12
CHAPTER 2
REVIEW ON MODELING AND STATE OF CHARGE
ESTIMATION OF LITHIUM-ION BATTERY IN
ELECTRIC VEHICLE
2.1 INTRODUCTION
Extension of driving range and battery run time optimization are
necessary key points in the modeling of EV. In this view, BMS plays a major
role to ensure a safe and trustworthy battery operation, especially when using
Lithium-ion (Li-ion) batteries in an electric vehicle. Key function of BMS is
SOC estimation. A well-parameterized battery model is required for an
accurate state estimation. Because, a good battery model is essential to
monitor & analyse the battery behavior, identify the faults and prevent a
battery from thermal runaway. Furthermore, State of Charge plays a crucial
role in regulating battery operation and it must consequently be monitored
using correct estimation methods. Therefore, a well-trained battery model
together with proper SOC estimation technique is used to estimate the SOC of
a battery precisely.
By focusing on these features, in this chapter, the well-known
battery models such as the electrochemical model, equivalent circuit model,
and data-driven model are comprehensively reviewed along with their
strengths and weaknesses. Further, the SOC estimation of a battery is also
discussed by using standard methodologies such as direct estimation methods
and model-based estimation methods. The comparisons of three most distinct
13
battery models and the classification of SOC estimation techniques to develop
a proper BMS for EV with the focus on accuracy, configuration effort,
computational complexity, ease of implementation, and real-time applications
are systematically reviewed. In this chapter, two key issues such as battery
modeling and SOC estimation of Li-ion battery cell are focused and the
relation of these issues is illustrated in Figure 2.1.
Figure 2.1 Relation between battery modeling and SOC estimation
methods
2.2 BATTERY MODELS
In EV applications, accurate SOC estimation reduces the range
anxiety of EV customers. A good battery model is essential for battery state
estimation. The starting point of BMS design is to build a proper battery
model. Till date, numerous battery models have been developed with various
accuracy levels. The most distinct battery models are the electrochemical
model, equivalent circuit model and data-driven model.
14
2.2.1 Electrochemical Model
Electrochemical models are used to describe battery behaviors
based on the electrolyte concentration, size of the electrodes (anode and
cathode) and electrochemical process inside the battery using partial
differential equations. Eventhough EM provides accurate battery parameters,
it needs more computation power and time to find numerous parameters such
as electrolyte potential, solid potential, open circuit potential, overpotential,
electrolyte concentration, solid concentration, battery cell current, temperature
and so on. Moreover, it is difficult to implement in real-time applications.
Comparison of various types of electrochemical models of battery is
summarized in Table 2.1. On the basis of its physical process, Doyle et al.
(1993) have suggested a Pseudo-2-D (P2D) electrochemical model. The
extended simulation time of the P2D model due to huge number of nonlinear
equations makes it computationally inefficient for BMS applications.
Domenico et al. (2008) have proposed a reduced order EM by
taking an average solid electrolyte concentration instead of the distribution
along the electrodes. Eventhough the model losses some information, it can
be implemented in real-time onboard buses. However, the identification of
parameters is a difficult task.
Ahmed et al. (2014); Ahmed et al. (2014a) have identified the
battery parameters using genetic algorithm and estimated SOC. For
identification, the model is run 10 times and a population size of 1000 is used
as well as identification of 18 parameters is done in 6 hours. Even this model
has the advantage of reducing computational power compared to the full order
model; accuracy is reduced due to some assumptions that it is made to reduce
the order of the model.
15
Table 2.1 Comparison of various types of electrochemical models of
battery
S. No Model Merits Demerits
Pseudo-2-D (P2D)
electrochemical model
or Doyle-Fuller- High computational
1 High accuracy
Newman complexity
electrochemical
battery model.
Radial-domain PDEs
in the SP model still
Simple model that
need to be solved in
Reduced order model reduces the electrode
order to obtain the
2 or Single particle (SP) to a single particle and
solid-phase Li-ion
model neglect the liquid
concentration.
phase.
Lack accuracy at high
C rates.
It simplifies the
electrode into a single
active particle. Curve
fitting or approximate
Extended Single Model complexity is
3 solution process is
Particle (ESP) Model. increased
used to solve PDEs.
Improves the
computational
efficiency.
Han et al. (2015) have proposed an approximate model that
monitors the diffusion process and the distribution of electrolyte
concentration inside the battery. Zou et al. (2016) have proposed a reduced-
order model based on a singular perturbation approach and an averaging
16
theory to estimate the SOC of Li-ion as well as to predict the discharging
capacity of the battery. This methodology of model simplification is suitable
for all batteries. However, developing a high-fidelity model concerning
temperature, aging, and capacity fading not only improves the accuracy but
also increases the complexity.
2.2.2 Equivalent Circuit Model
ECM uses electrical components such as voltage source, resistors
and capacitors that describe the electrical behavior of the battery. According
to circuit theory, model equations and parameters are specified in Table 2.2.
In ECM, a high valued capacitor (Sitterly et al. 2011) or controlled voltage
source (Chen & Rincón-Mora, 2006) is used to denote the Open Circuit
Voltage (OCV) of battery and it is the most important parameter for many
state estimation techniques. Hu et al. 2012 parative
study of twelve equivalent circuit models for Li-
twelve ECM models, Plett, 2004b has projected six battery models such as a
combined model, a simple model, zero-state hysteresis model, one state
hysteresis model, enhanced self-correcting model (with 2 state Low Pass
Filter, 4 state Low Pass Filter).
The simplest ECM is the Rint model (Nejad et al. 2016) which
considers the battery as a voltage source with series resistance. Although this
is simple to implement, it cannot accurately present the characteristics of the
battery used in EV. A single Resistance-Capacitance parallel network is
added to Rint model to form a Thevenin model (Xu et al. 2016). The
Thevenin model is widely used to capture the dynamic characteristics of a
battery. The modified version of Thevenin model is a Partnership for a New
Generation of Vehicle (PNGV) model (Siguang & Chengning, 2009) which
includes a fictive capacitor to reflect the effects of OCV variation (Id et al.
2018).
17
18
PNGV model also known as FreedomCar model (Omar et al. 2014)
comprises open circuit voltage, ohmic resistance, fictive capacitor,
polarization resistance and capacitor. PNGV model is an ideal choice for low
SOC area whereas it is not suitable for high SOC region (Lai et al. 2019).
Among the various ECM models available in the literature,
Resistance-Capacitance (RC) network-based models such as one RC network
ECM (Fotouhi et al. 2015; Fotouhi et al. 2018), two RC network ECM (Y. Hu
& Wang, 2015; K. Yang et al. 2021; Yao et al. 2013) and three RC network
ECM (Cao et al. 2016; A. Wang et al. 2017) are widely accepted models for
online applications (Thirugnanam et al. 2014; C. Zhang et al. 2018)..
Among these three, two RC network model has high accuracy to
predict the relation between the input current and the output voltage (I-V) of
the battery as well as the charging and discharging time of the battery (Chen
& Rincón-Mora, 2006). Since the battery is a nonlinear system, the battery
dynamics vary for different operating conditions such as SOC, temperature,
charging, and discharging rate of a battery. Therefore, the model
asured data. Hence, the
battery model is fixed, and the model parameters are to be identified. The
well-known Least Square (LS) technique (Hu & Wang, 2015; Nugroho et al.
2016; Qiu et al. 2020), and the Genetic Algorithm (Fotouhi et al. 2018; Malik
et al. 2014; Thirugnanam et al. 2014) are applied to identify the model
parameters with experimental data. Sitterly et al. 2011 have established that
the model must be identified. Without laboratory experimentation, the model
is identified with real-time operating data from EV for BMS development. An
automated real-time battery characterizer has been developed to capture and
update cell characteristics in real-time. For evaluation, the RC model that is
developed in the National Renewable Energy Laboratory (NREL) has been
19
taken. Real-time data are obtained from the Advanced Vehicle Simulator
(ADVISOR) tool. In addition to that, the identification of the system, the
convergence rate of the estimation algorithm, the bias in the identification and
the correction of the bias are thoroughly verified. However, ECM circuit
elements do not describe the proper physical representation of batteries.
Besides, the model parameters are updated at all times with the SOC,
temperature, and charge-discharge rate of a battery.
2.2.3 Data-Driven Model
DDM is significantly more efficient than ECM as well as EM but it
highly depends on the amount of data along with the training method. DDM
has the ability of computational intelligence to approximate highly non-linear
battery characteristics. Various data-driven models including ANN
(Boujoudar et al. 2019), Adaptive Neuro-Fuzzy Inference Systems (ANFIS)
(Awadallah & Venkatesh, 2016), Deep Neural Network (DNN)
(Khumprom & Yodo, 2019), Support Vector Machine (SVM) (Junping et al.
2006; Klass et al. 2015; Weng et al. 2015) describe the behavior of a battery
ANN is a powerful tool to develop a model for any non-linear
function when training of data is done properly (He et al. 2014). The accuracy
of this model demands large quantities of high-quality data with different load
conditions and also abundant training time. Feed Forward Neural Network
(FFNN) is the least complex among other data driven models because of its
simple learning process and non-intensive computation. Like FFNN, DNN
has high computational efficiency compared to other learning algorithms. But,
FFNN and DNN are one-way where the history of data is not considered.
Both are well trained, i.e., network biases and weights are learned through
offline and there is no need for a backward pass during online SOC
estimation. Therefore, the main challenge is that the models have to be trained
20
well in offline. In, the authors have pointed out that the time required to form
a DNN is from a few hours to 40-50 hours at different ambient temperatures;
although the training is performed on the Graphical Processing Unit (GPU),
and it has the advantage of parallel computing. In (Boujoudar et al. 2019), the
Nonlinear Auto Regressive model with eXogenous input (NARX) is used to
obtain the required battery voltage.
Unlike ANN, the Recurrent Neural Network (RNN) has the
advantage of using the history of data with recurrent units (Zhang et al.
2019a). The only problem is memory loss; that means it cannot capture the
information for a long-time because of the gradient vanishing phenomenon.
Hence, Gated Recurrent Unit (GRU)-RNN (Ardeshiri & Ma, 2021; C. Li et
al. 2019) and Long Short Term Memory (LSTM)-RNN (Heinrich et al. 2021;
Yang et al. 2020; Zhang et al. 2018) are established to capture long term
requirements. Since GRU-RNN is structured over time; another method of
backpropagation has to be applied, known as Back-Through-Time-
Propagation (BPTT) to build a network. Proper data normalization is essential
to improve the GRU-RNN training process and it increases the convergence
rate. In LSTM, the memory block is added to a hidden unit to store the
information for a longer time. ANFIS combines the advantages such as
flexibility, subjectivity of fuzzy systems and the learning capability of neural
networks (Khumprom & Yodo, 2019). The Takagi Sugeno (T-S) fuzzy
model (Samadi & Saif, 2017) has an inherent multiple-model structure that
can cope with the nonlinear dynamics of battery. But it is more rule-based and
the accuracy of the model varies with the number of rules, whereas the black-
box models can provide accurate results and they require more computational
complexity as well as it is hard to analyze their black box nature.
In SVM, a small number of samples has been found among huge
volume of data to describe the system dynamics with kernel trick
21
(Klass et al. 2015). Though SVM has a simple structure than ANN, a
computationally expensive optimization problem has been solved to
determine kernel parameters. Radial Basis Function (RBF) kernel is the more
popular choice for determining kernel parameters, due to its good
generalization capability (Hu et al. 2014). Since SVM fails to handle a large
amount of data, it is difficult to implement in SOC and perform SOH
estimation of battery packs.
ANN, DNN, and SVM methods use machine learning algorithms to
estimate the SOC of a battery based on statistical data and they are capable to
predict non-linear parameters of a battery. Among these, DNN performs
better than ANN and SVM. Eventhough Deep learning algorithms need high
computational time and more resources than others, they are well suited for
Remaining Useful Life (RUL) prediction of the battery (Kara, 2021).
However, DDM requires an intensive calculation to understand the properties
of battery in real-time based on the data acquisition and training process. Data
pre-processing and noise removal are also important in all the above-
mentioned methods. Extreme calculations could be another major obstacle for
the implementation of economic applications. Chen et al. 2008 have
introduced the Grey Model (GM) approach for describing a system as a "grey
box" and it can predict unknown parameters using little-known battery
information. D. Zheng et al. 2018 have stated that the Metabolic Grey Model
(MGM), a new methodology is used to track the dynamic data calculations
and the SOC estimation accuracy of a battery is enhanced through the
metabolic mechanism. Each model has its strength and some drawbacks.
Table 2.3 summarizes the advantages and disadvantages of various Data-
Driven models.
22
23
24
25
26
27
Data collection is critical in the development of a data-driven
model because these models require a large amount of data for training. With
the increased deployment of batteries in a variety of applications, batteries
experience varying degradation rates under various operating conditions. As a
result, application-specific data are required to develop an accurate battery
model (Dos Reis et al. 2021). Since data collection is a time-consuming and
costly process, researchers can use publicly available data instead of
conducting experiments on batteries under various operating conditions.
2
DM
1 ECM
EM
0
0-Very Low
1-Low
2-Medium
3-High
4-Very High
Figure 2.2 Comparison of Battery Models with indexed values
Table 2.4 summarizes different combinations of battery models
with parameter identification algorithms. Figure 2.2, shows the comparison of
various battery models with indexing values in terms of accuracy,
configuration effort, computational complexity, ease of implementation, real-
time applications and interpretability.
28
It can be concluded that EM provides high accuracy and
interpreting ability yet still poor in computational complexity, configuration
effort, ease of implementation and real-time applications. The performance of
ECM is most opposite to that of the electrochemical model. Data-driven
models show moderate performance.
The performance comparisons of various types of battery models,
including their strengths and weaknesses, are outlined in Table 2.5. As per the
above discussion, both accuracy and simplicity are the main concerns for
choosing a battery model to design a BMS. According to Table 2.4, care must
be taken to select the appropriate model, model parameter type and parameter
identification algorithm to enhance the accuracy as well as reduce the
complexity.
2.3 SOC ESTIMATION
According to the statement of Plett, 2004a, the SOC of battery is
the ratio of remaining charge to the total charge capacity of the battery.
More specifically, the reduction in SOC accuracy reflects lower
battery efficiency. Therefore, the accuracy of SOC estimation is important in
designing a BMS and it is suitable for electric vehicles.
The classification of SOC estimation methods is primarily divided
into two such as direct estimation methods and model-based estimation
methods.
29
30
31
32
33
34
35
2.3.1 Direct Estimation Methods
The SOC can be calculated by directly measuring the voltage and
current of the battery. For example, the Ampere hour (Ah) method is used to
estimate the state of a battery by integrating the charging and discharging
currents. This is a very simple and low computational method (Hannan et al.
2017). Despite the advantages of the Ah method, it is difficult to measure the
initial SOC in real-time EV applications. Moreover, the accuracy of SOC is
limited due to unknown initial capacity of the battery, the self-discharge rate
and the loss of battery capacity. In general, the estimation using Ah is done by
estimation when the battery is discharged at a constant current rate and
constant temperature. Since EV batteries experience dynamic variations in
load current and temperature during the charging and discharging processes,
they fail to estimate accurate SOC. Hence, some modifications have been
current rate and temperature (Leksono et al. 2013; Xie, Ma & Bai 2018).
In OCV method, the SOC is directly measured by one-to-one
relation between the SOC and the OCV of the battery. If this relationship is
linear, it gives accurate results. But, it is not exactly linear for all types of
batteries because it depends not only on size and type of electrode materials
but also on the electrolyte concentration of the battery (Hannan et al. 2017).
Also, this method needs too much rest time to estimate the accurate SOC of a
battery. Of course, this method is not practical to implement in a running
electric vehicle, since this is valid only for the unloaded condition of the battery
(Fathoni et al. 2017). As a result, only direct measurement methods are suitable
only for low power applications not for real-time EV. Thus, the model-based
estimation methods have been developed to save estimation time by predicting
the performance of battery with modeling (Premkumar et al. 2018).
36
2.3.2 Model based Estimation Methods
This method is further classified into two, such as Adaptive filter &
observer based SOC Estimation and computer intelligence algorithm based
SOC Estimation.
[Link] Adaptive filter & observer based SOC estimation
In general, well-known adaptive techniques of control theory are
combined with the battery model to achieve an online SOC estimate. The
familiar Adaptive filter & observer algorithms are KF (Asghar et al. 2016; W.
Wang & Mu, 2019), EKF (Al-Gabalawy et al. 2020; Plett, 2004a; Taborelli &
Onori, 2014), Unscented Kalman Filter (UKF) (Chen et al. 2019;
Santhanagopalan et al. 2016; Zhu et al. 2020), Fading Kalman Filter (FKF)
(Lim et al. 2016; Y. Yang et al. 2017), Cubature Kalman filter (W. Li et al.
2021; J. Peng et al. 2019; J. Shen et al. 2021), Particle filter (Lai et al. 2021;
Liu et al. 2019; Zhang et al. 2019a) (Xiong, Yu, et
al. 2017; Xue et al. 2014). In all these methods, each algorithm has some
advantages and disadvantages. For example, KF is an optimal estimator and is
often used to estimate the states of linear systems. In the case of non-linear
systems, KF requires complex calculations. Plett et al. 2014 has proposed an
EKF method for SOC estimation of non-linear battery model. However, if the
system is extremely non-linear, linearization error occurs due to the lack of
accuracy, and also it necessitates more computational effort.
Compared to EKF, the UKF may be implemented to provide better
results for highly non-linear model. Though it eliminates the linearization
error, Cholesky factorizations on every step and sigma point selection affect
the performance. FKF is capable to correct the modeling error using a fading
concept. But, it needs more computing power to implement the hardware.
However, filter parameters like noise covariance matrices affect the
37
estimation accuracy & convergence rate and those noises must be a Gaussian.
KF algorithms fail to handle the non-gaussian noises. In general, noise
covariance matrices are chosen by trial and error method based on statistical
knowledge (Liu et al. 2017; Meng et al. 2019; Pang et al. 2020). Besides,
AEKF (Yang et al. 2019; Yang et al. 2017), and AUKF (Meng et al. 2016;
Peng et al. 2017; Zhang et al. 2020) are developed to automatically update
noise covariance matrices. Despite accurate estimation, computational time
and complexity have been increased.
However, the calculation complexities of all KF algorithms are
identical. Besides, it requires a well-parameterized battery model but often the
parameters are varied drastically for temperature changes and battery aging.
For this purpose, the recursive least square algorithm is recommended by
many authors. If the system has parameter uncertainties and often errors occur
i
the same. But, it also exhibits similar problems as in KF based methods. Also,
Accuracy and convergence rate of this observer depend on gain. Determining
an observer gain is based on the adjustment coefficient which can be chosen
by an optimization technique. KF algorithms can handle initial SOC error due
to their self-correcting nature. Therefore, KF algorithms are suitable to
estimate the state of fast time-varying systems when a well-parameterized
model is available.
[Link] Computer intelligence algorithm based SOC estimation
According to the rapid development of computer intelligence
algorithms, some machine learning algorithms have been used to estimate the
SOC of the battery (Dineva et al. 2021; Ng et al. 2020). These algorithms can
use training data to estimate SOC without prior knowledge of battery
chemistry. SOC has been estimated using fuzzy logic (Singh et al. 2006). In
38
this fuzzy model, a maximum as well as a minimum voltage value is collected
with the cycle number used as an input to generate a fuzzy interference
system.
Due to self-learning, the ANN (Boujoudar et al. 2019) has been
familiar with the validation of complex nonlinear model. Although it is
extremely dependent on collective information by training, it requires a
smaller computational cost. But, the challenge is that the ANN has the risk of
overlearning. As a result, the optimization algorithm is necessary to find the
number of hidden layers and the number of neurons per layer. When large
real-world data are available, DNN can be used to estimate SOC without the
need of filtering algorithms (Chemali et al. 2018). DNN can estimate SOC
through online with low computational time, but it needs more time for an
offline training process, since it is a multi-layer architecture that increases the
number of neurons. SVM uses a regression algorithm to predict the
parameters of the battery and estimates the SOC of a battery. According to the
structural risk minimization principle, SVM works for state estimation
(Alvarez Anton et al. 2013).
Eventhough the standard SVM has a simple structure; it needs more
computational effort to find the kernel parameters. SVM provides a modular
framework that can adapt to any kind of problem with kernel functions. While
the ANN has a fixed structure and provides more outputs, the size of SVM
has increased much more and offers only one output.
Hu et al. (2016) and Shen et al. (2002) have estimated SOC using
ANFIS model. ANFIS utilizes the best features of ANNs and Rule-based
Fuzzy systems to learn the behavior of a battery from the dataset and adjusts
the parameters of a battery according to error generation (Singh et al. 2020).
Zahid et al. (2018) have proposed the SOC estimation method based on
39
Subtractive Clustering Neuro-Fuzzy architecture under ten different unseen
driving cycles.
Some of the researchers have proposed a combined model that uses
both the types of algorithms. For example, Charkhgard & Farrokhi (2010)
have created a battery model using neural network concepts and estimated the
capacity using EKF. Yang et al. (2020) have developed an estimation
technique using LSTM and UKF. Likewise, Du et al. (2014) have developed
a neural network model using an extreme learning machine and AUKF is used
for SOC estimation. But these are not more attractive and they also make the
calculations more complex as well as increase the computational cost. Some
joint estimation algorithms are developed like SOC and SOH estimation
(Li et al. (2021); Song et al. (2020); Zou et al. (2015), SOC and SOE
estimation (Xie et al. 2018; Zhang et al. 2017), and SOC and SOP estimation
(Hu et al. 2014; Verma et al. 2020; Xiong et al. 2013) using aforementioned
estimation techniques.
Both the algorithms have been implemented in real-time
applications and those methods include some advantages and certain
drawbacks. It can be concluded that direct estimation methods are simple and
low-cost but not suitable for real-time applications. Adaptive filter and
observer methods are the best in terms of accuracy, storage size and suitable
for real-time applications yet still poor in terms of computational complexity,
configuration effort, and ease of implementation. Computer intelligence
methods furnish superior performance in terms of computational complexity,
real-time applications and ease of implementation but lack accuracy, require
more training and high storage size than the others. Table 2.6 summarizes the
information about the combinational technique that may comprise the model,
identification technique and an algorithm used for SOC estimation.
40
41
42
43
44
45
46
47
48
Based on the review, it is concluded that an adaptive filter based
SOC estimation can give accurate results with the proper selection of battery
model (EM/ECM/DDM), parameter type (Time variant/Time Invariant), and
parameter identification algorithm (Least square/optimization algorithm). On
the other hand, computer intelligence algorithm based SOC estimation
requires an appropriate model selection, an algorithm for tuning
hyperparameters, a proper training algorithm and a large amount of data
collection & normalization to extend a trade-off between accuracy and
complexity.
2.4 SUMMARY
The battery management system plays a vital role in monitoring
and controlling battery energy storage systems in electric vehicles. Since the
development of battery technology has been increasing, the performance
enhancement of the battery management system in electric vehicles is an
important factor to be considered because it controls the life of the costly
battery. The most important facts of the battery management system have
been reviewed in this chapter, especially in the fields of battery modeling, and
state of charge estimation. To improve the accuracy while reducing the
complexity of battery modeling, selection of a suitable model type as well as a
parameter identification algorithm is required. Based on the review of SOC
estimation, it is stated that an adaptive filter-based state of charge estimation
may provide reliable results when the battery model, parameter type and
parameter identification algorithm are properly selected. On the other hand,
computer intelligence based state of charge estimation necessitates an
appropriate model selection, an algorithm for tuning hyperparameters, a
proper training approach and a considerable amount of data collection as well
as normalization to provide a trade-off between accuracy and complexity.
However, most of the key technologies in the battery management system are
49
achieved and validated in specific laboratory test conditions. But the real-
world performance is different from laboratory results and hence, it is difficult
to extend guarantee. In order to reduce the range anxiety of electric vehicle
customers and prevent cell inconsistency in the battery pack, an accurate and
simple algorithm has to be developed for state of charge estimation. The
algorithm must be validated for the dynamic conditions of an electric vehicle.