Robust Feedforward Neural Network-Based State of Charge

Estimation of Lithium-Ion Batteries

V Indragandhi a) and S Vedhanayaki

School of Electrical Engineering, Vellore Institute of Technology, Vellore, India.
a)
Corresponding author: [email protected]

Abstract. State-of-charge (SOC) estimation is essential for electric vehicles (EVs') optimal functioning, but it is a difficult
problem due to the very dynamic operating environment. By far, a lot of studies have been done on techniques and modelling to
precisely estimate SOC for Lithium-Ion Batteries (LiBs) used in electric vehicles. In this paper, Feedforward Neural Network
(FNN)-based SOC estimation has been proposed. This network uses discharge current, terminal voltage, and temperature as input
and provides SOC as an output. The proposed FNN estimator is applied to Lithium-ion batteries (LiBs) to check its validity in SOC
estimation. The proposed model is simulated in MATLAB software and the results show that the estimation of SOC by FNN is
better compared to conventional methods with reduced MAE.

INTRODUCTION
Nowadays, innovation in the field of renewable energy sources and electric vehicles has become a major concern,
to meet the increase in crude oil prices and achieve net zero carbon emission [1]. Research in the aforesaid area
includes research in battery technologies. Lithium-ion batteries are highly preferred in electric vehicles due to their
major advantages compared to other batteries, which include high specific density, high power density, no memory
effect, and small size [2]. Different types of batteries used by manufacturers of modern electric vehicle and their
characteristics are shown in table 1. Like other batteries, lithium-ion batteries also suffer from aging effects and
temperature effects. To ensure the battery , and to increase its lifetime, estimation of the
State of Charge (SOC) is important which defines the amount of energy available in the battery [3]. The nonlinear
behavior of factors such as temperature, charging/discharging, varying current profiles, discharge rate, and different
driving modes influences the accurate estimation of SOC. Researchers have proposed numerous methods for SOC
estimation to achieve reduced computation and ease of implementation. The classical methods are classified into four
types namely, conventional methods [4], model-based methods [5], observer-based methods [6], and filter-based
methods [7].
TABLE 1: Characteristics of batteries used in modern Electric Car

Energy Thermal
Total battery Weight Battery
Model density cells chemistry manageme
capacity (KWh) (Kg) manufacturer
Wh/Kg nt system
Passive
Volkswagen 264
358 349 103 NCM 333 air Samsung DI
E-golf 88 S &3 P
cooling
Passive
Volkswagen 168
368 248 148 NCM 622 air LG Chem
E-up 84 S & 2 P
cooling
Passive
Seat Mii 168
368 248 148 NCM 622 air LG Chem
electric 84 S & 2 P
cooling
Passive
Skoda Citigo 168
368 248 148 NCM 622 air LG Chem
EIV 84 S & 2 P
cooling
 Active
Renault
22 165 133 - NCM 712 liquid LG Chem
Twingo
cooling
Renault ZOE Active
192
(old 441 305 145 NCM 622 air LG Chem
96 S & 2 P
generation) cooling
Renault ZOE Active
192
(new 546 326 168 NCM 712 air LG Chem
96 S & 2 P
generation) cooling
Active
96
BMW I2 422 278 152 NCM 622 liquid Samsung DI
96 S & 1 P
cooling
Active
192
Peugeot e-208 50 326 164 NCM 712 air LG Chem
96 S & 2 P
cooling
Passive
288 Envision
Nissan leaf 62 410 151 NCM 523 air
96 S & 3 P AESC
cooling
Active
288
Chevrolet 68 430 158 NCM 712 liquid LG Chem
96 S & 3 P
cooling
Active
Hyundai Kona 294
67.5 452 149 NCM 622 liquid LG Chem
Electric 98 S & 3 P
cooling
Active
Hyundai 176
404 359 1124 NCM 622 liquid LG Chem
IQNIQ 88 S & 2 P
cooling
Active
432
Jaguar Ispace 90 603 149 NCM 622 liquid LG Chem
108S & 4P
cooling
Active
Mercedes 384
85 652 130 NCM 622 liquid LG Chem
Benz EQC 96 S & 4 P
cooling
Active
432
Audi E-Tron 95 700 136 NCM 622 liquid LG Chem
108S &4 P
cooling
Active
4416
Tesla Model 3 80.5 478 168 NCA liquid Panasonic
96S & 46P
cooling

Conventional methods include internal resistance, coulomb counting, look-up table, and impedance spectroscopy.
The coulomb counting method is highly preferred in electric vehicle applications but it requires accurate initial SOC
for precise SOC estimation. The internal resistance method, look-up table method, and impedance spectroscopy
methods are suitable for SOC estimation in a laboratory environment and hence not suitable for EV application.
Model-based methods include electric circuit models, empirical models, and electrochemical models. Model-based
methods have complexity in the identification of parameters and hence they are usually combined with filter-based
methods and observer-based methods to increase the accuracy of SOC estimation. Non-linear observer-based methods
such as sliding mode observer, non-linear observer, proportional integral observer, and Luenberger observer are
proposed to linearize the nonlinear behavior of the battery and estimate its SOC. Though observer-based methods
reduce the difference between the estimated state and measured state using a feedback system, it suffers complexity
in selecting a suitable gain matrix. Adaptive filter-based methods include particle filter, H infinity filter, and Kalman
filter. Filter-based method s accuracy is high when combined with ECM models, but extreme nonlinearity affects the
estimation accuracy [8]. To overcome the drawback of conventional methods data driven methods have been proposed
that estimates SOC based on fuzzy, artificial neural networks, ANFIS, SVM, PSO, and GA.
 In this paper, feedforward neural network-based SOC estimation is proposed to estimate the SOC of the battery.
This method increases the accuracy of estimation and reduces the computation time. The second section will explain
the state of charge of the battery, and the third section will explain the methodology of the proposed work. Simulation
and results will be discussed in the fourth section. The conclusion and future scope will be discussed in the fifth
section.

STATE OF CHARGE
Accurate estimation of SOC plays a significant role in battery technology. SOC is defined as the quantification
of the remaining energy available in the battery. Initially, SOC is called the ratio of the average concentration of
lithium-ion in the cathode to its maximum concentration. Theoretically, SOC value will range from 0 to1 but
practically attachment of excess lithium ion to the cathode leads to degradation of the battery. Hence SOCmin will be
greater than 0 and SOCmax will be less than 1. The United States Advanced Battery Consortium (USABC) has defined
SOC as the ratio of the remaining capacity of a battery under a specific condition to its rated capacity [9].

The rated capacity of the battery is affected by the manufacturing process, improper conduction of the experiment,
inaccurate calibration of test equipment, and environmental conditions. To overcome these issues, researchers
proposed the maximum capacity of the battery including kinetic effects as the replacement for the rated capacity [10].
Finally, SOC is called the ratio of the remaining capacity of the battery to its maximum capacity.

SOC estimation is the major function of the Battery Management System, since it displays the remaining energy
left in the battery, helps in cell balancing, provides prior knowledge on control required for the charging and
discharging process, and extends the battery life.

METHODOLOGY
Feedforward Neural Network (FNN) is one type of Artificial Neural Network (ANN). The traveling of
information is unidirectional in FNN since the cyclic connection of nodes is absent. It consists of a single-layer
perceptron. In the perceptron layer, series input enters the layer, and weight gets multiplied and added together to
produce weighted input. The flow chart of the proposed technology is shown in figure 1 and the block diagram of the
proposed model is given in figure 2.

FIGURE 1: Flowchart of the proposed system

 FIGURE 2: Block diagram of the proposed system

The collection of data is the premier step in ANN. Training the network with numerous data helps to
increase the estimation accuracy. In this process, accurate data that helps to estimate SOC is selected
from defective one. In the proposed work, the LG-HG2 dataset prepared by McMaster University in
2020 was used for training and testing.
To reduce the numerical differences existing between the various inputs of the battery pack, they can be
scaled between 0 and 1 by the normalization of pre-processed data. In the proposed work Min-Max
normalization is used to normalize the data.

The architecture of the proposed FNN network consists of three inputs namely discharge current,
terminal voltage, and temperature as shown in Figure 2. The hidden layer consist of ten neurons and the
output layer consists of one neuron which depicts the estimated SOC. The number of neurons for the
hidden layer is produced based on a trial-and-error basis. The number of neurons in the hidden layer is
given by equation 2.

Where,
= Total number of neurons in the hidden layer
= Input to the hidden layer (discharge current, terminal voltage, and temperature)
= Weight between neurons of the input layer (i) and hidden layer (j)
= Bias applied to the hidden layer

FIGURE 3: Architecture of proposed FNN

 The activation function otherwise called a transfer function is used to obtain the output of a node. The activation
function is classified into two types namely linear and non-linear activation functions. The non-linear activation
function is further classified as the sigmoid function, hyperbolic tangent Activation Function, and Rectified Linear
Unit activation function. In our proposed work, the hyperbolic tangent Activation Function is used for the hidden layer
and the sigmoidal function is used for the output layer. The equation for the hyperbolic tangent Activation Function
and sigmoidal function is given in equations 3 & 4.

(3)

(4)

Input to the output neuron is given as

(5)
Where,
= the output neuron from the hidden layer to the output layer
= Weight exists between the hidden layer and the output layer
= Bias to the output layer.

SIMULATION AND RESULT

The framed FNN is trained with a learning rate of 0.1 with a maximum of 200 iterations. Approximately
47500 data is fed to the network where 70% of the data is used for training and validation and 30% is used for
testing. The framed model is trained numerous times for provided input and output. Once the training reaches
2000 epochs, it will stop. The MATLAB simulation diagram of the proposed network is shown in Figure 4. The
model is trained with a 0.1 learning rate. The parameters of the battery considered for simulation are given in
table 2.

FIGURE 4: Simulink model of proposed NN-based SOC estimation

 TABLE 2: Parameter of lithium-ion battery

S.No. Parameters Rating

1. Battery type Lithium-ion
2. Nominal Voltage (V) 4.2 V
3. Rated capacity (Ah) 15Ah
4. Cut off voltage (V) 2.7 V
5. Fully charged voltage (V) 4.19 V
6. Battery response time 1 second
7. Nominal discharge current 1.7A
8. Internal resistance 0.009 ohm.
9. Capacity (Ah) at nominal voltage 3.6 V

The model is said to be trained when it reaches maximum gradient and hence that can be used for
future online SOC estimation. The SOC estimation by the proposed network for both charging and
discharging modes is shown in Figure 5 and their corresponding battery current, voltage, and temperature
are given in Figure 6-8. The trained network is tested for different current profiles. SOC estimation using
FNN has a better performance compared to existing methods as shown in table 3.

TABLE 3: Comparison of proposed method with existing method

S.no Estimation method Average error (%) Reference

1 Genetic Algorithm < 2% 11
2 Fuzzy Logic < 5% 12
3 Support vector Machine < 6% 13
4 Artificial Neural Network < 4.6% 14
5 Feedforward Neural Network < 1.5% Proposed
work

FIGURE 5: SOC estimation using FNN

 FIGURE 6: Current waveform of training data fed to FNN

FIGURE 7: Voltage waveform of training data fed to FNN

FIGURE 8: Temperature of training data fed to FNN

CONCLUSION
In this paper, SOC estimation based on FNN was proposed to estimate the SOC of Lithium-ion batteries used
in Electric vehicles. The offline trained model can be used to estimate the real-time SOC of the batteries. The
computational time of the proposed network is low compared to conventional methods. The trained models are
validated with new data sets and it is observed that the proposed method has a low MAE of <1.5%. Since FNN is
a good approach for the estimation of SOC, in future work it can be used to estimate other aspects of battery
including SOH and SOP.
 REFERENCES
1. Rietmann, Nele, Beatrice Hügler, and Theo Lieven. "Forecasting the trajectory of electric vehicle sales and
the consequences for worldwide CO2 emissions." Journal of Cleaner Production 261 (2020): 121038.
2. Zhang, Dong-Hua, Guo-Rong Zhu, Shao-Jia He, Shi Qiu, Yan Ma, Qin-Mu Wu, and Wei Chen. "Balancing
control strategy for li-ion batteries string based on dynamic balanced point." Energies 8, no. 3 (2015): 1830-
1847.
3. Chen, J.; Ouyang, Q.; Xu, C.; Su, H. Neural network-based state of charge observer design for lithium-ion
batteries. IEEE Trans. Control Syst. Technol. 2017, PP, 1 9.
4. Meng, Jinhao, Guangzhao Luo, Mattia Ricco, Maciej Swierczynski, Daniel-Ioan Stroe, and Remus
Teodorescu. "Overview of lithium-ion battery modeling methods for state-of-charge estimation in electrical
vehicles." Applied sciences 8, no. 5 (2018): 659.
5. Laadjal, Khaled, and Antonio J. Marques Cardoso. "Estimation of lithium-ion batteries state-condition in
electric vehicle applications: issues and state of the art." Electronics 10, no. 13 (2021): 1588.
6. Espedal, Ingvild B., Asanthi Jinasena, Odne S. Burheim, and Jacob J. Lamb. "Current trends for state-of-
charge (SoC) estimation in lithium-ion battery electric vehicles." Energies 14, no. 11 (2021): 3284.
7. Singirikonda, Srinivas, and Y. P. Obulesu. "Advanced SOC and SOH estimation methods for EV batteries
A review." Advances in Automation, Signal Processing, Instrumentation, and Control (2021): 1963-1977.
8. Chen, Zewang, Liwen Yang, Xiaobing Zhao, Youren Wang, and Zhijia He. "Online state of charge estimation
of Li-ion battery based on an improved unscented Kalman filter approach." Applied Mathematical
Modelling 70 (2019): 532-544.
9. USABC Electric Vehicle Battery Test Procedures Manual. Appendix F: Glossary of battery and battery
testing terminology for the USABC battery test procedures, USABC, 2009. p. 1 10.
10. Truchot, Cyril, Matthieu Dubarry, and Bor Yann Liaw. "State-of-charge estimation and uncertainty for
lithium-ion battery strings." Applied Energy 119 (2014): 218-227.
11. Zheng, Yuejiu, Languang Lu, Xuebing Han, Jianqiu Li, and Minggao Ouyang. "LiFePO4 battery pack
capacity estimation for electric vehicles based on charging cell voltage curve transformation." Journal of
power sources 226 (2013): 33-41.
12. Salkind, Alvin J., Craig Fennie, Pritpal Singh, Terrill Atwater, and David E. Reisner. "Determination of state-
of-charge and state-of-health of batteries by fuzzy logic methodology." Journal of Power sources 80, no. 1-
2 (1999): 293-300.
13. Lei, X., Q. Q. Chen, K. P. Liu, and L. Ma. "Support vector machine based SOC estimation for electric
vehicles." PRO. CSEE 28, no. 18 (2008): 114-118.
14. Affanni, A., A. Bellini, C. Concari, G. Franceschini, E. Lorenzani, and C. Tassoni. "EV battery state of
charge: Neural network based estimation." In IEEE International Electric Machines and Drives Conference,
2003. IEMDC'03. vol. 2, pp. 684-688. IEEE, 2003.