Comparing Four Model-Order Reduction Techniques, Applied To

eTransportation 1 (2019) 100009
Contents lists available at ScienceDirect
eTransportation
journal homepage: www.journals.elsevier.com/etransportation
Comparing four model-order reduction techniques, applied to

lithium-ion battery-cell internal electrochemical transfer functions
Albert Rodríguez, Gregory L. Plett*, M. Scott Trimboli
Department of Electrical and Computer Engineering, University of Colorado Colorado Springs, Colorado Springs, CO, 80918, United States
a r t i c l e i n f o a b s t r a c t
Article history: Physics-based models of lithium-ion battery dynamics are developed from fundamental electrochemical
Received 4 May 2019 principles and describe cell internal electrochemical variables in addition to terminal voltage. Real-time
Received in revised form estimates of the values taken on by internal cell variables provided by such models might be leveraged by
29 July 2019
future battery-management systems to control fast-charging and routine use of a battery pack to
Accepted 31 July 2019
maximize performance but minimize aging. These models are most naturally described as sets of coupled
Available online 6 August 2019
partial-differential equations (PDEs), and so the greatest obstacle to their adoption stems from the
computational complexity involved in finding solutions to the model equations. To make a feasible
Keywords:
Realization algorithm
physics-based model for battery management, we must construct reduced-order approximations to
Model-order reduction these PDE models. In this paper, we present four methods to find high-fidelity discrete-time state-space
Electrochemical variable reduced-order models (ROMs) that approximate infinite-order transcendental transfer functions that
Transfer function model the PDE relationships of all electrochemical variables of interest. These four methods are
Lithium-ion cell model compared for a single cell based on speed, memory usage, robustness, and accuracy of the predictions of
Reduced-order model the resulting reduced-order models with respect to precise numerical simulations of the PDEs. We find
that all four methods produce ROMs that match the linearized PDEs closely in the frequency domain and
that yield time-domain simulations that match those from the nonlinear PDEs as well, but that each xRA
method has distinct features so that different applications might prefer one method versus another.
© 2019 Elsevier B.V. All rights reserved.
1. Introduction so forth) [1]. The best methods of which we are aware do so using
models (sets of mathematical equations) that describe the dy-
The commercial feasibility of electrified transportation has a namics of the cells in the battery pack. These models may be
significant dependence on the practicality of the battery pack used empirical (e.g., equivalent-circuit models, ECMs), or physics-based
in the vehicle. Since the battery pack is often the single-most [2e6]. Empirical models describe only the voltage response of a
expensive component in the vehicle, it is essential to be able to battery cell to an input-current stimulus; physics-based models
use its capabilities to the maximum extent possible while also (PBMs) furthermore have the ability to describe accurately all the
maintaining as long a service life as possible. (These objectives are internal electrochemical dynamics that occur when lithium-ion
generally in opposition since aggressive usage tends to introduce cells are operated under different conditions. Therefore, physics-
physical and electrochemical stress factors that accelerate battery based models are required by any BMS that aims to predict the
aging.) A battery-management system (BMS) continuously moni- precursors to degradation, which depend directly on the internal
tors this critical vehicle component to maintain the safety of the electrochemical dynamics and not directly on cell terminal voltage.
driver, maintain the safety of the battery pack as a whole, and While present BMS rely almost exclusively on ECMs, we believe
jointly maximize battery performance and battery service life. that future BMS will instead require PBMs to be able jointly to
A BMS must be able to compute estimates of battery-cell state- maximize battery performance and service life through improved
of-charge, state-of-health, state-of-energy, state-of-function (and power and energy management, or advanced balancing systems,
controlling the internal electrochemical state away from conditions
that would tend to accelerate degradation. For example, in some
* Corresponding author. cases it has been shown that PBM-based BMS can safely authorize
E-mail addresses: arodrig7@uccs.edu (A. Rodríguez), gplett@uccs.edu 22% more power than would be authorized by ECM-based BMS,
(G.L. Plett), mtrimbol@uccs.edu (M.S. Trimboli).
https://doi.org/10.1016/j.etran.2019.100009
2590-1168/© 2019 Elsevier B.V. All rights reserved.
2 A. Rodríguez et al. / eTransportation 1 (2019) 100009
provide more than 212% more available energy, and that PBM- B2Rnm , C2Rqn , and D2Rqm . In this work, u½k corresponds to
based BMS can extend battery-pack life by more than 40% [7,8]. the electrical current applied to the cell (i.e., u½k ¼ iapp ½k and m ¼
The widely accepted PBM of a lithium-ion celldknown as the 1) and y½k corresponds to the set of electrochemical variables of
DoyleeFullereNewman (DFN) modeldconsists of four coupled interest. The state x½k will be selected by the model-order reduc-
partial differential equations (PDEs) and one algebraic equation tion techniques that we will describe.
[9,10]. The PDEs that govern the model arise from the fundamental This paper assumes that a set of continuous-time transfer
physical principles of charge and mass conservation in the solid and functions GðsÞ2Cq1 has already been selected for implementa-
electrolyte materials of the cell; the algebraic equation is a closure tion. Continuous-time models, however, are not directly useful to a
term that models the lithium flux from the solid electrode materials BMS since the BMS samples voltage and current at a periodic rate,
to the electrolyte. and therefore discrete-time models are of more direct use. So, we
The greatest obstacle to using PBMs in a BMS stems from their must use knowledge of GðsÞ in some way to produce the discrete-
computational complexity. The DFN model is too complex to be time model.
used directly in an embedded battery-management system, since In this paper, we present and compare four different “realization
doing so would require solving the coupled PDEs in real time. To algorithms” (generically referred to as “xRAs”) that convert GðsÞ to
make a feasible PBM for battery management, various methods are the discrete-time form of Eqs. (1) and (2):
proposed in the literature to construct approximate reduced-order
models (ROMs) of lithium-ion battery cell dynamics. These include The discrete-time realization algorithm (DRA) converts GðsÞ to
single-particle models (e.g., [11,12]) which tend not to perform well discrete-time unit-pulse response g½k and then uses either the
at high rates [13,14], and parabolic approximations of solid con- HoeKalman or eigensystem realization algorithm subspace
centrations to reduce complexity of PDE solutions (e.g., [15,16]). system-identification methods to produce a reduced-order
Other approaches such as Pade approximations and residue state-space model.
grouping have also been considered [17e19].1 The continuous-time realization algorithm (CRA) converts GðsÞ
We prefer model-order reduction methods based on finding to a reduced-order continuous-time state-space model using a
transfer functions of all electrochemical variables of interest and subspace system-identification method, and then converts the
then converting these transfer functions into low-order high-fi- continuous-time state-space model to a discrete-time state-
delity discrete-time state-space approximate models. An early space model using standard methods.
work by Smith et al. presented transfer functions for solid surface The hybrid realization algorithm (HRA) converts GðsÞ to
concentration, interfacial potential difference, and reaction flux discrete-time frequency response GðejuT Þ and then uses a sub-
[20]. Later, Lee et al. added transfer functions for solid and elec- space system-identification method to reduce the model's order.
trolyte potential, separately, and for electrolyte concentration [21]. The Lagrange-interpolated realization algorithm (LRA) first uses
Rodríguez et al. improved the computational complexity of the interpolation polynomials to create a high- but finite-order
electrolyte-concentration transfer function [22], and extended the continuous-time state-space model to approximate GðsÞ, and
transfer function models to model electrodes having multiple then uses a standard balanced model-order reduction technique
blended active materials [23]. Most recently, Rodríguez et al. to arrive at a low-order continuous-time state-space model,
improved all of these models by showing how to remove a limiting which is then converted to discrete time using standard
assumption in their derivation [24]. techniques.
A hinderance to using these transfer-function models in a BMS is
that they cannot be expressed as ratios of polynomials in the Lap- In the sequel, Sect. 2 first describes a procedure to convert a
lace variable “s”, as is most common for transfer functions continuous-time frequency response to a discrete-time frequency
considered in control-systems literature. Instead, they are infinite- response (directly needed by the HRA and used to validate all the
order transcendental functions of the Laplace variable. Standard xRA methods). Section 3 then shows how the four methods
controls methods are unable to work with them directly. Thus, for compute different reduced-order state-space A matrices; Sect. 4
practical use, these transcendental models need to be approxi- gives the assumed form of the B matrix; Sect. 5 shows how to solve
mated using low-order rational-polynomial function models that for the C matrix while constraining the dc gain of the ROM to agree
capture their most significant dynamics. with GðsÞ; and, Sect. 6 discusses the solution for the D matrix. Then,
The purpose of this paper is to propose and compare four Sect. 7 presents some simulation results to benchmark the methods
different techniques whose objective is to generate reduced-order against each other, and we conclude the paper with some summary
models from the infinite-order electrochemical transfer functions. comments.
These ROMs can then be used in advanced BMS controls algorithms. The details of the realization-algorithm derivations presented in
The different techniques presented here will provide reduced- Sects. 2 through 6 are important to be able to comprehend and
order discrete-time state-space models of the form implement the algorithms. However, on an initial reading of this
paper it may be beneficial first to skip the derivations and refer to
x½k þ 1 ¼ Ax½k þ Bu½k (1) the summary figure (Fig. 1) in combination with the results pre-
sented in Sect. 7 to form an initial understanding of how the al-
gorithms are used before returning to study the derivations of the
y½k ¼ Cx½k þ Du½k; (2)
methods.
where the state vector of the system is represented by x½k2 Rn1,
2. Discrete-time frequency response from a continuous-time
the input vector by u½k2Rm1 and the output vector by y½k2 Rq1.
frequency response
Note that n corresponds to the order of the model, m to the number
of inputs and q to the number of outputs. Furthermore, A2 Rnn ,
2.1. Deriving the method
1
The comparison between the various methods proposed in the literature to
The ultimate goal of this paper is to develop and compare
construct reduced-order approximations are out of the scope of this manuscript. methods that approximate infinite-order continuous-time transfer
Instead, our focus is on ways to produce ROMs directly from transfer functions. functions with reduced-order discrete-time models. The
A. Rodríguez et al. / eTransportation 1 (2019) 100009 3
Fig. 1. Realization algorithms summary flowchart.
continuous-time transfer functions may be converted readily to

continuous-time frequency responses via the transformation
GðjuÞ ¼ GðsÞjs¼ju ; however, it is less clear how to convert these
continuous-time frequency responses into a discrete-time fre- Fig. 2. Sampled-data system.
quency response as is needed by some of the realization methods
presented herein. If the transfer function happens to be of a
rational-polynomial form in the Laplace variable “s”, then well- period Ts . Then, we define an equivalent Laplace-domain repre-
known methods may be used (e.g., [25]). Otherwise, we need a sentation of the sampled signal
new approach. One such method is proposed in this section.
Fig. 2 illustrates the components of a sampled-data system that X∞
U ðsÞ ¼ L uðtÞdTs ðtÞ ¼ u½keskTs ;
are relevant to our discussion. A digital computer produces a k¼0
discrete-time signal u½k, which we might assume originated as
continuous-time uðtÞ, sampled every Ts seconds.2 The sampled where L f,g denotes the Laplace transform of its argument. With
signal u½k is then passed through a digital-to-analog converter this notation, we can write YðsÞ ¼ GðsÞU ðsÞ and can also find that

incorporating a zero-order hold (ZOH) circuit to produce a Y ðsÞ ¼ G ðsÞU ðsÞ.
continuous-time signal uðtÞ. The uðtÞ signal is then input to a The starred Laplace transform can be related to the standard
continuous-time system having transfer function GðsÞ, which pro- Laplace transform as [26].
duces the output signal yðtÞ. The output is then sampled, with

sampling period Ts , to produce a discrete-time signal y½k. 1 X∞
2p gð0Þ
Leveraging knowledge that the ZOH transfer function is ð1 G ðsÞ ¼ G sþj k þ : (5)
Ts Ts 2
expð sTs ÞÞ =s, we combine the continuous-time elements as k¼∞
The summand for k ¼ 0 corresponds to the original frequency

1 expðsTs Þ spectrum, which is copied into G ðsÞ. The summands for ks0
GðsÞ ¼ GðsÞ; (3) comprise aliases of the original frequency spectrum that are
s
introduced by the sampling operation. While the summation limits

which gives an overall continuous-time frequency response theoretically go from ∞ to ∞, quite good approximations of G ðsÞ
can be achieved using much fewer terms since we generally do not
1 expðjuTs Þ expect much aliasing to occur if we have sampled rapidly enough.
GðjuÞ ¼ GðjuÞ: (4) The value for gð0Þ can be found via the initial-value theorem as
ju
gð0Þ ¼ lims/∞ GðsÞ ¼ limu/∞ GðjuÞ. (This is the value of the “D”
The sampled-data system in Fig. 2 comprises both discrete-time term for both the continuous-time and discrete-time state-space
and continuous-time elements. It is customary to use the z-trans- models.) Note that it can be subtracted from GðjuÞ before con-
form or discrete-time Fourier transforms to analyze discrete-time verting from continuous-time to discrete-time, and then added
systems and to use the Laplace transform or continuous-time back in later. Therefore, we can assume gð0Þ ¼ 0 in Eq. (5) without
Fourier transforms to analyze continuous-time systems. To loss of generality.
analyze hybrid sampled-data systems using a common framework, The starred Laplace
transform can also be related to the z-
we can use the “starred Laplace transform” [26]. To do so, suppose
transform as YðzÞ ¼ Y ðsÞjs¼lnðzÞ=Ts [26]. Therefore, we have
that u½k is the result of sampling a continuous-time signal uðtÞ with

sample period Ts . A continuous-time representation of the sampled 1 X ∞
lnðzÞ 2p
signal is uðtÞdTs ðtÞ; where dTs ðtÞ is the Dirac comb function having YðzÞ ¼ G þj k UðzÞ:
Ts Ts Ts
k¼∞
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
GðzÞ
2
In this notation, square brackets are used to denote a discrete-time signal and k
is the (integer) discrete-time sampling index; parenthesis are used to denote a The discrete-time frequency response that we desire to compute
continuous-time signal. Therefore, u½k ¼ uðkTs Þ. The sampling period itself, Ts , is is GðejuTs Þ ¼ GðzÞjz¼expðjuTs Þ . Therefore,
largely determined by the application. According to the Nyquist sampling theorem,

the sampling frequency must be at least twice the highest frequency being
1 X∞
2p
modeled. Lithium-ion battery cells have quite slow dynamics, so sampling fre- G ejuTs ¼ G ju þ j k : (6)
quencies on the order of 1 Hz are reasonable. If the BMS requires sampling at a Ts Ts
k¼∞
higher rate (e.g., to capture the impact of inverter noise on cell response), the
higher sampling is reflected directly in smaller Ts. To be clear, GðejuTs Þ on the left-hand side of Eq. (6) is the
discrete-time frequency response of the sampled-data system that Fig. 3 compares the magnitude and phase responses of the
we have set out to approximate, and GðjuÞ on the right-hand side of proposed method and the exact result in Eq. (8). We see that a
Eq. (6) is the known continuous-time frequency response as visually perfect match has been achieved between actual and
calculated by Eq. (4). This result is valid for frequencies up to the estimated discrete-time system frequency responses, as hoped,
Nyquist rate: p=Ts < u < p=Ts . showing that the method described in this section produces accu-
rate results. We also note that the discrete-time and continuous-
2.2. Summarizing the frequency-response conversion method time frequency responses for this system are different from each
other, as expected.
Eq. (6) is the key result of this section. It shows how to convert a
general continuous-time frequency response into an equivalent 3. Solving for the state-space A matrix
discrete-time frequency response.3 It will be used as part of the
validation process for all four realization algorithms presented in 3.1. Modeling integration dynamics
this paper, and also plays an integral part in the derivation of the
HRA method itself. To summarize the method: With this background, we are ready to investigate the four
different realization algorithms. These algorithms differ primarily
1. Compute GðsÞ from GðsÞ using Eq. (3). in how they determine the “A” matrix of the final reduced-order
2. Compute D ¼ gð0Þ ¼ lims/∞ GðsÞ and replace GðsÞ : ¼ GðsÞ D. discrete-time state-space model. In all cases, however, this matrix
3. Compute the continuous-time frequency response GðjuÞ using is forced to be real (because PBM dynamics are nonoscillatory) and
Eq. (4). diagonal, without loss of generality. For stability, the diagonal en-
4. Compute the discrete-time approximate frequency response tries (termed the “poles” of the model) must have magnitude
b juTs Þ at N output frequency points using Eq. (6). As we have
Gðe strictly less than 1. However, some PBM transfer functions (e.g.,
noted, only a few terms from the summationdcentered around those involving changes in cell or electrode state-of-charge) have
k ¼ 0dare required in practice, assuming that the sampling rate integration dynamics, which must be implemented using a pole
is chosen to be fast enough to avoid a great deal of aliasing. That having value exactly equal to 1. Fortunately, it is quite simple to deal
is, for some kmax , we compute with this special case.
We begin by computing res ¼ lims/0 sGðsÞ, which is the res-
1 X
kmax
idue of the integrator pole in the original continuous-time transfer
b ejuTs
G ¼ Gðju þ j2pk=Ts Þ: (7)
Ts function. We then remove the integrator from the transfer function
k¼kmax
by replacing GðsÞ : ¼ GðsÞ res =s. We then invoke the DRA, CRA,
HRA, or LRA to find a discrete-time state-space model of order n for
5. Finally, add back the high-frequency gain by replacing integrator-removed GðsÞ, where
Gðe b juTs Þ þ D.
b juTs Þ : ¼ Gðe 2 3
6 a1 0 / 0 7
6 0 a2 / 0 7
A¼6
6 «
7; (9)
2.3. Illustrating the frequency-response conversion method 4 « 1 « 7
5
0 0 / an
This proposed method is able to convert continuous-time fre-
quency responses to discrete-time frequency responses for arbi- and jai j < 1 c i2½1…n for a stable model, ai 2R c i2½1…n for a
trary systems, unlike conventional methods that can be used only nonoscillatory system and a1 < a2 < … < an . Finally, we augment the
with finite-order rational-polynomial or state-space models [25]. identified model with the integration dynamics that were removed
However, we choose at this point to demonstrate the method on a previously. That is, we make the following replacements:
finite-order rational-polynomial transfer function, since that makes 2 3 2 3
it possible to compare the solution we obtain to one obtained from
A 05 B
the standard methods. In particular, we choose A : ¼4 ; B :¼ 4 5; C :¼ C res ;
0 1 1
0:1s2 þ s þ 1 1 0:1u2 þ ju
GðsÞ ¼ ; so GðjuÞ ¼ ;
s2 þ 3s þ 1 1 u2 þ 3ju and D is unchanged. The final augmented model is of order n þ 1.
Note that this first step of removing integration dynamics is com-
and sampling period Ts ¼ 0:1 s. For this simple case, we can mon to all of the xRAs we discuss next, so we will not mention it
compute the exact equivalent discrete-time system as specifically in any derivation. Also, the net effect of this step is that
the input transfer functions to all of the xRA methods are strictly
0:1ej0:2u 0:1088ej0:1u þ 0:0174 stable.
G ejuTs ¼ (8)
ej0:2u 1:732ej0:1u þ 0:7408
3.2. Discrete-time realization algorithm (DRA)
using MATLAB's c2d.m function, substituting z ¼ expðjuTs ). The
value of kmax ¼ 5 was chosen for Eq. (7), and the discrete-time
3.2.1. Deriving the DRA steps
frequency vector for building GðsÞ was chosen to comprise one
The next subsections describe four different techniques to
point at u ¼ 0 and 50 additional points spaced evenly on a loga-
obtain the reduced-order discrete-time A matrix. We first describe
rithmic scale between 105 and 100 times the Nyquist rate.
the DRA. The derivation carried out here is a simplified version of
one that appears in [2,27]; we also improve the method in this
3
paper to enforce correct dc gain.
While Jury [26] seems to hint at this approach for converting general
continuous-time to discrete-time frequency responses in his paper using the in-
The DRA first approximates discrete-time unit-pulse response
termediate starred Laplace transform, he does not develop the method. As far as we g½k from knowledge of GðsÞ; then, it uses a subspace system-
are aware, we have presented the first full derivation here. identification method to convert g½k to a state-space form. The
Fig. 3. Actual and estimated discrete-time magnitude and phase response.
unit-pulse response has discrete-time transfer function GðzÞ, which emulation sampling period, T1
we can write in terms of continuous-time transfer function GðsÞ by
assuming that the input to GðsÞ is piecewise constant with period Ts
as illustrated in Fig. 2. Hence, X
1 N1
gd ½k ¼ Gd ½f expðj2pkf =NÞ: (13)
N
f ¼0
GðsÞ
GðzÞ ¼ Z GðsÞ ¼ Z ð1 expð sTs ÞÞ
s The cumulative summation of this impulse response, g d ½n,

GðsÞ GðsÞ yields the approximation to the continuous-time step response
¼Z z1 Z ; (10)
s s
where the very compact notation Z ½ , means: “find the X

k1
gd;s ½k ¼ T1 gd ½i: (14)
continuous-time time-domain function corresponding to the
i¼0
Laplace-transform frequency-domain argument, then sample the
continuous-time time-domain function, then take the z-transform This result is interpolated with sample period Ts to give gs ½k.
of the samples”. Recalling that GðsÞ =s represents a step response, The difference between gs ½k and gs ½k 1 in Eq. (11) gives the
the unit-pulse response g½k corresponding to GðzÞ is a one- discrete-time unit-pulse response g½k.
timestep difference of the sampled continuous-time step Given the unit-pulse response g½k we can find discrete-time A,
response. That is, B, and C state-space matrices using the HoeKalman algorithm [30].
First, recall that the unit-pulse response can be written in terms of
g½k ¼ gs ½k gs ½k 1; (11) the discrete-time state-space matrices as
where gs ½k is the discrete-time sampled step response of the sys- 2 3 8

tem with Gs ðzÞ ¼ Z ½GðsÞ=s. To find gs ½k, we recall that the step <
D; k¼0
response of the system can be computed by integrating the impulse g4k5 ¼
: CAk1 B; k > 0:
response. We approximate the continuous-time impulse response
via a “discrete equivalent” or frequency-domain emulation Given the unit-pulse response values, we can form the block-
approach using a bilinear transformation [28]. This allows us to Hankel matrix
write discrete equivalent Gd ðzÞ as
2 3
Gd ðzÞzGðsÞjs¼ 2 z1 ;
T1 zþ1
6 g½1 g½2 / g½j 7
6 g½2 g½3 / g½j þ 1 7
where T1 is an emulation sampling period (different from and often H i;j ¼6
6 «
7;
7
4 « 1 « 5
significantly shorter than the final system sampling period Ts ). In g½i g½i þ 1 / g½i þ j 1
order to arrive at an accurate estimation of the continuous-time
transfer function, the sampling frequency f1 ¼ 1=T1 must be high
enough to capture the system dynamics. where the required length of the discrete-time unit-pulse response
We now recognize that the discrete Fourier transform (DFT) of a is N ¼ i þ j. Further, this approach requires H i;j to be of larger size
sequence is related to its z-transform via [29]. than we expect for a hypothesized system order, that is, i > n and
j > n. We usually choose i ¼ j, so N 2n.4
Gd ½f ¼ Gd ðzÞjz¼expðj2pf =NÞ An interesting feature of the block Hankel matrix, which is
critical for this development, is that it can be written as H i;j ¼
2 expðj2pf =NÞ 1 (12)
¼G ; 0f N O i C j , where the extended observability matrix, O i , and the
T1 expðj2pf =NÞ þ 1 extended controllability matrices, C j , are defined as
where N is the number of points chosen for the underlying

sequence, and is usually chosen to be a power of 2 for efficient 4
For transcendental transfer functions having an infinite number of poles, we
computation. The inverse DFT of Gd ½f gives gd ½k, which is the cannot choose N to be twice the number of actual poles. Hence, N is chosen such
approximation of the continuous-time impulse response at the that the approximate unit-pulse response used as input to the HoeKalman algo-
rithm settles to near its steady-state value.
" #T response data. Hence, we can delete parts of the long unit-pulse-
T
T T response that do not provide “much” information about the sys-
T 2 i1
O i ¼ C ðCAÞ CA ,,, CA (15) tem dynamics when building the Hankel matrix, reducing its size.
We begin by forming an r s generalized Hankel matrix
2 3
Cj¼ B AB A2 B / Aj1 B : (16) g½k g½k þ t1 / g½k þ ts1
6 7
6 g½j1 þ k g½j1 þ k þ t1 / g½j1 þ k þ ts1 7
6
H r;s ½k 1 ¼ 6 7
If we are able to factor H i;j into O i and C j , then we find the A « « 1 « 7
4 5
matrix after doing some processing. g½jr1 þ k g½jr1 þ k þ t1 / g½jr1 þ k þ ts1
To do the factoring, we rely on the singular value decomposition
(SVD). That is,
where ji ði ¼ 1; …; r 1Þ and ti ði ¼ 1; …; s 1Þ are arbitrary in-
2 32 3
tegers. We now define the generalized controllability and observ-
S 0 54 V T1 5
H i;j ¼ USV T ¼ U1 U2 4 1 ability matrices as
0 S2 V T2
" #T
T T
T j1 jr 1
where S contains the singular values of H i;j , which are ordered O r ¼ C ; CA ; ,,,; CA
from largest to smallest. S1 2Rnn contains the n largest (i.e.,
dominant) singular values and S2 is approximately zero. Hence, we

retain only H i;j zU1 S1 V T1 and compute
Cs¼ B; At1 B; /; Ats 1 B ;
1=2 1=2
O i ¼ U1 S1 and C j ¼ S1 V T1 : (17)
where we note that H r;s ½0 ¼ O r C s and H r;s ½1 ¼ O r AC s . We then
At this point the order of the reduced-order model, n, has been take the SVD of H r;s ½0. That is,
determined from the singular values. Then, the A matrix can be
found by noting the shift structure of the extended observability 2 32 3

matrix as, S 0 54 V T1 5
H r;s ½0 ¼ U1 U2 4 1 ;
0 S2 V T2
y [
A¼ OY
i
Oi; (18)
where S1 2Rnn and we think of S2 being essentially zero. Hence,
where y represents the matrix pseudo inverse, O [
i
is equal to O i 1=2
we retain only H r;s ½0zU1 S1 V T1 and compute O r ¼ U1 S1 and
with the top block row deleted, and O Y
i
is equal to O i with the 1=2
C s ¼ S1 V T1 : Then, we set
bottom block row deleted. That is,
" #T A ¼ O yr H y
r;s ½1C s : (19)
T T
Y T T 2 i2
O i
¼ C ðCAÞ CA ,,, CA ; Again, we use similarity transformations to diagonalize the
discrete-time reduced-order state-space model A matrix, we
replace any unstable poles in A by their reciprocals and complex
" #T
T T T poles by their magnitudes.
[
O i
¼ ðCAÞT CA2 CA3 ,,, CAi1 :
Finally, we use similarity transformations to diagonalize the 3.2.2. Summarizing the DRA
discrete-time reduced-order state-space model A matrix. Further, To summarize the steps of the DRA:
for our application, we expect A to have real, stable poles, which is
not guaranteed by the method as presented so far. So, we elect to 1. Select an emulation sampling period T1 such that 1 =T1 is
replace unstable poles in A by their reciprocals (maintaining the significantly greater than the bandwidth of GðsÞ, and the length
magnitude response) and complex poles by their magnitudes of the unit-pulse response N is at least twice as large as the
(largely maintaining the magnitude response).5 assumed reduced-order model dimension n. Compute Gd ½f us-
Improving memory efficiency. The algorithm explained above ing Eq. (12).
requires very long unit-pulse-response data in order to retain a 2. Compute the approximate continuous-time impulse response
wide range of frequency information, which is not memory effi- via Eq. (13), which is the inverse FFT of Gd ½f .
cient. Several steps from the eigensystem realization algorithm 3. Find the approximate continuous-time step response of GðsÞ via
(ERA) in [32] can replace the HoeKalman steps; this is of interest to Eq. (14) and interpolate at the final desired sample rate Ts to give
us because it allows us to work with incomplete unit-pulse- gs ½k.
4. Compute the unit-pulse response g½k from gs ½k via Eq. (11).
5
5. Populate the Hankel matrix H i;j (or H r;s ½k 1) with (part of)
This approach to guaranteeing stability and nonoscillatory responses is ad hoc
the discrete-time unit-pulse response values g½k and perform
but effective. A more rigorous alternative to guaranteeing stability and non-
oscillatory response is to enforce eigenvalue constraints on the subspace solution the SVD. Then, determine from the singular values the system
using linear matrix inequality (LMI) methods, such as presented in [31]. Our order n.
experience when applying the xRA methods in this paper to lithium-ion battery- 6. Partition the SVD to give the matrices U1 , S1 , and V T1 . From these,
cell transfer functions is that the ad hoc approach gives results of equivalent quality compute O i and C j (or O r and C s ) via Eq. (17).
to the more rigorous LMI approach. Both methods guarantee stability and non-
oscillatory response. Both methods result in ROMs having similar matches to the
7. Compute ROM matrix A via Eq. (18) (or Eq. (19)).
frequency-responses and time-response behaviors of the PDE models. Thus, we 8. Diagonalize A and replace any unstable poles by their reciprocals
present only the simpler method here. and any complex poles by their magnitudes.
3.3. Continuous-time realization algorithm (CRA) u_ ss ðtÞ ¼ ðjuÞuss ðtÞ; furthermore, the state equation gives x_ss ðtÞ ¼
Ac xss ðtÞ þ Bc uss ðtÞ. For multiple differentiations of the output, we
3.3.1. Deriving the CRA steps can then write
The continuous-time realization algorithm (CRA) was developed
by DePalma et al. [33,34], based on a subspace system-
identification approach presented in [35].6 The first goal of the ðjuÞi1 GðjuÞ ¼ CAi1
c XðjuÞ þ CAc Bc Im þ ðjuÞCAc Bc Im þ /
i2 i3
CRA is to derive a reduced-order continuous-time state-space þ ðjuÞi1 DIm :

realization of the form
Collecting multiple frequency responses and differentiated fre-
_ ¼ Ac xðtÞ þ Bc uðtÞ
xðtÞ quency responses as separate block rowsin one complex data ma-
yðtÞ ¼ CxðtÞ þ DuðtÞ; trix, we can write
where Ac 2Rnn , Bc 2Rnm , C2Rqn and D2Rqm . The CRA then 2 3

converts this continuous-time state-space realization to a discrete-
6 Gðju1 Þ Gðju2 Þ / GðjuN Þ 7
time realization using known techniques.7 6 ðju1 ÞGðju1 Þ ðju2 ÞGðju2 Þ / ðjuN ÞGðjuN Þ 7
c 6
G ¼6 7;
The derivation proceeds as follows. We define unit vector E1 ¼ « 1 7
4 5
½1 0 ,,, 0T , with similar E2, and so forth. In steady state, if the ðju1 Þi1 Gðju1 Þ ðju2 Þi1 Gðju2 Þ / ðjuN Þi1 GðjuN Þ
input to the system is uðtÞ ¼ E1 ejut , then xss ðtÞ ¼ M1 ejf1 ejut and we
have
with i being a user-defined index (called the “oversizing param-
ðjuÞM1 ejf1 ejut ¼ Ac M1 ejf1 ejut þ Bc E1 ejut : eter”) which is chosen to be larger than the order of the system:
|fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} i > n, and N is the number of frequency-response samples. We also
x_ss ðtÞ xss ðtÞ define
We can cancel ejut from both sides and write
2 3
6 Im Im / Im 7
ðjuÞM1 e ¼ Ac M1 e
jf1 jf1
þ Bc E1 : 6 ðju1 ÞIm ðju2 ÞIm / ðjuN ÞIm 7
|fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} Ic ¼ 6
6
7
7
X1 ðuÞ X1 ðuÞ 4 « 1 5
ðju1 Þ Im ðju2 Þ Im / ðjuN Þ Im
i1 i1 i1
Repeating with inputs uðtÞ ¼ E2 ejut , uðtÞ ¼ E3 ejut , and so forth,
we can assemble the following form Xc ¼ Xðju1 Þ Xðju2 Þ / XðjuN Þ

ðjuÞ X1 ðuÞ / Xm ðuÞ ¼ Ac X1 ðuÞ / Xm ðuÞ þBc E1 / Em Then, we can further write the inputeoutput equations
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}
XðjuÞ XðjuÞ Im
ðjuÞXðjuÞ ¼ Ac XðjuÞþBc Im ; Gc ¼ Gi Xc þ Qi Ic ;
where continuous-time extended-observability matrix Gi has the

where XðjuÞ2Cnm is a matrix of constant complex coefficients, same form as (but different values from) discrete-time O i given in
dependent on the forcing frequency u. Eq. (15) and Qi is a block-Toeplitz matrix:
When a system's input is a sinusoid, the steady-state output
yss ¼ Y1 ejq1 ejut is the system's frequency response G1 ðuÞ ¼ Y1 ejq1 " #
multiplied by the input ejut : T T T
T
T
Gi ¼ C ðCAc Þ CA2c ,,, CAi1
c
yss ðtÞ ¼ Cxss ðtÞ þ Duss ðtÞ 2 3
Y1 ejq1 ejut ¼ CM1 ejf1 ejut þ DE1 ejut 6 D 7
|fflfflffl{zfflfflffl} 6 CBc D 7
G1 ðuÞ Qi ¼ 6
6 « 1 1
7:
7
G1 ðuÞ ¼ CX1 ðuÞ þ DE1 : 4 5
CAi2
c Bc / CBc D
Collecting frequency responses resulting from every input
together, we have It can be more convenient to work with real data matrices rather
than complex data matrices, so define

G1 ðuÞ / Gm ðuÞ ¼ C X1 ðuÞ / Xm ðuÞ þ DIm ;
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
G¼ imag Gc 2Rqi2mN
real Gc ;
GðjuÞ XðjuÞ

where GðjuÞ2Clm . I ¼ real Ic ; imag Ic 2Rmi2mN

Now, notice that differentiating the output gives yðtÞ _ ¼ C xðtÞþ
_
_
DuðtÞ: With a frequency input, in steady state y_ ss ðtÞ ¼ ðjuÞyss ðtÞ and X ¼ real Xc ; imag Xc 2Rn2mN ;
resulting in
6
We include the full derivation here since the method has not previously been
published in a readily accessible archival format. We also improve the method in
this paper to enforce correct dc gain. G ¼ Gi X þ Qi I: (20)
7
Note that continuous-time Ac and Bc matrices have different values from the
corresponding discrete-time system's A and B matrices, while the continuous-time If we project Eq. (20) onto the orthogonal complement of I, we
C and D matrices are the same as the discrete-time system's C and D matrices. eliminate the second term and obtain
This definition allows for the determination of a well-

G I⊥ ¼ Gi X I⊥ þ Qi I I⊥ ¼ Gi X I⊥ ; conditioned basis GF , which is computed as
|ffl{zffl}
0 h iT
1=2 1=2 1=2
GcF ¼ Z 0 R0 ; Z 1 R1 ; ,,,; Z i1 Ri1
which has the same column space as Gi . So, by performing the SVD
GF ¼ real GcF ; imag GcF :
of G =I⊥ we can compute an estimate for Gi (up to an unknown but
unimportant similarity transformation). That is, The matrix IF is also computed using the Forsythe recursions but
2 32 3 in this case, the recursion is initialized with R0 ¼ ½Im ; Im ; /; Im
Then, we can implement the orthogonal projection as
S 0 54 V T1 5
G I ⊥ ¼ U1 U2 4 1
0 S2 V T2
GF I⊥ T
F ¼ GF GF IF IF IF
T 1
IF ¼ GF GF ITF IF ; (22)
|fflfflfflffl{zfflfflfflffl}
and, as with the DRA, we place the singular values in descending Imi
order and truncate to form S1 ; S2 can be considered essentially

zero. Hence, we retain only G =I⊥ zU1 S1 V T1 and compute and perform the SVD on GF =I⊥ F . As before, we only retain
GF =I⊥ T
F zU1 S1 V 1 and compute
Gi ¼ U1 S1=2
1 : 1=2
Yi ¼ U1 S1 :
The continuous-time Ac matrix can be obtained directly by
The Ac matrix can now be computed as
y
Ac ¼ GYi G[i (21) y [[
Ac ¼ D1 Yh
i
Yi D2 YYY
i
; (23)
where
where
" #
T T T
GYi ¼ C T ðCAc ÞT CA2c ,,, CAi2
c ; D1 ¼ diag Z 1=2 Z
1=2
; Z
1=2 1=2
Z ; /; Z
1=2 1=2
Z
1 2 2 3 i2 i1
"
T T #T
T D2 ¼ diag Z1 ðZ0 Z2 Þ1=2 ; /; Zi2 ðZi3 Zi1 Þ1=2 ;
G[i ¼ ðCAc Þ T
CA2c CA3c ,,, CAi1
c :
We convert the continuous-time Ac matrix to discrete time via a Yhi

is equal to Y i with the top block row and bottom block row
matrix-exponential operation: A ¼ expmðAc Ts Þ [25]. Then, we use deleted, Y[[
i
is equal to Y i with the top two block rows deleted, and
similarity transformations to diagonalize A. Finally, we replace Y YY
i
is equal to Y i with the bottom two block rows deleted.
unstable poles in A by their reciprocals and complex poles by their We find that the CRA is also very sensitive to the model order n
magnitudes. and therefore trying to reduce an infinite-order transfer function
(like the electrochemical transfer functions) can be very chal-
3.3.2. Improving conditioning lenging. A better approach is to increase the model order to n1 > n
It is worth mentioning that the performance of this algorithm such that we obtain a state-space model that closely approximates
for large model order n and oversizing parameter i is generally poor the infinite-order frequency response. Then, we perform a balanced
since the block Vandermonde matrices G and I become very ill- model order reduction (e.g., using balred.m in MATLAB [36]) to
conditioned. In order to make the CRA more robust numerically, obtain the final state-space reduced-order model of order n. Again,
the G and I data matrices can be modified using Forsythe re- for our application, we expect A to have real, stable poles, which is
cursions, which construct for them a well-conditioned basis. The not guaranteed by the method as presented so far, so we replace
improved method from [35] is repeated here for completeness, but unstable poles in A by their reciprocals and complex poles by their
without proof: magnitudes.
1. Initialization 3.3.3. Summarizing the CRA

To summarize the steps of the CRA:
R0 ¼ Gðju1 Þ; Gðju2 Þ; /; GðjuN Þ
1. Compute GF and IF using Gðjuk Þ and the Forsythe recursions (via
Z0 ¼ diag diag R0 R0 computing Rk and Zk ).
R1 ¼
R0 Du 2. Project GF onto the orthogonal complement of IF via Eq. (22).
Z1 ¼ diag diag R1 R1 3. Perform the SVD to GF =I⊥F . The model order n1 > n is selected
such that we accurately approximate the actual system.
where 4. Partition the SVD to give the matrices U1 , S1 , and V T1 and
1=2
compute Yi ¼ U1 S1 .
Du ¼ diag ðju1 Þ ðju2 Þ / ðjuN Þ 5 11m 5. Compute the Ac matrix via Eq. (23) and discretize it with sam-
pling time Ts to produce the discrete-time matrix A. Then,
compute the C matrix using the approach of Section 5, the B
and here, ½$ refers to the complex conjugate transpose of ½ $. matrix as described in Section 4 and the D via Eq. (31) as
described in Section 6.
2. Recursion (for k ¼ 2 to i 1): 6. Perform balanced model-order reduction on the discrete-time
1 state-space model of Step 5 in order to obtain a discrete-time
Rk ¼ Rk1Du þZk1 Z
k2 Rk2 state-space model with the desired order n. Then, keep the

Zk ¼ diag diag Rk Rk :
obtained A matrix.
7. Diagonalize A and replace unstable poles by their reciprocals We can further evaluate the terms of this equation at multiple
and complex poles by their magnitudes. frequency points u1 …uN , giving
G c ðu1 Þ / G c ðuN Þ ¼ O i X c ðu1 Þ / X c ðuN Þ

3.4. Hybrid realization algorithm (HRA) |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
c c
G X
3.4.1. Deriving the HRA steps þT i U ðu1 Þ / U ðuN Þ :
c c
The HRA determines the state-space A matrix from the discrete- |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
c
U
time frequency responses of the electrochemical variables, which
themselves are computed using the method explained in Sect. 2. A final step yields matrices containing only real values
The HRA derivation is based on a subspace system-identification

method from [37], and begins with the assumed discrete-time real G c imag G c ¼ O i real X c imag X c
state-space model of Eqs. (1) and (2). Taking z-transforms of |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
r r
these equations gives G X

þ T i real U c imag U c : (24)
zXðzÞ ¼ AXðzÞ þ BUðzÞ |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
YðzÞ ¼ CXðzÞ þ DUðzÞ: U r
We note that multiplying the output equation by z has the In this equation, we use values computed via the method from
following effect Sect. 2 to populate G r . U r can be computed from the chosen fre-
quency vector fu1 …uN g. X r is unknown but we will see that we do
zYðzÞ ¼ CðzXðzÞÞ þ DðzUðzÞÞ ¼ CAXðzÞ þ CBUðzÞ þ DðzUðzÞÞ: not require knowledge of its values.
We then perform the LQ decomposition
Repeating this process i 1 times (where i is again termed the
“oversizing parameter”) and combining the i modified output 2 3 2 32 3
r
equations gives 4 U r 5 ¼ LQ ¼ 4 L11 0 5 4 Q1 5
; (25)
2 3 2 3 2 3 G L21 L22 Q2
|fflfflfflffl{zfflfflfflffl}
6 YðzÞ 7 6 C 7 6 UðzÞ 7 data matrix
6 zYðzÞ 7 6 CA 7 6 zUðzÞ 7
6 7¼6 7 6 7;
6 7 6 « 7 XðzÞ þ T i6 7
4 i1« 5 4 i1 5 4 i1« 5 where L has the same dimension as the “data matrix” on the left
z YðzÞ CA z UðzÞ side of Eq. (25), and where Q1 and Q2 are orthonormal. Then,
|fflfflfflfflfflffl{zfflfflfflfflfflffl}
O i
U r ¼ L11 Q1
G r ¼ L21 Q1 þ L22 Q2 :
where T i is a block-Toeplitz matrix for the discrete-time system
that has the same form (but different values) from the continuous- Multiplying the second equation by Q T2 on the right gives
time block-Topelitz matrix Qi seen earlier, and O i is the extended G Q T2 ¼ L22 . Multiplying Eq. (24) by Q T2 and substituting this result
r
observability matrix of the system. we get

Suppose that a complex sinusoid of frequency u is applied to the
jth system input. That is, uj ½k ¼ Ej expðjukTs Þ. The system output in G r Q T2 ¼ L22 ¼ O i X r Q T2 :
steady state is then the inputeoutput frequency response corre-
sponding to frequency u applied to that input, which we can write The net effect of the LQ decomposition has been to project Eq.
as Gj ðejuTs Þ, multiplying the input. The state vector in steady state is (24) onto a subspace orthogonal to the input U r , thus removing its
then the inputestate frequency response Xi ðejuTs ) multiplying the effect from Eq. (24) and removing the need to know T i. The
input. We can write resulting equation states that L22 has the same column space as O i .
To recover O i (up to an unknown but unimportant similarity
2 3 2 3 transformation) from L22 , we perform the SVD
G ejuTs Uj ðzÞ ðzÞ
6 7 6 U j 7
6 7 6 zUj ðzÞ 7 2 32 3
6 zG ejuTs Uj ðzÞ 7
6 7 ¼ O i Xj ejuTs Uj ðzÞ þ T 6
i6
7;
7 S 0 54 V T1 5
6
4 « 7
5 4 i1 « 5 L22 ¼ U1 U2 4 1 ;
i1 juTs z Uj ðzÞ 0 S2 V T2
z G e Uj ðzÞ
where we think of S2 being essentially zero. L22 can be well

where we recognize that Uj ðzÞ ¼ Ej z=ðz expðjuTs ÞÞ. Notice that
approximated using the n most significant singular values in S1 , so
z =ðz expðjuTs ÞÞ is common to all terms so cancels out of the
we choose the number of poles in the final realization to be n. We
equation. Also note that we can write similar equations for j cor-
set our estimate of the extended observability matrix to O i ¼
responding to all inputs, and when we do so we can stack the re- 1=2
U1 S 1 .
sults horizontally. Evaluating that result at z ¼ expðjuTs Þ gives
There are various ways to recover the state-space matrices at
2 3 2 3 this point. The shift property of the extended observability matrix
G ejuTs
6 7 6 Im 7 allows computing A as the least-squares solution to
6 7 6 ejuTs Im 7
6 ejuTs G ejuTs 7 6 7;
6 7 ¼ O i X c ð uÞ þ T i6 7 y Y
6 « 7 4 juði1ÞT « 5 A¼ O[ Ob i : (26)
4 5 i
juði1ÞTs e s
Im
e G ejuTs
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} Again, for our application, we expect A to have real, stable poles,
G c ðuÞ U c ðuÞ which is not guaranteed by the method as presented so far. So, we
replace unstable poles in A by their reciprocals and complex poles
where X c ðuÞ ¼ X1 ðejuTs Þ X2 ðejuTs Þ / Xm ðejuTs Þ by their magnitudes
3.4.2. Summarizing of the HRA the parameters and standard least-squares methods can be used to
To summarize the steps of the HRA: solve for the vector of unknown ak and bk coefficients using the
system of equations describing all frequencies. But, while large-
1. Compute GðejuTs Þ from GðjuÞ using the method of Sect. 2. order n will produce a very accurate approximation of GðsÞ, the
2. Form the data matrices G r and U r in Eq. (24). computational requirements to solve the linear system of equations
3. Perform the LQ decomposition in Eq. (25) to find L22 . also increases.
4. Compute the SVD of L22 to give U1 and S1 . The model order n is Once we have obtained a high-order continuous-time rational
selected by counting the number of “significant” singular values transfer function HðsÞ that accurately approximates the true
in S1 . (infinite-order) frequency response, we can easily convert it to
1=2
5. Compute O i ¼ U1 S1 . state-space representation (e.g., using tf2ss.m in MATLAB), that is,
6. Compute A via Eq. (26). HðsÞ ¼ CðsI AÞB þ D. We convert this state-space model from
7. Diagonalize A and replace unstable poles by their reciprocals continuous-time to discrete time using c2d.m in MATLAB and we
and complex poles by their magnitudes. reduce model order using balred.m in MATLAB.
For our application, we expect A to have real, stable poles, which
is not guaranteed by the method as presented so far. So, we replace
3.5. Lagrange-interpolated realization algorithm (LRA)
unstable poles by their reciprocals and complex poles by their
magnitudes.
3.5.1. Deriving the LRA steps
Notice that the approach explained above is valid only for
The LRA is a realization procedure that approximates the actual
single-output systems and therefore needs to be extended to multi-
infinite-order transfer function frequency response using a
output systems. Since we are interested at this point only in a high-
continuous-time high-order model (HOM). The continuous-time
order approximation (i.e., we do not care about the model order),
HOM is discretized and reduced using balanced model-order-
we can easily extend the approach to this case by assembling the
reduction techniques, generating a final ROM of order n.
different single-output high-order models. That is,
A single-input single-output continuous-time rational transfer
function of order n > n has the form 2 3 2 3
6 A1 0 0 0 7 B
6 17
bn1 sn1 þ / þ b1 s þ b0 6 0 A2 0 0 7 6 B2 7
HðsÞ ¼ þ d0 ; (27) A¼6
6 0
7; B¼6 7
6 « 7; (28)
sn þ an1 sn1 þ / þ a1 s þ a0 4 0 1 0 75 4 5
0 0 0 Aq Bq
where d0 is a direct feed through term and the coefficients ai and bi
for i2½0…n 1 will determine the pole and zero locations of the 2 3 2 3
transfer function. Notice we can accurately approximate any given
6 C1 0 … 0 7 6 D1 7
transfer function GðsÞ (even an infinite-order transfer function) by 6 0 C2 0 0 7 6 D2 7
using Lagrange interpolation. C¼6
6 0
7; D¼6 7
6 « 7; (29)
4 0 1 0 75 4 5
The Lagrange interpolation approach consists of estimating
0 0 0 Cq Dq
values between known data points. In the frequency domain, we
would like Hðsi Þ ¼ Gðsi Þ at different frequency points si , i2 ½1…n.
Note that we have 2n degrees of freedom corresponding to the where the state-space matrices Ak , Bk , C k and Dk correspond to the
number of a and b coefficients to be estimated. However, we use continuous-time HOM approximation for the transfer function k.
only n different frequency points in order to obtain real valued Thus, the A, B, C and D matrices estimate a total of q continuous-
coefficients since the other n points will correspond to their time transfer functions. Then, the continuous-time state-space
negative frequencies, that is, Hðsi Þ ¼ Gðsi Þ and Hð si Þ ¼ Gð si Þ ¼ model can be discretized with a given sampling time Ts by
Gðsi Þ . Therefore, we need to solve a system of 2n equations to computing the discrete-time state-space matrices using known
determine the coefficients a and b from Eq. (27). The system of methods (e.g., using c2d.m in MATLAB). Finally, a balanced model-
equations to be solved looks like order reduction is performed to the high-order discrete-time state-
space model (e.g., using balred.m in MATLAB) to produce the
bn1 sn1
1 þ / þ b 1 s1 þ b 0 discrete-time reduced-order model of order n.
Hðs1 Þ ¼ þ d0 ¼ Gðs1 Þ
sn1 þ an1 sn1
1 þ / þ a1 s1 þ a0
bn1 ð s1 Þn1 þ / þ b0
Hðs1 Þ ¼ þ d0 ¼ Gðs1 Þ
ð s1 Þn þ an1 ð s1 Þn1 þ / þ a0 3.5.2. Summarizing the LRA
« To summarize the steps of the LRA:
bn1 sn1
n þ / þ b 1 sn þ b0 1. Approximate the actual continuous-time frequency response
Hðsn Þ ¼ þ d0 ¼ Gðsn Þ
sn þ an1 sn1
n
n þ / þ a 1 sn þ a0 GðsÞ using a rational continuous-time high-order model HðsÞ via
solving Gðsi Þ ¼ Hðsi Þ and Gðsi Þ ¼ Hðsi Þ for i2½1…n.
bn1 ð sn Þn1 þ / þ b0
Hðsn Þ ¼ þ d0 ¼ Gðsn Þ : 2. Repeat Step 1 for each transfer function to be estimated. Then,
ð sn Þn þ an1 ð sn Þn1 þ / þ a0 assemble the individual high-order models via Eqs. (28) and
(29) to build HðsÞ.
Note that the ith equation can be written as 3. Discretize the rational high-order model HðsÞ to obtain a
discrete-time high-order model HðzÞ using c2d.m in MATLAB.
ðGðsi Þ d0 Þ sni þ an1 sn1
i þ / þ a0 ¼ bn1 sn1
i þ / þ b 1 si 4. Perform a balanced realization of the discrete-time HOM HðzÞ to
þ b0 : obtain the discrete-time diagonal state-space A matrix. Replace
unstable poles by their reciprocals and complex poles by their
Assuming that d0 is known (cf. Sect. 6), this equation is linear in magnitudes.
4. Solving for the state-space B matrix matrix equation:

2 32 3 2 3
At this point, we have discussed four different ways to produce T
4 2M 2 Mð0Þ 54 C T 5 4 2M T1 5
reduced-order discrete-time state-space A matrices. State-space ¼ : (30)
T
M ð0Þ 0 lT GT ð0Þ
models have multiple degrees of freedom for choosing some
parameter values, after which the remaining values will be fixed.
We solve Eq. (30) for C (and l, which is not used) using least
For simplicity of a final implementation, we choose to assign B ¼
squares.
1n1 without loss of generality, since the C matrix will be computed
In summary, the approach presented in this section, computes
to scale outputs properly for this choice of the discrete-time B
the C matrix that minimizes the squares of the errors between the
matrix.
actual and approximated discrete-time frequency responses while
enforcing that the dc gain of the approximated system exactly
5. Solving for the state-space C matrix
matches the dc gain of the actual system.
When solving for the C matrix, we desire to minimize the
squares of the errors between the actual and approximate fre-
6. Solving for the state-space D matrix
quency responses while enforcing that the dc gain of the approxi-
mated system exactly matches the dc gain of the actual system. To
The final remaining unknown of Eqs. (1) and (2) is the D matrix,
do so, we formulate a constrained optimization cost function to be
which models the instantaneous change in output y½k when the
minimized as
system is excited with input u½k. For a PBM, the D matrix can be
found in closed-form since the electrochemical transfer functions
X
N
kGðuk Þ Hðuk Þk22 þ l ðGð0Þ Hð0ÞÞ
T are analytical. That is, we simply compute
J¼
k¼1
T
D ¼ lim G1 ðsÞ; lim G2 ðsÞ; …; lim Gq ðsÞ : (31)
s/∞ s/∞ s/∞
where G corresponds to the actual frequency response and H cor-
responds to the low-order approximation of G. For notational
simplicity, we write Gðuk Þ ¼ Gðejuk Ts Þ D and Mðuk Þ ¼
ðejuk Ts I AÞ1 B. Then, Hðuk Þ ¼ CMðuk Þ and
7. Results
N h
nX i h io
T
J¼ Gðuk Þ Hðuk Þ Gðuk Þ Hðuk Þ þ l ðGð0Þ Hð0ÞÞ
k¼1
nX
N h i h i h i h io h i
¼ Tr G uk G uk Tr G uk M uk C T Tr CM uk G uk þ Tr CM uk M uk C T þ Tr ðGð0Þ CMð0ÞÞlT
k¼1
This paper has presented four methods to convert a user-

Taking derivatives of J with respect to l and C gives selected set of an arbitrary number of continuous-time transcen-
vJ
¼ Gð0Þ CMð0Þ ¼ 0
vl
vJ XN X
N
Gðuk ÞM ðuk Þ þ ðMðuk ÞG ðuk ÞÞ þ C Mðuk ÞM ðuk Þ þ ðMðuk ÞM ðuk ÞÞT lMð0ÞT ¼ 0:
T
¼
vC
k¼1 k¼1
dental transfer functions describing the internal variables of a

lithium-ion cell into a single reduced-order discrete-time state-
Recognizing that if Z is a complex matrix, then [Zþ(Z*)T]¼2ℝ[Z], space model that can jointly approximate this set of variables
we let with a single low-order model. This kind of physics-based reduced-
order model will be required by future BMS control algorithms that
must be able to predict and place constraints not only on cell ter-
X
N
minal voltage (such as can be done already with any existing cell
M1 ¼ R½Gðuk ÞM ðuk Þ
model) but also on cell internal electrochemical variables (such as
k¼1
X
N the solideelectrolyte potential difference near the separator, which
M2 ¼ R½Mðuk ÞM ðuk Þ: is the most critical variable at the most critical location for pre-
k¼1 dicting the onset of lithium platingdwhich itself can lead to
dendrite formation and short circuits [38]dand for predicting the
Then, we can transpose both constraints and combine into one
magnitude of solideelectrolyte interphase (SEI) layer growth rate). (the rest are similar): we compare the true discrete-time frequency
We would now like to know whether the different xRA methods response of the solideelectrolyte potential difference at four cross-
produce ROMs of differing quality, and whether there are other sectional locations in the cell at 50 % state-of-charge and 25+ C and
considerations that might be important when selecting among at 100 % state-of-charge and 15+ C, computed from the transfer
them for the one to use for a particular application. functions using the method in Sect. 2, to the ROM's discrete-time
We assume that the parameter values of the PDE full-order frequency response.11 This variable is especially interesting as it is
model (FOM) have already been identified for a cell of interest.8 a predictor of the onset and magnitude of side reactions such as
Therefore, we assume that FOM predictions match true-cell vari- solideelectrolyte interphase growth or lithium plating [41,42]. If it
ables with acceptable accuracy. So, if the ROMs produced by the can be predicted well, BMS control algorithms might be designed
xRAs match the FOM in meaningful ways, then we can claim that that can regulate battery use to slow the rate of degradation due to
the xRAs have been successful in their objective. No physical cell side reaction.
tests or lab data are needed to do this evaluation. In this paper, we A ROM that can jointly simulate time-domain profiles from
choose to use FOM parameter values from the literature, already these 21 transfer functions yields (more than) enough information
validated on a physical cell (see the Appendix). When comparing to predict cell voltage. So, to illustrate the relative ability of the xRAs
the FOM to the ROMs, we can directly compare the frequency re- to produce ROMs that predict voltage well, we present root-mean-
sponses; we can also simulate both model types for the same input squared (rms) voltage modeling error for a simple experiment
stimulus to compare their time-domain responses. performed at different ambient temperatures and initial state-of-
Accordingly, this section presents a sample of time- and charge points. The “true” voltage is produced by a COMSOL simu-
frequency-domain results to illustrate the level of performance that lation of the full PDE model for a “pulse-test” input-current profile
can be expected of each method. We have chosen to implement the comprising a 3-min 0.5C constant-current discharge pulse followed
four xRA methods on a set of q ¼ 21 lithium-ion battery-cell by a 2-min rest.12 The estimated voltage was produced by each ROM
transfer functions, using five nonintegrator states plus one inte- for the same input-current profile. Voltage error was computed as
grator state for a final n þ 1 ¼ 6.9 The mathematical equations of “true” voltage minus estimated voltage.
the transfer functions are not reproduced here since the focus of We point out that the xRA process must be executed one time
this paper is on the xRAs and not on the transfer functions them- only to produce a ROM that can be executed any number of times
selves.10 However, it is still important to report the set of transfer after that. So, the computational requirements of the individual
functions being jointly estimated by a single 6-state ROM corre- xRAs is not as important as those of the ROMs they produce.13
sponding to a one-dimensional FOM, which are (where x ¼ 0 de- Therefore, we did not focus on optimizing our xRA codes. Howev-
notes the negative-electrode/current-collector interface, x ¼ 1 the er, it is of interest to have a rough idea of the relative computational
negative-electrode/separator interface, x ¼ 2 the positive- demand of each xRA so we present execution-time and memory
electrode/separator interface, and x ¼ 3 the positive-electrode/ requirements of each of the methods in the sequel.
current-collector interface): Also, note that all of the ROMs produced by all methods were
constrained to have only real poles with magnitude less than 1
~ s;e ðx; zÞ =Iapp ðzÞ, which is the integrator-removed debiased
½C (except for the integrator pole having magnitude exactly equal to
solid surface concentration of lithium, at x ¼ ½0; 1; 2; 3 1). Therefore, all ROMs were nonoscillatory and stable.
Jðx; zÞ =Iapp ðzÞ, which is the flux of lithium from solid to elec-
trolyte, at x ¼ ½0; 1; 2; 3 7.1. DRA
~ s;e ðx; zÞ =Iapp ðzÞ, which is the integrator-removed debiased
½F
solideelectrolyte potential difference, at x ¼ ½0; 1; 2; 3 The performance of the DRA is addressed in this section. It is
~ e ðx; zÞ =Iapp ðzÞ, which is the debiased electrolyte potential, at
F worth mentioning up front that the overall performance of the DRA
x ¼ ½1; 2; 3 (it is zero at x ¼ 0) depends strongly on its tuning parameters. The results shown in
~ s ðx;zÞ =Iapp ðzÞ, which is the debiased solid potential, at x ¼ ½1; 2
F this section correspond to the following set of tuning parameters:
(it is zero at x ¼ 0 and x ¼ 3)
~ e ðx; zÞ =Iapp ðzÞ, which is the debiased concentration of lithium
C
in the electrolyte, at x ¼ ½0; 1; 2; 3 11
Note that we have implemented the xRAs on transfer functions corresponding
to a 1-d FOM. For either a 1-d FOM or resulting 1-d ROM to produce physically
The PBM parameter values used for these results are the same as meaningful results, there must either be a high degree of uniformity in the 1-
those presented in [24] and are listed in Table 2. d dynamics experienced at different 3-d spatial locations in the cell, or models
corresponding to “worst-case” spatial locations must be used. In practice, unifor-
While we could examine the ability of ROMs produced by the mity can be enhanced by careful cell design (e.g., placing tabs of pouch cells on
xRAs to approximate the frequency responses of any or all of these opposite ends, or welding together many current-collector tabs along the length of
six variables at any set of locations fxg across the cell cross section, a prismatic or cylindrical cell). Note also that the limitation in the results is not due
we have chosen to illustrate only two representative examples here to the xRAs themselves but rather to the fact that we are unaware of existing 3-
d transfer functions. Should 3-d transfer functions be developed, the xRAs could
be applied to predict electrochemical variables at any 3-d spatial location in the cell.
This is an area open to future research.
8
This “system-identification” or “parameter-identification” process is outside the 12
This profile was chosen (1) since its SOC does not diverge far from the setpoint,
scope of this paper, but see [39]. allowing evaluation of time-domain results without the need to blend models as
9
The model order effects a tradeoff between ROM accuracy and computational discussed in [40] and (2) since it is relatively simpledthe results of 640 time-
complexity of ROM execution. In our experience, we find that models having 5e6 domain simulations are reported in summary form in this paper and we desired
nonintegrator poles tend to give good performance while retaining similar a FOM simulation that would execute in reasonable time. However, our experience
computational complexity to equivalent-circuit models being used in practice with other input-current profiles that might be more representative of automotive
today. However, the “best” model order might vary depending on the application/ applications is that they produce qualitatively similar results. The interested reader
purpose of the final model. can confer [24], where numerous other time-domain results are presented showing
10
The interested reader can find a full derivation of these transfer functions in how a DRA-ROM can predict internal electrochemical variables and external ter-
[24], along with a description of how to use the ROM produced using these transfer minal voltage (including multiple C-rates, UDDS profiles, etc.).
13
functions in a time-domain simulation; reference [40] shows how to use the ROMs The ROMs produced by each of the xRAs have exactly identical computational
to produce simulations spanning a wide range of cell SOC and temperature by requirements, roughly equivalent to an equivalent-circuit model having n
blending between models. resistorecapacitor branches.
Emulation sampling period T1 ¼ 2 Hz. discrete-time unit-pulse responses and since this step is shared by
Length of the impulse response 220 s; that is, N ¼ 221 . the DRA method presented here and the work in [43], we do not
expect significant speedup.
The unit-pulse-response data used to build the generalized Memory. The largest matrix size used by the DRA for this study
Hankel matrix H r;s corresponds to ti ¼ ji ¼ 0 to 4; 000; 5; 000 to was 3; 573:451 MB, which corresponds to H r;s ½0 and H r;s ½1. The
5; 100; and 51 values starting at 6; 000; 7; 000; 8; 000; 10; 000; matrices H r;s ½0 and H r;s ½1 have as many rows as number of
12; 000; 14; 000; 16; 000; 18; 000; 19; 000 and 20.000 (e.g., 6; 000 to outputs times the length of ti and as many columns as the length of
6; 050; 7; 000 to 7; 050; etc.). ji . In our example, the size of H r;s ½0 and H r;s ½1 is 96; 852 4; 612.
These parameters were selected to give good overall perfor- Hence, increasing the size of ti and ji will increase the size of the
mance, but by no means represent an “optimal” choice. largest matrix used by the DRA. Again, memory requirements of the
Speed. The DRA average run time was 214:07 s, which is quite DRA might be improved using the refinements presented in [43].
slow considering we are using the ERA submethod which does not Accuracy. We are most interested in the time-domain perfor-
require processing the complete unit-pulse response. In order to mance of physics-based ROMs generated using the DRA. The top-
generate individual ROMs from 15+ C to 55+ C in 10 C+ increments left frame of Fig. 4 shows in a “heatmap” format the cell voltage
and from 0 % to 100 % state-of-charge in 5 % increments (as required rms error between the full-order model and the physics-based
to compute the results presented in Fig. 4), we need to generate a ROM generated using the DRA for the pulse tests starting at
total of 168 models, which takes approximately 10 h. The speed of different states-of-charge and temperatures. The smallest cell
the DRA might be improved using the refined numerical methods voltage root-mean squared error is 0:23 mV corresponding to 90 %
in [43]; however, the bulk of the time is spent computing the state-of-charge and 55+ C. The top row of Figs. 5 and 6 show the
Fig. 4. Pulse-response voltage-prediction rms error for ROMs produced by all four xRA methods at different temperature and SOC setpoints.
~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in

Fig. 5. True discrete-time frequency-responses for the integrator-removed solideelectrolyte potential difference transfer function ½F
the cell compared to frequency response approximated by ROMs created using the four proposed xRA methods (linearized around a 50 % SOC and 25+ C setpoint).
frequency-domain performance of the DRA when approximating Speed. The HRA average run time is 32.07 seconds, which is
the interphase potential difference transfer function at the cell's much faster than the DRA but slower than the CRA. It takes
internal region boundaries at two different setpoints (50 % SOC and approximately 90 min to generate 168 models.
25+ C, and 100 % SOC and 15+ C, respectively). The ROM generated Memory. The largest matrix size used by the HRA is 3.696 MB,
using the DRA is able to approximate the actual transfer-function which corresponds to the data matrix in Eq. (25). The data matrix
frequency response very accurately. (We have chosen not to dimension is 2ði þ 1Þq 2N where N corresponds to the total fre-
include figures plotting the time-domain results directly, since the quency points used to build GðsÞ. In our example, the data matrix
voltage difference between the FOM and ROM is not visually dimension is 462 1; 000. Notice the largest matrix size of the HRA
discernible for these low levels of error, and since it would not be is the smallest of all realization algorithms introduced.
reasonable to include time-domain plots for all 640 simulations.) Accuracy. The bottom-left frame of Fig. 4 shows the cell-voltage
rms error between the full-order model and the physics-based
7.2. CRA ROM generated using the HRA for the pulse tests starting at
different states-of-charge and temperatures. In general, we see the
The performance of the CRA is addressed in this section. Again, HRA obtains better results than both the DRA and CRA for most of
the overall performance of the CRA depends on its tuning param- the hot temperatures. The smallest cell-voltage rms error is
eters. The results shown correspond to the following set of tuning 0:40 mV corresponding to 30 % state-of-charge and 55+ C. The third
parameters: row of Figs. 5 and 6 show the frequency-domain performance of the
HRA when approximating the phase potential difference transfer
The continuous-time frequency vector for building GðjuÞ was function at the cell boundaries at two different setpoints (50 % SOC
chosen to comprise one point at u ¼ 0 and 140 additional points and 25+ C, and 100 % SOC and 15+ C, respectively). Notice the HRA
spaced evenly on a logarithmic scale between 106 and produces different results from the DRA and the CRA but it is still
103 rad s1 . capable of approximating the actual transfer function frequency
The oversizing parameter i was chosen to be i ¼ 40. response very accurately.
The high fidelity model order n1 was chosen to be equal to the
number of singular values greater than 1010 .
7.4. LRA
Speed. The CRA average run time is 6.15 seconds: it is able to
produce a ROM much faster than the DRA. It takes approximately The performance of the LRA is addressed in this section. The
18 min to generate 168 models. overall performance of the LRA depends on its tuning parameters.
Memory. The largest matrix size used by the CRA is 5.094 MB, The results shown correspond to the following set of tuning
which corresponds to D1 and D2 . The matrices D1 and D2 are square parameters:
of size equal to i 2 times the number of outputs q. In our example,
the size of D1 and D2 is equal to 798 798. Hence, increasing the The continuous-time frequency vector for building GðsÞ was
oversizing parameter i or the number of transfer functions q to be chosen to comprise one point at u ¼ 0, 30 additional points
estimated will increase the size of the largest matrix used by the spaced evenly on a logarithmic scale between 1012 and 106
CRA. (to capture slow dynamics) and 500 additional points spaced
Accuracy. The top-right frame of Fig. 4 shows the cell-voltage evenly on a logarithmic scale between 106 and 107 (to capture
rms error between the full-order model and the physics-based faster dynamics).
ROM generated using the CRA for the pulse tests starting at The order of each estimated transfer function was chosen to be
different state-of-charge and temperatures. The smallest cell ni ¼ 40 for i2½1…q. Hence, the order of the complete high-
voltage rms error for the CRA is 0:26 mV corresponding to 90 % order system is n ¼ ni q.
state-of-charge and 55+ C. The second row of Figs. 5 and 6 show the
frequency-domain performance of the CRA when approximating Speed. The LRA average run time is 66.00 seconds where most of
the phase potential difference transfer function at the cell bound- the time is used to solve the linear system of equations. The LRA is
aries at two different setpoints (50 % SOC and 25+ C, and 100 % SOC much faster than the DRA but slower than the CRA and HRA. It takes
and 15+ C, respectively). We see that the CRA generates a approximately 3 h to generate 168 models.
different approximation from the DRA but it is still able to Memory. The largest matrix size used by LRA is 5.645 MB, which
approximate the actual transfer function frequency response corresponds to the high-order model A matrix. The A matrix
accurately. dimension is ni q ni q.
Accuracy. The bottom-right frame of Fig. 4 shows the cell-voltage
7.3. HRA rms error between the full-order model and the physics-based
ROM generated using LRA for the pulse tests starting at different
The performance of the HRA is addressed in this section. Again, states-of-charge and temperatures. In general, we see the LRA ob-
the overall performance of the HRA depends on its tuning param- tains better results than the DRA, CRA and HRA, especially at hot
eters. The results shown correspond to the following set of tuning temperatures. The smallest cell-voltage rms error is 0:14 mV cor-
parameters: responding to 85 % state-of-charge and 45+ C. The bottom row of
Figs. 5 and 6 show the frequency-domain performance of the LRA
The oversizing parameter was chosen to be i ¼ 10. when approximating the phase potential difference transfer func-
The discrete-time frequency vector for building GðzÞ was chosen tion at the cell boundaries at two different setpoints (50 % SOC and
to comprise one point at u ¼ 0 and 499 additional points spaced 25+ C, and 100 % SOC and 15+ C, respectively). We see that the
evenly on a logarithmic scale between 105 and 100 times the ROM generated using the LRA approach is in very close agreement
Nyquist rate. with the actual transfer function. Recall that the actual transfer
The value of kmax to estimate the discrete-time frequency function has an infinite number of poles and zeros and the reduced-
response from the continuous-time infinite-order transfer order model is approximating a total of 21 transfer functions using
function was chosen to be kmax ¼ 5. only 6 poles.
~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in

Fig. 6. True discrete-time frequency-responses for the integrator-removed solideelectrolyte potential difference transfer function ½F
the cell compared to frequency response approximated by ROMs created using the four proposed xRA methods (linearized around a 100 % SOC and 15+ C setpoint).
8. Summary
This paper has introduced four different realization algorithms

whose objective is to produce a discrete-time state-space ROM that
approximates the continuous-time electrochemical transfer func-
tions derived in [24]. However, these methods can be applied to any
transfer function of interest regardless of the derivation approach.
These methods do not “fit” a model to inputeoutput data; instead,
they directly compute a reduced-order model from the parameter
values of the DFN model (via intermediate transfer functions),
retaining the significance of all internal modeled electrochemical
variables in the final ROM. In summary, the different realization
algorithms (xRA) consist of:
DRA: Produces a discrete-time reduced-order state-space model
from discrete-time unit-pulse-response data.
CRA: Produces a continuous-time reduced-order state-space
model from continuous-time frequency-response data. Then, the
continuous-time reduced-order model is discretized.
HRA: Produces a discrete-time reduced-order model from
discrete-time frequency-response data, which is obtained using the
approach described in Sect. 2.
LRA: Produces a continuous-time high-order model from
continuous-time frequency-response data. Then, this model is
discretized and undergoes balanced model-order reduction using
standard techniques.
Fig. 1 shows a flowchart summarizing the different realization- Fig. 7. The xRA producing the best model at each temperature and SOC setpoint.
algorithm steps in order to generate a discrete-time reduced-or-
der model. The HRA and CRA use the orthogonal projection method
whereas the CRA and the LRA use balanced reduction techniques to
reduce the high-fidelity model order.
The computational requirements for the different realization
algorithms are listed in Table 1. Table 1
Summary of xRA computational requirements.
In terms of speed, the CRA turns out to be the fastest, followed
by the HRA, the LRA and the DRA, which is the slowest by far. In DRA CRA HRA LRA
terms of memory, the HRA is the most efficient, followed by the Run time ðsÞ 214.06 6.15 32.07 66.00
LRA, the CRA and the DRA, which requires much more memory Largest matrix size (MB) 3573.451 5.094 3.696 4.548
than the rest. In terms of robustness, all four methods produce
ROMs that are stable and nonoscillatory. In terms of accuracy, the
LRA produces the best time-domain results, followed by the HRA,
the DRA and the CRA, which is very sensitive to its tuning param- Declaration of interests
eters (this sensitivity is largely alleviated by the Forsythe recursions
and starting with a large initial reduced-order model order which is The authors declare that they have no known competing
later reduced using balred.m). Fig. 7 shows the xRA producing financial interests or personal relationships that could have
lowest cell voltage rms error between the full-order PDE model and appeared to influence the work reported in this paper.
the physics-based ROM for the pulse tests starting at different
states-of-charge and temperatures. We see the HRA and LRA al-
Acknowledgement
gorithms usually produce the best and more robust results and
therefore, we prefer these model-order reduction methods for
The information, data, or work presented herein was funded in
approximating the electrochemical transfer functions of interest.
part by General Motors as part of the “Physics-based reduced-order
However, other realization algorithms might perform better at
model for vehicle BSE” project. Financial support for the research
certain states-of-charge and temperature.
reported in this paper was also received from by the Advanced
All of the ROMs produce poorer voltage estimates at low tem-
Research Projects Agency-Energy (ARPA-E, U.S. Department of
peratures. This is not due to the xRA algorithmsdwhich capture
Energyy, under award number DE-AR0000271. This work used the
model frequency responses accurately even at low temper-
EAS Data Center or EAS Cloud, which is supported by the College of
aturesdbut rather due to the dynamics of the cells becoming more
Engineering and Applied Science at the University of Colorado
nonlinear as temperature decreases. The reader can verify this
Colorado Springs. The authors would like to express their gratitude
assertion by comparing the frequency-response approximations
to each of these sponsors.
made by the ROMs to the frequency responses of the desired
transfer functions in Figs. 5 and 6: the quality of the frequency-
response approximations is not noticeably different between the AppendixCell parameter values
two linearization setpoints. Instead, a largely linear ROM loses ac-
curacy in cases where the underlying FOM is highly nonlinear, such The parameters of the cell used in this publication are listed in
as at cold temperatures and extreme SOCs. Improving low- Table 2 and were those published by Lee et al. [21], except for the
temperature modeling is a focus of current research and is a mean molar activity coefficient, which is from [24], and the acti-
planned topic of future publications. vation energies, which were extracted from [44].
Table 2 [20] Smith KA. Electrochemical modeling, estimation and control of lithium ion
Cell parameters for simulation. batteries. Ph.D. thesis. Pennsylvania State University; 2006.
[21] Lee JL, Chemistruck A, Plett GL. J Power Sources 2012;220:430e48.
Negative Positive [22] Rodríguez A, Plett GL, Trimboli MS. J Power Sources 2017;360:642e5.
Symbol Units electrode Separator electrode [23] Rodríguez A, Plett GL. J Energy Storage 2017;11:219e36.
[24] Rodríguez A, Plett GL, Trimboli MS. J Energy Storage 2018;20:560e75.
L mm 128 76 190
[25] Hespanha JP. Linear systems theory. Princeton University Press; 2018.
Rs mm 12.5 e 8.5 [26] Jury EI. Transactions of the American institute of electrical engineers, Part I:
A m2 1 1 1 communication and electronics 73. 1954. p. 332e46.
s Sm1 100 e 3.8 [27] Lee JL, Chemistruck A, Plett GL. J Power Sources 2012;206:367e77.
Eas kJ mol1 0 e 0 [28] Franklin GF, Workman ML, Powell D. Digital control of dynamic systems. third
ed. Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc.; 1998.
εs m3 m3 0.471 e 0.297
[29] Oppenheim AV, Schafer RW. Discrete-time signal processing. third ed. Upper
εe m3 m3 0.357 0.724 0.444 Saddle River, NJ, USA: Prentice Hall Press; 2010.
brug e 1.5 e 1.5 [30] Ho B, Kalman R. Regelungstechnik 1966;14:545e8.
cs;max mol m3 26 390 e 22 860 [31] Miller DN, de Callafon RA. Automatica 2013;49:2468e73.
ce;0 mol m3 2000 2000 2000 [32] Juang JN, N J, Pappa RS. J Guid Control Dyn 1985;8:620e7.
[33] DePalma D, Plett GL, Trimboli MS. Generating high-fidelity reduced-order
q0 e 0.05 e 0.78
physics-based lithium-ion cell models, Energy Challenges and Mechanics
q100 e 0.53 e 0.17
(ECM4). Working on Small Scales 2015.
Ds m2 s1 3:9 1014 e 1:0 1013 [34] DePalma D. Efficient computation of reduced-order models of lithium-ion
EaDs kJ mol1 4 e 20 cells, master's thesis. University of Colorado Colorado Springs; 2014.
De m2 s1 7:5 1011 7:5 1011 7:5 1011 [35] Overschee PV, Moor BD. Signal processing 52. 1996. p. 179e94 [Subspace
Methods, Part II: System Identification].
EaDe kJ mol1 10 10 10
[36] A. Varga, in: Proceedings of the 30th IEEE conference on decision and control,
Eak kJ mol1 20 20 20 volume vol. 2, pp. 1062e1065
0
tþ e 0.363 0.363 0.363 [37] Liu K, Jacques RN, Miller DW. J Dyn Syst Meas Control 1996;118:211e20.
[38] Liu B, Jia Y, Li J, Yin S, Yuan C, Hu Z, Wang L, Li Y, Xu J. Journal of materials
k0 mol m2 s1 2:29 105 e 2:21 105
1
chemistry A, vol. 6; 2018. p. 21475e84.
Eak0 kJ mol 30 e 30
[39] Jobman R, Trimboli MS, Plett GL. J Energy Chall Mech 2015;2. article 3.
a e 0.5 e 0.5 [40] Lee JL, Aldrich LL, Stetzel KD, Plett GL. J Power Sources 2014;255:85e100.
Rfilm U m2 0.0 e 0.0 [41] Arora P, Doyle M, White RE. J Electrochem Soc 1999;146:3543e53.
dln f± e 3 3 3 [42] Ramadass P, Haran B, Gomadam PM, White R, Popov BN. J Electrochem Soc
dln ce 2004;151:A196e203.
[43] Gopalakrishnan K, Zhang T, Offer GJ. J Electrochem Energy Convers Storage
We compute seff ¼ sεs , keff ¼ kεbrug
e , Deff
e ¼ De εbrug
e . 2017;14. 011001e011001e8.
In the electrolyte, conductivity is a function of concentration: [44] W. Gu, C.-Y. Wang, in: Proceedings of the ECSssss, volume vol. 99, pp.
kðce Þ ¼ 4:1253 102 þ 5:007 104 ce 4:7212 107 c2e 748e762
þ1:5094 1010 c3e 1:6018 1014 c4e :
For the negative electrode, the open-circuit potential function is: Nomenclature
Uocp ðqÞ ¼ 0:16 þ 1:32 expð 3:0qÞ þ 10:0 expð 2000:0qÞ:
For the positive electrode, the open-circuit potential function is: A: state transition matrix of the state-space model
Uocp ðqÞ ¼ 4:19829 þ 0:0565661 tanhð14:5546q þ 8:60942Þ B: input matrix of the state-space model
" #
1 Cx : concentration of lithium transfer function in phase indicated by subscript x,
0:0275479 1:90111 mol m3
ð0:998432 qÞ0:492465 C: output matrix of the state-space model
0:157123 expð0:04738q8 Þ C : controllability matrix
D: input-output coupling matrix of the state-space model
þ0:810239 exp½40ðq 0:133875Þ: g: time-domain response of a system
G: transfer function of a system
H: approximated transfer function of a system
H : block Hankel matrix
References iapp : applied cell current, A
I: identity matrix
[1] Plett GL. Battery management systems: volume 2, equivalent-circuit methods. J: optimization cost function
Artech House; 2015a. Jx : reaction flux transfer function of phase indicated by subscript x, mol m2 s1
[2] Plett GL. Battery management systems: volume 1, battery modeling. Artech n: model order
House; 2015b. O : observability matrix
[3] Newman J, Thomas-Alyea KE. Electrochemical systems. third ed. John Wiley res0 : residue of the integrator pole
Sons, Inc.; 2004. 3rd edition. s: complex Laplace variable
[4] Wang L, Yin S, Xu J. J Power Sources 2019;413:284e92. t: time, s
[5] Liu B, Wang X, Chen H-S, Chen S, Yang H, Xu J, Jiang H, Fang D-N. J Appl Mech Ts : sample period, s
2019;86:041005. u: input of state-space model
[6] Berrueta A, Urtasun A, Ursúa A, Sanchis P. Energy 2018;144:286e300. x: lD linear coordinate across the cell, m
[7] Smith KA, Wang C-Y. J Power Sources 2006;160:662e73. x: state of state-space model
[8] R. Anderson, R. Zane, G. Plett, D. Maksimovic, K. Smith, M. Trimboli, in: SAE y: linear output of state-space model
technical paper 2017-01-1210 z: complex Z transform variable
[9] Doyle M, Fuller TF, Newman J. J Electrochem Soc 1993;140:1526e33.
[10] Fuller TF, Doyle M, Newman J. J Electrochem Soc 1994;141:1e10. Greek
[11] Haran BS, Popov BN, White RE. J Power Sources 1998;75:56e63.
[12] Ning G, Popov BN. J Electrochem Soc 2004;151:A1584e91.
l: Lagrange multiplier
[13] Santhanagopalan S, Guo Q, Ramadass P, White RE. J Power Sources 2006;156:
u: angular frequency, rads
620e8.
Fx : potential transfer function of the phase indicated by subscript x, V
[14] Romero-Becerril A, Alvarez-Icaza L. J Power Sources 2011;196:10267e79.
[15] Gu WB, Wang C, Liaw B. J Electrochem Soc 1998;145:3418e27.
Subscript/superscript
[16] Subramanian VR, Diwakar VD, Tapriyal D. J Electrochem Soc 2005;152:
A2002e8.
[17] Smith KA, Rahn CD, Wang C-Y. J Dyn Syst Meas Control 2008;130. 011012: e: pertaining to the electrolyte phase
1e011012:8. s: pertaining to the solid phase
[18] Forman JC, Bashash S, Stein JL, Fathy HK. J Electrochem Soc 2011;158: [: matrix with top block row deleted
A93e101. Y: matrix with bottom block row deleted
[19] Ramadesigan V, Boovaragavan V, Carl Pirkle JJ, Subramanian VR. J Electrochem ⊥: orthogonal projection
Soc 2010;157:A854e60. y: Moore-Penrose pseudo inverse

Comparing Four Model-Order Reduction Techniques, Applied To

Uploaded by

Copyright:

Available Formats

Comparing Four Model-Order Reduction Techniques, Applied To

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

Comparing Four Model-Order Reduction Techniques, Applied To

Uploaded by

Copyright:

Available Formats

eTransportation 1 (2019) 100009

Contents lists available at ScienceDirect

Comparing four model-order reduction techniques, applied to

Fig. 1. Realization algorithms summary ﬂowchart.

continuous-time transfer functions may be converted readily to

The summand for k ¼ 0 corresponds to the original frequency

Fig. 3. Actual and estimated discrete-time magnitude and phase response.

where the very compact notation Z ½ , means: “ﬁnd the X

where gs ½k is the discrete-time sampled step response of the sys- 2 3 8

where N is the number of points chosen for the underlying

CRA is to derive a reduced-order continuous-time state-space þ ðjuÞi1 DIm :

where Ac 2Rnn , Bc 2Rnm , C2Rqn and D2Rqm . The CRA then 2 3

where continuous-time extended-observability matrix Gi has the

    This deﬁnition allows for the determination of a well-

order and truncate to form S1 ; S2 can be considered essentially

We convert the continuous-time Ac matrix to discrete time via a Yhi

1. Initialization 3.3.3. Summarizing the CRA

G c ðu1 Þ / G c ðuN Þ ¼ O i X c ðu1 Þ / X c ðuN Þ

observability matrix of the system. we get

where we think of S2 being essentially zero. L22 can be well

4. Solving for the state-space B matrix matrix equation:

This paper has presented four methods to convert a user-

dental transfer functions describing the internal variables of a

~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in

~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in

This paper has introduced four different realization algorithms

You might also like

CRA is to derive a reduced-order continuous-time state-space þ ðjuÞi1 DIm :

This deﬁnition allows for the determination of a well-

~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in

~ s;e ðx; zÞ =Iapp ðzÞ at four different locations in