Electrochimica Acta 295 (2019) 1057e1066

Contents lists available at ScienceDirect

Electrochimica Acta
journal homepage: www.elsevier.com/locate/electacta

A comparative study of global optimization methods for parameter

identification of different equivalent circuit models for Li-ion batteries
Xin Lai a, Wenkai Gao a, Yuejiu Zheng a, *, Minggao Ouyang b, Jianqiu Li b, Xuebing Han b,
Long Zhou a
a
College of Mechanical Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, PR China
b
State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing, 100084, PR China

a r t i c l e i n f o a b s t r a c t

Article history: A suitable model structure and matched model parameters are prerequisites for the precise estimation of
Received 29 April 2018 the battery states. Previous studies pay little attention to whether a parameter identification method is
Received in revised form suitable for a model. In this study, a comparative study is conducted by implementing model parameter
28 October 2018
optimization for nine equivalent circuit models using nine optimizers in the entire SOC area. The
Accepted 19 November 2018
Available online 23 November 2018
following conclusions are drawn: (1) PNGV and the exact algorithms are an ideal combination in the low
SOC area (0%e20%). (2) In the high SOC area (20e100%), exact algorithms are an ideal choice for the first-
order RC models, and PSO is an ideal identification algorithm for second-order RC models. For the third-
Keywords:
Equivalent circuit model
and fourth-order RC models, firefly algorithm has the highest accuracy with longer identification time.
Parameter identification (3) Firefly algorithm has the superior capacity to identify the accurate model parameters and PSO has the
Optimization algorithm best comprehensive performance for on-line parameter identification.
Metaheuristic algorithm © 2018 Elsevier Ltd. All rights reserved.
Li-ion battery

1. Introduction which can only be indirectly estimated via measurable external

characteristic parameters (e.g., voltage, current, and temperature).
Although conventional automobiles are highly convenient in Owing to the high nonlinearity of the battery system itself, accurate
daily life, they contribute to two major global concerns, namely estimation of the battery state is fairly difficult.
environmental pollution and energy depletion. Therefore, electric
vehicles (EVs) have witnessed rapid development in recent years 1.1. Review of the equivalent circuit model and parameter
and have emerged as the focus of R&D institutes and car manu- identification
facturers globally [1,2]. As the core component of EVs, batteries are
critical limiting factor for the large-scale development of EVs [3]. In The state estimation of LIB is essentially based on the battery
recent years, Li-ion batteries (LIBs) have gained many people's model. Therefore, an accurate battery model and matched model
attention with advantages of high energy density, long lifespan, and parameters are the prerequisite for the precise estimation of the
high efficiency [4,5]. To ensure the good performance of LIBs, a battery states. In practice, the equivalent circuit models (ECMs) are
battery management system (BMS) is required for the efficient, the most common battery models with critical advantages of
reliable, and durable battery operations. Estimation of the battery simple model structure and small computation amount. ECMs are
state is one of the important tasks of a BMS. Battery state estimation derived from the empirical knowledge by applying idealized circuit
involves the estimation of state of charge (SOC), state of health elements such as the resistance, capacitance and voltage source to
(SOH), state of function (SOF), state of energy (SOE), etc. Accurate replicate the electrical characteristics of LIBs [8,9]. The popular and
battery state estimation can contribute to reasonable management widely utilized ECMs contain the Rint model [10], the Thevenin
of LIBs so as to fully exploit them and prolong their lifespan [6,7]. model [11], the partnership for a new generation of vehicles (PNGV)
However, in practice, the battery state is a non-measurable variable, model [12], and the general nonlinear (GNL) model [13]. In fact,
these ECMs are based on the resistance-capacitance (RC) networks
with different orders. Depending on the dynamics of the load
* Corresponding author. profile and the required modelling accuracy, the order of the par-
E-mail address: [email protected] (Y. Zheng). allel RC networks may vary from first to fifth [3]. Moreover, RC

https://doi.org/10.1016/j.electacta.2018.11.134
0013-4686/© 2018 Elsevier Ltd. All rights reserved.
1058 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066

models combining with one-state hysteresis are reported by a se- identification algorithms to complex models or complex parameter
ries of studies [9,14,15]. Ref. [14] conducted a comparative study of identification algorithms to simple models. In other words, high
twelve ECMs for LIBs, and concluded that the first-order RC model model accuracy can only be achieved by using a suitable battery
is preferred for LiNMC batteries. Ref. [9] performed a comparison of model and appropriate model parameters. In this regard, model
eleven typical ECMs in terms of accuracy, stability, and robustness, parameter identification approach and model structure are equally
and the comparison results indicated that the first- and second- important. The aforementioned optimization approaches have
order RC models are the best choice owing to their balance of ac- various structure patterns and advantages, making them popular in
curacy and reliability for LiNMC batteries. Moreover, to further practical engineering applications. However, according to the no-
improve the dynamic performance of battery models, fractional- free-lunch (NFL) theorem [34], there is no algorithm for solving
order models were been explored for LIBs by simply replacing the all optimization problems. This means that an optimizer may
ideal capacitor in the ECM to a fractional element [16]. Ref. [5] perform well for a certain set of problems but fail to address
indicated that the complex ECMs with more model parameters can another set. ECMs have multiple structure patterns and the math-
lead to over-fitting problems with a certain probability, resulting in ematical equations of each model are different. Obviously, it is not
poor accuracy. Obviously, the accuracy of ECM is highly depends on reasonable to evaluate the model accuracy by using the same
the model structure and model parameters, and the improvement method to implement parameter identification and optimization
of model accuracy can only be achieved by appropriate model for all the models. In this work, to comprehensively determine the
structure and matched model parameters. For parameter identifi- most suitable parameter identification algorithm for each model,
cation of ECM, the most popular existing approaches include the nine representative optimization algorithms (EAs and MAs) are
genetic algorithm (GA) [17e20], particle swarm optimization (PSO) selected to implement parameter identification and optimization
algorithm [21e23], and the least-squares method [20,24]. To the for the ECMs of LIBs based on experimental data of a LiNMC under
best of our knowledge, there are few studies comparing these the new European driving cycle (NEDC). The most suitable
commonly used parameter identification algorithms for ECMs in parameter optimization algorithm for each model is determined
the literature. from the aspects of model accuracy, reliability, and computation
Parameter identification of ECMs is conducted to address the time, and the results can provide a reference for the selection of the
problem of parameter optimization, for which the exact algorithms model and parameter optimization algorithm for the state esti-
(EAs) [25] and metaheuristic algorithms (MAs) [26] are commonly mation of LIBs in the BMS.
adopted. EAs are deterministic solving methods that yield the same The accuracy, reliability and complexity of the battery model are
results for every calculation over a short time, but they are highly very important for a BMS. Due to the limitation of the micro-
vulnerable to falling into local optimal solutions and thus gener- controller's computational ability, the online identification of
ating relatively large solution errors. MAs are random solving model parameters requires higher accuracy and stability on the
methods that search for global optimal solutions. In this method, basis of lowest complexity, and complexity is the most critical. For
every calculation produces different results but the local optimal offline identification, accuracy is the most critical. Therefore, this
solutions can be avoided efficiently. Therefore, this method is paper aims to find out the best model and parameter identification
extensively used to solve practical engineering problems. MAs are method for the online identification and offline identification by
categorized into three main types [27,28]: evolution-, physics-, and comparative study, respectively.
swarm-based methods. Evolution-based methods are inspired by
the laws of natural evolution. The search process starts with a 1.3. Organization of the paper
randomly generated population that evolved over subsequent
generations. The best individuals are always combined together to The remainder of this paper is organized as follows. Section 2
form the next generation of individuals. At present, the most introduces nine ECMs and nine parameter identification and opti-
popular evolution-based method is GA [18], which simulates Dar- mization approaches in four categories. Section 3 describes the
win's theory of evolution. Physics-based methods basically imitate experimental equipment and the curves of current and voltage
the physical laws of nature; the most typical algorithm is the variation under the NEDC in the entire SOC area. Nine ECMs and
simulated annealing algorithm (SA) [29]. Swarm-based methods nine optimization approaches are compared for model accuracy,
are developed by imitating the social behaviors of animal groups, stability and computing time in Section 4. The optimization results
which have emerged as a research hotspot in recent years. The most are compared to determine the most suitable parameter identifi-
widely adopted swarm-based algorithms include the PSO algo- cation approach for each model. Finally, Section 5 states some
rithm [30], ant colony optimization algorithm (ACO) [31], artificial conclusions.
fish swarm algorithm (AFSA) [32], and firefly algorithm (FA) [33].

1.2. Contribution of the study 2. Model and method

In the most existing studies, model parameter identification of 2.1. Equivalent circuit models
LIBs has been conducted using a certain method; however, the
specific reasons for selecting the method have not been explained. A typical ECM generally used the RC network comprising re-
Obviously, the ECM model accuracy can be enhanced only by sistors and capacitors to simulate the dynamic characteristics of the
simultaneously modifying the model structure (to reduce the battery. A commonly used ECM structure with an n-RC network
model error) and enhancing the parameter identification accuracy called the nRC model hereafter, is shown in Fig. 1.
(to reduce the identification error). Previous studies on LIB models The output voltage of the battery can be expressed as:
have paid little attention to whether a parameter identification " #
X
n  
approach is suitable for a certain model. Moreover, comparisons of Vi ¼ OCV  Ii R0 þ Ri 1  et=Ri Ci (1)
such identification methods and algorithms have scarcely been i¼1
reported in the literature. Therefore, a suitable model and matched
model parameters are the basic prerequisites for accurate battery where Vi denotes the battery terminal voltage, OCV denotes the
state estimation. It is not appropriate to apply simple parameter open circuit voltage, R0 denotes ohmic resistance, Ri and Ci are the
 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066 1059

discharging, respectively.
From Eq. (2), it is seen that the number of parameters to be
identified increases with the order of ECM. The numbers of pa-
rameters to be identified for each model are listed in Table 1. For
fifth-order RC models, the number of parameters to be identified
can reach up to 12. In the process of identification and optimization,
the closer the model terminal voltage to the measured terminal
voltage, the more accurate are the model parameters. Therefore,
the root-mean-square error (RMSE) between the model terminal
voltage and the measured terminal voltage can be employed as the
fitness value to assess the model parameters and acquire the
optimal model parameters that make the model voltage closest to
the measured voltage. The objective function for the optimization
Fig. 1. Schematic diagram of the equivalent circuit model. can be expressed as:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n
u1 X  
i-th polarization resistance and i-th polarization capacitance,
Rðqn Þ ¼ t ui;k ðqn Þ  u b i;k ðqn Þ 2 (3)
respectively. n
k¼1
Some of the frequently used ECMs are included in the nRC.
Actually, 1RC model is Thevenin model, and 2RC model is dual
where Rðqn Þ is the RMSE, ui;k is the model voltage, and ub i;k is the
polarization (DP) model [14]. As the number of RC networks in-
measured voltage.
creases, the mathematical representation of the model becomes
During parameter optimization, the upper and lower bounds of
more complex, which making model parameter identification and
the parameters can be obtained on the basis of the experimental
state estimation more difficult and not conducive to the calculation
results. The bounds of the same parameters for different models are
of the BMS. Therefore, the number of RC network in the ECM is
kept the same for fair comparison.
limited to four in this paper. Moreover, the nRC model with hys-
teresis (nRCH) [14] and the PNGV model [35] are also widely re-
ported. Based on the above nRC model, nRCH model and PNGV 2.3. Optimization algorithms
model, nine ECMs are constructed, as shown in Table 1. In this work,
we use nine optimization algorithms to conduct parameter iden- Based on the objective functions and constraint conditions
tification and optimization on nine ECMs and obtain the suitable mentioned above, parameter identification is conducted for the
parameters optimization method for each ECM. nine ECMs introduced in Table 1 using nine representative opti-
mization algorithms in four categories, as shown in Table 2.

2.2. Optimization variables and objective function for ECMs

3. Experiments
For the ECMs, the model parameters to be identified and opti-
The basic parameters of LiNCM are listed in Table 3. As shown in
mized can be expresses as:
Fig. 2, the experiments are performed on a test bench made by
h Rþ R t1 R1 t2 R2 / tn Rn kp M i DIGATRON with a current range of 100 to þ100 A and a voltage
0 0
qn ¼ |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} (2) range of 0e60 V. The voltage accuracy is 1 mV and the current
1st RC 2nd RC nth RC hysteresis accuracy is ±0.1% of the full scale. The tested cells are placed in a
temperature chamber (TEMI580, Dongguan Bell Company) to
where Rþ 
0 and R0 represent the ohmic resistance of charging and maintain the ambient temperature.

Table 1
Discretization equations of the nine ECMs.

Model Model equations and description Number of

parameters

1RC (Thevenin) Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k u1;k ¼ expðDt=t1 Þu1;k1 þ R1 ½1  expðDt=t1 ÞIk 4

where u1;k and t1 ¼ R1 C1 are the voltage and time constant of the RC network. zk is the SOC.

1RCH M I  0 6
Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ uh;k uh;k ¼ Hð1  ejkp Ik j Þ,H ¼ where uh;k is the hysteresis voltage, kp is a
M I>0
decaying factor and M is the maximum amount of hysteresis voltage
PNGV 1 5
Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ ucb;k ucb;k ¼ ucb;k1 þ R1 ½1  expð  Dt=t1 ÞIk
Cb
where Cb is the equivalent capacitance, ucb;k is the voltage of the equivalent capacitance.
2RC (DP) Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k u2;k ¼ expðDt=t2 Þu2;k1 þ R2 ½1  expðDt=t2 ÞIk 6
where u2;k and t2 ¼ R2 C2 are the voltage and time constant of the second RC network.
2RCH Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k þ uh;k 8
3RC Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k þ u3;k u3;k ¼ expðDt=t3 Þu3;k1 þ R3 ½1  expðDt=t3 ÞIk 8
where u3;k and t3 ¼ R3 C3 are the voltage and time constant of the third RC network.
3RCH Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k þ u3;k þ uh;k 10
4RC Vk ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k þ u3;k þ u4;k u4;k ¼ expðDt=t4 Þu4;k1 þ R4 ½1  expðDt=t4 ÞIk 10
where u4;k and t4 ¼ R4 C4 are the voltage and time constant of the fourth RC network.
4RCH Vi ¼ OCVðzk Þ þ Ik R0 þ u1;k þ u2;k þ u3;k þ u4;k þ uh;k 12
1060 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066

Table 2
Optimization algorithms.

Method type Algorithm name Inspiration Year of proposal

Nonlinear programming Find minimum of constrained nonlinear (FMIN) e 1951

Evolution-based Genetic algorithm (GA) [18] Biological evolution 1992
Physics-based Simulated annealing algorithm (SA) [29] Solid annealing 1983
Swarm-based Particle swarm optimization (PSO) [30] Bird flock 1995
Ant colony optimization (ACO) [31] Ant colony 1991
Firefly algorithm (FA) [33] Firefly 2010
Grey wolf optimization (GWO) [36] Grey wolf 2013
Whale optimization algorithm (WOA) [37] Whale 2016
Multi-verse optimization (MVO) [38] Cosmology 2016

Table 3
Main Parameters of LiNMC.

Nominal Capacity(Ah) Nominal Voltage(V) Lower cut-off Voltage(V) Upper cut-off Voltage(V) Maximum charge current(A)

32.5 3.75 2.5 4.15 65

Fig. 2. Schematic of the test bench.

The capacity and hybrid pulse power characterization (HPPC) 4. Results and discussion
tests were first performed to determine the capacity and open-
circuit voltage (OCV) of the LIB, and the battery was subsequently To adequately compare and explore the characteristics of each
fully charged. The discharge experiment was then performed until optimization algorithm and also eliminate the random errors, the
full discharge of the battery under the NEDC in the entire SOC area. optimization algorithms listed in Table 2 were adopted to identify
All the experiments were performed under a controlled ambient the parameters of each ECM listed in Table 1 for 100 times in the
temperature of 25  C. The current and voltage curves obtained from entire SOC area. The statistical results (8100 identification results
the experimental results are shown in Fig. 3. The experimental data under 81 cases in total) were obtained. In practical BMSs, the
will be used for estimating the model parameter. computation time for the identification of the model parameters is
a crucial index. The computation time for parameter identification

Fig. 3. Experimental results under NEDC working cycles in the entire SOC area. (a) Current profile; (b) Voltage profile.
 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066 1061

under 81 cases was also recorded, and the mean computation time the mean value of the 100 identification results excluding the
of each condition was calculated, as shown in Table 4. The algo- highest and lowest 5% of data. From Fig. 4 and Table 4, we can draw
rithms were executed in MATLAB 2016a on a computer with a the following conclusions:
3.1 GHz processor, 8 GB memory, and 64-bit operating system. By
analyzing the size and distribution of these data, the ECMs and (a). It is very obvious that PNGV is an ideal ECM in the low SOC
their parameter identification and optimization approaches were area (0%e20%), and its RMSE is less than half of the errors of
comprehensively evaluated. For objective comparison, the critical other models. The voltage error of PNGV model is shown in
parameters of the swarm-based algorithms were set to the same Fig. 5, it can be seen that the accuracy advantage of PNGV in
values: the maximum number of generations was 500, the number the low SOC area is very obvious. Moreover, the accuracy of
of search agents was 50, and the upper and lower bounds of the FMIN, FA and GWO is almost the same, which is higher than
objective functions and parameters were equal to the values set in that of other algorithms. However, FMIN has the shortest
Section 2.2. identification time among these algorithms. Therefore, PNGV
model and FMIN algorithm are the best choice for model
4.1. Optimal identification accuracy and time analysis of ECMs selection and parameter identification in the low SOC area.
(b). In the high SOC area (20%e100%), PNGV is not an ideal
In this study, the identification accuracy of each ECM is repre- choice, and the model accuracy improves as the order of the
sented by the mean RMSE of the identification results obtained RC network increases. However, this does not mean that a
using the nine optimization algorithms in the entire SOC area, as higher-order RC network will lead to higher model accuracy.
shown in Fig. 4. It needs to be pointed out that the mean RMSE is In fact, the model accuracy is closely related to optimization
algorithms: FMIN algorithm produces a relatively high model
error that remains nearly constant regardless of the order of
Table 4 the RC model, implying that this type of algorithm, based on
Computing time of various algorithms (s).
accurate solving, has a relatively low solving precision that
Model SA PSO GA FMIN ACO FA GWO MVO WOA cannot be improved even by using higher-order RC models.
PNGV 245 31 238 8 371 314 317 272 282 We also find out that in the parameter identification for
1RC 187 26 172 7 325 278 275 244 258 third-order and higher-order models using GA and SA, the
1RCH 199 27 185 7 334 285 283 249 263 model accuracy no longer improved, especially for GA, the
2RC 326 38 332 9 436 363 376 312 316 accuracy decreases when the order of the RC model is greater
2RCH 440 45 527 14 520 416 456 371 363
3RC 497 49 626 16 562 442 497 401 386
than second. However, good identification results can be
3RCH 543 51 629 17 590 451 499 412 387 obtained by using swarm-based algorithms. Furthermore, as
4RC 611 54 633 20 631 464 505 429 388 the order of the model increases, the model becomes more
4RCH 671 55 878 21 680 474 577 447 415 and more complex, and each optimization algorithm has its

Fig. 4. Optimal identification accuracy. (a) RMSE in the entire SOC range; (b) RMSE in the low SOC range; (c) RMSE in the high SOC range.
1062 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066

identification. Besides, neither GA nor SA is an ideal algorithm for

these RC models.

4.2. Error distribution and reliability analysis of ECMs

By implementing parameter identification for each ECM using

the nine optimization algorithms respectively for 100 times in the
entire SOC area, we can acquire the statistical distributions of the
model errors, and the boxplots of various ECMs under different
optimization algorithms are shown in Fig. 6. As shown in Fig. 6(j),
the spacings between the different parts of the box indicate the
degree of dispersion (spread) and skewness in the data, and show
outliers. We can make the following conclusions: for multiple times
of model parameter identification, GA, SA, and PSO show results
with relatively significant fluctuations; FMIN shows the smallest
fluctuations, whereas other swarm-based algorithms show rela-
tively slight fluctuations. More specifically regarding each model
respectively:

(a) For the parameter identification of the first-order RC models

(PNGV, 1RC and 1RCH), GA and SA demonstrate results with
significant fluctuations (as shown in Fig. 6(a)-(c)). Superior
identification performance is achieved using FMIN or
swarm-based algorithms. The minimal model errors of FMIN
and swarm-based algorithms are nearly the same. However,
FMIN requires a shorter identification time. Hence, FMIN is
an ideal choice for identifying the parameters of PNGV and
1RC models. For the 1RCH, FA is an ideal choice to pursue the
highest accuracy, and FMIN is a good choice for compre-
hensive consideration of accuracy and identification time.
Moreover, GA and SA are non-ideal algorithms for the first-
Fig. 5. Voltage errors in the entire SOC area. order RC models.
(b) For the parameter identification of second-order RC models
(2RC and 2RCH), FMIN produces nearly constant identifica-
own maximum identification ability. Take PSO for example, tion results but with lower accuracy than other algorithms.
the best model accuracy is reached by identification 2RC, and The errors of GA and SA are distributed in a relatively wide
the accuracy is kept constant or even reduced for other high- range with high probability. FA shows the highest identifi-
order models. Therefore, the largest identification ability of cation accuracy with no outlier for 2RC and 2RCH. However,
PSO is 2RC. Similarly, the largest identification ability of ACO the identification time of FA is much longer than that of PSO
is 3RC, and FA has the best identification ability in the above and FMIN. Therefore, FA is suitable for offline identification.
nine optimization algorithms. The distribution of outliers of PSO demonstrate wide-range
(c). We can see that FMIN has the shortest identification time fluctuations but with an occurrence probability of only 1%.
among the nine algorithms. All the MAs show a much longer Therefore, PSO is a good choice for parameter identification
identification time than the EAs. PSO has the shortest iden- of second-order RC models in online application, with higher
tification time among the swarm-based algorithms, which is identification accuracy and shorter computation time.
approximately one-sixth of the time taken by the other (c) For the parameter identification of third-order RC models
swarm-based algorithms. (3RC and 3RCH), GA and PSO have the large number of out-
(d). As shown in Fig. 4 (b) and (c), FA has the highest accuracy in liers in the identification results, and SA shows relatively
the low SOC range for the second- and higher-order RC large error with significant fluctuations, indicating that these
models, and FA and GWO have the highest accuracy in the three optimization algorithms are unsuitable for parameter
high SOC range. However, it can be seen from Table 4 that FA identification of third-order models. FA and GWO have
has a longer identification time. Therefore, FA is the best identification results with only slight fluctuations and fairly
choice for offline identification for high-order RC models. high accuracy, but they require a long time for identification.
Therefore, FA and GWO should be selected if the objective is
According to the analysis presented above, higher-order RC to obtain the optimal identification accuracy.
models and FA should be selected if the objective is to achieve the (d) For the parameter identification of fourth-order RC models
minimal model error in the high SOC area. The most suitable (4RC and 4RCH), PSO shows the largest number of outliers in
optimization algorithms for each model can be obtained by the identification results, and GA and SA show large error
comprehensively considering the model accuracy and optimization with considerable fluctuations. FA, GWO and MVO show the
time. In the above nine algorithms, FMIN is the best choice for the highest accuracy for 4RC, and FA shows the highest accuracy
PNGV, 1RC and 1RCH. For the second-order models, PSO is the best for 4RCH, which are suitable for the offline parameter
choice with the higher accuracy and the shortest computation time identification.
in all above swarm-based algorithms. For the higher-order RC
models, FA has the highest accuracy with much longer identifica- Overall, from the model error distribution, we can conclude that
tion time. Therefore, FA is the best choice for the offline accurate optimization techniques such as FMIN are the ideal choice
 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066 1063

Fig. 6. RMSE distribution of various ECMs under different optimization algorithms. (a) PNGV; (b) 1RC; (c) 1RCH; (d) 2RC; (e) 2RCH; (f) 3RC; (g) 3RCH; (h) 4RC; (i) 4RCH; (j)
Schematic of the box-plot.
1064 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066

for the parameters identification of first-order RC models (PNGV, presented by the minimal times of the reduplicate identification
1RC and 1RCH). PSO is an ideal choice for second-order RC model. processes listed in Table 5. The efficiency is described by the
With the increase of the order of the ECMs, the number of outliers identification time. The values of the three items are normalized to
in the identification results of PSO increase. For the 2RCH and 1e5, where 5 represents the best and 1 represents the worst. The
higher-order RC models, FA is an ideal choice for the offline iden- radar chart shown in Fig. 7 can be plotted based on the computation
tification. Neither GA nor SA is an ideal approach for RC models of results. It is noted that FMIN is a simple algorithm with weak
any order. identification ability, and the first-order RC models have low ac-
In the practical parameter identification of ECMs, MAs may fall curacy with simple structure and few parameters. According to the
into local minimum, and result in large RMSE error. Hence, several previous calculation results show that the FMIN has almost the
reduplicate identification processes are essential to have a higher same accuracy as other algorithms for the first-order RC models
probability to get the global minimum. With the help of the 8100 (PNGV and 1RC). However, FMIN has the shortest identification
identification data sets, we can estimate the minimal times of time (<10 s) and the best stability. Therefore, FMIN has absolute
reduplicate identification required for each algorithm under certain predominance for the parameter identification of the first-order RC
accuracy and probability requirements. The minimal times of models in the aspects of accuracy, reliability and identification
identification directly show the reliability of the optimization al- time. Moreover, SA algorithm with poor accuracy and long identi-
gorithm. The larger the minimal times, the worse its reliability. fication time is unsuitable model parameter identification
However, as analyzed in Section 4.1, each optimization algorithm approach. Therefore, they are not compared here.
has its own identification ability. Therefore, the reliability of the There is no doubt that accuracy is the most important evaluation
optimization algorithm should be calculated within the scope of its factor for offline application in a BMS. And the product of compu-
ability. In our study, the identification ability of each model is tational time and minimal number of identification is the actual
represented by the mean RMSE of the 100 identification results cost to achieve the accuracy. To comprehensively evaluate the al-
excluding the highest and lowest 5% of data. gorithms, weighted factor comparison method is adopted:
Suppose the identification ability of the ECM is u mV, and to
ensure that the RMSE is less than u mV with a probability of p, the
minimum time for model parameters identification x can be 1
calculated as follows： f ¼ w1 A þ w2 CR (5)
5

Na x where w1, w2 are the weighted factors. A, C and R represent the

1 1 ¼p (4)
Nb accuracy, computational time and reliability classified in Fig. 7,
respectively. f is the score of the algorithms.
where Nb is the total number of identifications (100 in this case), Hence, we take w1 ¼ 0.9, w2 ¼ 0.1 to evaluate the algorithms for
and Na is the number of identifications at which the model error offline application. Whereas the actual cost to achieve the desirable
meets the error requirement (less than u mV) during the 100 accuracy is comparatively important for online application.
identification times. Therefore, we take w1 ¼ 0.5, w2 ¼ 0.5 to evaluate the algorithms for
The minimal times of identifications of each optimization al- on-line application. The algorithm scores of the algorithms for
gorithm for different ECMs can be calculated using Eq. (4). The different applications are listed in Table 6 and Table 7.
results obtained are listed in Table 5. In the table, the numbers It is obvious that FA has the superior capacity to identify the
represent the computed minimal times of identification, and accurate ECM parameters. For the 1RCH, the improved accuracy
numbers in parentheses represent the practical minimal times of with FA is relatively not obvious owning to the limited ability of
identification. It is obvious that the more times of identification 1RCH (as shown in Fig. 4). As illustrated before, FMIN is the best
requires, the lower the reliability has. From Table 5, we can see that choice for 1RCH parameters identification. For the second- and
FMIN has the highest reliability. It should be noted the reliability in higher-order RC model, FA is the ideal choice for offline
each case is based on the self-accuracy of each model (represented identification.
by the mean RMSE). The scores in Table 7 indicate that PSO has the best compre-
hensive performance for the on-line parameter identification with
4.3. Comprehensive comparison the superior reliability and comparatively low computational cost.
As aforementioned, the best model accuracy is reached by identi-
Based on the analysis presented above, various model param- fication 2RC for PSO, and the accuracy is kept constant or even
eter optimization algorithms can be comprehensively compared reduced for other high-order models. In addition, PSO has the large
from accuracy, stability, and identification time. Accuracy can be number of outliers in the identification results for 2RCH and higher-
described using the mean RMSE of the 100 identification results order RC model. Therefore, 2RCH model and higher-order RC
excluding the highest and lowest 5% of data. Reliability can be models are not suitable for online identification.

Table 5
Minimal number of identification: p ¼ 99%.

Model SA PSO GA FMIN ACO FA GWO MVO WOA

PNGV 7.5 (8) 5.8(6) 9.0 (9) 1.0 (1) 28.3(29) 11.1(12) 7.5 (8) 3.9 (4) 6.6 (7)
1RC 5.8 (6) 3.8 (4) 6.3 (7) 1.0 (1) 4.5 (5) 2.3 (3) 9.3 (10) 4.0 (4) 6.5 (7)
1RCH 5.5 (6) 4.4 (5) 5.0 (5) 1.0 (1) 6.5 (7) 9.3 (10) 16.8(17) 14.6(15) 6.3 (7)
2RC 5.8 (6) 5.2 (6) 5.3 (6) 1.0 (1) 6.5 (7) 5.5 (6) 13.4(14) 9.3 (10) 7.3 (8)
2RCH 5.9 (6) 5.8 (7) 5.6 (6) 1.0 (1) 7.3 (8) 6.8 (7) 10.0(10) 11.1(12) 8.2 (9)
3RC 6.5 (7) 5.0 (5) 5.9 (6) 1.0 (1) 6.8 (7) 3.8 (4) 7.5 (8) 14.0(14) 7.5 (8)
3RCH 7.7 (8) 4.4 (5) 5.6 (6) 1.0 (1) 8.5 (9) 4.9 (5) 5.6 (6) 6.8 (7) 7.0 (7)
4RC 6.8 (7) 4.6 (5) 5.6 (6) 1.0 (1) 8.7 (9) 3.1 (4) 7.3 (8) 8.5 (9) 7.5 (8)
4RCH 5.9 (6) 3.1 (4) 5.5 (6) 1.0 (1) 7.3 (8) 5.6 (6) 6.1 (7) 8.7 (9) 7.3 (8)
 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066 1065

Fig. 7. Comprehensive comparison of various ECMs under different optimization algorithms (5 represents the best and 1 represents the worst). (a) 1RCH; (b) 2RC; (c) 2RCH; (d) 3RC;
(e) 3RCH; (f) 4RC; (g) 4RCH.

Table 6 Table 7
Algorithm scores for offline identification. Algorithm scores for on-line identification.

Model PSO GA ACO FA GWO MVO WOA Model PSO GA ACO FA GWO MVO WOA

1RCH 4.25 4.04 4.09 4.12 3.90 3.97 4.08 1RCH 4.42 3.93 3.54 3.36 2.50 2.78 3.68
2RC 4.41 4.06 4.19 4.41 4.14 4.19 4.10 2RC 4.39 3.64 3.43 3.76 2.89 3.36 3.50
2RCH 4.36 3.93 4.14 4.39 4.23 4.18 4.07 2RCH 4.27 3.19 3.20 3.54 3.14 3.13 3.33
3RC 4.41 3.49 4.15 4.48 4.33 4.15 4.09 3RC 4.39 2.75 3.17 3.79 3.33 2.83 3.37
3RCH 4.38 3.55 4.10 4.51 4.39 4.33 4.08 3RCH 4.44 2.80 3.00 3.70 3.50 3.52 3.40
4RC 4.41 1.18 4.07 4.52 4.36 4.30 4.11 4RC 4.43 1.48 2.91 3.82 3.35 3.34 3.38
4RCH 4.40 0.99 4.09 4.75 4.40 4.35 4.12 4RCH 4.61 0.93 2.92 3.74 3.33 3.32 3.36

5. Conclusions
(1) In the low SOC area (0%e20%), the accuracy of PNGV is much
Appropriate ECMs and matched model parameters serve as the higher than other models, and the complexity of its model
basis of accurate battery state estimation. In this study, a compar- structure is less than 2RC. Therefore, PNGV is an ideal choice
ative study is conducted by implementing model parameter opti- for the ECMs in the low SOC area.
mization for nine equivalent circuit models using nine optimizers (2) The calculation results show that FMIN is an ideal choice for
in the entire SOC area. The following conclusions can be made: parameter identification of PNGV and 1RC. For 1RCH, FMIN is
1066 X. Lai et al. / Electrochimica Acta 295 (2019) 1057e1066

a good choice for comprehensive consideration of accuracy MVO multi-verse optimization

and identification time. NEDC new European driving cycle
(3) The model accuracy improves as the order of the RC network OCV open circuit voltage
increases. However, this does not mean that a higher-order PNGV partnership for a new generation of vehicle
RC network can lead to a higher model accuracy. In fact, PSO particle swarm optimization
the model accuracy is closely related to optimization algo- RC Resistance-capacitance
rithms. The calculation results show that FA is an ideal choice RMSE root mean square error
for second-and higher-order RC models with the highest SA simulated annealing algorithm
identification accuracy and longer identification time. SOC state of charge
Therefore, FA is very suitable for the offline model parame- WOA whale optimization algorithm
ters identification, and PSO is a good choice for compre-
hensive consideration of accuracy and identification time.
(4) For the offline identification of the model parameters of
ECMs, FMIN is an ideal choice for PNGV, 1RC, and FA is an References
ideal choice for 1RCH and higher-order models. For online [1] Y.J. Xing, E.W.M. Ma, K.L. Tsui, M. Pecht, Energies 4 (2011) 1840e1857.
identification, FMIN is an ideal choice for PNGV, 1RC, and PSO [2] M.M. Thackeray, C. Wolverton, E.D. Isaacs, Energy Environ. Sci. 5 (2012)
is the ideal choice for 1RCH and higher-order models. 7854e7863.
[3] S. Nejad, D.T. Gladwin, D.A. Stone, J. Power Sources 316 (2016) 183e196.
Moreover, 2RCH model and higher-order RC models are not
[4] A. Szumanowski, Y.H. Chang, IEEE Trans. Veh. Technol. 57 (2008) 1425e1432.
suitable for online identification. [5] Y.J. Zheng, M.G. Ouyang, X.B. Han, L.G. Lu, J.Q. Li, J. Power Sources 377 (2018)
161e188.
Online identification with the characteristics of real-time vari- [6] R. Kaiser, J. Power Sources 168 (2007) 58e65.
[7] X. Lai, C. Qin, W.K. Gao, Y.J. Zheng, W. Yi, Appl. Sci. Basel 8 (2018).
able parameters can improve the accuracy and robustness of ECMs. [8] L.F. Zheng, J.G. Zhu, G.X. Wang, D.D.C. Lu, T.T. He, Energy 158 (2018)
However, the platforms for implementing the identification algo- 1028e1037.
rithms are generally low cost microcontrollers, the computational [9] X. Lai, Y.J. Zheng, T. Sun, Electrochim. Acta 259 (2018) 566e577.
[10] H.W. He, R. Xiong, J.X. Fan, Energies 4 (2011) 582e598.
power and resource are usually limited. Therefore, accurate online [11] C.Z. Liu, W.Q. Liu, L.Y. Wang, G.D. Hu, L.P. Ma, B.Y. Ren, J. Power Sources 320
identification with less computational burden is a huge challenge (2016) 1e12.
for BMSs. In addition, the existing parameter identification [12] H.W. He, R. Xiong, H.Q. Guo, S.C. Li, Energy Convers. Manag. 64 (2012)
113e121.
methods mostly rely on complex calculation process. The further [13] C. Lin, H. Mu, R. Xiong, J.Y. Cao, Appl. Energy 194 (2017) 560e568.
works mainly include (1) accurate identification of model param- [14] X.S. Hu, S.B. Li, H. Peng, J. Power Sources 198 (2012) 359e367.
eters for different fractional-order models, (2) the design of online [15] J.L. Xie, J.C. Ma, K. Bai, Int. J. Energy Res. 42 (2018) 2710e2727.
[16] C.F. Zou, L. Zhang, X.S. Hu, Z.P. Wang, T. Wik, M. Pecht, J. Power Sources 390
model parameter identification method with good accuracy and
(2018) 286e296.
less computational burden for ECMs, (3) the development of ECMs [17] X.M. Cheng, L.G. Yao, Y.J. Xing, M. Pecht, Energies 9 (2016).
considering partial electrochemical characteristics. [18] J.H. Holland, Evol. Comput. 8 (2000) 373e391.
[19] C.P. Zhang, J.C. Jiang, Y. Gao, W.G. Zhang, Q.J. Liu, X.S. Hu, Appl. Energy 194
(2017) 569e577.
Acknowledgments [20] K. Lim, H.A. Bastawrous, V.H. Duong, K.W. See, P. Zhang, S.X. Dou, Appl. Energy
169 (2016) 40e48.
This work is supported by the National Natural Science Foun- [21] T. Mesbahi, F. Khenfri, N. Rizoug, K. Chaaban, P. Bartholomeus, P. Le Moigne,
Elec. Power Syst. Res. 131 (2016) 195e204.
dation of China (NSFC) under the grant number of 51505290 and [22] Y. Tian, D. Li, J.D. Tian, B.Z. Xia, Electrochim. Acta 225 (2017) 225e234.
5187138. [23] W. Gao, Y. Zheng, M. Ouyang, J. Li, X. Lai, X. Hu, IEEE Trans. Ind. Electron., doi:
10.1109/TIE.2018.2838109.
[24] H.F. Dai, T.J. Xu, L.T. Zhu, X.Z. Wei, Z.C. Sun, Appl. Energy 184 (2016) 119e131.
Nomenclature [25] J. Kinable, B. Smeulders, E. Delcour, F.C.R. Spieksma, Eur. J. Oper. Res. 261
(2017) 475e485.
ACO ant colony optimization [26] S. Akyol, B. Alatas, Artif. Intell. Rev. 47 (2017) 417e462.
[27] H. Abedinpourshotorban, S.M. Shamsuddin, Z. Beheshti, D.N.A. Jawawi, Swarm
BMS battery management system Evol. Comput. 26 (2016) 8e22.
EA exact algorithm [28] P.J.F.C.M. Fonseca, Evol. Comput. 3 (1995) 1e16.
ECM equivalent circuit model [29] K. S, J. Stat. Phys. 34 (1984) 975e986.
[30] D. Bratton, J. Kennedy, in: 2007 Ieee Swarm Intelligence Symposium, 2007,
EV electric vehicle p. 120.
FA firefly algorithm [31] V. Maniezzo, A. Carbonaro, Oper. Res. Comput. Sci. 15 (2002) 469e492.
FMIN find minimum of constrained nonlinear [32] Y. Tian, C. Lu, Z.L. Wang, L.F. Tao, Math. Probl Eng. 2014 (2014) 1e10.
[33] X.S. Yang, Int. J. Bio Inspired Comput. 2 (2010) 78e84.
GA genetic algorithm
[34] T.C. Service, Inf. Process. Lett. 110 (2010) 917e923.
GNL general nonlinear [35] Q.Q. Wang, J. Wang, P.J. Zhao, J.Q. Kang, F. Yan, C.Q. Du, Electrochim. Acta 228
GWO grey wolf optimization (2017) 146e159.
HPPC hybrid pulse power characterization [36] S. Mirjalili, S.M. Mirjalili, A. Lewis, Adv. Eng. Software 69 (2014) 46e61.
[37] S. Mirjalili, A. Lewis, Adv. Eng. Software 95 (2016) 51e67.
LIB Li-ion battery [38] S. Mirjalili, S.M. Mirjalili, A. Hatamlou, Neural Comput. Appl. 27 (2016)
MA metaheuristic algorithm 495e513.