Computers and Electrical Engineering 112 (2023) 108983

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

Adaptive energy management strategy of plug-in hybrid

electric bus
Juanying Zhou a, b, *, Lufeng Wang c, Lei Wang d, Jianyou Zhao a
a
School of Automobile, Chang’an University, Xi’an, Shaanxi 710064, China
b
College of Automotive Engineering and General Aviation, Shaanxi Vocational &Technical College, Xi’an, Shaanxi 710038, China
c
College of Automobile, Shaanxi College of Communication Technology, Xi’an, Shaanxi 710018, China
d
School of Automobile & Rail Transportation, Tianjin Sino-German University of Applied Sciences, Tianjin 300350, China

A R T I C L E I N F O A B S T R A C T

Keywords: In this study, the online application of a new energy management strategy (EMS) is proposed for a
Automotive engineering power-split plug-in hybrid electric bus (PHEB) to improve its system’s fuel economy and reduce
Plug-in hybrid electric bus power battery degradation. The new EMS employs Pontryagin’s minimum principle (PMP) and
Energy management strategy
the power battery state of charge (SOC) feedback method to adjust the covariates online. A
Pontryagin’s minimum principle
model-free adaptive control (MFAC) EMS was designed and constructed according to the compact
Fuel consumption
Model-free adaptive control form dynamic linearization model to further obtain accurate SOC feedback; the time–domain
cumulative as well as covariate upper and lower limit constraints were used to obtain a real-time
MFAC-III strategy for online applications. The average reduction in battery capacity loss on all ten
routes was 8.19 %, resulting from the correlation between the MFAC-III EMS and MFAC EMS;
evaluations conducted on both strategies indicated that the MFAC-III outperforms the MFAC
strategy in hardware-in-the-loop.

1. Introduction

The inability of conventional fuel vehicles to adapt to real-time driving situations affects their fuel economy and corresponding
emissions [1]. To solve this problem effectively, recent research in transportation has focused on developing new power systems and
advanced energy-saving control technologies to achieve high efficiency, safety, and cleanliness. Hybrid power systems, through the use
of advanced technologies, are loaded with at least two power sources of different operating characteristics and one clean power source,
thereby improving energy conservation and reducing emissions in vehicles [2]. The energy management strategy (EMS) in these
vehicles distributes power or torque output from each power source according to the vehicle’s power requirements, the operational
state of the power system, and the differences in efficiency, transient response characteristics, and load energy of each power source.
This conserves power, improves the system’s fuel economy, and reduces emission as well as battery degradation through single or
multiple optimizations. This is the core technology of hybrid vehicle development and design [3]. According to different control
methods and the effects of EMSs, recent EMSs are mainly rule- and optimization-based. Rule-based EMSs include deterministic rules
and fuzzy control strategies while optimization-based EMSs can employ various strategies such as global optimization, online or
dynamic optimization, and intelligent optimization techniques [4].
Rule-based EMSs may not perform optimally under certain conditions, thereby reducing their adaptability and effectiveness [5–7].

* Corresponding author at: School of Automobile, Chang’an University, Xi’an, Shaanxi 710064, China.
E-mail address: zhoujy@chd.edu.cn (J. Zhou).

https://doi.org/10.1016/j.compeleceng.2023.108983
Received 18 May 2023; Received in revised form 9 October 2023; Accepted 10 October 2023
0045-7906/© 2023 Elsevier Ltd. All rights reserved.
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

They often need to be combined with other intelligent control algorithms to further improve and enhance their control effects [8,9].
The global optimization algorithms can obtain global optimal solutions but are problematic for creating challenges during online
applications. The EMS that relies solely on fixed regulations may combine uncomplicated and instantaneous algorithms but lacks
adaptability to changing operating conditions, which does not guarantee the most efficient control outcomes [6], and the algorithm
often needs to be combined with other intelligent control algorithms to further improve and enhance its control effects [8]. Pon­
tryagin’s maximum principle (PMP) is a mathematical optimization method that is useful for tackling optimal control problems. The
algorithm obtains the optimal global solution by modeling a Hamiltonian function with the best costate variable [10,11]. Hence, the
PMP algorithm can be transformed into an instantaneous optimal control algorithm to obtain optimal global solutions by dynamically
adjusting the costate variable [12,13]. Enhancing the ability of the PMP-based approach to adapt to online applications has become a
major research focus. An online EMS is optimized by adjusting the costate variables of the PMP-based strategy with the power battery
SOC feedback [14]. Therefore, a PMP-based strategy tailored to the online adaptation of the costate variable by power battery SOC
feedback is proposed. This strategy is capable of adaptively updating and balancing key hybrid electric vehicle (HEV) control pa­
rameters [15]. The proposed adaptive PMP-based strategy containing closed solutions of costate variables does not depend on known
operating conditions [16]. A new adaptive PMP-based approach was suggested for HEVs by using the particle swarm optimization
(PSO) algorithm to classify the operating modes and the Markov model for forecasting velocity and driving behavior across various
operational conditions [17].
To sum up, researchers have conducted several studies on the energy management of hybrid power systems. Adaptive PMP-based
strategies become the focus of recent energy management research; developing a more adaptive online EMS and determining an
optimal costate variable has been the priority.
This paper is organized as follows: Section 2 describes the powertrain system of the model in detail. The PMP-based approach is
presented as a proposed strategy in Section 3. In Section 4, an initial value selection scheme of the adaptive PMP-based strategy is
introduced. In Section 5, an adaptive PMP-based strategy with improved strategies is discussed. Section 6 highlights the efficiency of
the approach utilizing adaptive PMP. Finally, Section 7 presents the discussion and conclusions.

2. Overall PHEB model

The primary aim of this study was to consider a power-split plug-in hybrid electric bus (PHEB). Fig. 1 shows the schematic diagram
of the PHEB powertrain structure; its main components include a drive and generator motor, an engine, a power battery pack, and
double planetary gears (PG1 and PG2).

2.1. Longitudinal dynamics model of entire vehicle

This study mainly analyzed the EMS of HEVs with supercapacitors in their hybrid power systems; only the kinematic characteristics
along the longitudinal driving direction were considered. The equation of longitudinal force state while driving was adopted as the
required power model for the entire vehicle [18]. This study utilized the backward simulation method to estimate the amount of
energy needed for vehicular propulsion according to the velocity and acceleration patterns in driving cycles. The resulting equation is
expressed as:
( )
u(t) ACD u(t)2 du
Pdri (t) = mgf cosθ + + mδcosθ + mgsinθ , (1)
3600ηm 21.15 dt

where Pdri is the required power, (kW); u is the vehicle velocity, (km/h); m is the overall vehicle weight, (kg); ηm is the efficiency of the
drive powertrain; g is gravitational acceleration, (m/s2); f is the rolling friction coefficient; θ is the road ramp angle, (◦ ); A is the

Fig. 1. Configuration of powertrain structure of power-split hybrid.

2
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

windward area, (m2); CD is the wind resistance coefficient; and δ is the rotational inertia conversion factor.
Table 1 illustrates the main parameters of the PHEB model.

2.2. PHEB model

This study employs the forward simulation method to construct the PHEB model. The amount of power required by the vehicle is
determined by simulating the driver’s velocity. The Hybrid Control Unit (HCU) regulates torque distribution; it drives the powertrain
components to calculate the torque while driving, which is then iterated through the entire vehicle model to determine the current
velocity. This method serves as the basis for developing and optimizing the PHEB EMS as well as the foundation for developing and
optimizing energy.

2.2.1. Driver model

The driver model controls vehicle velocity by using the error obtained through expected and actual velocities. This can be expressed
as:
⎧ ∫
⎪
⎪ dev (t)
⎪ ΔTdri = Kp ev (t) + Ki ev (t)dt + Kd
⎪ ; ev (t) ≤ 0
⎪
⎪ dt
⎨ ∫
dev (t) , (2)
⎪
⎪ ΔT = Kp ev (t) + Ki ev (t)dt + Kd ; ev (t) > 0
⎪ brk
⎪ dt
⎪
⎪
⎩
ev (t) = vtar (t) − v(t)

where ΔTdri and ΔTbrk are the drive torque and brake torque adjustments, respectively; Kp is the proportionality coefficient; Ki is the
integration coefficient; Kd is the difference coefficient; ev(t) is the difference in vehicle velocity during the driving time t; vtar(t) is the
vehicle target velocity at moment t; and v(t) is the vehicle velocity at moment t.

2.2.2. Drive motor and generator motor models

A quasi-transient motor MAP model based on experimental data is used to construct the motor model and can be expressed as:
⎧
⎪
⎪
⎪ Tmotor out = Tmotor − Imotor ⋅ ω̇motor
⎪
⎪
⎪
⎪
⎪ ηmotor = f (Tmotor , ωmotor )
⎪
⎪ /
⎪
⎪
⎨ Pmotor = Tmotor ⋅ ωmotor ⋅ ηnl motor 9550, if Tmotor ≤ 0, nl = 1; Tmotor > 0, nl = − 1
Tgen = Ts1 − Igen ⋅ ω̇gen (3)
⎪
⎪ ( )
⎪
⎪
⎪
⎪ η
⎪ gen
= f T gen , ωgen
⎪
⎪ /
⎪
⎪ ω gl
⎪ P
⎩ gen = T gen ⋅ gen ⋅ ηgen 9550, if Tgen ≥ 0, gl = 1; Tgen < 0, gl = − 1,

where Tmotor, Pmotor, ωmotor, and Imotor are the torque, power, velocity, and rotational inertia of the drive motor, respectively; Tmotor _out is
the drive motor output torque; Tgen, Pgen, ωgen, and Igen are the torque, power, velocity, and rotational inertia of the generator motor,
respectively; ηmotor and ηgen are the drive motor and generator motor efficiencies, respectively; ω̇motor is the drive motor instantaneous
velocity; and gl and nl are the values under certain conditions. Fig. 2 shows the drive motor and generator motor characteristic map.

2.2.3. Engine model

The engine model is developed using the experimental modeling approach and can be expressed as:

Table 1
Main parameters of PHEB model.
Parameter Value

Vehicle body mass (kg) 13,050

Full load mass (kg) 18,000
Vehicle dimensions (m3) 12 × 2.55 × 3.2
Windward area (m2) 6.00
Wind resistance coefficient 0.55
Rolling friction coefficient 0.015 ~ 0.02
Wheelbase (m) 6.05
Rotational inertia conversion factor 1.1
Wheel rolling radius 0.512

3
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 2. Motor characteristic maps: (a) Drive motor characteristic map; (b) Generator motor characteristic map.

⎧ /
⎪
⎪ Peng = Teng ⋅ ωeng 9550
⎪
⎪
( − Ieng ⋅ ω̇)eng
⎨ Tc1 = Teng
ṁeng = f Teng , ωeng , (4)
⎪
⎪
⎪ Ceng = ṁfuel ⋅ Peng/(
⎪ )
⎩
ηeng = 3.6 × 106 ṁfuel ⋅ HLHV

where Peng is the engine power; Teng is the engine torque; ωeng is the engine velocity; Ieng is the engine rotational inertia; Tc1 is the
planetary gearbox torque of the PG1; ω̇eng is the instantaneous engine velocity; ṁfuel is the instantaneous engine fuel consumption rate;
Ceng is the instantaneous engine fuel consumption; ηeng is engine thermal efficiency; and HLHV is the lower heating value of the fuel.
Fig. 3 shows the engine characteristic map.

2.2.4. Battery model

1 Battery Circuit Model

The Rint model of the internal resistance voltage is adopted for the battery cell model; the consistency between the battery cells
should be considered during research. Fig. 4 presents the battery model.
The open circuit voltage is expressed as:

Fig. 3. Engine characteristic map.

4
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 4. Battery circuit model.

Voc (SOC(t)) = Rb (SOC(t))Ib (t) + Ub (t). (5)

The battery power is expressed as:

Pb (t) = Rb (SOC(t))I 2b (t) + Ub (t)Ib (t), (6)

where Voc is the overall voltage rating of the battery pack, (V); SOC is the current battery charge level; Rb is the equivalent internal
resistance of the battery, (Ω); Ib is the battery pack current, (kW); Ub is the battery output voltage, (V); and Pb is the power of the battery
pack, (kW).
The battery current can be expressed by combining Eqs. (5) and (6) as:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Voc (SOC(t)) − V 2oc (SOC(t)) − 4Rb (SOC(t))Pb (t)
Ib (t) = . (7)
2Rb (SOC(t))
The present SOC of the power battery is determined using a calculation method and utilizing the formula:
∫ tf
Ib (τ)
SOC(t) = SOC(t0 ) − ηℓb (t) dτ, if Ib (t) > 0, ℓ = 1; Ib (t) ≤ 0, ℓ = 1, (8)
t0 Qbat

where t0 and tf are the current cycle start and end times; ηb is the battery efficiency; ℓ is the value under certain conditions; and Qbat is
the battery capacity.

1 Battery Aging Model

The battery aging model mathematically expresses the battery aging process by loss in percentage of battery capacity over time.
This is often expressed as:
( )
− Ea
Qloss = Bexp (Ah)z , (9)
R(T + 273.15)

where Qloss is the percentage of power battery capacity loss, (%); B is the pre-exponential coefficient relating to the battery SOC and
takes the value of 31,630; Ea is the activation energy, relating linearly to the battery charge/discharge multiplier; R is the gas molar
constant, which takes the value of 8.314 J/(mol⋅K); Ah is the total ampere-hour throughput, (A⋅h); and z is the power coefficient.
According to the battery’s lifespan experimental data, the parameter identification of Eq. (8) is expanded, and the obtained battery
capacity loss is expressed as:
( )
− 31700 + 163.3C
Qloss = Bexp (Ah)z , (10)
R(T + 273.15)

where C is the battery discharge multiplier. When the battery capacity loss reaches 20 % below standard, the battery life is deemed to
have reached its end of life (EOL). In common driving scenarios, when the battery is charged or discharged at 0.3 times its normal
capacity per hour and is operating under standard temperature conditions of 25 ◦ C (298.15 K), the total ampere-hour throughput of the
power battery when it reaches EOL is:
∫ EOL
|Inom (t)|dt
Γnom = , (11)
0 3600

where Γnom is the power battery termination throughput, (A⋅h), and Inom is the battery current in standard driving cycles, which takes
the value of 2.5 A. The actual total operating ampere-hour throughput of the power battery, Γ real, is expressed as:
∫ EOL
|Ireal (t)|
Γreal = dt. (12)
0 3600
The introduction of the aging factor, σ, helps determine how ambient temperature and charge/discharge rates affect battery life

5
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

degradation: this is the velocity of battery life degradation in actual driving cycles compared to the same in standard driving cycles.
This is expressed as Eq. (13):
∫ EOL
Γ |Inom (t)dt|
σ(I(t), T(t)) = nom = ∫0EOL , (13)
Γreal |Ireal (t)dt|
0

where Inom is the standard number of charging battery and takes the value of 0.35. expressed as: The power battery degradation degree
is
∫ tf
Ah = σ (Ib , T, SOC)|Ib (t)|dt. (14)
t0

This study considered the general technical requirements and other similar standards of a DB31T 306-2020 public bus based on the
matching power system parameters of the China city bus cycles (CCBC). The PHEB powertrain parameters are outlined in Table 2; the
double planetary gear row-coupling mechanism and driveline parameters are selected empirically [18].

2.3. Driving cycle

A driving cycle is a standard technology that describes the velocity and time of a vehicle driving in a specific traffic environment; it
is used to develop and evaluate new models as well as improve fuel economy and reduce emissions [19]. The common standard driving
cycles as seen in Europe, the United States, and Japan used to evaluate the performance of new energy vehicles need to be more
optimal [20]. This study investigates PHEB EMS-specific bus route operating conditions.
Following the line strength and area coverage theory [21], a bus line in a city was selected as the test line; this is illustrated in Fig. 5.
This representative route connects the northeastern and southern suburbs of the city, basically covering the traffic-congested parts of
the city, the low-velocity sections of secondary roads, and the medium-velocity sections of the main roads.
The entire bus line operating route was defined once a round trip was completed to obtain the data. The route has a one-way travel
distance of 24.33 km and 44 stations. The bus velocity data of its daily operation on this testing route were collected by CAN-BUS and
sampled continuously for 30 days and the wavelet filtering method was used to solve the noise issue with the obtained data.

2.3.1. Synthetic driving cycle

The data on vehicle velocity are segmented at stations throughout the vehicles’ return routes, and the information from each
interval station is kept in a database. The Markov Chain Montecarlo algorithm [22] helps express a stochastic simulation by syn­
thesizing the vehicle velocity data for each storage unit and determining the operating conditions by connecting that of neighboring
stations in sequential order. The synthetic driving cycle performance is then validated, and the data are presented in Fig. 6.

2.3.2. Actual driving cycle

Ten sets of actual driving cycle routes were randomly selected for testing to validate the impact of the EMSs. Fig. 7 presents the
actual driving cycle data.

3. Pontryagin’s minimum principle

The approximate global optimization of the EMSs is achieved by the PMP-based strategy [23]. The objective is to reduce the total
energy consumption of the engine and power battery throughout the entire journey. This can be expressed as:

Table 2
Basic parameters of PHEB powertrain.
No. Power component Parameter Value

1 Engine Maximum power 102 kW

2 Drive motor Rated power 68 kW
Peak power 106 kW
3 Generator motor Rated power 60 kW
Peak power 105 kW
Single unit nominal capacity 50 Ah
4 Power battery Rated voltage 502 V
Continuous discharge capacity 3C
Instantaneous discharge capacity 5C
5 Battery pack Voltage 502.4 V
Number of series connection 157
Number of parallel connection 1
6 Double planetary gear PG1 characteristic parameters 2.6
PG2 characteristic parameters 2.6
7 Driveline parameter Main gear ratio 3.41

6
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 5. Testing route.

Fig. 6. Synthetic driving cycle.

Fig. 7. Actual driving cycle.

7
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

∫ tf ( )
Jmin = H Peng (τ), Pb (τ), τ dτ, (15)
t0

where Jmin is the objective function and H (⋅) is the Hamilton function. Concerning hybrid power system strategies, the goal of EMSs is
to create a power distribution trajectory, which reduces the engine’s fuel consumption to the barest minimum in the hybrid system
while providing the right amount of power required by the vehicle. The Hamilton function is expressed as:
Pb (x(t), u(t), t)
H(x(t), u(t), t) = ṁfuel (u(t), t) + λT (t) ⋅ , (16)
Qbat ⋅ Voc

where x(t) is the control variable; λT is the status variable system covariance; and u(t) is the status variable.
With the introduction of the relevant costate variable, ̃ λ and H (⋅) can be expressed as:
⎧
⎪
⎪ ̃Pb (x(t), u(t), t)
⎨ H(x(t), u(t), t) = ṁfuel (x(t), u(t), t) + λ
⎪
HLHV
( ) T
( ) . (17)
⎪
⎪ λ t ⋅ HLHV
⎪̃
⎩ λ t =
Qbat ⋅ Voc

Then, the best control input is achieved through the mathematical formulation below:
u∗ (t) = argmin{H[x∗ (t)], u∗ (t), λ∗ (t), t]} (18)
u∈U

Furthermore, the constraint conditions are as follows:

⎧
⎪
⎪ Teng min ≤ Teng ≤ Teng max
⎪
⎪ ωeng min ≤ ωeng ≤ ωeng max
⎪
⎪
⎪
⎪
⎨ Tmotor min ≤ Tmotor ≤ Tmotor max
ωmotor min ≤ ωmotor ≤ ωmotor max , (19)
⎪
⎪
⎪
⎪ Tgen min ≤ Tgen ≤ Tgen max
⎪
⎪
⎪
⎪ ω ≤ ωgen ≤ ωgen max
⎩ gen min
SOCmin ≤ SOC ≤ SOCmax

where the subscripts _min and _max are the smallest and largest values of the variables being represented, respectively. A multi-island
genetic algorithm can be utilized to solve the appropriate, relevant costate variable known as driving cycles. However, the online
performance of the PMP-based strategy remains problematic; thus, an optimally constant costate variable determined by the
dichotomous method is used to improve the solution. The constant costate variable of the PMP-based strategy can be approximately
equivalent to the optimal global algorithm [24]. However, the optimal covariates were solved while considering known working
conditions. The comparison analysis of the performance of the PMP-based and dynamic programming (DP)-based strategies under
actual driving cycles is presented in Table 3. The optimal covariates are different for different road conditions of the same route.
Therefore, it is crucial to design an adaptive PMP-based strategy to overcome the limitations of constant costate variables to adjust
them dynamically.

4. Adaptive PMP-based strategy

The adaptive adjustment of the costate variables demands predicting working conditions, recognizing driving cycles, and SOC
feedback [25]. This study uses the principle of adaptive costate variable control based on SOC feedback, which is illustrated in Fig. 8.
This can be expressed in an equation as:

Table 3
Performance comparison of PMP-based and DP-based strategies under actual driving cycles.
No. Optimal costate variable SOC final value Fuel consumption/(L) Energy consumption (%)

PMP-based strategy DP-based strategy PMP-based strategy DP-based strategy

CYC1 2.338 0.299 0.301 3.749 3.673 2.07

CYC2 2.319 0.297 0.301 4.089 3.943 3.70
CYC3 2.325 0.305 0.302 3.853 3.719 3.60
CYC4 2.325 0.300 0.302 4.206 4.114 2.24
CYC5 2.303 0.301 0.301 4.036 3.919 2.99
CYC6 2.331 0.299 0.301 4.059 3.885 4.48
CYC7 2.331 0.298 0.301 3.863 3.730 3.57
CYC8 2.294 0.304 0.302 4.183 4.005 4.44
CYC9 2.275 0.305 0.301 4.356 4.181 4.19
CYC10 2.319 0.302 0.302 4.357 4.222 3.20

8
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 8. Adaptive control principle based on SOC feedback.

∫ t
λ(t) = λ0 + Kp (SOCtar − SOC(t)) + Ki (SOCtar − SOC(τ))dτ, (20)
0

where λ0 is the initial value of the costate variable and SOCtar is the target value of the battery SOC. Accurate values of the initial
costate variable can improve the practicality of the feedback control and reduce computational difficulty. The costate curves for
different values of λ0 are shown in Fig. 9. If the proportional integral (PI) controller’s settings remain constant, different numerical
values of λ0 will have a greater impact on the costate trajectory. When λ0 = 0, the controller increases the costate from 0, and the
costate trajectory fluctuates even more; when λ0 = 2.373 (the optimal costate value), the controller adjusts the costate around its
optimal value within a short range, and the costate trajectory fluctuates less.
The performance of the whole vehicle with different values of λ0 is shown in Fig. 10. The figure captures the initial values of λ0 and
its impact on the battery SOC trajectory, vehicle fuel consumption, and power battery capacity loss. The figure shows that using the
optimal costate variable as the initial values of λ0 can reduce fuel consumption and battery life loss. However, solving the optimal
costate variable here depends on offline conditions with known driving cycles. Therefore, it is required to investigate further and
ascertain the numerical amount of λ0 for dynamic conditions.
Eq. (16) shows the costate variable as a weighting factor for fuel–electricity and electricity–energy conversion efficiency. The
fuel–electricity conversion efficiency can be approximated by the weighted cumulative conversion efficiency, which is expressed as:
∑
t
Pgen (t)
(21)
0
ηme (t) = / , Pmotor (t) < Pgen (t)&Peng (t) > 0,
∑
t
Ps1 (t) ηeng (t)
0

where ηme(t) is the fuel–electricity conversion efficiency at a given moment, t, and Ps1 is the input power of the sun gear of PG1. By
introducing the equivalence coefficient, wme, as the alternative to the weighted cumulative efficiency, the equation can be expressed as:
wme (t) = 1/ηme (t). (22)
The trajectory of the wme value as a function of the driving operating conditions is presented in Fig. 11. Its value trajectory is
between 2.355 and 2.357, which is close to that of the optimal costate variable. ωme can be considered as the approximation of the
optimal costate variable.
Fig. 12 presents the effects of the different λ0 values; the comparison of the effects of λ0 taking values of 2.373 as the optimal
constant costate variable and wme are also shown. The figure shows that the performance of λ0 = wme (green line) is slightly better than
that of λ0 = 2.373 (blue line) before a 33 km trip.
Conclusively, the effect of λ0 = wme is similar to that of the optimal constant costate variable and does not depend on predicted
complete driving cycles. It has a good online application to adjust according to the vehicle’s demanded power. Thus, this is the
preferred initial value for SOC feedback and for the adaptive control strategy.

Fig. 9. Costate curves at different values of λ0.

9
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

(caption on next page)

10
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 10. Performance of entire vehicle with different λ0 values: (a) Battery SOC trajectory for different λ0 values; (b) Fuel consumption for different
λ0 values; (c) Battery capacity loss for different λ0 values.

Fig. 11. Equivalence coefficient.

Fig. 12. Effects for different values of λ0.

5. Model-free adaptive control-based strategy

The model-free adaptive control (MFAC) is an algorithm proposed by Professor Han Zhongsheng based on data-driven control
applying online data. This study adopts the compact form dynamic linearization (CFDL) method as the mathematical representation of
data structure because of its simplicity and wide applicability [24].

5.1. MFAC control law based on CFDL design

The MFAC law for the CFDL data model is expressed as:
ρk φ(k)
u(k) = u(k − 1) + (y∗ (k + 1) − y(k)), (23)
μk + φ(k)φT (k)

where ρk is the step size factor and it falls under the range of (0,1]. The expression of the criterion function using the pseudo-gradient
estimation algorithm is:

J(̂ ̂ (k)Δu(k − 1)|2 + μk |̂

φ (k)) = |y(k) − y(k − 1) − φ φ (k) − φ
2
̂ (k − 1)| . (24)
The expression above takes the partial derivative of φ
̂ (k) in Eq. (24) to form ∂J(̂ ̂ (t) = 0 and obtain:
φ (t))/∂ φ
ηk Δu(k − 1)
φ
̂ (k) = φ
̂ (k − 1) + (Δy(k) − φ
̂ (k − 1)Δu(k − 1)). (25)
μk + |Δu(k − 1)|2
To introduce an estimated reset mechanism, Eq. (25) can be expressed as:

11
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

{
φ
̂ (k) = φ φ (k)| ≤ εor|Δu(k − 1)| ≤ ε
̂ (0), if |̂
, (26)
φ
̂ (k) = φ
̂ (0), ifsign(̂φ (k)) ∕
=φ̂ (0)

where ε is a sufficiently small positive number.

5.2. MFAC adaptive strategy design

The MFAC model is expressed as:

{
SOC(k + 1) = SOC(k) + φ(k)Δλ(k)
, (27)
Δλ(k) = λ(k) − λ(k − 1)

where SOC(k) is the battery charge level at the present moment; SOC(k + 1) is the battery charge level (SOC) at the next moment; Δλ(t)
is the increment of the costate at the present moment; λ(k) is the costate at the present moment; λ(k-1) is the costate at the last moment.
The expression for the MFAC by a CFDL-based algorithm is:
ρk φ(k)
λ(k) = λ(k − 1) + (SOC∗ (k + 1) − SOC(k)), (28)
μk + φ(k)φT (k)

where SOC*(k + 1) is the target battery SOC values at the next moment; μk is the weight factor: μk is greater than 0; Φ(k) is the pseudo-
partial derivative. The expression to introduce the initial costate variable, λ0(k), is:
̃
λ(k) = λ0 (k) + λ(k). (29)
The expression for the estimation algorithm of the pseudo-partial derivative is:

ηk Δ̃λ(k)
φ
̂ (k) = φ
̂ (k − 1) + ̂ (k − 1)Δ̃
(SOC(k) − φ λ(k − 1)). (30)
μk + |Δ̃λ(k − 1)|2
The principle of the MFAC adaptive strategy is shown in Fig. 13.
The algorithm flow of the MFAC adaptive strategy is as follows:

(1) Determine the system’s initial value; set the initial values of SOC0, λ0, and Φ(k); and set µk > 0 and µk ∈(0,1).
(2) Acquire the error value; determine the error by acquiring the difference between SOC*(k + 1) and SOC(k).
(3) Perform an online estimation; perform online analysis of φ ̂ (k) according to Eq. (25) and introduce the estimation reset
mechanism as in Eq. (26).
(4) Solve for control variables; replace φ
̂ (k) in Eq. (30) with Φ(k) from Eq. (23); and use the solved λ(k) as the control variable of the
PMP algorithm.
(5) Update k = k + 1. If k ≤ tf, skip stepping (2). If k > tf, the loop ends.

5.3. MFAC adaptive strategy performance analysis

Any Xi’an bus route can be selected for testing the proposed strategy. The PHEB fuel consumption analysis for ten randomly
selected actual driving cycles using various methods is shown in Fig. 14. The analysis shows that the optimal constant costate variable
best improves fuel efficiency and enables the PMP to allocate power-system coupling more efficiently. The rule-based CDCS strategy
shows the worst performance because although it keeps the engine working on the optimal curve, it does not effectively consider the
motor system’s efficacy; the MFAC adaptive strategy adjusts the costate variable to make it close to the optimal value, signifying the

Fig. 13. Principle of MFAC adaptive strategy.

12
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 14. Fuel consumption analysis using different strategies.

best fuel efficiency. CDCS and MFAC manifest the PI-based adaptive approach.
Concerning fuel economy analysis, a gap exists between the MFAC adaptive strategy and the optimal constant of the PMP-based
strategy. Therefore, in subsequent sections, improvements will be made to the MFAC adaptive strategy.

5.4. Improvement of MFAC adaptive strategy

5.4.1. MFAC-II adaptive strategy

Buses on urban routes park often because of having to stop at red lights, giving way to pedestrians, making station stops, and having
passengers scan their payment codes in turns. Fig. 15 depicts the parking period battery SOC trajectory of a PHEB under various costate
factors. If the costate variable is less than 2.4, then the engine does not generate electricity during parking, and the battery SOC does
not change. If the costate variable is greater than 2.4, the increased costate variable amplifies the equivalent fuel gained from battery
charging, increases the power output of the engine, then the generator charges the power battery while driving, and the battery SOC
rapidly increases.
To avoid problematic large increments of the costate variable during parking, Eq. (27) is changed and expressed as:
⎧
̃ ̃
⎨ λ(k) = λ(k − 1), v = 0&ΔSOC = 0
⎪
ρk φ(k) . (31)
⎪
⎩̃ λ(k) = λ0 (k) + λ(k − 1) + (SOC∗ (k + 1) − SOC(k)), else.
T
μk + φ(k)φ (k)

When the vehicle is parked (v = 0) and ΔSOC = 0 at a present time, the costate variable will remain in the pre-parking state to avoid
time–domain accumulation; this is the MFAC-II adaptive strategy.
Fig. 16 shows the effectiveness of the MFAC-II adaptive strategy in achieving better control performance across smooth, normal,
and congested road conditions. Under the three road conditions, the costate variable under the MFAC-II adaptive strategy hardly
changes, and the overall convergence time is shorter than in the MFAC adaptive strategy.

Fig. 15. Various battery SOC with different costate variables during PHEB parking.

13
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 16. Comparison of adjusted costate variables under three road conditions.

5.4.2. MFAC-III adaptive strategy

Changes in the costate variable affect battery charging and discharging as well as fuel efficiency, as shown in Fig. 17. Fig. 17(a)
shows that battery SOC changes with time and because of several factors, and the battery SOC is depicted with the ideal costate
variable represented by the red curve. Fig. 17(b) shows the gasoline usage for the whole vehicle and that the closer the costate variable
is to the ideal value, the less gasoline is used. This is because the PHEB uses less electrical energy while running, and the power battery
is strong enough to run at an inexpensive coupling ratio with the engine. This study shows that determining the ideal covariance is
impossible, but limiting the covariance to the red box, as in Fig. 17(b), improves fuel economy.
Fig. 18 shows the optimal costate variables for each selected route, and the minimum and maximum constant costate variables are
λmin = 2.273 and λmax = 2.387, respectively. Setting the adjustment boundary for the costate variable not only ensures that the
bounded range of the battery SOC is satisfied under different road conditions, but also manages fuel economy because of the many
changes to the costate variable.
If a constraint is added to Eq. (27), it will be expressed as:
{
̃
λ(k) = λmax − wme (k − 1)if ̃
λ(k) > λmax − wme (k − 1)
. (32)
̃
λ(k) = λmin − wme (k − 1)if ̃
λ(k) < λmin − wme (k − 1)
This is the MFAC-III adaptive strategy; the control performance of the MFAC-III adaptive strategy is verified in three typical road
conditions. Fig. 19 presents the performance comparison of the different strategies under the three typical road conditions.
Fig. 19(a) shows that the MFAC-III adaptive strategy effectively addresses the undesired charging and discharging of the power
battery during operation, resulting from excessive changes in the MFAC costate variable. Using the MFAC-III adaptive strategy, battery
SOC decreases slowly, local power increase is complemented by energy feedback or optimal coupling shunt, and the final battery SOC
values at the end of driving operations are all within the desired error range.
Fig. 19(b) shows that the MFAC-III adaptive strategy constrains the costate variable within the highest and lowest boundaries of
costate variable optimization determined for the selected route conditions; thus, the final fuel consumption is around the optimal
costate variable fuel consumption value. The fuel consumption rate is more efficient with the MFAC-III adaptive strategy than with the
optimal covariate under the three road conditions, and the improvement is more prominent.
Fig. 19(c) shows that the power battery life in the MFAC-III adaptive strategy is effectively improved. The MFAC-III adaptive
strategy hinders high-power charging and discharging and effectively reduces battery capacity loss.
Table 4 presents the performance evaluation of the three different strategies. The fuel consumption with the MFAC-III adaptive
strategy under the three road conditions is reduced by 3.11 %, 4.07 %, and 4.31 %, respectively. Unlike the MFAC adaptive strategy,
which enhances the system’s overall energy, resulting from the large amplitude of the costate variables, the MFAC-III adaptive strategy
ensures sustainable charging and discharging of the power battery, and the weighted circulation penalty is also reduced. The charge
and discharge power battery stability with the MFAC-III adaptive strategy reduces the weighted circulation penalty coefficient, and the
battery life is subsequently improved by lowering the ampere-hour throughput. Finally, with this strategy, the battery capacity losses

14
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 17. Effect of costate variables on battery SOC and fuel consumption: (a) Battery SOC trajectory with different costate variables; (b) Fuel
consumption variation for different costate variables.

Fig. 18. Optimal costate variables for selected routes.

15
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 19. Battery SOC, fuel consumption, and battery capacity loss of different strategies under three typical road conditions: (a) smooth road
conditions; (b) normal road conditions; (c) congested road conditions.

16
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Table 4
Performance comparison of three strategies.
Strategy Fuel consumption (L) Battery capacity loss (%)

Smooth Normal Congested Smooth Normal Congested

condition condition condition condition condition condition

Optimal constant costate variable 3.4972 4.2311 3.9377 0.0183 0.0252 0.0225
MFAC 3.6765 4.4833 4.1049 0.0206 0.0294 0.0249
MFAC-III 3.5622 4.3010 3.9758 0.0187 0.0261 0.0228
MFAC-III vs MFAC -3.11 % -4.07 % -4.31 % -9.22 -11.22 -8.43

under the three road conditions are reduced by 9.22 %, 11.22 %, and 8.43 %, respectively.
The MFAC-III adaptive strategy can ensure the discharge target is met for a PHEB, optimize fuel–electricity coupling depth, and
protect battery life sufficiently. It is also feasible as an online EMS for PHEB.

6. Hardware-in-loop simulation

The hardware-in-the-loop (HIL) simulation testbed is used to validate the proposed adaptive EMS and confirm its online efficiency
and efficacy under stochastic driving cycles.
Fig. 20 presents the HIL test system, primarily comprising a PC host computer, digital signal processing(DSP) controller, Speedgoat
baseline target machine, and target machine interface PC. The energy management system built by applying the MATLAB/Simulink
platform is divided into two parts: the PHEB model and EMS. The PHEB model is converted to the C++ code using the Simulink Coder
module of MATLAB/Simulink and compiled into an executable mldatx file, then downloaded into Speedgoat baseline. The control
strategy is generated in C code by MATLAB, and the Code Composer Speedgoat online simulation target machine and DSP controller
interact with the upper PC through CAN protocols to receive online signal interactions from the model and control strategy to transmit
the energy management system data running on the DSP to the PC. The host PC displays the energy management system data running
on the DSP to the PC and uses Simulink Data Inspector to observe and record the HIL results. The target machine interface can detect
the operating status of the Speedgoat target machine and adjust the model parameters online.

Fig. 20. HIL test platform.

17
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

The MFAC-III adaptive strategy was tested and validated using the constructed HIL testbed while considering ten randomly selected
operating conditions and bus routes.
The fuel consumption comparison for the MFAC-III and MFAC adaptive strategies is shown in Fig. 21. The MFAC-III adaptive
strategy improved fuel consumption than the MFAC did on all ten routes, especially under the CYC1, CYC4, and CYC6 road conditions.
The battery capacity losses for the MFAC-III and MFAC adaptive strategies are shown in Fig. 22. The MFAC-III adaptive strategy
showed more improvement compared to the MFAC adaptive strategy regarding battery capacity loss on all ten routes, especially under
the CYC4 and CYC6 road conditions.
Table 5 shows the performance comparison of the MFAC-III and MFAC adaptive strategies in actual road conditions.
When comparing the performance of the MFAC-III and MFAC adaptive strategies on fuel consumption, the highest improvement is
5.3 % in CYC1, and the lowest is 1.08 % in CYC10. A slight decrease is evident in the average fuel consumption at 3.3 %. Comparing the
MFAC-III and MFAC adaptive strategies on battery capacity loss shows an improvement of up to 25.49 % in CYC4 and a relatively
stable performance in CYC10, resulting in an average reduction of battery capacity loss by 8.19 % across all ten routes. The HIL testing
results show that the MFAC-III provides more significant improvement for both indicators than the MFAC adaptive strategy.

7. Conclusion

To improve fuel economy for PHEBs and determine a working online application model of PMP-based EMS, a model-free adaptive
PMP EMS is proposed in this paper. Comprehensive simulation analyses and HIL simulation tests are conducted. The key conclusions
can be summarized as follows:

(1) The optimal constant costate variable of the PMP-based strategy is obtained using an offline method in CCBCs. The PMP-based
strategy’s costate variables are adjusted using the battery SOC feedback-PI control approach. The equivalence coefficients are
employed to acquire the initial value selection scheme of the adaptive costate variables online. The CFDL method is used to
build the MFAC adaptive strategy to improve the online costate variables. The findings indicate that fuel efficiency under the
MFAC adaptive strategy exceeds the same under the CDCS and PI adaptive strategies but is slightly lower under the PMP-based
strategy with the optimal constant costate.
(2) The MFAC adaptive strategy is improved by time–domain accumulation and covariance upper and lower boundary constraints
to obtain the MFAC-III adaptive strategy. The comprehensive simulation analysis and HIL simulation result indicate that the rate
of fuel consumption and battery capacity loss using the MFAC-III adaptive strategy is reduced compared to the same while using
the MFAC adaptive strategy. Implementing the MFAC-III adaptive strategy has yielded positive results, with increased energy
efficiency and reduced battery ampere-hour throughput; this ultimately leads to battery longevity.
(3) The proposed MFAC-III adaptive strategy on battery life considers the expected discharge target of the hybrid system, the
fuel–electricity coupling depth, and sufficient battery protection characteristics; it is a feasible online EMS for PHEB.

Funding

This work was supported by the National Key Research and Development Program of China under grant number 2020YFB1600400,
the Science Program of Tianjin Municipal Education Commission under grant number 2019ZD20, and the Scientific Research Project
of Department of Transport of Shaanxi Province under grant number 18-27R.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be

Fig. 21. Fuel consumption comparison for MFAC-III and MFAC.

18
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

Fig. 22. Battery capacity losses for MFAC-III and MFAC.

Table 5
Performance comparison of MFAC-III and MFAC in actual road conditions.
Indicators Fuel consumption (L) Battery capacity losses (%)
Actual
MFAC-III MFAC MFAC-III vs MFAC (%) MFAC-III MFAC MFAC-III vs MFAC (%)
conditions

CYC1 3.893 4.111 -5.30 0.022 0.025 -10.48

CYC2 4.131 4.215 -1.99 0.022 0.024 -6.67
CYC3 3.913 3.990 -1.93 0.024 0.025 -3.98
CYC4 4.240 4.466 -5.06 0.023 0.031 -25.49
CYC5 4.166 4.311 -3.36 0.027 0.028 -5.36
CYC6 4.122 4.318 -4.54 0.025 0.032 -20.44
CYC7 3.983 4.132 -3.60 0.022 0.023 -5.13
CYC8 4.392 4.535 -3.15 0.027 0.027 -1.12
CYC9 4.441 4.607 -3.60 0.026 0.027 -2.26
CYC10 4.592 4.642 -1.07 0.024 0.024 0.00

directed to the corresponding authors.

CRediT authorship contribution statement

Juanying Zhou: Conceptualization, Validation, Writing – review & editing, Funding acquisition. Lufeng Wang: Software, Formal
analysis, Investigation, Resources, Writing – original draft. Lei Wang: Data curation, Visualization, Supervision. Jianyou Zhao:
Conceptualization, Methodology, Project administration, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

References

[1] Chen X, Liu Y, Wang Q. Pathway toward carbon-neutral electrical systems in China by mid-century with negative CO2 abatement costs informed by high-
resolution modeling. Joule 2021;5(10):2715–41.
[2] Martine CM, Hu X, Cao D. Energy management in plug-in hybrid electric vehicles: recent progress and a connected vehicles perspective. IEEE Trans Veh Technol
2016;66:4634–49.
[3] Sulaiman N, Hannan MA, Mohamed A. A review on energy management system for fuel cell hybrid electric vehicle: issues and challenges. Renew Sustain Energy
Rev 2015;52:802–14.
[4] Chen Z, Xiong R, Cao J. Particle swarm optimization-based optimal power management of plug-in hybrid electric vehicles considering uncertain driving
conditions. Energy 2016;96:197–208.
[5] Serrao L, Onori S, Rizzoni G. A comparative analysis of energy management strategies for hybrid electric vehicles. J Dyn Syst Meas Control 2011;133(3):200–5.
[6] Clara MM, Hu X, Cao D. Energy management in plug-in hybrid electric vehicles: recent progress and a connected vehicles perspective. IEEE Trans Veh Technol
2016;66(6):1–16.

19
J. Zhou et al. Computers and Electrical Engineering 112 (2023) 108983

[7] Chen Z, Xia B, You C. A novel energy management method for series plug-in hybrid electric vehicles. Appl Energy 2015;145:172–9.
[8] Jurik T, Cela A, Hamouche R. Energy optimal online navigation system. IEEE Intell Transp Syst Mag 2014;6(3):66–79.
[9] Park S, Ahn C. Power management controller for a hybrid electric vehicle with predicted future acceleration. IEEE Trans Veh Technol 2019;68:10477–88.
[10] Zeng YP, Huang ZK, Cai Y. A control strategy for driving mode switches of plug-in hybrid electric vehicles. Sustainability 2018;10:4237.
[11] Liu T, Hu X, Li SE. Reinforcement learning optimized look-ahead energy management of a parallel hybrid electric vehicle. IEEE ASME Trans Mechatron 2017;22
(4):1497–507.
[12] Deng T, Lu RZ, Li YN. Adaptive energy control strategy of HEV based on driving recognition by LVQ algorithm. Chin J Mech Eng 2016;27(03):420–5.
[13] Yin AD, Zhang LM. Study of energy management strategy for plug-in hybrid electric bus based on driving cycle recognition. J Hefel Univ Technol 2020;(02):
145–50 (Natural Science).
[14] Guo H, Sun Q, Wang C. A systematic design and optimization method of transmission system and power management for a plug-in hybrid electric vehicle.
Energy 2018;148:1006–17.
[15] Lee W, Jeoung H, Park D. An adaptive concept of PMP-based control for saving operating costs of extended-range electric vehicles. IEEE Trans Veh Technol
2019;68:11505–12.
[16] Nguyen BH, German R, Trovao JPF. on-line energy management of battery/supercapacitor electric vehicles based on an adaptation of Pontryagin’s minimum
principle. IEEE Trans Veh Technol 2018;68:203–12.
[17] He ZC, Yang G, Lu LG. Battery DC internal resistance test method based on the constant current external characteristic and SOC. J Tsinghua Univ 2015;5:9–18
(Sci & Technol).
[18] Yan NN. Historical data based online optimal control strategy for plug-in hybrid electric bus. Jilin University; 2019.
[19] Weiss M, Bonnel P, Hummel R. On-road emissions of light-duty vehicles in Europe. Environ Sci Technol 2011;45(19):8575–81.
[20] Liu X, Guo H, Cheng X. A robust design of the model-free-adaptive-control-based energy management for plug-in hybrid electric vehicle. Energies 2022;15(20):
7467.
[21] Jichao H, Tiezhu Z. Investigation of energy management strategy for a novel electric-hydraulic hybrid vehicle: self-adaptive electric-hydraulic ratio. Energy
2023;278:127582.
[22] Ronghui Z, Na W. Constrained hybrid optimal model predictive control for intelligent electric vehicle adaptive cruise using energy storage management
strategy. J Energy Storage 2023;65:107383.
[23] Vinot, E., Trigui, R. (2023). Rule-based energy management of hybrid electric vehicles focus on load following strategy. Encyclopedia of electrical and electronic
power engineering, 15, (pp. 529–41).
[24] Yanli Y, Xuejiang H. Hierarchical energy management control based on different communication topologies for hybrid electric vehicle platoon. J Clean Prod
2023;(137414):412.
[25] Shengwei Q, Hongwen H. Health-aware model predictive energy management for fuel cell electric vehicle based on hybrid modeling method. Energy 2023;278.
127919.

Juanying Zhou received her M.S. degree in Vehicle Engineering from Chang’an University, Xi’an, China, in 2009. She is currently pursuing her PhD in Vehicle Engi­
neering at Chang’an University, Xi’an, China. She has applied for six patents, among which three were authorized. She has published over 10 papers featuring her
research interest in theory and optimal energy control strategies for HEVs.

Lufeng Wang received his B.S. degree in Management from Chang’an University, Xi’an, China, in 2007. He also received his M.S. degree in Vehicle Engineering from the
same university in 2017. He has published seven papers, and his research interest is in theory and optimal energy control strategies for HEVs.

Lei Wang received his B.S. degree from Chang’an University in 2004 and his M.S. degree from the same University in 2008. He is pursuing a PhD in Vehicle Application
Engineering at Chang’an University. He has been granted eight patents and has published 10 papers. His-research interest is in new energy vehicles and intelligent
transportation.

Jianyou Zhao received his PhD degree from Chang’an University. He has published over 160 papers and 70 authorized inventions. Furthermore, he has participated in
more than 20 scientific research projects from the Natural Science Foundation, State 863 Plan Project, and the like. His-research interest is in theory and optimal energy
control strategies for HEVs.

20