Energy 138 (2017) 118e132

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

A unified modeling framework for lithium-ion batteries: An artificial

neural network based thermal coupled equivalent circuit model
approach
Qian-Kun Wang a, Yi-Jun He a, *, Jia-Ni Shen a, Zi-Feng Ma a, b, Guo-Bin Zhong c
a
Dept. of Chemical Engineering, Shanghai Electrochemical Energy Devices Research Center, Shanghai Jiao Tong University, Shanghai 200240, China
b
Sinopoly Battery Research Center, Shanghai 200241, China
c
Electric Power Research Institute of Guangdong Power Grid Co., Ltd., Guangzhou 510080, China

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

Article history: The thermal coupled equivalent circuit model provides a vital role not only in accurate and reliable state
Received 10 February 2017 monitoring, but also in effective thermal management of lithium-ion batteries. However, it lacks
Received in revised form appropriate modeling strategies for including both the temperature and state of charge effects into the
4 July 2017
thermal coupled equivalent circuit model. In this study, a unified artificial neural network based thermal
Accepted 7 July 2017
Available online 9 July 2017
coupled equivalent circuit model approach is proposed to accurately and reliably capture the electrical
and thermal dynamics of lithium-ion batteries. Both reversible and irreversible heat generation mech-
anisms are introduced in the thermal model. The quantitative relationship between circuit parameters
Keywords:
Lithium-ion batteries
and temperature/state of charge in equivalent circuit model is modeled by artificial neural network. Both
Equivalent circuit model electrical and thermal related parameters are simultaneously identified by means of least square strategy
Thermal model with l1-norm penalty on output weights in artificial neural network and positive constraints on circuit
Artificial neural network parameters. The effectiveness of the proposed artificial neural network based thermal coupled equivalent
State of charge circuit model approach is validated by the experimental constant current discharge, pulse current
discharge test and hybrid pulse power characterization test of a commercial large-format pouch-type
lithium-ion battery. It implies that the proposed hybrid modeling strategy can provide a general
framework for the inclusion of other effects such as health state and current into battery models and can
be easily extended to more complicated models such as first-principle electrochemical-thermal model.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction significant attentions for guaranteeing the safe operation of batte-

ries [5]. Note that accurate and reliable state monitoring is one of
Lithium-ion batteries (LIBs), as one of the most promising en- the most crucial functions in BMS, which is greatly helpful for
ergy storage technologies, are increasingly used for portable elec- providing the basic parameters to assist in carrying out other
tronic devices, electric vehicle, hybrid electric vehicles and smart functions in BMS such as overcharge and overdischarge protection
grid [1]. The safe, reliable and efficient operation of LIBs is depen- [6]. However, the internal states of LIBs such as SOC and SOH cannot
dent not only on inherently safe material design and manufacture be directly online measured. Hence, it requires efficient techniques
technology [2], but also on a high-fidelity battery management to estimate these internal states from available measured variables,
system (BMS) [3]. A properly designed BMS should usually include i.e. current, voltage and temperature. Being a residual battery ca-
many functionalities [4], such as state of charge (SOC) estimation, pacity indicator, SOC servers as an essential parameter for many
state of health (SOH) prognostics, fault detection and charge/ functionalities of BMS such as over-charge/over-discharge control
discharge rate control. In addition, proper thermal management, as and SOH prognostics. Accurate and reliable SOC estimation could be
another essential functionality of BMS, has recently attracted greatly helpful for efficient, safe and long-life operation of LIBs.
Except for traditional Coulomb counting and open circuit voltage
based SOC estimation methods, battery model based SOC estima-
tion methods have attracted significant attentions and various
* Corresponding author. advanced methods have been developed in the last decade [7]. The
E-mail address: [email protected] (Y.-J. He).

http://dx.doi.org/10.1016/j.energy.2017.07.035
0360-5442/© 2017 Elsevier Ltd. All rights reserved.
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 119

performance of SOC estimation is usually affected by the accuracy between circuit parameters and nonlinear factors. Chen et al. [21]
and reliability of battery model. Various battery models such as proposed a runtime-based model to predict the IeV performance,
electrochemical model, data-driven model and equivalent circuit in which the circuit parameters were identified as constants within
model (ECM) have been developed to describe the nonlinear dy- 20e100% SOC and exponentially changed with respect to SOC
namics of LIBs. Pseudo two-dimensional model developed by Doyle within 0e20% SOC. Kroeze and Krein [22] proposed a polynomial
et al. [8] is one of the most well-known electrochemical models for function to model the nonlinear effect of SOC on circuit parameters.
predicting the detailed dynamics of LIBs. Although the electro- Cheng et al. [23] introduced the trigonometric function polynomial
chemical model has relatively high generalization performance and to model the SOC dependence of circuit parameters. However, the
has been widely used for battery simulation and design, the high functional forms such as polynomial order were usually empirically
computational cost hinders its online application. The data-driven determined in these studies [24], which would not be easily
model, as a black-box modeling approach, has been used for extended to other types of LIBs. Hence, in our previous study [25], a
directly describe the voltage response dynamics [9]. However, the simultaneous model structure selection and parameter estimation
performance of the data-driven model is strongly dependent on the approach was proposed to optimally select the appropriate poly-
quality of samples and its generalization performance cannot be nomial functional forms of circuit parameters with respect to SOC.
guaranteed for unseen operation conditions. ECMs could be con- On the other hand, Arrhenius-like function [26] is usually proposed
structed either in the frequency-domain or in the time-domain. For to model the nonlinear relationship between temperature and
example, Moss et al. [10] first constructed a frequency-domain ECM circuit parameters such as paralleled resistance. Recently, poly-
using electrochemical impedance spectroscopy (EIS) data and then nomial function and Arrhenius-like function were combined to
transformed the frequency-domain ECM into the time-domain model the SOC and temperature dependency of circuit parameters
ECM for predicting the battery voltage response and power per- in thermo coupled ECM model [27], respectively. In the study of Du
formance at room temperature. Hariharan and Kumar [11] directly et al. [28], general fuzzy logic approach was developed to model the
constructed a time-domain ECM with SOC dependent circuit pa- parametric dependency on both SOC and temperature. In addition,
rameters to fit battery hybrid pulse power characterization (HPPC) Hu et al. [29] used a linear spline function to model the parametric
testing protocol. Recently, Hu et al. [12] performed a systematical dependence. However, note that the studies in both Du et al. [28]
comparison of different types of ECM for different batteries, in and Hu et al. [29] lacked thermal model validation.
which the model complexity, accuracy and robustness are thor- To conclude, it is important to explicitly embed the nonlinear
oughly evaluated. To conclude, ECM at present is the most prom- factors such as temperature and SOC into circuit parameters for
ising choice for performing state estimation in practical BMS due to improving the prediction accuracy of both electrical and thermal
its relatively high accuracy and low computational cost. characteristics. However, it is worth noting that current developed
The understanding of the thermal characteristics of LIBs plays a thermo-ECM models not only lack a universal modeling approach
significant role in optimal design of thermal management system for simply embedding various effects such as temperature, SOC,
to prevent the thermal runaway during the charging/discharging health state and current into circuit parameters, but also lack an
process [13]. The thermal model, providing an efficient way to automatic model selection approach for optimally reducing the
predict the temperature evolution during battery operation, has model complexity for meeting the low computational requirement
been attracted increasingly attentions [14]. There are two heat in practical BMS. Therefore, in this study, we propose a unified
generation mechanisms in LIBs, i.e. reversible heat generation due modeling approach for the thermo-ECM models with both tem-
to entropy change and irreversible heat generation due to elec- perature and SOC effects. An artificial neural network (ANN) is used
trochemical polarization [15]. Various multi-physics battery to quantitatively capture the relationship between circuit param-
models have been developed to capture both the electrical and eters and temperature/SOC. Both electrical and thermal related
thermal dynamics in LIBs [16]. Although three-dimensional elec- parameters are simultaneously identified using the battery testing
trochemical-thermal model can provide the detailed evolution in- data of a commercial large-format pouch-type LIB. The effect of the
formation about potential, current density and temperature number of hidden neurons in ANN on the model performance will
distributions during battery operation, the high computational cost be investigated. Moreover, an l1-norm penalty on output weights in
makes it unavailable to be used in BMS. Hence, lumped-parameter ANN is added to the objective function for optimally selecting the
ECM and thermal coupled models are nowadays widely used due to hidden neurons for circuit parameters, which consequently results
their computational efficiency [17]. In general, the ECM and ther- in improving the prediction performance and reducing the model
mal model are coupled by means of two parameters, i.e. tempera- complexity.
ture and heat generation rate. The heat generation rate is calculated The remainder of this article is structured as follows. The
by the ECM and is used for the heat source term in the thermal thermo-ECM model is first described. Then, the ANN based thermo-
model; while the temperature is predicted by the thermal model ECM model is introduced to include the nonlinear relationship
and is used for update the temperature-related circuit parameters between circuit parameters and temperature/SOC and the param-
in ECM [18]. For the sake of simplicity, the ECM and thermal eter identification method is described. The effectiveness of the
coupled model is called as the thermo-ECM model. proposed methods is next validated by the experimental datasets.
It is well known that the circuit parameters in ECM such as open Finally, conclusions are presented.
circuit voltage (OCV), resistance and capacitance are dependent on
temperature and SOC. Generally, model-free look-up table is used 2. Methods
to describe the nonlinear relationship between circuit parameters
and temperature/SOC [19]. However, look-up table approach is The equivalent circuit model for LIBs is first introduced. Then,
often inefficient and inaccurate to some extent when applied in the artificial neural network used for modeling the nonlinear
real-time BMS applications. In addition, although look-up table is relationship between circuit parameters and temperature/SOC is
relatively simple to be constructed in one- or two-dimensional described. Based on the description of heat generation rate, the
factors, it is difficult to handle with the case of relatively high- thermal model is then established. The electrical and thermal
dimensional factors such as temperature, SOC, health state and coupling strategy is next introduced to establish the ANN based
current [20]. In the last decade, different analytical expression thermo-ECM. Finally, the parameter identification method is
based models have been proposed to capture the relationship proposed.
120 Q.-K. Wang et al. / Energy 138 (2017) 118e132

2.1. Equivalent circuit model

V1
I2 ¼ (4)
ECMs could be constructed either in the time-domain using R1
current-voltage (IeV) curves [30], or in the frequency-domain us- From Eqs. (2)e(4), the voltage V1 can be derived as,
ing the EIS data [31]. Although EIS data can provide more detailed
kinetic information of LIBs than traditional IeV curves, it needs to dV1 I V
be further developed to be available in practical BMS due to its ¼  1 (5)
dt C1 R1 C1
relatively high cost and testing time [32]. The time-domain ECM is
mainly composed of a controlled voltage source in series with The terminal voltage VT is calculated by
resistance and resistance-capacitance (RC) parallel networks.
VT ¼ VOC  IR0  V1 (6)
Generally, a voltage source, resistor and RC networks are used to
simulate the open circuit voltage (OCV), the internal ohmic resis-
where IR0 is voltage drop through the ohmic resistance.
tance, and the electrochemical polarization related transient
Hence, the ECM described by Eqs. (1), (5) and (6) is constructed
response characteristics, respectively. The number of RC network
to predict the nonlinear charging/discharging dynamics of LIBs.
could often affect the performance of ECM and a detailed perfor-
Since current and voltage are usually measured at discrete time
mance comparison between different structures of ECMs can be
with a specific sampling frequency, it is reasonable to assume that
referred to the study of Hu et al. [12].
during the time interval [tk-1, tk], the applied current I and four
As in our previous study [25], to describe the nonlinear dy-
circuit parameters, i.e. VOC, R0, R1 and C1, keep constant. Therefore,
namics during battery operation, an ECM with one-pair RC branch
Eqs. (1), (5) and (6) can be analytically reformulated in discrete time
shown in Fig. 1 is used. In Fig. 1, VOC, R0, R1 and C1 denote the OCV,
as Eqs. (7), (8) and (9) respectively.
internal resistance, polarization resistance and capacitance
respectively. I is total current with a positive value at discharging Ik1 Dt
and a negative value at charging. I1 and I2 are the currents through SOCk ¼ SOCk1 þ (7)
Cn
resistance R1 and capacitance C1, respectively. It is well known that
all circuit parameters (e.g. VOC, R0, R1 and C1) are usually dependent
 

Dt Dt
on SOC and temperature. Since the OCV is often independently V1;k ¼ V1;k1 exp  þ Ik1 R1;k1 1  exp 
measured at different SOC and temperatures and shows an
tk1 tk1
increasing monotonicity with respect to SOC, interpolation method (8)
is used to construct the OCV model. For other circuit parameters, i.e.
R0, R1 and C1, an ANN is used to establish the empirical model be- VT;k ¼ VOC;k  Ik R0;k  V1;k (9)
tween these three parameters and SOC/temperature.
The change of SOC is described as, where Dt ¼ tk  tk1 is the time interval between two sampling
points, and tk1 ¼ R1;k1 C1;k1 denotes the time constant of the
parallel RC network during the time interval [tk-1, tk]. It is worth
dSOC I noting that the above four time-varying circuit parameters are
¼ (1)
dt Cn dependent on both temperature and SOC and three of them, i.e. R0,
R1 and C1, will be modeled by means of ANN described in the
where Cn is the nominal capacity of battery. following subsection.
Based on Kirchhoff's current law, the relationship between the
current I1 and current I2 can be described as 2.2. Artificial neural network

ANNs are widely used in data-driving modeling, in this work, a

I ¼ I1 þ I2 (2) multi-input-multi-output (MIMO) single layer feed-forward net-
The relationship between the voltage V1 and current I1 in the works (SLFNs) is used to construct the model of three circuit pa-
capacitance C1 can be described as rameters with respect to temperature and SOC. Note that this study
only uses the random generation of input weights and hidden
biases in ANN, while the output weights are simultaneously esti-
dV1 mated with the parameter in thermal model by means of least
I1 ¼ C1 (3) square optimization problem. A detailed description of proposed
dt
ANN is shown as follows.
The current I2 related to V1 is given by The multi-output SLNF with L hidden neurons is given as,

X
L  
yi ¼ bij g aj ; bj ; x ; i ¼ 1; 2; /; p (10)
j¼1

where x2Rn and y2Rp are the input and output vectors respec-
tively, aj 2Rn and bj 2R are input weights and bias of the jth hidden
neuron respectively, and bij 2R is the weight connecting the jth
hidden neuron to the ith output neuron. gðaj ; bj ; xÞ is the output of
the jth hidden neuron with respect to the input x. In this study, the
sigmoid activation function is chosen shown as,

1
gða; b; xÞ ¼ (11)
1 þ expð  ða$x þ bÞÞ
Fig. 1. Schematic of one-pair RC equivalent circuit model.
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 121

The weights a ¼ ½a1 ; a2 ; /; aL T and biases b ¼ ½b1 ; b2 ; /; bL T of Note that the entropic heat coefficient vVOC =vT is also depen-
hidden neurons are randomly selected and the output weights b dent on both temperature and SOC. In this study, the entropic heat
need to be estimated. In this study, the input x is ½SOC; TT and the coefficient is first experimentally tested and calculated based on
output y is ½R0 ; R1 ; C1 T . Based on the above definitions, three circuit OCV measurements at different temperatures and SOC. Then, an
parameters for a specific temperature and SOC at time interval [tk, entropy model can be constructed by means of interpolation
tkþ1] can be written as, method.

X
L 
2.4. ANN based Thermo-ECM
R0;k ¼ b1j g aj ; bj ; ½SOCk ; Tk T (12)
j¼1
Until now, the complete model of ANN based thermo-ECM is
obtained to predict the coupled electrical and thermal dynamics
X
L 
during battery operation. The model is described by Eqs. (7)e(9),
R1;k ¼ b2j g aj ; bj ; ½SOCk ; Tk T (13)
j¼1
(12)e(14) and (17). Fig. 2 illustrates the schematic diagram of the
coupling strategy of ANN, equivalent circuit model and thermal
model. First, the ECM calculates SOC and terminal voltage VT based
X
L 
on the applied current I and circuit parameters, including VOC, R0, R1
C1;k ¼ b3j g aj ; bj ; ½SOCk ; Tk T (14)
j¼1
and C1. The circuit parameter VOC is calculated by the OCV model,
and R0, R1 and C1 are calculated by the ANN based on the SOC and
temperature T. Then, VOC, I and VT are used to calculate the irre-
versible heat generation rate, and entropic heat coefficient vVOC =vT,
2.3. Thermal model
T and I are used to calculate the reversible heat generation rate. The
summation of both irreversible and reversible heat generation rates
The optimal design and thermal management of LIBs relies on
is then transferred to the thermal model for calculating the
reliable predictions of cell temperature and heat generation rate.
temperature.
The complex heat generation in LIBs is mainly composed of the
contributions of electrochemical reactions, phase changes, heat
capacity changes, mixing effects and electrical work. Generally, the 2.5. Parameter identification method
heat generation rate can be expressed as a complex nonlinear
function of concentrations, potentials, current density, tempera- The total 3Lþ2 parameters, including 3L output weights b in Eqs.
ture, and so on. It is worth noting that most of these variables are (12)e(14), and the heat capacity Cp and heat transfer coefficient h in
not measurable in practical BMS. However, as the voltage and the thermal model, needs to be estimated from the battery testing
temperature profiles contain abundant internal electrical and data. A weighted normalized sum square error (WNSSE) between
thermal dynamics information in LIBs, it is more suitable to obtain the model outputs (i.e. terminal voltage and temperature) and the
the heat generation rate based on the measurable variables such as experimental data is used for parameter estimation. The WNSSE is
applied current, terminal voltage and temperature in BMS. In this defined by,
study, the heat generation rate Q is calculated based on the thermal 00 12 1
0 12
modeling study of Bernardi et al. [15] and is shown as, K B b
X b
BB
V T;k  VT;k C T k  Tk A C
WNSSE ¼ @@ A þw@ C
A (18)
vV V T;k T k
Q ¼ IðVOC  VT Þ  IT OC (15) k¼1
vT
where VT;k and V b
where the first and second terms denote the irreversible and T;k are the measured and predicted terminal
reversible heat generation rates respectively. The irreversible heat, voltages respectively; Tk and T b are the measured and predicted
k
generated due to electrochemical polarization, is always temperatures respectively; and w is weight characterizing the
exothermic, while the reversible heat, generated due to entropy relative importance of temperature with respect to terminal
changes, can be either exothermic or endothermic depending on voltage in parameter estimation.
the entropy of reaction and direction of current. It is worth noting that three circuit parameters, i.e. R0, R1 and C1,
In this study, the temperature is assumed to be uniform should be greater than zeros in the whole investigated SOC and
throughout the cell. The governing equation of energy conservation temperature ranges. Hence, the parameter estimation problem
for this lumped-parameter cell is described as, should introduce the positive value constraints on the circuit pa-

 rameters. Consequently, the parameter estimation problem would
dT vV become a nonlinear semi-infinite programming (NSIP), which is
mCp ¼ I VOC  VT  T OC  hAðT  Tamb Þ (16)
dt vT difficult to be efficiently solved. In this study, a simple discretization
strategy is used to transfer the NSIP problem into a NLP problem
where m is the cell weight, Cp is the heat capacity, A is the cell with finite constraints. The Latin hypercube sampling (LHS) method
surface area and h is the convective heat transfer coefficient. Both [33] is used to uniformly generate a set of samples over the SOC and
thermal parameters, i.e. Cp and h, need to be estimated based on the temperature ranges. Hence, the following constraints are added
battery experimental testing data. during the parameter estimation procedure.
Similar with ECM, Eq. (16) is analytically reformulated in
discrete time as, X
L 
T
R0;p ¼ b1j g aj ; bj ; SOCp ; Tp > 0; p ¼ 1; /; P (19)


I Dt  vVOC  j¼1
Tk ¼ Tk1 þ k1 VOC;k1  VT;k1  Tk1
mCp vT k1
X
L 
hADt   T
 T  Tamb;k1 (17) R1;p ¼ b2j g aj ; bj ; SOCp ; Tp > 0; p ¼ 1; /; P (20)
mCp k1 j¼1
122 Q.-K. Wang et al. / Energy 138 (2017) 118e132

Fig. 2. Schematic diagram of the coupling strategy of neuron networks, equivalent circuit model and thermal model.

 100 X
K b
y k  yk
X
L
T MAPEð%Þ ¼ (24)
C1;p ¼ b3j g aj ; bj ; SOCp ; Tp > 0; p ¼ 1; /; P (21) K yk
k¼1
j¼1

where yk and by k denote the measured and fitted/predicted terminal

where SOCp and Tp are the pth SOC and temperature sample voltages or temperatures respectively.
respectively, and P is the prescribed number of samples generated
by LHS method.
In general, with the increasing of hidden neurons in ANN, the 3. Experimental
fitting ability could be improved, but the prediction performance
might be deteriorated and the computational cost would become To test the battery electrical and thermal dynamics, a commer-
higher. Hence, the effect of number of hidden neurons on both cial lithium ion phosphate (LiFePO4) pouch-type battery with
fitting and prediction performance of ANN based thermo-ECM nominal capacity of 20 Ah and nominal voltage of 3.2 V was used.
model will be thoroughly investigated. On the other hand, the l1- All the experiments were conducted with a NEWARE CT-3002-5 V-
penalized objective function is also defined to realize the optimal 500 A battery test system and a HARDY HLT4005P programmable
pruning of hidden neurons, which is shown as, temperature test chamber with a volume of 30  40  45 cm3 . The
data acquisition system has a logging frequency of 1 Hz, and the
00 1 measurement precisions of current, voltage and temperature are
12 0 12 0.05%, 0.05% and 0.4% respectively. To measure the temperature
K B b
X b X
BB
V T;k  VT;k C T k  Tk A C 3

@@ A þw@ Cþ
A li kbi k1 (22) distribution during battery operation, eleven thermocouples were
V T;k T k i¼1 placed on the surface. Based on the experimental results, it was
k¼1
found that the temperature at the center location of battery is
almost the same as the average battery temperature. Hence, the
where li  0; i ¼ 1; 2; 3 are the regularization parameters, and bi ¼ measured battery center temperature was used to perform
PL
j¼1 bij denotes the l1 norm of a vector bi. As the values of R0, R1 parameter estimation in the ANN based thermo-ECM.
and C1 usually have different orders, the regularization values of l2 Figs. 3a and b show the schematic diagram of battery charac-
and l3 are set equal to l1 and 106 l1 respectively as in our previous teristic tests and temperature control procedure respectively. The
study [25]. To efficiently handle the objective function with l1 norm battery characteristic tests include OCV, entropic heat coefficients,
penalty, the big-M method is used and the parameter estimation pulse current discharge (PCD), HPPC and constant current
problem can be reformulated as mixed-integer nonlinear pro- discharge (CCD) tests. Before each PCD, HPPC or CCD test, the
gramming (MINLP) model. The detailed description of big-M temperature control procedure is shown as: (1) change the cham-
method can be referred to our previous study [25]. ber temperature to 293 K; (2) the battery rests at 293 K for 5 h to
To evaluate the fitting and prediction performance of the ANN achieve thermal equalization; (3) charge the battery to its full ca-
based thermo-ECM model, two criteria, i.e. the root mean square pacity through constant current constant voltage (CCCV) with 1/3C
error (RMSE) and mean absolute percentage error (MAPE), are rate; (4) change the chamber temperature to target temperature;
used. RMSE and MAPE are defined by Eqs. (23) and (24) (5) the battery rests at target temperature for 5 h; and (6) shut
respectively. down the chamber to simulate the natural convection environ-
ment. Several thermocouples were placed near the battery to
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi measure the ambient temperatures and the average ambient
u
2
uPK temperature is calculated to be used in thermal model.
u b
y  y
t k¼1 k k
It should be also noted that for the OCV and entropic heat co-
RMSE ¼ (23)
K efficients tests, the chamber will be kept work to maintain the
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 123

Fig. 3. Schematic diagram of battery test protocol. (a) battery characteristic tests; (b)
temperature control procedure.

battery at specific temperature. The OCV was measured at different

SOCs from 0.0 to 1.0, and the detailed testing procedure can be
referred to our previous study [34]. To obtain the entropic heat
coefficients, the OCV at all evaluated SOCs was first measured at a
series of temperatures, i.e. 283 K, 293 K, 303 K and 313 K. Then the
slope of the OCV with respect to temperature at a specific SOC was
calculated as the entropic heat coefficient.
The PCD, HPPC and CCD tests were performed for parameter
estimation and model validation. During the PCD test, the battery
was pulse discharge at 1C to reach 95% and 90% SOC, and then to
reduce 10% SOC of battery until cut off voltage, between each pulse
discharge the battery was rest for 100s. The HPPC test was carried
out by repeating the following test procedure until the battery
reached the cut off voltage: firstly the battery was discharged at 1C
for 360 s to reduce 10% SOC, and then rest for 100 s, followed by
discharged at 2.5C for 10 s, then rest for 40 s and finally charged at
2C for 10s. During each CCD test, the battery was discharge at
Fig. 4. Typical terminal voltage response, temperature evolution and heat generation
constant current until the cut off voltage 2.5 V. The CCD tests with rate curves in PCD test: (a) terminal voltage; (b) temperature; (c) heat generation rate.
1C-, 1.5C- and 2C-rate are implemented at 263, 293 and 323 K, and
the PCD and HPPC implemented at 293 K. The detailed testing
protocols of PCD and HPPC can also be referred to our previous generation rate in the late period of PCD test is ascribed to a greater
studies [35]. battery polarization, which is consistent with the definition of
Figs. 4aec show the terminal voltage response, temperature irreversible heat generation.
evolution and heat generation rate curves in PCD test, respectively. Figs. 5aec show the terminal voltage response, temperature
It is observed from Fig. 4b that the average ambient temperature evolution and heat generation rate curves in HPPC test, respec-
only shows a small increase from 293.27 to 294.33 K during the tively. Compared to Fig. 4b, the battery temperature profile in
whole PCD test because of the heat generation of battery and the Fig. 5b also shows a similar variation characteristic. The sharp
heat exchange between ambient and external environment. The temperature increasing is mainly in the late period of HPPC test
battery temperature evolution profile roughly shows three distinct started from 2700 s, in which the temperature increases from
characteristic regions, namely a slow increasing region from 293.21 294.56 to 297.83 K. The main reason is that similar with the PCD
to 294.27 K, an almost flat region from 294.27 to 294.42 K and a test, a greater battery polarization is observed in the late period of
sharp increasing region from 294.42 to 298.96 K during the time HPPC test, which could result in higher irreversible heat generation
intervals [0, 1000] s, [1000, 2500] s and [2500, 4360] s respectively. rate. From Figs. 4b and 5b, it is also found that there exists a small
Correspondingly, the variation characteristic of total heat genera- temperature decreasing during the rest period of both PCD and
tion rate profile in Fig. 4c is almost consistent with that of tem- HPPC tests, which is ascribed to the convective heat transfer be-
perature during the whole PCD test. The higher irreversible heat tween battery and surrounding.
124 Q.-K. Wang et al. / Energy 138 (2017) 118e132

Fig. 6. Typical terminal voltage response, temperature evolution and heat generation
rate curves in CCD test at 293 K and 2 C-rate: (a) terminal voltage; (b) temperature; (c)
heat generation rate.

Fig. 5. Typical terminal voltage response, temperature evolution and heat generation
rate curves in HPPC test: (a) terminal voltage; (b) temperature; (c) heat generation respectively. The parameter identification method is implemented
rate. in GAMS 24.7.1 and the corresponding NLP and MINLP models are
solved by a branch and reduce algorithm based global solver
BARON 16.3.4 [36]. All computations are carried out on a PC with
The terminal voltage response, temperature evolution and heat
2.40 GHz processor and 8 GB of RAM. The maximum CPU time is set
generation rate curves in CCD test with 2C-rate and the tempera-
to 50,000 s in all calculations.
ture of 293 K are shown in Figs. 6aec, respectively. Compared to
Figs. 4b and 5b, it is clearly found that the CCD test with 2C-rate
would show a significant higher battery temperature increasing 4.1. Model validation of ANN based thermo-ECM
from 293.18 K to 206.78 K. That is due to a significant higher irre-
versible heat generation rate shown in Fig. 6c for CCD test with 2C- Based on the PCD test dataset, the effect of number of hidden
rate, which is ascribed to a greater battery polarization. neurons on model fitting performance is first investigated. Figs. 7a
and b show the fitting RMSE and MAPE of terminal voltage and
temperature with respect to number of hidden neurons using ANN
4. Results and discussion based thermo-ECM, respectively. From Fig. 7a, it is clearly observed
that both fitting RMSE and MAPE of voltage first sharply decrease
The PCD and HPPC test datasets are used for parameter identi- with the increasing of hidden neurons and then keep almost con-
fication and prediction performance validation of ANN based stant as the number of hidden neurons exceeds 6. The fitting RMSE
thermo-ECM and ANN based thermo-ECM with l1-norm penalty, and MAPE of voltage can significantly decrease from 53.4 to 8.2 mV
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 125

Fig. 7. Fitting RMSE and MAPE of ANN based thermo-ECM with respect to number of
hidden neurons in PCD test: (a) terminal voltage; (b) temperature.

Fig. 8. Comparison of terminal voltage fitting results using ANN based thermo-ECM
with 3, 6 and 9 hidden neurons in PCD test: (a) 5e100% SOC; (b) 20e30% SOC.

and from 1.11% to 0.20% when the number of hidden neurons in-
creases from 1 to 6, respectively. Similar with Fig. 7a, the fitting
addition, the introduction of more RC network would significantly
RMSE and MAPE of temperature in Fig. 7b also show a sharp
increase the computational complexity and might deteriorate the
decreasing trend with the increasing of hidden neuron from 1 to 6
voltage prediction performance.
and then keep almost constant. The fitting RMSE and MAPE of
Fig. 9 compares the temperature fitting performance using ANN
temperature can decrease from 0.57 to 0.12 K and from 0.15% to
based thermo-ECM with 3, 6, and 9 hidden neurons. As shown in
0.03% when the number of hidden neurons increases from 1 to 6,
Fig. 9, the model with 3 hidden neurons clearly underestimates the
respectively. Although these results indicate that increasing the
battery temperature during the earlier stage of PCD test, namely the
model complexity can improve the model fitting performance, it
would also increase the computational cost and the over-fitting
risk. Hence, the number of hidden neurons should be optimally
selected in terms of the fitting and prediction performance as well
as computational cost.
The comparison of voltage fitting performance using ANN based
thermo-ECM with 3, 6, and 9 hidden neurons is shown in Fig. 8.
From Fig. 8a, it can be seen that the proposed model with 3 hidden
neurons poorly captures the voltage response during the whole
PCD test, especially at the discharge period. The ANN based
thermo-ECM with 6 and 9 hidden neurons give almost the same
fitting performance, and match well with the measured terminal
voltage over the whole SOC range. Fig. 8b shows the enlarged view
of fitting performance at the SOC range of 20e30%, it can be found
that the ANN based thermo-ECM with 3 hidden neurons could
overestimate or underestimate the voltage response. Increasing the
model complexity from 6 to 9 hidden neurons slightly improves the
fitting performance, however, meanwhile results in higher
computational cost. It can also be found that even the model with 9
hidden neurons cannot perfectly match the experimental data. The
possible reason is that the voltage response shows a multiple time-
scale characteristic and it might require a complicated ECM with
more RC network. However, it should be noted that the fitting Fig. 9. Comparison of temperature fitting results using ANN based thermo-ECM with
RMSE of 8.0 mV is sufficiently low for practical BMS application. In 3, 6 and 9 hidden neurons in PCD test.
126 Q.-K. Wang et al. / Energy 138 (2017) 118e132

time interval [500, 2400] s, and then overestimate the battery and b, it can be thus concluded that the ANN based thermo-ECM
temperature with a gap about 0.2 K between the measured and with 6 hidden neurons is accurate and complex enough for prac-
fitted temperature data. The 6 and 9 hidden neurons ANN based tical BMS application in terms of prediction performance and
thermo-ECM have almost the same fitting performance, with RMSE computational cost.
of 0.12 and 0.11 K, respectively. From Figs. 8 and 9, it is, therefore, Fig. 11 compares the terminal voltage prediction results using 3,
reasonable to conclude that the ANN based thermo-ECM with 6 6 and 9 hidden neurons ANN based thermo-ECM. From Fig. 11a, the
hidden neurons is complex enough to capture the nonlinear electric model with 3 hidden neurons shows the worst voltage prediction
and thermal characteristics of the selected battery in PCD test. performance especially at medium and low SOC range, with a RMSE
The unseen HPPC test dataset is then used to validate the pre- of 23.0 mV and a MAPE of 0.56%. The ANN based thermo-ECM with
diction performance of the proposed ANN based thermo-ECM with 6 and 9 hidden neurons can almost match the experimental voltage
different number of hidden neurons. Figs. 10a and b show the responses over the entire SOC range. It can be observed that the
prediction RMSE and MAPE of battery terminal voltage and tem- voltage prediction error during the charge period is larger than that
perature using ANN based thermo-ECM with respect to number of during the discharge and rest periods. The main reason is that the
hidden neurons, respectively. As shown in Fig. 10a, the prediction PCD test does not include the charge process. It is expected that the
RMSE of voltage first sharply decreases from 43.5 to 10.8 mV with voltage prediction performance during the charge period can be
increasing hidden neurons from 1 to 6. Unlike the continued significantly improved if the voltage response characteristic data of
improvement of fitting performance with the increasing of hidden charge process is included into the model establishment. In addi-
neurons shown in Fig. 7a, the minimum value of prediction RMSE is tion, it is obviously seen that the voltage prediction error during the
obtained with 6 hidden neurons. It indicates that a higher model charge region increases with the decreasing of SOC. The possible
complexity would result in higher risk of over-fitting and conse- reason is due to the accumulation of prediction error with the
quently reduce the prediction performance. Therefore, it again implementation of HPPC test. Fig. 11b shows the enlarged view of
implies that the number of hidden neurons should be finely Fig. 11a at the 20e30% SOC. It is observed that the 3 hidden neurons
selected to balance the trade-off among fitting and prediction ANN based thermo-ECM clearly underestimates the voltage
performance as well as computational cost. The prediction MAPE of response during discharge and rest periods. It is also found that the
voltage shows similar trend with RMSE, and the minimum MAPE of model with 6 hidden neurons matches very well with the experi-
0.26% is also obtained with 6 hidden neurons. From Fig. 10b, it is mental voltage response and outperforms the other two models. In
found that the prediction RMSE and MAPE of temperature can addition, all these three models clearly overestimate the voltage
sharply decreases from 0.65 to 0.13 K and from 0.19% to 0.03% response in the charge period.
respectively, as the number of hidden neurons increases from 1 to The comparison of temperature prediction results using 3, 6 and
6. In addition, the RMSE and MAPE are almost the same for the 9 hidden neurons ANN based thermo-ECM is shown in Fig. 12. It is
model with 6, 7 and 8 hidden neurons. The minimum values of
RMSE and MAPE are 0.12 K and 0.03% respectively. From Figs. 10a

Fig. 10. Prediction RMSE and MAPE of ANN based thermo-ECM with respect to Fig. 11. Comparison of terminal voltage prediction results using ANN based thermo-
number of hidden neurons in HPPC test: (a) terminal voltage; (b) temperature. ECM with 3, 6 and 9 hidden neurons in HPPC test: (a) 10e100% SOC; (b) 20e30% SOC.
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 127

found that the 3 hidden neurons model clearly first underestimates

the experimental temperature and then overestimates the experi-
mental temperature. Based on the voltage prediction result ob-
tained with the 3 neurons model in Fig. 11a, it is found that the
predicted voltages in the early and later periods of HPPC test are
higher and lower than the experimental voltage respectively, which
correspondingly could underestimate and overestimate the battery
polarization. Generally, the underestimation and overestimation of
the battery polarization could result in underestimating and over-
estimating the irreversible heat generation rate respectively, which
would consequently underestimate and overestimate the experi-
mental temperature using the 3 hidden neurons model shown in
Fig. 12. Moreover, it can be easily observed that the 9 hidden neu-
rons model also underestimates the experimental temperature in
the early period of HPPC test and the 6 hidden neurons model
shows an almost perfect match with experimental temperature
over the whole SOC range. The prediction RMSE of 3, 6 and 9 hidden
neurons models are 0.50, 0.13 and 0.25 K, respectively, and the
MAPE are 0.14%, 0.03% and 0.06%, respectively. From Figs. 11 and 12,
it is, therefore, reasonable to conclude that the proposed ANN Fig. 13. Total term curve of ANN with l1-norm penalty with respect to logl in PCD test.
based thermo-ECM with 6 hidden neurons is validated to be able to
effectively capture the nonlinear voltage response and temperature
evolution of LIBs and can be ready for applying to the practical BMS. For example, the number of terms can increase from 3 to 22 as logl
decreases from 1 to 6. However, it should be noted that the l1-
norm penalty approach cannot guarantee to find all possible
4.2. Model validation of ANN based thermo-ECM with l1-norm
models with all model complexities without proper selection of the
penalty
sequentially changed regularization parameters. Therefore, a series
of regularization parameter should be properly predetermined for
To further investigate whether there are redundant in the hid-
avoiding missing possible optimal models.
den neurons and how to optimal selection of hidden neurons for
Figs. 14a and b show the fitting RMSE and MAPE of battery
circuit parameters, the ANN based thermo-ECM with l1-norm
terminal voltage and temperature with respect to logl, respectively.
penalty are sequentially solved for eight different regularization
Similar as Figs. 7a and b, both fitting RMSE and MAPE of voltage and
parameters, i.e. logl ¼ 1, 0, 1, 2, 3, 4, 5 and 6. The number
temperature decrease with the increasing of the model complexity.
of hidden neurons is set equal to 10 in this study, and consequently
Fig. 14a shows that the fitting RMSE and MAPE of voltage can
there are total 30 output weights in ANN. Fig. 13 shows the model
decrease from 43.8 to 6.5 mV and from 0.92% to 0.15% respectively
complexity with respect to the logarithm of regularization
as logl decreases from 1 to 6. Similarly, as shown Fig. 14b, the
parameter. When the l1-norm penalty is added into objective
fitting RMSE and MAPE of temperature can decrease from 0.35 to
function, some output weights between hidden neurons and circuit
0.11 K and from 0.09% to 0.03% respectively as logl decreases from 1
parameters would be zeros and the number of terms in ANN can
to 6. It is also found that the l1-norm penalty approach can reduce
consequently denotes the number of non-zero output weight.
the total model complexity without deterioration of the fitting
Hence, the number of terms in ANN can be used to characterize the
performance. For example, the fitting RMSE of 6.9 mV obtained by
model complexity. As shown in Fig. 13, the number of terms in ANN
the model with logl ¼ 5 is smaller than that of 8.3 mV obtained by
usually increases with the decreasing of regularization parameter.
the 6 hidden neurons model in Fig. 7a. Correspondingly, there are
total 16 terms in ANN using the l1-norm penalty approach with
logl ¼ 5, in which 6, 7 and 3 hidden neurons are optimally
selected for modeling the circuit parameters of R0, R1 and C1
respectively. It is, therefore, obviously found that compared to the 6
hidden neurons model, the l1-norm penalty approach could not
only reduce two terms in ANN but also improve the fitting per-
formance. The main reason is that different circuit parameters have
different complexity characteristics with respect to SOC and tem-
perature, and consequently it is suitable to model them with
different complexities. Based on the l1-norm penalty approach re-
sults with logl ¼ 5, it is observed that the 7 non-zero terms in
ANN is selected to model the most complex circuit parameter R1.
The results thus validate the effectiveness of the l1-norm penalty
approach, which could be helpful for adaptively determining the
best possible non-zero output weights in ANN and consequently
reducing the model complexity without sacrificing the fitting
performance.
Fig. 15 compares the fitting results of battery voltage response in
PCD test using the l1-norm penalty approach with logl ¼ 1, 2,
and 5, in which the corresponding non-zero terms in ANN are 3,
Fig. 12. Comparison of temperature prediction results using ANN based thermo-ECM
12 and 16 terms respectively. From Fig. 15a, it can be seen that the
with 3, 6 and 9 hidden neurons number in HPPC test. model obtained with logl ¼ 1 poorly captures the voltage response
128 Q.-K. Wang et al. / Energy 138 (2017) 118e132

Fig. 14. Fitting RMSE and MAPE of ANN based thermo-ECM with l1-norm penalty with
respect to logl in PCD test: (a) terminal voltage; (b) temperature.

Fig. 15. Comparison of terminal voltage fitting results using ANN based thermo-ECM
during the whole PCD test, and the corresponding values of RMSE with l1-norm penalty with logl ¼ 1, 2 and 5 in PCD test: (a) 5e100% SOC; (b)
20e30% SOC.
and MAPE are 33.5 mV and 0.76% respectively. The model with
logl ¼ 5 shows the best fitting performance and can match very
well with the measured voltage over the entire SOC range, with a
RMSE of 6.9 mV and a MAPE of 0.16%. From the enlarged section at
the 20e30% SOC shown in Fig. 15b, it can be seen that the models
with logl ¼ 1 and 2 overestimate and underestimate the voltage
response at this SOC range respectively, while the model with
logl ¼ 5 could almost accurately capture the nonlinear voltage
response in both discharge and rest periods.
The fitting results of battery temperature in PCD test using the
l1-norm penalty approach with logl ¼ 1, 2, and 5 are shown in
Fig. 16. It is found that the model with logl ¼ 1 has the maximum
fitting RMSE and MAPE 0.35 K and 0.09% respectively, and the
models with logl ¼ 2 and 5 have almost the same fitting per-
formance and almost accurately match the experimental temper-
ature, with RMSE of 0.16 and 0.11 K respectively. The model with
logl ¼ 1 show an obvious overestimation and underestimation of
experimental temperature at medium and low SOC range respec-
tively. The main reason is also ascribed to the poor estimation of the
voltage as explained in Fig. 12. From Figs. 15 and 14, the satisfied
fitting performance of both voltage and temperature using the
model with logl ¼ 5 thus indicates that the proposed l1-norm
Fig. 16. Comparison of temperature fitting results using ANN based thermo-ECM with
penalty approach can be used to effectively capture both electrical l1-norm penalty with logl ¼ 1, 2 and 5 in PCD test.
and thermal characteristics of LIBs by means of proper selection of
regularization parameters.
The prediction performance of the ANN based thermo-ECM for both terminal voltage and temperature. The corresponding
with l1-norm penalty is also validated by predicting the unseen minimum values of voltage prediction RMSE and MAPE are 11.1 mV
HPPC test dataset. Figs. 17a and 15b show the prediction RMSE and and 0.27% respectively, and the minimum values of temperature
MAPE of terminal voltage and temperature results with respect to prediction RMSE and MAPE are 0.13 K and 0.03% respectively. The
logl, respectively. As shown Figs. 17a and 15b, the minimum pre- RMSE and MAPE curves with respect to the logarithm of regulari-
diction RMSE and MAPE are obtained by the model with logl ¼ 5 zation parameter also shows that the prediction performance
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 129

Fig. 17. Prediction RMSE and MAPE of ANN based thermo-ECM with l1-norm penalty
with respect to logl in HPPC test: (a) terminal voltage; (b) temperature.

Fig. 18. Comparison of terminal voltage prediction results using ANN based thermo-
ECM with l1-norm penalty with logl ¼ 1, 2 and 5 in HPPC test: (a) 10e100% SOC;
usually first increases with the increasing of model complexity and (b) 20e30% SOC.
then decreases with further increasing the model complexity. The
possible reason is that the relatively simple and complex models
would increase the risks of under-fitting and over-fitting respec- logl ¼ 5 almost perfectly matches with the experimental tem-
tively. Therefore, selection of a reasonable model with proper perature. As explained in Fig. 12, the poor prediction performance
complexity would play a significant role in obtaining satisfactory of temperature with the models with logl ¼ 1 and 2 is mainly due
prediction performance. From Figs. 17a and b, it can be reasonable to the large prediction errors in voltage, which results in poor
to conclude that the model with 16 non-zero terms in ANN, which irreversible heat generation rate prediction. Compared the voltage
is obtained with logl ¼ 5, can be selected for practical BMS and temperature prediction results in Figs. 10e12 with that in
application in terms of prediction performance and computational
cost.
Fig. 18 compares the prediction results of battery voltage
response in HPPC test using the l1-norm penalty approach with
logl ¼ 1, 2 and 5. It is found that the model with logl ¼ 1 shows
the worst voltage prediction performance, with a RMSE of 35.5 mV
and a MAPE of 0.81%. At the 40e90% SOC range, a significant un-
derestimation of voltage is observed during the low-rate discharge
period and an overestimation of voltage is found during the rest
and charge periods. In addition, the model with logl ¼ 1 fails to
capture the voltage drop during the end of HPPC test. The models
with logl ¼ 2 and 5 show similar accurate prediction perfor-
mance at the 50e100% SOC range, while the model with logl ¼ 2
shows a larger prediction error at the low SOC range. From the
enlarged section of 20e30% SOC shown in Fig. 18b, it is found that
both models with logl ¼ 1 and 2 obviously show poor matches
with the experimental voltage and the model with logl ¼ 5 could
almost accurately predict the voltage response except for an over-
estimation of voltage during the charge period.
Fig. 19 shows the prediction results of battery temperature using
the l1-norm penalty approach with logl ¼ 1, 2 and 5. It is
obviously observed that both models with logl ¼ 1 and 2 over-
Fig. 19. Comparison of temperature prediction results using ANN based thermo-ECM
estimate battery temperature after 1500 s, while the model with with l1-norm penalty with logl ¼ 1, 2 and 5 in HPPC test.
130 Q.-K. Wang et al. / Energy 138 (2017) 118e132

Figs. 17e19, although the 6 hidden neurons model and the model only takes a relatively small effect on discharge capacity. For
with logl ¼ 5 have almost comparable voltage and temperature example, the maximal discharge capacity difference between 1C-
prediction performance, the obtained model with logl ¼ 5 is and 2C-rate at the same temperature is only about 0.3 Ah. From
simpler than the 6 hidden neurons model in terms of the number of Fig. 20b, it is seen that battery temperature rising during CCD test at
non-zeros terms in ANN. Therefore, it further indicates that the l1- low temperature is significantly higher than that during CCD test at
norm penalty approach provides a more convenient way to adap- high temperature. For example, the temperature rising during CCD
tively determine the optimal terms for different circuit parameters, test with 2C-rate at 263, 293 and 323 K are 25.92, 13.60 and 1.35 K
which would be greatly helpful for constructing a simpler model respectively. The main reason is that the battery resistance would
without sacrificing the prediction performance. increase with the decreasing of temperature, especially at relatively
lower temperature, which could correspondingly result in higher
4.3. Model validation by constant current discharge test heat generation rate. Although both discharge rate and tempera-
ture would take complex effects on the battery voltage and tem-
The ANN based thermo-ECM is further validated by battery perature responses, the small fitting RMSE under all discharge rates
discharge test with different C-rate and temperatures. The dataset and temperature validates that the proposed ANN method is able to
of CCD tests with 1C- and 2C-rate at 263, 293 and 323 K are used to capture such complex nonlinear relationship between tempera-
carry out the parameter identification for constructing the ANN ture/SOC and circuit parameters.
based thermo-ECM model, and the dataset of CCD tests with 1.5C- The prediction performance of proposed ANN based thermo-
rate at 263, 293 and 323 K are used to verify the prediction per- ECM for CCD test data with 1.5C-rate at different temperatures is
formance of constructed model. shown in Fig. 21. It can be found that the predicted terminal voltage
Fig. 20 shows the fitting performance of proposed ANN based and battery temperature coincide well with experimental data. The
thermo-ECM for battery terminal voltage and temperature during prediction RMSE of battery terminal for 263, 293 and 323 K are
CCD tests with 1C- and 2C-rare at different temperatures. It is 10.2, 11.8 and 9.6 mV respectively, and the prediction RMSE of
clearly observed that the proposed method could almost perfectly battery temperature for 263, 293 and 323 K are 0.54, 0.48 and 0.14 K
capture both the voltage and temperature responses with the respectively. Both the fitting and prediction performance indicate
fitting RMSE of battery terminal voltage and temperature is 9.8 mV that the proposed ANN base thermo-ECM could be used to accu-
and 0.58 K, respectively. The discharge capacity would usually in- rately capture battery electrical and thermal dynamic response.
crease with the increasing of temperature and decrease with the
increasing of discharge rate. It is found that the temperature has a 5. Conclusions
significant effect on discharge capacity, especially for low temper-
ature discharge. For example, the discharge capacities for CCD tests A unified modeling framework is proposed for describing the
with 1C-rate at 263, 293 and 323 K are 16.5, 18.8 and 19.3 Ah electrical and thermal dynamic behaviors of LIBs. A lumped-
respectively. While in the normal discharge rate region, the rate parameter thermo-ECM model is constructed and the
temperature/SOC-related circuit parameters are modeled by means
of ANN. The coupling between the thermal model and ECM is

Fig. 21. Prediction performance of battery terminal voltage and temperature for CCD
Fig. 20. Fitting performance of battery terminal voltage and temperature for CCD tests tests with different C-rates and temperature. (a) battery terminal voltage; (b)
with different C-rates and temperatures. (a) battery terminal voltage; (b) temperature. temperature.
 Q.-K. Wang et al. / Energy 138 (2017) 118e132 131

thoroughly explained. A weighted normalized sum square error is LHS Latin hypercube sampling
used as optimization objective and the parameter identification MINLP mixed-integer nonlinear programming
problems with and without l1-norm penalty are efficiently solved RMSE root mean square error
using BARON. Three types of characteristic tests, i.e. PCD, HPPC and MAPE mean absolute percentage error
CCD, of a commercial LIB are used to validate the effectiveness of CCCV constant current constant voltage
the proposed approach. PCD pulse current discharge
The fitting and prediction performance of ANN based thermo- HPPC hybrid pulse power characterization
ECM with different number of hidden neurons are first systemati- CCD constant current discharge
cally studied. It is found that the fitting performance can be
continuously improved with the increasing of hidden neurons. Nomenclature
However, the prediction performance can be deteriorated as the
number of hidden neurons is above 6. The voltage and temperature
prediction RMSE using the 6 hidden neurons model are 10.8 mV Thermo-ECM
and 0.13 K in HPPC test respectively. Then, the l1-norm penalty VOC open circuit voltage [V]
approach is applied to adaptively select the best possible terms in R0 internal resistance [U]
ANN for different circuit parameters. It is found that fitting per- R1 polarization resistance [U]
formance usually increases with the decreasing of regularization C1 capacitance [F]
parameter, while the best prediction performance is obtained by I total current [A]
the model with logl ¼ 5, in which the corresponding voltage and I1 currents through resistance R1 [A]
temperature RMSE are 11.1 mV and 0.13 K respectively. The fitted I2 currents through capacitance C1 [A]
and predicted curves of both voltage and temperature illustrates t time [s]
that both strategies could almost accurately capture the nonlinear Cn battery norminal capacity [Ah]
electrical and thermal characteristics of LIBs. Moreover, these re- V0 voltage drop across R0 [V]
sults indicate that the l1-norm penalty approach is more convenient V1 voltage drop across parallel RC network [V]
to construct a simpler model with satisfactory fitting and predic- VT terminal voltage [V]
tion accuracy. Moreover, CCD tests with different C-rate at different t time constant of the parallel RC network [s]
temperatures are also used to further validate the effectiveness of Q heat generation rate [W]
proposed ANN based thermo-ECM. It is found that the fitting RMSE T battery temperature [K]
of terminal voltage and temperature are 9.8 mV and 0.58 K Tamb ambient temperature [K]
respectively, and the prediction RMSE of terminal voltage and Cp heat capacity [J/(kg  K)]
temperature for CCD test with 1.5 C-rate at different temperatures m cell weight [kg]
are less than 12 mV and 0.6 K respectively. To conclude, the pro- A cell surface area [m2]
posed ANN based thermo-ECM provides a general modeling h convective heat transfer coefficient [J/(m2*K)]
framework for the inclusion of both temperature and SOC effects,
and can be easily extended to the inclusion of other effects such as Subscript
health state and current. Furthermore, the accurate prediction k number of time sampling
ability of the proposed model would be greatly helpful for p prescribed number of samples generated by Latin
achieving accurate and reliable state monitoring in practical BMS. hypercube sampling

Acknowledgments Artificial neural network

x input vectors
The authors gratefully acknowledge financial support from the y output vectors
National Key R&D Program of China (2016YFB0901505), the Major aj input weights of the jth hidden neuron
State Basic Research Development Program of China bj input bias of the jth hidden neuron
(2014CB239703), the National Natural Science Foundation of China bij weight connecting the jth hidden neuron to the ith output
(21576163, 21336003), the Science and Technology Commission of neuron
Shanghai Municipality (14DZ2250800), and the Project-sponsored g activation function
by SRF for ROCS, SEM.
Optimization problem
Abbreviation bT
V predicted terminal voltages [V]
b
T predicted temperatures [K]
ECM equivalent circuit model k

LIB lithium ion battery w weight characterizing the relative importance

SOC state of charge
li regularization parameters
BMS battery management system
SOH state of health References
OCV open circuit voltage [1] Gao DWZ, Mi C, Emadi A. Modeling and simulation of electric and hybrid
ANN artificial neural network vehicles. Proc IEEE 2007;95(4):729e45.
EIS electrochemical impedance spectroscopy [2] Lu L, Han X, Li J, Hua J, Ouyang M. A review on the key issues for lithium-ion
battery management in electric vehicles. J Power Sources 2013;226:272e88.
RC resistance-capacitance
[3] Waag W, Fleischer C, Sauer DU. Critical review of the methods for monitoring
MIMO multi-input-multi-output of lithium-ion batteries in electric and hybrid vehicles. J Power Sources
SLFN single layer feed-forward network 2014;258:321e39.
WNSSE weighted normalized sum square error [4] Pattipati B, Sankavaram C, Pattipati KR. System identification and estimation
framework for pivotal automotive battery management system characteris-
NSIP nonlinear semi-infinite programming tics. IEEE Trans Syst Man Cybern C Appl Rev 2011;41(6):869e84.
NLP nonlinear programming [5] Feng X, Lu L, Ouyang M, Li J, He XA. 3D thermal runaway propagation model
132 Q.-K. Wang et al. / Energy 138 (2017) 118e132

for a large format lithium ion battery module. Energy 2016;115:194e208. [22] Kroeze RC, Krein PT. Electrical battery model for use in dynamic electric
[6] Cheng KWE, Divakar BP, Wu HJ, Ding K, Ho FH. Battery-management system vehicle simulations. In: PESC record - IEEE annual power electronics spe-
(BMS) and SOC development for electrical vehicles. IEEE Trans Veh Technol cialists conference; 2008. p. 1336e42.
2011;60(1):76e88. [23] Cheng XM, Yao LG, Xing YJ, Pecht M. Novel parametric circuit modeling for Li-
[7] Li Z, Huang J, Liaw BY, Zhang J. On state-of-charge determination for lithium- Ion batteries. Energies 2016;9(7):539e53.
ion batteries. J Power Sources 2017;348:281e301. [24] Castano S, Gauchia L, Voncila E, Sanz J. Dynamical modeling procedure of a Li-
[8] Doyle M, Fuller T, Newman J. Modeling of galvanostatic charge and discharge ion battery pack suitable for real-time applications. Energ Convers Manage
of the lithium/polymer/insertion cell. J Electrochem Soc 1993;140(6): 2015;92:396e405.
1526e33. [25] Shen JN, He YJ, Ma ZF. Simultaneous model selection and parameter estima-
[9] Klass V, Behm M, Lindbergh G. Capturing lithium-ion battery dynamics with tion for lithium-ion batteries: a sequential MINLP solution approach. AIChE J
support vector machine-based battery model. J Power Sources 2015;298: 2016;62(1):78e89.
92e101. [26] Hariharan KS. A coupled nonlinear equivalent circuit-thermal model for
[10] Moss PL, Au G, Plichta EJ, Zheng JP. An electrical circuit for modeling the lithium ion cells. J Power Sources 2013;227:171e6.
dynamic response of Li-ion polymer batteries. J Electrochem Soc [27] Miranda AG, Hong CW. Integrated modeling for the cyclic behavior of high
2008;155(12). A986eA94. power Li-ion batteries under extended operating conditions. Appl Energy
[11] Hariharan KS, Kumar VS. A nonlinear equivalent circuit model for lithium ion 2013;111:681e9.
cells. J Power Sources 2013;222:210e7. [28] Du J, Liu Z, Wang Y, Wen C. A fuzzy logic-based model for Li-ion battery with
[12] Hu XS, Li SB, Peng H. A comparative study of equivalent circuit models for Li- SOC and temperature effect. In: IEEE international conference on control and
ion batteries. J Power Sources 2012;198:359e67. automation; 2014. p. 1333e8.
[13] Bandhauer TM, Garimella S, Fuller TF. A critical review of thermal issues in [29] Hu Y, Yurkovich S, Guezennec Y, Yurkovich BJ. Electro-thermal battery model
lithium-ion batteries. J Electrochem Soc 2011;158(3):R1e25. identification for automotive applications. J Power Sources 2011;196(1):
[14] Rao Z, Wang S. A review of power battery thermal energy management. 449e57.
Renew Sust Energ Rev 2011;15(9):4554e71. [30] Malik A, Zhang Z, Agarwal RK. Extraction of battery parameters using a multi-
[15] Bernardi D, Pawlikowski E, Newman J. A general energy balance for battery objective genetic algorithm with a non-linear circuit model. J Power Sources
systems. J Electrochem Soc 1985;132(1):5e12. 2014;259:76e86.
[16] Wu B, Li Z, Zhang J. Thermal design for the pouch-type large-format lithium- [31] Buller S, Thele M, De Doncker RWAA, Karden E. Impedance-based simulation
ion batteries I. Thermo-electrical modeling and origins of temperature non- models of supercapacitors and li-ion batteries for power electronic applica-
uniformity. J Electrochem Soc 2015;162(1). A181eA91. tions. IEEE Trans Ind Appl 2005;41(3):742e7.
[17] Lin X, Perez HE, Mohan S, Siegel JB, Stefanopoulou AG, Ding Y, et al. A lumped- [32] Dong TK, Kirchev A, Mattera F, Kowal J, Bultel Y. Dynamic modeling of Li-ion
parameter electro-thermal model for cylindrical batteries. J Power Sources batteries using an equivalent electrical circuit. J Electrochem Soc 2011;158(3).
2014;257:1e11. A326eA36.
[18] Veth C, Dragicevic D, Pfister R, Arakkan S, Merten C. 3D electro-thermal model [33] Stein M. Large sample properties of simulations using Latin hypercube sam-
approach for the prediction of internal state values in large-format lithium ion pling. Technometrics 1987;29(2):143e51.
cells and its validation. J Electrochem Soc 2014;161(14). A1943eA52. [34] He YJ, Shen JN, Shen JF, Ma ZF. Embedding monotonicity in the construction of
[19] Damay N, Forgez C, Bichat MP, Friedrich G. Thermal modeling of large pris- polynomial open-circuit voltage model for lithium-ion batteries: a semi-
matic LiFePO4/graphite battery. Coupled thermal and heat generation models infinite programming formulation approach. Ind Eng Chem Res
for characterization and simulation. J Power Sources 2015;283:37e45. 2015;54(12):3167e74.
[20] Saw LH, Ye Y, Tay AAO. Electro-thermal characterization of Lithium Iron [35] Shen JN, He YJ, Ma ZF, Luo HB, Zhang ZF. Online state of charge estimation of
Phosphate cell with equivalent circuit modeling. Energ Convers Manage lithium-ion batteries: a moving horizon estimation approach. Chem Eng Sci
2014;87:367e77. 2016;154:42e53.
[21] Chen M, Rinco  n-Mora GA. Accurate electrical battery model capable of pre- [36] Tawarmalani M, Sahinidis NV. A polyhedral branch-and-cut approach to
dicting runtime and I-V performance. IEEE Trans Energy Convers 2006;21(2): global optimization. Math Program 2005;103(2):225e49.
504e11.