Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)

IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

Artificial Intelligence based State of Charge

estimation of Li-ion battery for EV applications
Tejaswini P 1 Sivraj P 2
1 2
Department of Electronics and Communication Department of Electrical and Electronics
Engineering, Engineering,
Amrita School of Engineering, Coimbatore, Amrita Amrita School of Engineering, Coimbatore, Amrita
Vishwa Vidyapeetham, India Vishwa Vidyapeetham, India
tejaswinir9@gmail.co m p_sivraj@cb.amrita.edu

Abstract—Battery technologies and advanced battery important role in control and monitoring the battery operations
management systems are amongst the most trending research in in an EV [4]. BMS is one single functional module solely
automotive sectors as a result of the unprecedented push for responsible for cell balancing, monitoring cell level
electric vehicles. This paper, among existing methods for S tate of temperature, estimating state of health and SoC of a battery
Charge (S oC) estimation of Lithium-Ion batteries used in Electric
Vehicles (EV), explores various artificial intelligence-based and pack consisting of battery module with a specific combination
direct measurement techniques. A performance comparison of of cells (e.g. Nissan leaf has 192 cells, Tesla Model S has
S oC estimation using the coulomb counting approach, S upport 7104 cells and so on). SoC is a crucial parameter as it aids EV
Vector Machine (S VM) methods, and an optimal feed-forward driver to estimate the equivalent distance that can be covered
artificial neural network (ANN) for different storage based on the remaining capacity. Accurate SoC estimation
temperature, initial conditions, and stress tests have been helps to avoid battery failure in scenarios like, middle of busy
presented for a Lithium-Ion battery for a variety of standard traffic day or during a travel that is scheduled with limited
data sets. The stated models are trained using to predict S oC access to charging stations.
when voltage and current are given as inputs. Both the models
are tuned and trained in a cloud-based open-source jupyter
environment, collaboration. The results obtained post- SoC estimation is a challenging task; it’s a non-linear
performance analysis depicts the potential of ANN for accurate function of temperature and current. Xiong et al. has divided
S oC estimation of battery used in EV. ANN has achieved a Mean SoC estimation methods into four sub categories viz., d irect
Absolute Error (MAE) in a range of 0.5-1.4% over one complete measurement, model-based, book-keeping estimation and
cycle. This work can be further extended to validate the real-time computer intelligence based methods [3]. Data driven
performance of ANN with data collected from a hardware setup. techniques for SoC estimation have comparatively higher
prediction accuracy when compared to traditional methods [4]
Keywords—Artificial Neural Network, Multi-Layer Perceptron, as they are based on data science and machine learning
State of Charge, Support Vector Machine.
algorithms. With advancements in Artificial Intelligence (AI),
machine learning is being used extensively in different
I. INT RODUCT ION domains like infotainment, driver assistance systems,
The current paradigm shift of the automotive industry autonomous vehicles etc. AI is ruling industries due to its
from Internal Combustion (IC) engines towards electric extraordinary learning capabilities [5]. Researchers are aiming
mobility demands reliable electric vehicle in terms of mileage to make use of AI in battery management systems for accurate
alike IC engines [1]. This can be achieved by using battery estimation of SoC. This encourages determining an efficient
with high power density and energy density. However, prediction strategy that is best suited, accurate, adaptive and is
choosing a battery pack of highest capacity would increase the capable of estimating SoC precisely for an EV application [6].
weight of Electric Vehicle (EV) which in turn affects the
overall performance. Hence there is a trade-off between AI techniques for regression are categorized as, linear
battery capacity and EV mileage [2]. Lithium-ion battery pack regression, decision trees, Support Vector Machine (SVM)
manages to meet the needs of EV and Hybrid Electric and neural networks [7]. The former two techniques are
Vehicles (HEV), although from functionality point of view, its efficient for linear data. Battery exhibits non-linear
durability, safety, and lifespan are still at stake. State of characteristics; hence the work focuses on using Artificial
Charge (SoC), temperature and number of charge/discharge Neural Networks (ANN) and SVM models for accurate SoC
cycles are major parameters to estimate lifetime of lithium ion estimation. The work objectives have been divided into 3
battery [2]. subparts. First task is to study the existing SoC estimation
techniques and calculate SoC using traditional coulomb
Battery management system (BMS) plays an counting approach. Second task is, pre-process battery data

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1356

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

considering the battery state based on current, charging, on direct measurement but, since it’s an open loop estimation
discharging, initial capacity and train both models with method, is susceptible to error and integrates with time [6].
voltage and current as inputs to predict SoC. Final task is a This reason inspires researchers to look for better approaches
performance study i.e. comparison of two vigorous AI for accurate SoC estimation.
techniques for accurate SoC prediction. Comparison is
brought out based on the Mean Absolute Error (MAE), Root Model based techniques are designed to adapt
Mean Square Error (RMSE) and R squared. automatically based on system dynamics. These methods are
complex and are uncertain. In model based methods, battery
(1) model is designed and expressed in terms of a state equation.
A nonlinear state equation along with an adaptive filter is used
to estimate battery SoC. Most commonly used algorithms are,
Where x is the experimentally measured value for n number of Kalman filter and its derivatives [9] [10] [11] [12], Luenberger
observations. observer [13], PI observer [14], sliding mode observer. Table
I, shows the approaches being followed by competitive battery
The next section of the paper gives an insight of pros suppliers [15].
and cons of traditional techniques currently used for SoC
estimation. Section 3 describes the system overview, TABLE I. ALGORITHMS P ROP OSED BY BATTERY SUP P LIERS [20]
methodology and data wrangling process. Section 4 elaborates
on implementation steps achievable to build the model and the
procedure carried out to train ANN and SVM models. Section Battery S uppliers Algorithm Proposed
5 describes the results and discussion based on performance
analysis of both the AI based models. In section 5 ANN is General M otors Coulomb Counting and OCV based method
presented with concluding remarks and feasible future scope. called startup SoC and running SoC

II. BACKGROUND ST UDY

Tesla Weighing function to update current
SoC is a measure of usable energy in a battery. capacity
Accurate SoC estimation has been an underlying challenge in
all the battery based applications [4]. Battery an energy
storage system is prone to different electrochemical, thermal Ford Non Linear observer with first order
and mechanical conditions. SoC of battery is a very subtle equivalent circuit
descriptor which is not easily measurable like other
descriptors such as voltage and current. In quintessence, SoC LG Chem Dual Kalman filter to determine time
should be estimated when the system is under thermodynamic varying parameters of Electrochemical cell
equilibrium for precise results [8]. In most of the battery model
operated devices, SoC is predominantly estimated using
empirical methods [1], whereas in case of an EV, battery is The data-driven techniques used for SoC estimation
charged at different charging stations having charging modes include artificial neural networks [13] [14], fuzzy logic based
like slow, moderate, fast charging and in contrast while dis - [16], support vector machine [17]. Researchers are attempting
charging it is subjected to different road tractions because of to extend AI for real-time control and monitoring of BMS
change in terrain. In addition, accurate SoC estimation aids in functionalities.
extending the usable range which indicates that EV can be
driven longer [5]. Most popular techniques for regression are lasso net,
ridge net, elastic net, random forest, gradient boost methods.
One amongst the most widely used empirical The 5 training techniques commonly used for ANN are –
methods for SoC estimation is – Coulomb counting or Ampere gradient descent, Newton’s, conjugate gradient, quasi Newton,
hour method or Current Integration method. This method is Levenberg Marquardt methods. Similarly, SVM makes use of
used to estimate SoC using charging/discharging current kernel functions viz., linear and radial based kernels to
integrated over time. The mathematical representation for SoC transform non-linear data to linear data of higher dimension.
at a given instant (t+1) is as mentioned below [6]:
Charkhgard et al. used a hybrid topology, viz, ANN
with an adaptive extended Kalman filter and obtained an
(2) RMSE of 2% [18]. Meng et al, fused unscented KF with SVM
to predict the SOC with the MAE less than 2% [19]. Chemli et
Where, SoC(t) denotes the initial SoC, I represent al. used a deep feedforward neural network to map battery
charge/discharge current in Amperes (A), and QRated represents drive cycle measurements to SoC and achieved MAE of
nominal capacity of the battery measured in Ampere Hours 1.10% for 25°C dataset and 2.17% at −20°C[19]. However
(Ah). Commercialized battery based applications use different these methods are suspected to possess unknown noise
variants of Coulomb counting algorithm as the method focuses parameters as battery is a time dependent entity.

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1357

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

III. SYST EM OVERVIEW , M ET HODOLOGY AND DAT A C. Data Wrangling

W RANGLING Data wrangling, also called as data munging, is a data
pre-processing strategy which transforms and maps data from
A. System Overview its raw format to a desired format for faster analysis.
The data flow is as shown in the Fig2. Regression
methods used for SoC estimation and performance comparison Battery datasets are made available by chemical
are explained below. researchers in public domain consisting of variety of battery
types, tested across different current profiles at different
temperatures based on the application. The dataset referred
consists of samples tested for standard current profiles viz.,
Dynamic Stress Test (DST), Federal Urban Driving Schedule
(FUDS), and US06 Highway Driving Schedule. These test
Fig. 2.System Overview conditions were performed at 0°, 25° and 45° Celsius and
charging and discharging is obtained by using constant
B. Methodology current/constant voltage method [20].
Following Fig 3.shows the flow followed to obtain a best From the data samples it is observed that battery
suited AI algorithm for SoC estimation.
undergoes different stress environments during its life-cycle.
Sudden spikes in current, voltage dips, voltage going above
nominal voltage, battery capacity exceeding the rated capacity
at different temperatures due to the fact that the EV traverses
different terrains like uphill, downhill, sudden jerks, braking
etc. which are major challenges for accurate SoC estimation.

The dataset consists of time series data of around

70,000 to 100,000 data points logged at an average step time
of 10 seconds for four C – ratings, 0.25C, 0.5C, 1.25C, and
2C, for a voltage range of 2.18V to 4.3V, and varying initial
storage conditions.

Feeding coarse data to an accurate model might still

not emphasize the input-output dependency. The data must be
closely studied, cleansed and unified by means of Exploratory
Data Analysis (EDA) using summary based statistics and
visualizations to relate the quality and tendency of the data
points before they are fed to the neural network.

Dataset consists of duplicates, imbalanced, corrupt

samples that can lead to inconsistency during training and
prediction, so such samples are excluded. Data is normalized
prior to training, to fit within a range of 0-1. This process is
carried out using numpy, matplotlib, seaborn and pandas
libraries. Current, Voltage and SoC for dataset consisting of
Fig. 3.Data flow to obtain best model for accurate SoC estimation 9,308 samples is visualized as shown in Fig 4.
AI techniques require large datasets for learning,
mapping and predicting. Only appropriate and relevant data
produce accurate predictions. The dataset chosen doesn’t
consist of SoC parameter. It is calculated mathematically
based on battery state, current and time interval. Once the
dataset is analyzed, inputs are fed to the AI models for
training, testing and validation purpose. Performance of both
models for versatile conditions is evaluated based on metrics
and best suited model is chosen for prediction.

Fig. 4.Current, Voltage and SoC surface plot using matplotlib

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1358

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

IV. IMPLEMENT AT ION Network parameters are configured based on trial and
error method for least losses based on loss curve as shown in
4.1 ANN and SVM models Fig. 6 for ANN are –
Regression techniques like lasso, ridge, elastic net,  Activation function – logistic
random forest and gradient boost methods are used for  Output Layer Activation function – ‘Identity’
comparison. Test dataset of around 1,000 samples at 25 degree  Hidden Layer size – 100 neurons in 1 hidden
Celsius, are preprocessed to estimate SoC using Coulomb layer
counting technique. All above mentioned algorithms were  Solver – Adam
trained using test dataset to estimate best learning algorithm to  Learning Rate – 0.1
be used as part of neural networks .  Maximum Iterations – 200
Hyperparameters for these techniques are as follows –
 Lasso - Alpha = [0.0001, 0.001, 0.01, 0.1, 1, 5, 10] During training, dataset is split as, 70% of data-points
 Ridge - Alpha = [0.0001, 0.001, 0.01, 0.1, 1, 5, 10] for training, 15% for cross-validation and remaining 15% for
 Elastic net - Alpha = [0.0001, 0.001, 0.01, 0.1, 1, 5, testing to reduce sampling bias effects. Datasets of different
10] and L1 ratio = [0.1, 0.3, 0.5, 0.7, 0.9] current profiles for different temperature levels is
 Random Forest-n_estimators = [10,20] and preprocessed and fed to both the models to evalu ate their
max_features= ['auto', 'sqrt', 0.33] performance for SoC estimation.
 Gradient Boost=n_estimators = [10,20] learning rate
= [0.05, 0.1, 0.2] and max_depth = [1,3,5] Identically, SVM involves number of learning tasks
termed as kernels. Kernels provide a modular framework that
Results obtained are as mentioned in Table II. R2 can be adapted to different tasks by us ing various kernel
score provides the measure of best fitted regression line. functions [16]. SVM parameters are configured using trial and
Random forest is the winning model with an R2 score of error method as below–
98.5% and least MAE of 1.7%.  Kernels - Linear and Radial based function
 gamma = 0.125
TABLE II. COMP ARISON OF DIFFERENT REGRESSION TECHNIQUES  Regularization parameter = 100
Elastic Random Gradient
Lasso Ridge
Net Forest Boost 4.2 Implementation Platform
The proposed SoC estimation method can be
M AE(%) 4.7 4.8 4.5 1.7 1.8
extended as a functional module in BMS. The EV is equipp ed
with battery pack which feeds the input to ANN model. The
R2 score 0.036 0.039 0.161 0.985 0.854 output from the model is processed and control signal is fed to
the Vehicle Display Unit (VDU) to assist driver on the
SoC is calculated mathematically based on coulomb remaining SoC. The implementation overview is as shown in
counting method. Data is preprocessed as mentioned in section below figure. The Intel i5 system with 8GB of RAM was used
3.C. Voltage, current are fed as inputs to ANN and SVM as the computational platform for analyzing the trained model.
models to learn and predict SoC. ANN model is built based on
random forest learning technique and Levenberg-Marquardt
training technique. MLP is a feed forward neural network is as
shown in Fig5 below.

Fig. 6. Implementation Overview

V. RESULT S AND DISCUSSION

The proposed techniques are quantified with
performance metrics viz., MAE and RMSE, helps to estimate
the average model prediction error to analyze both the fitted
models. Performance of ANN model, SVM models for DST,
FUDS and US06 at different temperatures is as shown in
Fig. 5.M LP architecture with inputs as Voltage, Current to predict Table III.
SoC

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1359

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

TABLE III. P ERFORMANCE METRICS FOR ANN MODEL , SVM MODELS AT Cost function or Loss function is used to configure
0,25,45 DEGREE CELCIUS
the ANN parameters to minimize the error gradient to make
Te st T ANN SVM Linear SVM RBF accurate predictions. Loss curve for training, testing and
MAE RMSE MAE RMSE MAE RMSE validating ANN is obtained using cross entropy (log loss),
MAE and RMSE as below for 1000 data points.
0 1 0.0034 9.2 0.0096 9.6 0.0099
25 1.3 0.0010 7.5 0.2283 6.9 1.0622

DST 45 0.4 0.0000 11.6 0.0136 12.6 0.0136

0 1.4 0.0005 8.6 0.0124 7.3 0.0068
25 0.5 0.0001 4.3 0.0022 4.8 0.0026

FUDS 45 0.9 0.0002 7.2 0.0095 7.1 0.0064

0 1.2 0.0002 6.9 0.0069 6.2 0.0051
25 1.2 0.0005 7 0.0134 7.4 0.0139

US06 45 1 0.0004 12.4 0.0297 4.1 0.0049

From the results it can be seen that MAE of ANN

varies in the range of 0.5-1.4%, whereas for SVM models
MAE lies between 4-12%. RMSE for ANN is converging to
zero however for SVM, there is an inconsistency for different Fig. 7.Loss curve obtained during training, testing and validating
conditions. This signifies that ANN is more accurate in ANN model for 1000 data points.
estimating SoC as compared to SVM methods.

Fig. 8.Comparison of Actual SoC, Predicted SoC verses time

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1360

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

R2 score is a metric to determine the accuracy of the [11] Wang, Y., Zhang, C. and Chen, Z., 2015. A method for state-of-charge
model based on the variance. ANN is having an average estimation of Li-ion batteries based on multi-model switching strategy.
Applied energy, 137, pp.427-434.
accuracy of 99.1% obtained using mean R2 score. A graph of [12] Chang, W.Y., 2013. The state of charge estimating methods for battery:
actual SoC and predicted SoC verses time for one test case i.e. A review. ISRN Applied Mathematics, 2013.
during charging with current as 0.5C, initial capacity as fully [13] Dong, G., Zhang, X., Zhang, C. and Chen, Z., 2015. A method for state
discharged, time interval of 10 seconds is as shown in Fig 7. of energy estimation of lithium-ion batteries based on neural network
model. Energy, 90, pp.879-888.
VI. CONCLUSION [14] Gokulachandran, J. and Mohandas, K., 2011. Tool life prediction model
using regression and artificial neural network analysis. International
Accurate SoC estimation enables BMS to improvise Journal of Production and Quality Engineering, 3(1), pp.9-16.
system reliability, integrity, accuracy and lifespan of a battery [15] Van Schalkwijk, W. and Scrosati, B., 2002. Advances in lithium ion
when trained considering multiple rugged scenarios. It has batteries introduction. In Advances in Lithium-Ion Batteries (pp. 1-5).
been observed that multi-layer perceptron based ANN model Springer, Boston, MA.
can be used for accurate runtime SoC estimation for Lithium- [16] Li, J., Barillas, J.K., Guenther, C. and Danzer, M.A., 2013. A
Ion batteries used in EV’s. ANN can capture trends in SoC comparative study of state of charge estimation algorithms for LiFePO4
batteries used in electric vehicles. Journal of power sources, 230,
variation for different current profiles. The accuracy of the pp.244-250.
model is 99% and MAE is in the range of 0.5-1.4%. The [17] Kashkooli, A.G., Fathiannasab, H., Mao, Z. and Chen, Z., 2019.
network is capable of estimating SoC for different Application of Artificial Intelligence to State-of-Charge and State-of-
temperatures, stress tests and initial conditions. Health Estimation of Calendar-Aged Lithium-Ion Pouch Cells. Journal
of T he Electrochemical Society, 166(4), pp.A605 -A615.
[18] Charkhgard, M. and Farrokhi, M., 2010. State-of-charge estimation for
This paper portrays a proof of concept of a high lithium-ion batteries using neural networks and EKF. IEEE transactions
performance based ANN. This method does not depend on on industrial electronics, 57(12), pp.4178-4187.
battery physics as ANN is purely a data driven approach. [19] Gholizade-Narm, H. and Charkhgard, M., 2013. Lithium-ion battery
Hence this network can be applied to different battery state of charge estimation based on square-root unscented Kalman filter.
variants. The work can further be extended to collect battery IET Power Electronics, 6(9), pp.1833-1841.
data for different drive cycles considering real-time scenarios [20] Meng, J., Luo, G. and Gao, F., 2015. Lithium polymer battery state-of-
charge estimation based on adaptive unscented Kalman filter and
like uphill, downhill, negative temperatures for precise SoC support vector machine. IEEE Transactions on Power Electronics, 31(3),
prediction. pp.2226-2238.
[21] Calce Battery Research group, Data,
REFERENCES https://web.calce.umd.edu/batteries/data.htm, last accessed, 2019/12/19.
[1] Li, Z., Huang, J., Liaw, B.Y. and Zhang, J., 2017. On state-of-charge
determination for lithium-ion batteries. Journal of Power Sources, 348,
pp.281-301.
[2] Landi, B.J., Ganter, M.J., Cress, C.D., DiLeo, R.A. and Raffaelle, R.P.,
2009. Carbon nanotubes for lithium ion batteries. Energy &
Environmental Science, 2(6), pp.638-654.
[3] Li, J., Cheng, H., Guo, H. and Qiu, S., 2018. Survey on artificial
intelligence for vehicles. Automotive Innovation, 1(1), pp.2 -14.
[4] Lim, D.J., Ahn, J.H., Kim, D.H. and Lee, B.K., 2016. A mixed SOC
estimation algorithm with high accuracy in various driving patterns of
EVs. Journal of Power Electronics, 16(1), pp.27 -37.
[5] Xiong, R., Cao, J., Yu, Q., He, H. and Sun, F., 2017. Critical review on
the battery state of charge estimation methods for electric vehicles. IEEE
Access, 6, pp.1832-1843.
[6] Pai, S. and Sindhu, M.R., 2019, October. Intelligent driving range
predictor for green transport. In IOP Conference Series: Materials
Science and Engineering (Vol. 561, No. 1, p. 012110). IOP Publishing.
[7] Huria, T ., Ceraolo, M., Gazzarri, J. and Jackey, R., 2013. Simplified
extended kalman filter observer for soc estimation of commercial power-
oriented lfp lithium battery cells (No. 2013-01-1544). SAE T echnical
Paper.
[8] Ismail, M., Dlyma, R., Elrakaybi, A., Ahmed, R. and Habibi, S., 2017,
June. Battery state of charge estimation using an Artificial Neural
Network. In 2017 IEEE T ransportation Electrification Conference and
Expo (IT EC) (pp. 342-349). IEEE.
[9] Sato, S. and Kawamura, A., 2002, April. A new estimation method of
state of charge using terminal voltage and internal resistance for lead
acid battery. In Proceedings of the Power Conversion Conference-Osaka
2002 (Cat. No. 02T H8579) (Vol. 2, pp. 565 -570). IEEE.
[10] Hammouche, A., Karden, E. and De Doncker, R.W., 2004. Monitoring
state-of-charge of Ni–MH and Ni–Cd batteries using impedance
spectroscopy. Journal of Power Sources, 127(1-2), pp.105-111.

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1361

Authorized licensed use limited to: University of Canberra. Downloaded on July 12,2020 at 17:20:53 UTC from IEEE Xplore. Restrictions apply.