and current [4].
creativecommons.org/licenses/by/ Different techniques to estimate the LiB’s SOC have been proposed in the literature.
4.0/). Some techniques such as neural networks (NN) have been used with great success [5];

218
Energies 2024, 17, 536
however, NN make use of extensive data that must be collected beforehand and are
computationally expensive compared to other solutions [5]. Other techniques make use of
electrochemical impedance spectroscopy (EIS) data, which requires special instrumentation
to be installed in the system [6].
One popular SOC estimation solution is the ampere-hour counting method. The
ampere-hour counting method determines SOC based on current measurements and the
remaining capacity of the battery [3]. This method’s popularity relies on its simplistic
approach. If the initial SOC is known, the previous SOC value can be subtracted or added
based on the current profile. However, this method comes with many drawbacks. Its
accuracy is highly dependent on the initial SOC value, correct current measurements,
and accurate battery capacity readings [3]. To ensure proper estimates of the SOC, this
method must be frequently calibrated; some calibration techniques include voltage-based
corrections using lookup tables [3]. Another method was presented in [7], where the
authors were able to jointly estimate the SOC and temperature at the same time. The ability
to track the temperature in conjunction with the SOC provides useful insights in terms of
battery life management and operational safety.
Furthermore, Kalman filters (KFs) present other estimation techniques that, when
combined with the ampere-hour counting method, have proven to be accurate at estimat-
ing SOC. KFs provide an accurate and computationally inexpensive solution, but require
an accurate battery model for their successful implementation [3]. A linear KF provides
an optimal solution to the linear discrete estimation problem. However, due to the battery’s
nonlinear nature, only modified versions of the KF have been used for SOC estimation.
Some KF variations include the extended Kalman filter (EKF) and the unscented Kalman
filter (UKF), among others [8,9]. Between these two strategies, the EKF is known to intro-
duce instability in the estimation process due to the linearization process embedded in the
algorithm [10]. On the other hand, the UKF has proven to be a more robust strategy [11,12].
Another robust strategy, known as robust fixed-lag smoothing, attempts to overcome
model uncertainties or mismatch by utilizing the least favorable model over a finite time
horizon [13]. This method is characterized by a dynamic game between two players: one
player selects the least favorable model in a prescribed ambiguity set, while the other
player selects the fixed-lag smoother, minimizing the smoothing error with respect to the
least favorable model. Efficient implementation of the robust fixed-lag smoother may
reduce computational burdens and avoid numerical instabilities, which may be helpful for
battery applications.
Electrochemical and equivalent circuit models (ECMs) are among the most popu-
lar models for batteries. Electrochemical models are based on the underlying physics of
the battery using 10–14 partial differential equations, resulting in highly complex and
computationally demanding models, but providing high-accuracy information about
the battery’s state. These types of models are often used for laboratory and battery
development research [14–17].
On the other hand, ECMs represent the battery as an electric circuit using voltage
sources, resistors, and capacitors. These types of models require low computational power
and have low complexity, but are less accurate and yield little information about the
battery [18]. Nevertheless, these traits allow for their implementation online.
Some ECMs studied include Rint model, Thevenin model, PNGV model, and Dual
Polarity (DP) model [19]. These models can be differentiated by the number of Resistor–
Capacitor (RC) branches in the circuit. Adding more RC branches allows the capture of
higher-order nonlinearities, resulting in a more accurate model [19]. However, adding
more RC branches increases the complexity and computational time of the algorithms.
In summary, a battery monitoring system (BMS) should be implemented to ensure safe
operation of LiB. The BMS’s main function is the accurate estimation of the battery’s current
SOC and operating temperature. In addition, the BMS can also track other parameters
such as the battery’s state of health (SOH); SOH is a measurement of the current health
of the battery and is sometimes calculated based on its available maximum capacity [20].
219
Energies 2024, 17, 536
As the battery is subject to aggressive current profiles, excessive cycling, or regular use,
its maximum battery capacity degrades over time [20]. Moreover, accurate estimation of
the battery’s SOH can significantly increase the accuracy of the ampere-hour counting
method, since it is dependent on the battery’s capacity [20]. Lastly, accurate tracking of the
battery’s SOH allows for an effective planned retirement of the battery, which ensures that
the system continues to operate optimally.
A battery is referred to as due for retirement once its SOH is at 80%, or in other
words, when the battery’s maximum available capacity is at 80% or less of its designed
capacity [20]. Battery retirement can be presented as a fault diagnosis problem, where
a SOH value of 80% or lower signals a fault in the battery [21]. A recent paper presented
a degradation empirical model-free battery end-of-life prediction framework [22]. This
method utilized the KF and Gaussian process regression. It is important to note that the
SOH should be rapidly tracked and updated for improved performance and reliability. The
authors in [23] introduced a fast capacity estimation method as well as a fast accelerated
degradation fault diagnosis strategy for SOH estimation. This article offers insights into
the importance of tracking micro-health parameters in batteries, which directly correspond
to the overall SOH of the battery or set of batteries.
The multiple model (MM) strategy has been used to detect faults in batteries [24]. In the
MM strategy, several models representing different behaviors of the system are generated
to make the algorithm resilient against uncertainty [25]. Moreover, [25] presented an
interacting multiple model (IMM) strategy where the IMM was combined with the EKF to
accurately estimate the SOC of a LiB. The IMM was given allowed two different variations
of noise in the battery model to account for the different degrees of parameter shift during
the estimation process. Lastly, in [26], a multiple model adaptive estimation (MMAE)
technique was used for fault diagnosis. The proposed strategy made use of EIS data and
EKFs to generate residual signals that were fed into an MMAE block to detect a fault in
the battery.
This paper focuses on the implementation of a MM strategy, i.e., the IMM strategy,
to estimate the battery’s capacity degradation while accurately estimating the SOC of
a battery under cycling conditions [27]. This is a unique contribution to the field of battery
monitoring, particularly when utilizing the relatively new sliding innovation filter (SIF).
The experimental dataset was partitioned into sections representing a 100% SOH, 75%
SOH, 50% SOH, 25% SOH, and 0% SOH, where each section can be identified as a mode to
be used within the IMM algorithm. The motivation behind this partition is that the IMM
would yield the best matching mode, thus identifying the current SOH of the battery.
The main contribution of this paper is the development of the SIF in conjunction with
the IMM (the so-called SIF-IMM) for determining the SOC and SOH of a battery. The
IMM algorithm is used for SOH estimation by partitioning the experimental dataset into
several SOH modes. This strategy has not been presented in the literature. In addition to
introducing this method, the paper compares the performances of SIF-IMM and KF-IMM
in estimating SOH.
The remainder of the paper is structured as follows: Section 2 presents the battery
and parameter models. Section 3 details the experimental data and estimation algorithms.
Section 4 covers the artificial measurements. Section 5 describes the model parameter
identification results. Section 6 presents the experimental setup and details the results of
the proposed strategy. Section 7 presents the concluding arguments of the work.
2. Battery Models
This section presents the battery model used for the experiment.
Dual Polarity Model

The DP model is an ECM composed of a voltage source, a resistor, and two RC-
branches. These elements represent the battery’s output voltage, internal resistance, and
short-term and long-term transient behaviors, respectively [16]. The battery’s transient
220
Energies 2024, 17, 536
behaviors are most noticeable at the end of a discharge charge, once the current is cut
off. It is evident that there is a quick rise in voltage followed by a slower rise in voltage.
These two phenomena have been attributed to the battery’s concentration polarization
and electrochemical polarization, respectively. In the literature, the DP model has been
shown to capture these behaviors by implementing a two RC-branch model [16]. The DP
model was selected due to its high accuracy and ability to capture more nonlinearities while
remaining computationally efficient. Figure 1 depicts the circuit diagram of the DP model.
Figure 1. DP model circuit architecture [16].
The circuit can be analyzed by breaking it down into three parts: OCV, resistance,
and capacitance. OCV represents the open circuit voltage of the battery, Ro represents the
internal resistance of the battery, and R pa and R pc represent the electrochemical polarization
resistance and concentration polarization resistance, respectively. Lastly, C pa and C pc
characterize the transient response during the transfer of power to/from the battery during
the electrochemical and concentration polarization [16].
Lastly, the system’s state space representation is described by the following equations:
⎡ ⎤
⎡ ⎤ − Δt
− R paΔtC pa R pa 1 − e R pa C pa
U pa, k+1 e 0 ⎢ ⎥
=⎣ ⎦ U pa, k + ⎢ ⎥ IL , (1)
U pc, k+1 − R pcΔtC pc U pc, k ⎣ − R pcΔtC pc ⎦
0 e R pc 1 − e
UL, k+1 = Uocv (SOCk+1 ) − U pa,k+1 − U pc,k+1 − IL,k R0 . (2)
3. Experimental Data and Estimation Algorithms

This section presents the selected experimental data and the estimation algorithms
used for the experiments.
3.1. B005 DATASET

The B005 dataset is part of several datasets released by the Prognostic Center of
Excellence (PCoE) and published by the Prognostic Data Repository for the advancement
of prognostic algorithms [28]. This dataset provides cycling data of a 2 Ah battery up to
a capacity of ~70%. The cycling data includes time, current, voltage, and temperature
measurements for each cycle [28]. The B005 dataset was selected to test the proposed
strategy as it provides data for each cycle. If the impedance data is counted, there are
615 cycles in total in this dataset. In this study, this dataset provides 340 discharge and
charge cycles. The battery has an initial capacity of 1.856 Ah and is cycled to 1.303 Ah.
Finally, the data was resampled from 3 s to 0.6 s to allow for improved algorithm adherence.
Figure 2 illustrates the battery capacity degradation over the 615 cycles.
221
Energies 2024, 17, 536
Figure 2. Measured battery capacity at the end of each discharge cycle [29].
3.2. Ampere-Hour Counting

This method is the most popular for estimating SOC, where the SOC is estimated
based on an initial value, i.e., SOC0 , and this value is decreased/increased depending on
the demanded/supplied current. SOC estimated using this method is given by [3]:
t
1
SOC = SOC0 − I dτ, (3)
Cn t0
where SOCo is the initial SOC, Cn is the nominal capacity of the battery, and I is the
discharge current.
3.3. Kalman Filter

The Kalman filter (KF) yields the optimal solution by minimizing the state estima-
tion error for a known linear system that is subject to white Gaussian noise [30]. The
linear system dynamics and the measurement model are described by the following
two equations [31]:
xk+1 = Axk + Buk + wk , (4)
zk+1 = Cxk+1 + vk+1 , (5)

where A is the dynamics matrix, B is the input matrix, C is the output matrix, x is the
system states, z is the measurement output, u is the input, w is the system noise, and v is
the measurement noise.
The KF algorithm is summarized in two stages: prediction and update [30].
(i) Prediction Stage:
x̂k+1|k = A x̂k|k + Buk (6)
Pk+1|k = APk|k A T + Q (7)
w ∼ N (0, Q) (8)
v ∼ N (0, R) (9)
(ii) Update Stage: −1
Kk+1 = Pk+1|k C T CPk+1|k C T + R (10)

x̂k+1|k+1 = x̂k+1|k + Kk+1 zk+1 − C x̂k+1|k (11)
Pk+1|k+1 = [ I − Kk+1 C ] Pk+1|k ( I − Kk+1 C ) T + Kk+1 RKkT+1 (12)
222
Energies 2024, 17, 536
where Q and R are the system and measurement noise covariance matrices, respectively.
3.4. Sliding Innovation Filter

A recently proposed estimation strategy is the novel sliding innovation filter (SIF). The
SIF was formulated based on a predictor–corrector estimation method [32]. Like the KF, the
SIF first makes predictions of the state estimates and state error covariances using values
from previous time steps, and then updates the predictions based on the measurements
and a correction term referred to as the SIF gain [32].
The main difference between the SIF and KF algorithms is how the gain was con-
structed. The SIF gain was built using the measurement matrix, the innovation, and
a sliding boundary layer term [32]. On the other hand, the KF makes use of the state error
covariance. In other words, the SIF makes use of the gain to drive the state estimates
within the defined boundary layer and forces the state estimates to switch about the true
trajectory [32]. The following equations describe the SIF algorithm [32].
(i) Update Stage: ∼

Kk+1 = C + sat z k+1|k /δ x̂k+1|k+1 (13)
∼
x̂k+1|k+1 = x̂k+1|k + Kk+1 z k+1|k Pk+1|k+1 (14)
Pk+1|k+1 = ( I − Kk+1 Ck+1 ) Pk+1|k ( I − Kk+1 Ck+1 ) T + Kk+1 R KkT+1 (15)
Note that C+refers to the pseudoinverse of C, sat refers to the diagonal of the satura-
tion term (value between −1 and +1), and δ is the sliding boundary layer width.
3.5. Interacting Multiple Model (IMM)

The IMM method is most useful for systems with more than one operating mode.
The IMM algorithm utilizes a number of models and computes the likelihood values for
each model based on the state estimates and the corresponding state error covariance. The
likelihood value is also based on the innovation (or measurement error) and provides an
indication of how similar the actual system’s behavior is compared to the filter model.
These likelihood values are then used to determine the operating mode [33]. Note that the
IMM is essentially a set of filters that run in parallel, with each filter utilizing a different
dynamic system and/or measurement model. The IMM algorithm can be described in
a set of five stages. These five stages are presented below [33].
(i) Calculation of the mixing probabilities μi| j,k|k :
1
μi| j,k|k = p μ (16)
c j ij i,k
r
cj = ∑ pij μi,k (17)
i =1
where μi| j,k|k refers to the mixing probabilities between modes i and j at time k, pij refers
to the mixing parameter (predefined value), and μi,k refers to the mode probability.
(ii) Mixing Stage: r
x̂0j,k|k = ∑ x̂i,k|k μi| j,k|k (18)
i =1
r T )
P0j,k|k = ∑ i| j,k|k i,k|k
μ P + x̂ i,k |k − x̂ 0j, k |k x̂ i,k|k − x̂ 0j, k |k (19)
i =1
where x̂0j,k|k and P0j,k|k are the state estimates and state error covariances used as
initial conditions for the filters (KF or SIF).
(iii) Mode-Matched Fitlering:

Λj,k+1 = N zk+1 ; ẑ j,k+1|k , S j,k+1 (20)
223
Energies 2024, 17, 536

1 − 12 e Tj,z,k+1 e j,z,k+1|k
Λj,k+1 = exp (21)
S j,k+1
2πS j,k+1
Abs
where Λj,k+1 refers to the likelihood value generated based on the measurement,
estimate measurement (from the corresponding filter), and the innovation covariance
matrix (from the corresponding filter). The likelihood value is then used to update
the mode probability (described in the next stage).
(iv) Mode Probability Update:
r
1
μ j,k = Λj,k+1 ∑ pij μi,k (22)
c i =1
r r
c= ∑ Λj,k+1 ∑ pij μi,k (23)
j =1 i =1
(v) State Estimate and Covariance Combination

r
x̂k+1|k+1 = ∑ μ j,k+1 x̂ j,k+1|k+1 (24)
j =1
r T )
Pk+1|k+1 = ∑ μ j,k+1 Pj,k+1|k+1 + x̂ j,k+1|k+1 − x̂ k+1|k+1 x̂ j,k+1|k+1 − x̂ k+1|k+1 (25)
j =1
Note that for this final stage (v), the overall state estimates and state error covariance
are used outside of the IMM process, and are used for output purposes only.
In the proposed strategy, the IMM is used to track the correct battery capacity, thus
estimating the SOH of the battery. Meanwhile, the KF and SIF are used to estimate the SOC
of the battery during cycling. The integration of these algorithms with the IMM resulted in
strategies referred to as KF-IMM and SIF-IMM. The figure shown in Appendix A helps to
further illustrate the overall architecture of the proposed strategies.
4. Artificial Measurements
As mentioned in Section 3.4, the SIF gain was derived based on the measurement ma-
trix, meaning that it is dependent on the availability of individual measurements for each
state estimate [32]. When individual measurements for each state are not available, which
is often the case for LiBs and most types of batteries, generating artificial measurements
is necessary to ensure that the SIF is effective [32]. The SIF strategy relies on the measure-
ments in order to formulate good estimates of the states through the use of a full identity
measurement matrix. This is one of the main disadvantages of this method. This section
presents how artificial measurements were generated for the battery and parameter model.
4.1. State Measurement Equations

LiBs only offer measurements of current, voltage, and temperature. To generate
artificial measurements for each state of the battery model, the model equations were
rearranged as followed:
* pa,k+1 = OCV (SOCk ) − UL,k+1 − R0,k Is,k − U pc,k
U (26)
U* pc,k+1 = OCV (SOC ) − UL,k+1 − Ro,k Is,k − U pa,k (27)

k+1 = OCV −1 UL,k+1 + U pa,k + U pc,k + Ro,k Is,k
SOC (28)
* pa , U
where U are the measurements for each state of the battery model, and
* pc , and SOC
OCV −1 (.) is the inverse function of OCV (SOC ) found in (27).
224
Energies 2024, 17, 536
4.2. Parameter Measurement Equations

To generate artificial measurements for the parameters of interest, Battcap and R0 ,
Equations (2) and (3) were rearranged as follows:
! "
* o = 1 OCV (SOC ) − UL − U pa − U pc
R (29)
Is,k
+ Cap = Δt × Is,k
Batt (30)
abs(3.6 × ΔSOCk )
*0 and Batt
where R + cap represent the artificial measurements for R0 and Battcap .
5. Model Parameter Identification

As mentioned before, the IMM makes use of several models. These models were
identified by breaking the B005 dataset into four regions, resulting in five different modes
to be identified corresponding to 100%, 75%, 50%, 25%, and 0% SOH. Furthermore, the
B005 dataset contains 340 discharge and charge cycles; however, only the discharge cy-
cles provide battery capacity measurements. Since this paper focuses on battery capac-
ity estimation, the models were derived using discharge cycles. To derive each battery
model, the following parameters needed to be identified (for each of the selected cycles):
OCV (SOC ), R0 , R pa , C pa , R pc , C pc . This section presents the model’s parameter identifi-
cation results using the nonlinear least squares (NLLS) algorithm for each of the modes
selected: 100%, 75%, 50%, 25%, and 0% SOH.
5.1. Least Squares Setup

To use the NNLS algorithm, a relationship between the measurable data and the
parameters must be established. Making use of the battery model, the following relationship
can be derived [29]:
OCV (SOC ) = α0 + α1 SOC + α2 SOC2 + α3 SOC3 + α4 SOC4 + α5 SOC5 (31)

− t
U pa = IL R pa 1 − e R pa C pa (32)

− t
U pc = IL R pc 1 − e R pc C pc (33)

− t − t
UL = OCV (SOC ) − IL R0 − IL R pa 1 − e R pa C pa − IL R pc 1 − e R pc C pc (34)

1 1
θ = α0 , α1 , α2 , α3 , α4 , α5 , R0 , R pa , , R pc , (35)
R pa C pa R pc C pc
where t represents the time vector, OCV (SOC ) is the OCV curve approximated to a 5th
order polynomial, and θ represents the parameter vector, i.e., a vector consisting of all the
parameters. Here, the capacitance values are estimated using a fraction to account for their
higher magnitude [29,34].
Lastly, MATLAB® ’s R2023a ‘lsqcurvefit’ function was used to perform the parameter
estimation process. This function makes use of boundaries and initial conditions which are
detailed in Table 1.
Table 1. NLLS boundaries and initial guess.
Parameters R0 Rpa Cpa Rpc Cpc

Unit Ω Ω 1/(ΩF) Ω 1/(ΩF)
LB 0.001 0.01 0.0001 0.01 0.01
UB 0.500 0.500 0.002 0.500 0.100
Guess 0.020 0.100 0.001 0.100 0.010
225
Energies 2024, 17, 536
5.2. Least Squares Results

5.2.1. The 100% SOH Model
As indicated before, 100% SOH represents a battery with an available capacity equal
to its designed capacity. Table 2 shows the results of the parameter identification process
for the 100% SOH mode. The terminal voltage’s RMSE was 0.063 V. Figure 3 depicts
the terminal voltage’s error plot between the model and the measured terminal voltage.
Finally, Figure 4 illustrates the generated model’s terminal voltage plot vs. the measured
terminal voltage.
Table 2. NLLS parameter estimation results for 100% SOH model.
RC Parameters Value OCV (SOC) Value

R0 0.0700 α1 0.6996
R pa 0.1070 α2 17.4679
C pa 1329.29 α3 −62.4061
R pc 0.0401 α4 76.7998
C pc 19325 α5 −31.7285
Figure 3. 100% SOH: LS model error—terminal voltage.
Figure 4. Terminal voltage: model vs. measured voltage for 100% SOH model.

The 75% SOH of the battery was determined to occur when the battery reaches
a capacity of 1.83 Ah which, based on the data, happens at the ~98th cycle. Table 3
226
Energies 2024, 17, 536
illustrates the results of the parameter identification process for the 75% SOH mode. The
terminal voltage’s RMSE was 0.032 V. Figure 5 depicts the terminal voltage’s error plot
between the model and the measured terminal voltage. Figure 6 shows the generated
model’s terminal voltage plot vs. the measured terminal voltage.

R0 0.07 α1 5.8933
R pa 0.10 α2 −23.1629
C pa 1428.60 α3 42.4898
R pc 0.04 α4 −35.0472
C pc 53333 α5 10.7280

a capacity of 1.65 Ah which, based on the data, happens at the ~225th cycle. Table 4 il-
lustrates the results of the parameter identification process for the 50% SOH mode. The
terminal voltage’s RMSE was 0.032 V. Figure 7 shows the terminal voltage’s error plot be-
tween the model and the measured terminal voltage, and Figure 8 illustrates the generated
model’s terminal voltage plot vs. the measured terminal voltage.
227
Energies 2024, 17, 536

R0 0.08 α1 5.0727
R pa 0.137 α2 −17.8196
C pa 428.730 α3 29.3507
R pc 0.078 α4 −21.1457
C pc 9500.530 α5 5.3372

a capacity of 1.48 Ah which, based on the data, happens at the ~375th cycle. Table 5
illustrates the results of the parameter identification process for the 25% SOH mode. The
terminal voltage’s RMSE was 0.034 V. The terminal voltage’s error plot between the model
and the measured terminal voltage is depicted in Figure 9. The generated model’s terminal
voltage plot vs. the measured terminal voltage in shown in Figure 10.
228
Energies 2024, 17, 536

R0 0.08 α1 3.1583
R pa 0.19 α2 −5.2627
C pa 401.82 α3 −2.0509
R pc 0.0277 α4 12.8807
C pc 86,473.20 α5 −7.9551

The 0% SOH of the battery occurs when the battery reaches a capacity of 1.30 Ah
which, based on the data, happens at the ~588th cycle. Table 6 illustrates the results of
the parameter identification process for the 0% SOH mode. The terminal voltage’s RMSE
was 0.042 V. Figure 11 shows the terminal voltage’s error plot between the model and the
measured terminal voltage. The generated model’s terminal voltage plot vs. the measured
terminal voltage is illustrated in Figure 12.
229
Energies 2024, 17, 536

R0 0.080 α1 2.4549
R pa 0.202 α2 −2.8347
C pa 401.510 α3 −5.1815
R pc 0.001 α4 14.5322
C pc 86,963.20 α5 −8.2310
6. Simulation Setup and Results

6.1. Simulation Setup
Mode Matching
In this experiment, the IMM was used for tracking the battery capacity of the LiB. The
IMM was combined with the KF and SIF, resulting in the KF-IMM and SIF-IMM algorithms.
The initial mode probabilities were
set to 95% normal and 5% faulty. The IMM mixing
0.95 0.05
parameter was defined as pij = . Initial estimates were set to zero.
0.05 0.95
The IMM has access to five different battery models representing different stages of
the SOH of the battery. As the battery ages, the IMM selects the mode that best matches
the battery’s current state and by doing so, the battery’s current SOH can be determined.
Table 7 summarizes the initial conditions used for this experiment.
230
Energies 2024, 17, 536
Table 7. Mode matching experiment: initial conditions.
Variables Values
Vpa 0
Vpc 0
SOC 100%
R0 0.07
BattCap 2
Q100,75,25%SOH diag(5 × 10−9 , 5 × 10−9 , 1 × 10−8 )
Q50% SOH diag(5 × 10−9 , 5 × 10−9 , 1 × 10−7 )
Q0% SOH diag(5 × 10−8 , 5 × 10−8 , 1 × 10−6 )
R100−0% SOH diag(5 × 10−2 , 5 × 10−2 , 5 × 10−2 )
Delta100% SOH diag(5 × 101 , 5 × 101 , 1 × 102 )
Delta75,25,0% SOH diag(1 × 101 , 1 × 101 , 1 × 102 )
Delta50% SOH diag(9, 9, 1 × 102 )
p 0.9999
μ 0.2
6.2. Simulation Results

This section presents the results obtained from the proposed strategy.
Mode Matching
Five different modes were identified in the B005 dataset and served as the database for
the IMM algorithm. At the end of each time step of each cycle, the parameter values of each
mode were multiplied by the probability of each mode and combined to determine the esti-
mated parameter values of the proposed IMM strategy. The following cycles were chosen
to demonstrate the results of the proposed strategy at each stage of SOH: 106, 278, 441, 596.
These cycles correspond to 75% SOH, 50% SOH, 25% SOH, and 0% SOH, respectively.
At cycle 106, the battery’s measured capacity is ~1.81 Ah. Therefore, the KF-IMM and
SIF-IMM should select, for the most part, the 75% SOH mode to better match the current
state of the battery. Figure 13 illustrates the estimation results of the KF-IMM and SIF-IMM.
Part (a) shows the terminal voltage estimation. Part (b) depicts the SOC estimation and
part (c) illustrates the battery capacity estimation at the 106th cycle.
The terminal voltage’s RMSE values for the KF-IMM and SIF-IMM were 0.0249 and 0.0255,
respectively; both of these values show strong accuracy. Furthermore, both SOC estimation
results showed a good profile. Moreover, the estimated battery capacity for both algorithms
remained, for the most part, at ~1.83 Ah.
Lastly, Figure 14 illustrates the mode probability of each algorithm at cycle 106.
Both algorithms switch between various modes to account for the sudden drop in voltage
at the beginning of the discharge process. This switching can also be observed when
the current is cut off, and the battery starts to recover after a period of ~53 min. Most
importantly, it is evident that both filters select the 75% SOH mode as the better match,
which reflects the measured battery capacity in the B005 dataset.
SIF-IMM should select, for the most part, the 50% SOH mode. Figure 15 illustrates the
terminal voltage, SOC, and battery capacity estimation results of the KF-IMM and SIF-IMM
at the 278th cycle.
The terminal voltage’s RMSE values for the KF-IMM and SIF-IMM were 0.0377 and
0.0257, respectively; both values suggest high accuracy. Furthermore, both SOC estimation
results showed a good constant discharge profile, with the KF-IMM fully discharging the
battery. In terms of battery capacity estimates, the KF-IMM had a mean value of 1.56 Ah,
which matches the measured value from the dataset. The SIF-IMM has a higher mean
value at 1.65 Ah. This higher value in battery capacity enforces the discrepancy in the SOC
discharge profiles between both algorithms.
231
Energies 2024, 17, 536
(a)
(b) (c)
Figure 13. (a) Measured terminal voltage vs. KF-IMM and SIF-IMM estimation at cycle 106.
(b) KF-IMM and SIF-IMM SOC estimation at cycle 106. (c) KF-IMM and SIF-IMM battery capacity
estimation at cycle 106.
(a) (b)
Figure 14. (a) KF-IMM mode probability at cycle 106. (b) SIF-IMM mode probability at cycle 106.
232
Energies 2024, 17, 536
(a)
(b) (c)
Figure 16 demonstrates the mode probability of each algorithm at cycle 278. In part
(a), the KF-IMM selects the 25% SOH model as the predominant mode for most of the cycle.
Looking at the same window of time in part (b) of Figure 15, the lower battery capacity
allows for a faster discharge rate, which is more advantageous towards the end of the cycle
when compared to the SIF-IMM results. However, the sudden jump in voltage near the
40 min mark, and the selection of the 75% SOH mode afterwards, can be a consequence of
the selection of the 25% SOH mode.
On the other hand, the SIF-IMM method chooses the 50% SOH mode for most of the
cycle. This mode selection reflects a better estimate of the current SOH of the battery based
on the measured battery capacity.
SIF-IMM should select, for the most part, the 25% SOH mode to match the current state of
the battery. The estimation results of the KF-IMM and SIF-IMM are shown in Figure 17.
Part (a) shows the terminal voltage results. Part (b) depicts the SOC estimation and part (c)
illustrates the battery capacity estimation at the 411th cycle.
233
Energies 2024, 17, 536
(a) (b)
(a)
(b) (c)
respectively, confirming their accuracy. In part (b), both SOC estimation curves have a good
overall profile that reflects a full discharge of the battery. Moreover, in terms of battery
234
Energies 2024, 17, 536
capacity, the mean value for both algorithms was ~1.51 Ah. These values are close to the
measured value of 1.42 Ah in the dataset. The corresponding mode probabilities for this
case are shown in Figure 18.
(a) (b)
At cycle 596, the battery’s measured capacity is ~1.3 Ah. Therefore, the KF-IMM
and SIF-IMM should select, for the most part, the 0% SOH mode. Figure 19 illustrates
the estimation results of the KF-IMM and SIF-IMM. Part (a) shows the terminal voltage
estimation. Part (b) depicts the SOC estimation and part (c) illustrates the battery capacity
estimation at the 596th cycle.
(a)
(b) (c)
235
Energies 2024, 17, 536
respectively; both values show strong accuracy. Both algorithms’ SOC curves show
a good full discharge profile. Moreover, the estimated battery capacity for both algorithms
remained, for the most part, at ~1.33 Ah.
The mode probability of each algorithm at cycle 596 is shown in Figure 20. Both
algorithms select the correct mode, which reflects the current SOH of the battery in the
B005 dataset.
(a) (b)
To analyze the overall estimation results of the KF-IMM and SIF-IMM, the mean
battery capacity for each cycle was generated, and the mode chosen by the IMM-selected
models was taken at the end of each cycle. Figure 21 shows the most frequently selected
mode for each cycle. Part (a) depicts the most-selected mode of the KF-IMM and part
(b) illustrates the most-selected mode of the SIF-IMM. Figure 22 depicts the mean battery
capacity of both algorithms for each cycle compared to the measured battery capacity.
(a) (b)
Figure 21. (a) KF-IMM’s most selected model for each cycle. (b) SIF-IMM’s most selected model for
each cycle.
From Figure 21 it is evident that neither algorithm made use of the 100% SOH model
at the beginning of the experiment. The reason is evident if one were to look at the starting
capacity of the battery, which is closer to the 75% SOH model. Based on the break points
for each identified SOH region, the algorithms should ideally switch between modes after
cycles 98, 225, 375, and 588. Since the battery’s capacity starts at 75% SOH, the first switch
should occur after the 225th cycle. However, the KF-IMM switches to the 50% SOH model
earlier than the SIF-IMM. This early jump reinforces the results from Figures 15 and 16.
236
Energies 2024, 17, 536
Furthermore, neither algorithm makes consistent use of the 50% SOH model and both
algorithms make an early jump to the 25% SOH model. Finally, both algorithms switch to
the 0% SOH model before the expected 588th cycle. Taking a holistic view on Figure 21,
both algorithms show the desired downward step trend.
Figure 22. Mean battery capacity across all cycles.
It can be seen from Figure 22 that both algorithms show a good battery capacity trend.
The battery capacity RMSE values for the KF-IMM and SIF-IMM were 0.060 and 0.065,
respectively. When compared to the difference between the healthy and retired battery
capacity values, the RMSE values correspond to an error of ~9% in determining the current
SOH of the battery. After further inspection of Figure 22, the early switches between modes
at cycles ~190 and ~440 can be justified.
7. Conclusions
This paper presented a strategy that utilizes the interacting multiple model (IMM)
algorithm integrated with the Kalman filter (KF) and sliding innovation filter (SIF) methods
for mode-matched filtering. As demonstrated by the results, the proposed strategy accu-
rately estimates the state of charge (SOC) and state of health (SOH) of a lithium-ion battery
(LiB) under cycling conditions. More specifically, the KF-based methods worked more
accurately under ‘normal’ operating conditions (e.g., when the dynamic models closely
resembled the actual operating models). However, when the SOH started to reduce or
the battery began to degrade (essentially when there was model mismatch), the SIF-based
methods yielded better estimates.
strategies called KF-IMM and SIF-IMM. Moreover, the proposed strategy was evaluated
using experimental data and was found to be successful in tracking the SOH of the battery.
In the future, the proposed strategy will be implemented on a real-world BMS for further
testing and verification. Lastly, this strategy has the potential to be used for predicting
battery retirement, the results of which will be explored further in future studies.
Author Contributions: Investigation, R.B. and S.A.G.; writing—original draft preparation, R.B.;
writing—review and editing, S.A.G., M.B., M.A. and S.M. All authors have read and agreed to the
published version of the manuscript.
Data Availability Statement: Publicly available datasets were analyzed in this study. The B005
dataset can be found here: https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/,
accessed on 20 August 2022. B. Saha and K. Goebel (2007). The “Battery Data Set” is provided
237
Energies 2024, 17, 536
by the NASA Ames Prognostics Data Repository (http://ti.arc.nasa.gov/project/prognostic-data-

repository, accessed on 20 August 2022), NASA Ames Research Center, Moffett Field, CA.
Appendix A
strategies referred to as KF-IMM and SIF-IMM. The following figure helps further illustrate
the overall architecture of the proposed strategies.
Figure A1. Overall architecture of IMM-KF and IMM-SIF strategies.
References
1. Ren, G.; Ma, G.; Cong, N. Review of Electrical Energy Storage System for Vehicular Applications. Renew. Sustain. Energy Rev.
2015, 41, 225–236. [CrossRef]
2. Thackeray, M.M.; Wolverton, C.; Isaacs, E.D. Electrical Energy Storage for Transportation—Approaching the Limits of, and Going
beyond, Lithium-Ion Batteries. Energy Environ. Sci. 2012, 5, 7854. [CrossRef]
3. Lu, L.; Han, X.; Li, J.; Hua, J.; Ouyang, M. A Review on the Key Issues for Lithium-Ion Battery Management in Electric Vehicles. J.
4. Zhang, S.; Sun, H.; Lyu, C. A Method of SOC Estimation for Power Li-Ion Batteries Based on Equivalent Circuit Model and
Extended Kalman Filter. In Proceedings of the 2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA),
Wuhan, China, 31 May–2 June 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 2683–2687.
5. Vidal, C.; Malysz, P.; Kollmeyer, P.; Emadi, A. Machine Learning Applied to Electrified Vehicle Battery State of Charge and State
of Health Estimation: State-of-the-Art. IEEE Access 2020, 8, 52796–52814. [CrossRef]
6. Meddings, N.; Heinrich, M.; Overney, F.; Lee, J.-S.; Ruiz, V.; Napolitano, E.; Seitz, S.; Hinds, G.; Raccichini, R.; Gaberšček, M.; et al.
Application of electrochemical impedance spectroscopy to commercial Li-ion cells: A review. J. Power Sources 2020, 480, 228742.
[CrossRef]
7. Zhang, R.; Li, X.; Sun, C.; Yang, S.; Tian, Y.; Tian, J. State of Charge and Temperature Joint Estimation Based on Ultrasonic
Reflection Waves for Lithium-Ion Battery Applications. Batteries 2023, 9, 335. [CrossRef]
8. Kandepu, R.; Foss, B.; Imsland, L. Applying the Unscented Kalman Filter for Nonlinear State Estimation. J. Process Control 2008,
18, 753–768. [CrossRef]
9. Partovibakhsh, M.; Liu, G. An Adaptive Unscented Kalman Filtering Approach for Online Estimation of Model Parameters and
State-of-Charge of Lithium-Ion Batteries for Autonomous Mobile Robots. IEEE Trans. Control Syst. Technol. 2015, 23, 357–363.
[CrossRef]
10. Yadkuri, F.F.; Khosrowjerdi, M.J. Methods for Improving the Linearization Problem of Extended Kalman Filter. J. Intell. Robot.
Syst. 2015, 78, 485–497. [CrossRef]
238
Energies 2024, 17, 536
11. Sanjurjo, E.; Naya, M.Á.; Blanco-Claraco, J.L.; Torres-Moreno, J.L.; Giménez-Fernández, A. Accuracy and Efficiency Comparison
of Various Nonlinear Kalman Filters Applied to Multibody Models. Nonlinear Dyn. 2017, 88, 1935–1951. [CrossRef]
12. Foletto, T.C.; Moreno, U.F.; Lopes, A.M. Comparison of Packet Loss Compensation in Nonlinear Networked Control Systems
Using the Methods Extended Kalman Filter and Unscented Kalman Filter. In Proceedings of the ETFA2011, Toulouse, France,
5–9 September 2011; IEEE: Piscataway, NJ, USA, 2011; pp. 1–7.
13. Yi, S.; Zorzi, M. Robust fixed-lag smoothing under model perturbations. J. Frankl. Inst. 2023, 360, 458–483. [CrossRef]
14. Wang, D.; Huang, H.; Tang, Z.; Zhang, Q.; Yang, B.; Zhang, B. A Lithium-Ion Battery Electrochemical–Thermal Model for a Wide
Temperature Range Applications. Electrochim. Acta 2020, 362, 137118. [CrossRef]
15. Tang, S.-X.; Camacho-Solorio, L.; Wang, Y.; Krstic, M. State-of-Charge Estimation from a Thermal–Electrochemical Model of
Lithium-Ion Batteries. Automatica 2017, 83, 206–219. [CrossRef]
16. Ghalkhani, M.; Bahiraei, F.; Nazri, G.-A.; Saif, M. Electrochemical–Thermal Model of Pouch-Type Lithium-Ion Batteries. Elec-
trochim. Acta 2017, 247, 569–587. [CrossRef]
17. Wang, L.; Niu, J.; Zhao, W.; Li, G.; Zhao, X. Study on Electrochemical and Thermal Characteristics of Lithium-Ion Battery Using
the Electrochemical-Thermal Coupled Model. Int. J. Energy Res. 2019, 43, 2086–2107. [CrossRef]
18. Barcellona, S.; Piegari, L. Lithium Ion Battery Models and Parameter Identification Techniques. Energies 2017, 10, 2007. [CrossRef]
19. Zhang, X.; Zhang, W.; Lei, G. A Review of Li-Ion Battery Equivalent Circuit Models. Trans. Electr. Electron. Mater. 2016,
17, 311–316. [CrossRef]
20. Davide, A. Battery Management Systems for Large Lithium-Ion Battery Packs; Artech: Morristown, NJ, USA, 2010.
21. Marcicki, J.; Onori, S.; Rizzoni, G. Nonlinear Fault Detection and Isolation for a Lithium-Ion Battery Management System. In
Proceedings of the ASME 2010 Dynamic Systems and Control Conference, Cambridge, MA, USA, 1 January 2010; ASMEDC: New
York, NY, USA, 2010; Volume 1, pp. 607–614.
22. Meng, J.; Yue, M.; Diallo, D. A Degradation Empirical-Model-Free Battery End-Of-Life Prediction Framework Based on Gaussian
Process Regression and Kalman Filter. IEEE Trans. Transp. Electrif. 2022, 9, 4898–4908. [CrossRef]
23. Xu, J.; Sun, C.; Ni, Y.; Lyu, C.; Wu, C.; Zhang, H.; Yang, Q.; Feng, F. Fast Identification of Micro-Health Parameters for Retired
Batteries Based on a Simplified P2D Model by Using Padé Approximation. Batteries 2023, 9, 64. [CrossRef]
24. Singh, A.; Izadian, A.; Anwar, S. Nonlinear Model Based Fault Detection of Lithium Ion Battery Using Multiple Model Adaptive
Estimation. IFAC Proc. Vol. 2014, 47, 8546–8551. [CrossRef]
25. Xia, X.; Wei, Y. Lithium-Ion Batteries State-of-Charge Estimation Based on Interactive Multiple-Model Extended Kalman Filter. In
Proceedings of the 2016 22nd International Conference on Automation and Computing (ICAC), Colchester, UK, 7–8 September
2016; IEEE: Piscataway, NJ, USA, 2016; pp. 204–207.
26. Singh, A.; Izadian, A.; Anwar, S. Fault Diagnosis of Li-Ion Batteries Using Multiple-Model Adaptive Estimation. In Proceedings
of the IECON 2013—39th Annual Conference of the IEEE Industrial Electronics Society, Vienna, Austria, 10–13 November 2013;
IEEE: Piscataway, NJ, USA, 2013; pp. 3524–3529.
27. Bustos, R. State of Health Estimation of Lithium-Ion Batteries Using Dual Filters and the IMM Strategy. Ph.D. Thesis, University
of Guelph, Guelph, ON, Canada, 2022.
28. Hasib, S.A.; Islam, S.; Chakrabortty, R.K.; Ryan, M.J.; Saha, D.K.; Ahamed, M.H.; Moyeen, S.I.; Das, S.K.; Ali, M.F.; Islam, M.R.; et al. A
Comprehensive Review of Available Battery Datasets, RUL Prediction Approaches, and Advanced Battery Management. IEEE
Access 2021, 9, 86166–86193. [CrossRef]
29. Tian, N.; Wang, Y.; Chen, J.; Fang, H. On Parameter Identification of an Equivalent Circuit Model for Lithium-Ion Batteries. In
Proceedings of the 2017 IEEE Conference on Control Technology and Applications (CCTA), Mauna Lani Resort, HI, USA, 27–30
August 2017; IEEE: Piscataway, NJ, USA, 2017; pp. 187–192.
30. Kalman, R.E. A New Approach to Linear Filtering and Prediction Problems. J. Basic Eng. Mar 1960, 82, 35–45. [CrossRef]
31. Gadsden, S.A.; Habibi, S.; Kirubarajan, T. Kalman and Smooth Variable Structure Filters for Robust Estimation. IEEE Trans. Aerosp.
Electron. Syst. 2014, 50, 1038–1050. [CrossRef]
32. Gadsden, S.A.; Al-Shabi, M. The Sliding Innovation Filter. IEEE Access 2020, 8, 96129–96138. [CrossRef]
33. Bustos, R.; Gadsden, S.A.; Al-Shabi, M.; Mahmud, S. Lithium-Ion Battery Health Estimation Using an Adaptive Dual Interacting
Model Algorithm for Electric Vehicles. Appl. Sci. 2023, 13, 1132. [CrossRef]
34. Wei, X.; Yimin, M.; Feng, Z. Lithium-Ion Battery Modeling and State of Charge Estimation. Integr. Ferroelectr. 2019, 200, 59–72.
[CrossRef]
239
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
www.mdpi.com
Energies Editorial Office

E-mail: energies@mdpi.com
www.mdpi.com/journal/energies
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are
solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s).
MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from
any ideas, methods, instructions or products referred to in the content.
Academic Open
Access Publishing
mdpi.com ISBN 978-3-7258-0606-5

Battery Modelling

Uploaded by

Copyright:

Available Formats

Battery Modelling

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Battery Modelling

Uploaded by

Copyright:

Available Formats

Special Issue Reprint

Basel • Beijing • Wuhan • Barcelona • Belgrade • Novi Sad • Cluj • Manchester

ISBN 978-3-7258-0605-8 (Hbk)

About the Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Emanuele Fedele, Luigi Pio Di Noia and Renato Rizzo

Iacopo Marri, Emil Petkovski, Loredana Cristaldi and Marco Faifer

Simone Barcellona, Lorenzo Codecasa, Silvia Colnago and Luigi Piegari

Javier Rey and Lázaro V. Cremades

Tadeusz Białoń, Roman Niestrój, Wojciech Skarka and Wojciech Korski

Emil Petkovski, Iacopo Marri, Loredana Cristaldi and Marco Faifer

Richard Bustos, Stephen Andrew Gadsden, Mohammad Biglarbegian, Mohammad AlShabi

Citation: Fedele, E.; Di Noia, L.P.;

Energies 2023, 16, 2963. https://doi.org/10.3390/en16072963 https://www.mdpi.com/journal/energies

2. Battery Electric Vehicles Cabling System

GRID AC/DC + DC/DC

MOTOR DC/AC DC/DC BATTERY

For high-voltage EV applications, single-core multi-stranded high-temperature ca-

3. Modeling of Cable Losses

The geometry of a multi-stranded DC cable is schematized in Figure 2. Each strand is

Figure 2. Multi-stranded cable geometry.

the AC to DC resistance ratio k R is equal to

3.2. Generic Current Waveform

4. Validation with Magnetic FEM Solver

Table 1. Cable parameters.

The AC equivalent resistance of the cable is evaluated in the post-processing phase

Figure 4. Comparison between analytic and FEM-calculated AC resistances at different frequencies

5. Application to Thermal Transient Analysis

where Δi is the peak-to-peak current ripple and d is the duty ratio.

Definition of cable geometry

P(k) Calculation of Jh(k)

Energies 2023, 16, 3855. https://doi.org/10.3390/en16093855 https://www.mdpi.com/journal/energies

1.1. Literature Review

Figure 1. The trade-off between computational cost and model accuracy.

Conversely, the data-driven approach, especially machine-learning-based techniques,

1.1.2. Cloud-Based Progress on Battery State Prediction

1.2. Major Challenges Involved

Figure 2. Synergy between onboard BMS and cloud BMS.

(1) Although data-driven machine learning techniques introduce considerable savings in

1.3. Contributions of the Work

(b) A specialized attention-based transformer neural network model is designed to learn

2. Key Components of Cloud BMS

Figure 3. Cloud-based framework for commercial EV applications.

2.1. Physical Entity

3.1. Transformer Neural Networks for Co-Estimation

Figure 5. Architecture of the deep learning for co-estimation of battery states.

3.1.1. Data Processing

The Transformer is an encoder-decoder structured sequence-to-sequence model, de-

Figure 6. Sliding window mechanism of the transformer model.

3.1.3. Dual Encoder

Multi-head_Attention(qh, kh, vh ) = Concat(head1 , headh )ωo (4)

3.2. Data Generation

Table 1. Cell chemistry and operating windows.

3.3. Evaluation Criteria

MAE can be given by:

4. Performance of Cloud-Based BMS

Subsequently, the self-attention transformer model is calibrated using another Cell_2,

4.2. SOH Estimation Results

Table 2. Hyperparameters of the transformer model.

Figures 9 and 10 illustrates the performance of the transformer model in estimating