Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)

IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

Machine Learning Algorithms for Estimation of

State-of-Charge of Li-ion Batteries
2023 7th International Conference on Electronics, Communication and Aerospace Technology (ICECA) | 979-8-3503-4060-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICECA58529.2023.10395268

B. Pushpavanam S.Kalyani P. Naveen

Department of Electrical and Electronics L & T Edu Tech, Department of EEE, KPR Institute of
Engineering, Larsen & Toubro Limited, Engineering and Technology, Coimbatore.
PSNA College of Engineering and Chennai , India naveenamp88@gmail .com
Technology, Dindigul , India. [email protected]
[email protected]

T. Akilan C.Swedheetha P.Manikandan,

Department of IQAC, Amity University, Department of CSE, Velammal College of Department of ECE, Kalasalingam Academy of
Kolkata. Engineering and Technology, Madurai. Research and Education, Krishnankoil.
[email protected] [email protected] [email protected]

Abstract—Accurately predicting the SoC of the battery is the conditions, maximizing the pack capacity by balancing and
main function of any Battery Management System (BMS). BMS redistributing, and thermal management by maintaining
must carefully read the battery parameters, to safe guard the temperature in safe range. BMS monitors the battery
battery, to predict accurately the battery’s state, so as to continuously in a periodic manner and aims in predicting the
improve the battery’s performance and to alert to users/external SoC and SoH of the battery. BMS masters the energy system
devices. The objective of this study is to offer a comprehensive in an Electric car. Through the acquisition of data, the Battery
overview of Machine Learning (ML) algorithms for predicting Management System (BMS) diligently observes the charging
the State-of-Charge (SoC) of Lithium-ion batteries for enabling and discharging patterns of batteries, with a specific focus on
an accurate and online estimation, mandatory for ensuring
Lithium-ion batteries. This is crucial as the charging and
battery as well as user safety and also aims to provide efficient
BMS for the advancing the development of EVs (Electric
discharging behaviors in Lithium-ion batteries exhibit distinct
Vehicles) and HEVs (Hybrid Electric Vehicles). characteristics. The sample battery management system is
shown in figure 1.
Keywords— Battery Management System (BMS), Machine
Learning (ML), State-of-Charge (SoC), Electric Vehicles (EVs)
and Hybrid Electric Vehicles (HEVs).

I. INTRODUCTION
Renewable energy sources are fastest growing electricity
sources. Energy storage has become one of the prime sectors
and it is critical in an Electric vehicle. Lithium batteries are
employed in various applications such as electric vehicles.
These are portable devices, and are predominant energy
storage systems (ESS) as the reasons are, its high energy
density, long life, and less weight[1]. Machine Learning
techniques that are based on data, are become increasingly
widespread for predicting the State-of-Charge(SoC) and Figure 1: BMS-Courtesy: ST Microelectronics Inc.
State-of-Health(SoH) as there is a state-of-art computing
techniques and availability of open source battery data.[2] B. SoC and its Importance:
A. Why BMS? In a battery or cell, State of Charge at a given instant is
defined as the ratio between charge available at that instant to
The essential parameters for battery monitoring the charge available when it is fully charged. It is expressed in
encompass Voltage, Current, and Temperature. The foremost percent, from 0% when empty to 100% when full. Analogous
aim of the BMS is to monitor the parameters and to keep them to how much fuel is available in car’s tank. SoC of the battery
in the safe operating regions. The change in these parameters is expressed by
occurs during charging and discharging or during open circuit.
BMS plays a critical role in overseeing each individual cell SoC= Qavailable/Qrating
within a battery pack, ensuring the overall stability of the where Qavailable is the leftover battery charge, and Qrating is
system. Its primary responsibility is to uphold the safety of the the maximum capacity of the battery. Though, the rated
battery, safeguarding both users and the environment from capacity of the battery Qrating is not constant throughout its life-
potential harm. BMS manages the battery pack in three ways, time. The maximum or the rated capacity of the battery Qrating
protection of the battery without causing any damages by not varies throughout its lifetime relying upon various external
allowing parameters outside the standard safe operating aspects like temperature of environment, current discharge,

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 1

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)
IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

aging factor of battery, SoH, the no. of discharge & charge B. Long short Term Memory (LSTM):[5]
cycles, and other reasons such as manufacturing defects
(agglomeration, small circular pinholes, irregular coatings, In their scholarly investigation, Fangfang Yang et al.
and contaminated metal) [3]. introduced a stacked LSTM network as a methodology for
deriving State-of-Charge estimations from measurements of
C. Importance of SoC: battery parameters. This approach involves modeling the
1) SoC is an indicator of sufficient charge availability for batteries considering its intricate charging/discharging
carrying out a particular operation. dynamics.
2) SoC estimation helps in operation of battery in its limits
of charge and discharge.
3) Lower or higher values of SoC tends to damage to
battery permanently. It is difficult to estimate SoC as it
represents internal state of the battery.
II. ML ALGORITHMS OF SOC ESTIMATION: A REVIEW
A. Extreme Learning Machine (ELM) algorithm:[4]
According to the statements of the researchers prime
factors that decide the efficiency of the ELM algorithm are the
training accurateness and the neuron count in the hidden layer.
Gravitational search algorithm(GSA) aids the ELM in
increasing the computational efficiency. The researchers also
confirm that there is no need of any mathematical model and
battery knowledge in the estimation of SoC using GSA. At
different temperatures model robustness is validated, and it is
verified using back-propagation and radial basis function
neural networks.

Fig. 3. Results of ELM models under different Test cycle

data:(a) (b) DST(Dynamic Stress Test)cycle, (c) (d)
FUDS, (e) (f) US06
From well-known driving cycles like the aggressive
US06 , DST and FUDS, the researchers trained and tested the
data’s as shown in figure 3. The test results are compared with
the model-based filtering approach (unscented Kalman filter)
in consideration with the accuracy of tracking, time for
computation, robustness.

Figure 2: one-layer ELM model structure.

𝐼𝑘
They have performed ELM execution by arbitrarily assigning
the weights of the input and biases of the hidden layers. The
𝑉𝑘 𝑆𝑂𝐶𝑘
hidden layers neurons were fixed suitably. Fig.2 shows the
structure of a single layer configuration of ELM. This method
finally results in good accuracy and lesser error rate and also 𝑇𝑘
reducing cost for computation.
For example, in the Federal urban driving schedule
cycle (FUDS), using Back Propagation Neural Network with
the back tracking search algorithm (BSA) model prediction Input Layer Stacked LSTM Layer Fully Connected Layer Output Layer
results less than 2%(RMSE). In, Federal Urban Driving
Schedule cycle ELM-based GSA model results in SoC Figure 4: Architecture of proposed network
estimation with better accuracy, with RMSE being less than
1.5%. This model demonstrates superior performance when The proposed architecture of stacked LSTM is shown in
compared to Deep Learning, fuzzy methods, Kalman, H∞ the figure 4. It is inferred from their proposed network that it
filters, and Support Vector Machine (SVM) approaches. revealed the non-linear correlation of the State-of-Charge and
input parameters and resulted in better estimation than the
conventional filtering methods. Also the inexact initial SoCs
converges with 2% of RMSE(Root Mean Square Error) and

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 2

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)
IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

1% of MAE(Mean Average Errors) for the suggested network network are shown with starting SoC 60% with accurate and
which converges faster to attain the true value of SoC. stable results as shown in figure 7.
At each time step the SoC estimation time is reduced in The proposed model learned in a perfect way with normal
milliseconds, which makes it appropriate for real-time environmental conditions and gives satisfying estimation of
implementations and present Electric Vehicles which State-of-Charge under varying temperatures, with MAEs
requires a suitable Battery management system. The within 1.5%. and RMSEs within 2%.
researchers suggest that their proposed methodology is one of
the dominant methods for predicting the convergence of SoC
in lithium-ion batteries. This method shall be enhanced by
considering other situational factors like, degradation,
humidity and battery materials. Figures 5a and 5b shows the
results for FUDS dataset when the network is trained using
DST dataset with starting SoC 80% and 60% respectively.

Fig. 7. Sample output for CNN-LSTM model

(a)% SoC and (b) Estimation error
D. Deep Neural Network (DNN):[7]

Fig.5. Sample output for LSTM model Dickshon et al. introduced an advanced deep neural
a) Starting SoC 80% and b) Starting SoC 60% network (DNN) for predicting State-of-Charge (SoC) in Li-
ion batteries specifically designed for Electric Vehicle (EV)
C. Combined CNN-LSTM Network: [6] applications. Their research affirms that, even for unfamiliar
Xiangbao et al., in their study, suggested a predictive drive cycles, an accurate SoC estimation can be achieved by
approach for estimating battery State-of-Charge (SoC) using augmenting the count of hidden layers in the neural network.
Convolutional Neural Network (CNN) and type of recurrent The authors created multiple DNN models incorporating
neural network (RNN) called Long Short-Term Memory varying numbers of hidden layers. The output of the DNN is
(LSTM) network, leveraging the inherent advantages of each. represented by the predicted or estimated SoC, which is a
By utilizing parameters with established values such as scalar value. The architecture of proposed DNN architecture
current (I), voltage (V), and temperature (T), the model is shown in figure 8.
extracts time sequence and space attributes from the input
data, combining the strengths of both networks. The
recommended architecture is presented in Fig.6.

Fig. 6. Architecture of CNN-LSTM Network

Fig 8: Layers of proposed DNN architecture
Their experimental results show nonlinear
correlations between State-Of-Charge and input parameters, By using server with two NVIDIA 1080Ti graphical
such as V, I, T, Iav, and Vav. With unknown initial values the processing units (GPUs) the researchers trained the DNN
network converges very fast and then obtained an accurate models with deep learning algorithms such as TensorFlow
estimation of SoC with overall RMSEs and MAEs under 2% with keras deep learning algorithms to validate their proposed
and 1% .In addition estimation results of their proposed methodology. Sufficient count of hidden layers in Deep

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 3

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)
IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

Neural Network was capable of generalizing the tendency of optimizes parameters of the model, sliding window method
estimation of SoC for various drive cycles. SoC estimation updates online the number of input vectors
with true value for US06 data is shown in figure 9.
It was proved that by increasing hidden layers count
of the DNN decreases the error. In comparison with the other
machine learning techniques like LSTM and gated recurrent
unit (GRU) it is clearly visible that the computation cost and
complexity are very much reduced as suggested by the
researchers. The requirement of resources is more, and also
memory requirement, time to train and test are more.
Comparison of error rates indicate that the suggested method Fig.10 LSSVM Battery model
is equal with existing methods.
SoC error is controlled below 1%. The above two
proposed models enhance the prediction and MAE for model
1 by is 0.57%-1.74% and for model 2 by 0.38%-0.87%. The
comparison of the results with the LSSVM and the proposed
models are as shown in figure 11.They confirm that their
techniques are superior to RNN model and conventional
Kalman Filter model at various temperatures.

Fig. 9 SoC estimation Vs US06 dataset true value at

various temperatures. - DNN
E. Least squares support vector machine (LSSVM):[8]
The researchers in their publication employed Least-
squares support-vector machines, a suite of interconnected
methods in supervised learning designed for data analysis,
pattern recognition, and addressing classification and
regression challenges. Utilizing the Grey Wolf Optimizer
(GWO) for parameter determination in LSSVM, they
enhanced the model's capability to achieve global optimal
estimation. Jiabo et al. introduced an LSSVM based on a
sliding window methodology, contributing to improved Fig. 11. At 25ºC (a) Estimation of SoC (b) Error of SoC
accuracy in State-of-Charge (SoC) estimation. III. PERFORMANCE ANALYSIS
Figure10 depicts the LSSVM model sampled at time t. Table 1 shows the Performance analysis of estimation of
The prediction and analysis of performance for the GWO- SoC using various AI techniques. Various algorithms used
LSSVM (Grey wolf optimizer) and CV-LSSVM (Cross for SoC estimation, novelty used in various techniques, type
Validation method) are also analyzed for better of battery used, validation profile, and the error rates at
understanding. The inferences from their paper are, GWO various temperature are tabulated for an easier understanding.

Table 1: Performance analysis of estimation of SoC using various AI techniques

s. AI Objective Novelty Validation Temp. Error rate Test Battery

No. Technique Profile
Enhance
Root Mean
computational
DST Square Errors: 2Ah 18650
intelligence by New ELM based
1. ELM [4] FUDS 25ºC 1.1% (DST), LiNiMnCoO2
attaining the GSA model
US06 1.4% (FUDS), /NMC
optimal value of
1.8% (US06)
hidden neurons.

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 4

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)
IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

Votsch Root Mean

Enhance tracking Multiple training
DST, temperature Square Errors:
and estimation and testing A123 18650 LFP
2. LSTM [5] US06, chamber - 1.39% (DST),
performance for datasets are Battery (1.1Ah)
FUDS Ambient 1.07% (FUDS),
improved results. generated.
Temperature 1.29% (US06)
Votsch 1.39%, 0.92%
DST temperature RMSE and MAE
CNN- Enhanced tracking combined CNN- A123 18650 LFP
3. FUDS chamber - respectively at
LSTM [6] efficiency LSTM network Battery (1.1Ah)
US06 Room room
Temperature Temperature
RNN Combined- CNN Red
Provide improved Twenty
(Recurrent & DL based Star X1
4. SoC prediction different MAE 0.135 192 Ah LIB
Neural Net) model for SoC electric
performance values
CNN [9] prediction vehicles
To improve the
DST 0ºC, 25ºC MAE < 1% 2.4Ah NMC
5. LSSVM [8] estimation GWO
FUDS and 45ºC. RMSE< 0.3% Type
accuracy of SoC
Improve error
SWPSO (Self
DRNN convergence speed 2 Ah 18650-
adaptive weight Experimental 4ºC, 24ºC,
6. SWPSO and to improve MAE <2% typed lithium-ion
particle swarm dataset and 44ºC
[10] estimation battery
optimization)
accuracy
Learning rate RMSE: 0.85%
25ºC
DL – FCN Improve the error optimization (MAE: 0.7%),
7. US06 -20ºC 2.9 Ah 18650PF
[11] rate strategies RMSE: 2%
- 25ºC
(MAE: 1.55%)
Minimized
Non-linear
Akaike GDX-optimizer:
Auto
Information Various C- RMSE 0.1323%,
Regressive
Criterion rates MAE 0.0888%,
with external To improve overall Panasonic-Sanyo
Algorithm - (C/2,C/4, Ambient MSE 0.0175%
8. input predicting 18650: 3.7V,
time series model C/6, C/8 Temperature rprop-optimizer:
u-NARX - performance 3.5Ah Li-NCA
developed u- and C/10 RMSE 0.2046%,
min-AIC
MLP with u- rates) MAE 0.1666%,
built NARX
NARX NN MSE 0.0419%
model [12]
models
Prediction with
satisfactory level of
UDDS −20ºC
accuracy, ensuring 1D-CNN
1D CNN HWFET -25ºC Panasonic 18650
9. rapid learning and & Transfer MAE < 2%
[13] LA92 −20ºC PF,
strong learning
US06 -40ºC. LG 18650 HG2
generalization
capabilities.
GRU–OCP To develop an INR 18650-20R
US06 0.52%(RMSE)
(One Cycle inexpensive, Type:
Stacked two BJDST 0ºC, 25ºC on training,
10. learning precise, and LiNiMnCoO2/N
hidden layers FUDS and 45ºC. 0.65% on test
rate Policy) resilient model for MC Li-ion
DST dataset.
[14] estimating SoC. battery
NARX – RMSE of 0.34%
To obtain accurate NN and moving DST 2400mAh
RNN 0ºC, 25ºC MAE of 0.21%
11. battery model for window UDDS 18650 lithium-ion
Moving and 45ºC. Execution Time –
accurate estimation technique. FUDS NMC
Window[15] 1.48217 s
Increasing the
DNN US06
To minimize errors count of hidden
(Four BJDST 0ºC, 25ºC 3.68%(RMSE) 2.0 Ah 18650
12. and improve the layers in the Deep
hidden FUDS and 45ºC. 0.13%( MSE) NMC
precision. Neural Network
layers) [7] DST
(up to 4).
TCN [16] LA92
To estimate SoC Special structure
Temporal UDDS 0ºC, 10ºC MAE of 0.67% 2.9Ah Panasonic
13. correctly under TCN with self-
Convolution HWFET and 25ºC RMSE of 0.87% 18650PF
different conditions learning
Network US06

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 5

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Seventh International Conference on Electronics, Communication and Aerospace Technology (ICECA 2023)
IEEE Xplore Part Number : CFP23J88-ART ; ISBN : 979-8-3503-4060-0

Applications, Vol. 56, No. 5, pp.52796 – 52814, September/October

IV. CONCLUSION 2020.
[8] Jiabo Li, Min Ye, Wei Meng1, Xinxin Xu, and Shengjie Jiao1 “A
This paper provides a comprehensive survey of various Novel State of Charge Approach of Lithium -Ion Battery Using Least
ML algorithms for predicting SoC of Li-ion batteries. This Squares Support Vector Machine” IEEE Access, October 23, 2020, pp.
review paper has presented in brief various algorithms used 195398 – 195410.
for SoC estimation, novelty used in various techniques, type [9] Fen Zhao, Yinguo Li, Xinheng Wang, Ling Bai, and Tailin Liu,”
of battery used, validation profile, and the error rates at Lithium-Ion Batteries State of Charge Prediction of Electric Vehicles
Using RNNs-CNNs Neural Networks”, pp.98168 – 98180, May 21,
various temperature for quick understanding. The key 2020.
challenges in various techniques are also discussed. The
DOI: 10.1109/ACCESS.2020.2996225
performance analysis of various AI techniques is tabulated.
[10] Yanbo Che, Yushu Liu, Ze Cheng, and Ji’ang Zhang, “SOC and SOH
Finally, from this review it is inferred that all the methods Identification Method of Li-Ion Battery Based on SWPSO-DRNN”,
strive to achieve an estimation error below 2% with minimal IEEE Journal of Emerging and Selected Topics in Power Electronics,
complexity and economical implementation. Vol. 9, No. 4, August 2021, pp. 4050 – 4061.
[11] M. A. Hannan, D. N. T. How, M. S. Hossain Lipu, Pin Jern Ker, Z. Y.
REFERENCES Dong, M. Mansur, and Frede Blaabjerg, “SOC Estimation of Li-ion
Batteries with Learning Rate-Optimized Deep Fully Convolutional
[1] P. J. Hall and E. J. Bain, ``Energy-storage technologies and electricity
Network”, IEEE Transactions on Power Electronics, Vol. 36, No. 7,
generation,'' Energy Policy, vol. 36, no. 12, Dec. 2008,pp. 4352_4355.
pp. 7349 – 7353, July 2021.
[2] Carlos Vidal, Pawel Malysz, Phillip Kollmeyer, And Ali Emadi,”
[12] Asadullah Khalid, and Arif I. Sarwat, “Unified Univariate-Neural
Machine Learning Applied to Electrified Vehicle Battery State of
Network Models for Lithium-Ion Battery State-of-Charge Forecasting
Charge and State of Health Estimation: State-of-the-Art”, IEEE Access,
Using Minimized Akaike Information Criterion Algorithm”, pp.39154
pp.52796 – 52814, March 16, 2020.
– 39170, Vol 9, February 23, 2021.
[3] D. Mohanty, E. Hockaday, J. Li, D. K. Hensley, C. Daniel, and D.
DOI: 10.1109/ACCESS.2021.3061478
L.Wood III, “Effect of electrode manufacturing defects on
electrochemical performance of lithium-ion batteries: Cognizance of [13] Arnab Bhattacharjee, Ashu Verma, Sukumar Mishra and Tapan K.
the battery failure sources,” J. Power Sources, vol. 312, 2016,pp. 70– Saha, “Estimating State of Charge for xEV Batteries Using 1D
79. Convolutional Neural Networks and Transfer Learning”, IEEE
Transactions on Vehicular Technology, Vol. 70, No. 4, pp. 3123 –
[4] Molla S. Hossain Lipu,Mahammad A. Hannan,Aini Hussain,Mohamad
3134, April 2021.
H. Saad, Afida Ayob, and Mohammad Nasir Uddin,” Extreme
Learning Machine Model for State-of-Charge Estimation of [14] Mahammad A. Hannan, Dickson N. T. How, Muhamad Bin Mansor,
Lithium-Ion Battery Using Gravitational Search Algorithm”, IEEE Md S. Hossain Lipu, Pin Jern Ker, and Kashem M. Muttaqi, “State-of-
Transactions on Industry Applications, Vol. 55, No. 4, July/August Charge Estimation of Li-ion Battery Using Gated Recurrent Unit with
2019, pp. 4225 – 4234. One-Cycle Learning Rate Policy, IEEE Transactions on Industry
Applications, Vol. 57, No. 3, pp.2964 -2971, May/June 2021.
[5] Fangfang Yang , Xiangbao Song, Fan Xu , and Kwok-Leung Tsui,”
State-of-Charge Estimation of Lithium-Ion, Batteries via Long Short- DOI: 10.1109/ACCESS.2021.3081619
Term Memory Network”, IEEE Access ,April 2019, Vol. 7, pp.53793 [15] Qiao Wang, Hairong Gu, Min Ye, Meng Wei, And Xinxin Xu,” State
– 53799. of Charge Estimation for Lithium-Ion Battery Based on NARX
[6] Xiangbao Song, Fangfang Yang, Dong Wang, and Kwok-Leung Tsui,” Recurrent Neural Network and Moving Window Method”, IEEE
Combined CNN-LSTM Network for State-of-Charge Estimation of Access, Vol. 9, pp.83364 -83375, June 4, 2021.
Lithium-Ion Batteries”, IEEE Access, July 3, 2019, pp.88894 – 88902. [16] Yuefeng Liu, Jiaqi Li, Gong Zhang1, Bin Hua, and Neal Xiong,” State
[7] Dickshon N. T. How, Mahammad A. Hannan, Molla S. Hossain Lipu, of Charge Estimation of Lithium-Ion Batteries Based on Temporal
Khairul S. M. Sahari, Pin Jern Ker, and Kashem M. Muttaqi,” State-of- Convolutional Network and Transfer Learning”, IEEE Access, Vol. 9,
Charge Estimation of Li-Ion Battery in Electric Vehicles: A Deep February 5, 2021, pp.34177 – 34187.
Neural Network Approach” IEEE Transactions on Industry

979-8-3503-4060-0/23/$31.00 ©2023 IEEE 6

Authorized licensed use limited to: AMITY University. Downloaded on September 10,2024 at 10:52:51 UTC from IEEE Xplore. Restrictions apply.