See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/356698394

State of Charge (SoC) Estimation of Battery Energy Storage System (BESS)

Using Artificial Neural Network (ANN) Based on IoT- Enabled Embedded
System

Conference Paper · August 2021

DOI: 10.1109/ICA52848.2021.9625697

CITATIONS READS

2 132

5 authors, including:

Putu Handre Kertha Utama Irsyad Nashirul Haq

Bandung Institute of Technology Bandung Institute of Technology
5 PUBLICATIONS   3 CITATIONS    43 PUBLICATIONS   204 CITATIONS   

SEE PROFILE SEE PROFILE

Edi Leksono
Bandung Institute of Technology
88 PUBLICATIONS   419 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Errors Modeling and Compensation of Inertial Navigation System/Long Baseline Acoustic Positioning System for Autonomous Underwater Vehicle View project

Peningkatan Technology Readiness Level (TRL) dari Purwarupa Sistem Monitoring & Kontrol Energi Listrik menggunakan Multipurpose Remote Control Berbasis
Internet of Things (IoT) untuk Aplikasi Energy Management System View project

All content following this page was uploaded by Irsyad Nashirul Haq on 17 January 2022.

The user has requested enhancement of the downloaded file.

 2021 International Conference on Instrumentation, Control, and Automation (ICA)
Bandung, Indonesia, August 25th – August 27th, 2021

State of Charge (SoC) Estimation of Battery Energy

Storage System (BESS) Using Artificial Neural
Network (ANN) Based on IoT- Enabled Embedded
2021 International Conference on Instrumentation, Control, and Automation (ICA) | 978-1-6654-3295-5/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICA52848.2021.9625697

System
Putu Handre Kertha Utama Hilda Hamdah Husniyyah Irsyad Nashirul Haq
Engineering Physics, Faculty of Engineering Physics, Faculty of Engineering Physics, Faculty of
Industrial Technology Industrial Technology Industrial Technology
Institut Teknologi Bandung Institut Teknologi Bandung Institut Teknologi Bandung
Bandung, Indonesia Bandung, Indonesia Bandung, Indonesia
[email protected] [email protected] [email protected]

Justin Pradipta Edi Leksono

Engineering Physics, Faculty of Engineering Physics, Faculty of
Industrial Technology Industrial Technology
Institut Teknologi Bandung Institut Teknologi Bandung
Bandung, Indonesia Bandung, Indonesia
[email protected] [email protected]

Abstract—The battery energy storage system (BESS) plays a system (BMS) is required to protect and monitor the battery
significant role in the microgrid system to harness renewable during operation.
energy sources. BESS generally consists of battery modules
connecting in series or parallel configurations to achieve BESS consists of battery modules connecting in series or
operational voltage and capacity. In such a complex system, a parallel configurations to achieve operational voltage and
battery management system (BMS) is necessary to guarantee capacity. In such a complex system, a battery management
safety, reliability, and efficiency while in operation. One critical system (BMS) is necessary to guarantee the safety, reliability,
function of BMS is the state of charge (SoC) estimation of the and efficiency of a Li-ion battery in operation [3]. One critical
battery system. It is necessary to have a battery monitoring function of BMS is the state of charge (SoC) estimation of the
system based on the internet of things (IoT) enabled devices that battery system[4]. The SoC is like the fuel gauges in fuel
can transmit SoC data in real-time. This paper proposed SOC storage that show the remaining usable fuel [5]. Hence, it is
estimation using an artificial neural network (ANN) to reduce necessary to have a battery monitoring system based on the
the estimation error due to physical parameters and reduce internet of things (IoT) enabled devices that can transmit data
computation cost using an IoT-enabled embedded system. The in real-time[6] [7].
experimental setup is set using 15 cells of high capacity of
lithium ferro phosphate (LFP) prismatic batteries with nominal In recent years, the estimation of battery SoC intensively
voltage 3.2VDC and 100Ah. The battery cells are connected in studied. SoC estimation method categorizes into four groups:
series to achieve a BESS nominal voltage of 48VDC. The result direct measurement, book-keeping, model-based, and
of SoC estimation shows the ANN model provides better computer intelligence [8]. The direct measurement approach,
accuracy than the support vector machine method. Both namely, the Current Integral method [9], the Open Circuit
qualitative perspective (curve plot) and quantitative perspective Voltage (OCV) method [10]. The model-based approach such
(model metrics) justify the accuracy of the ANN model. The as Equivalent Circuit Model [11], Incremental Capacity
ANN model also successfully deploys on MCU as an IoT-enabled Analysis (ICA)[12], Electrochemical Model [13] combined
embedded system. with Advanced Observers such as extended Kalman Filter
(EKF) [14], Unscented Kalman Filters [15]. And the computer
Keywords—BESS, SOC estimation, artificial neural network,
intelligence approach like Fuzzy Logic, Support Vector
internet of things, BMS
Machine (SVM), Sparse Learning Machine, and Artificial
I. INTRODUCTION Neural Networks (ANN) [15].
The battery energy storage system (BESS) plays a Finding the best techniques to estimate the SoC of the
significant role in the microgrid system to harness renewable battery system is necessary but deploy it in an embedded
energy sources [1]. BESS can also help provide a stable and system is another issue. The high computing cost and limited
reliable power supply in conventional grids. The typical storage capacity of the embedded system still become a matter
choice for this application is lithium-ion batteries [2] due to in SoC implementation. Despite this, the study in [16]
their excellent performances in self-discharge rate, lifespan, successfully implements real‐time battery SoC estimation
high energy, and power capability [3]. To enhance the using a physics-based model square root cubature Kalman
performance of the lithium-ion battery, a battery management filter on a microcontroller unit (MCU).
978-1-6654-3295-5/21/$31.00 ©2021 IEEE
77

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
 The physics-based model in [16] shows a satisfying result,
but this method is prone to modeling error due to physical
parameters uncertainty[15]. The advancement of digital
technology shifts the SoC estimation trend toward less
demanding physical parameters, data-driven computer
intelligence, or machine learning (ML) methods such as
support vector regression (SVR)[17, 18]. In [19], the SOC
estimation is based on the SVR algorithm with a root mean
square error (RMSE) of 2.3%. This paper addresses the
physics-based model issue and significant RMSE error of
SVR by utilizing the artificial neural network (ANN)
algorithm. The proposed method maximizes the potential of
ML to reduce the estimation error due to physical parameters,
reduce computation cost[15], and provide better accuracy
compared to SVR.
II. METHODOLOGY
The study runs in five steps (Fig. 1). First, battery
charging-discharging data was collected using a single-cell
tester device (battery data). The most reliable test to obtain the
battery SoC is the discharge test at a particular current rate
under controlled conditions[20]. The battery should be kept at
its rest time during the test so that the terminal voltage
approaches its open-circuit voltage (OCV). The rest time is
necessary to get the accurate OCV-SoC curve[21]. The OCV–
SoC data use as data training for developing the battery
model[19]. SoC is defined as the ratio of the current remaining
battery capacity to the battery capacity before discharge [6].
Mathematically, the SoC of the battery can be obtained as in
(1), where SoC represents the SoC of the battery,
represents the current remaining capacity of the battery in
Fig. 1. SoC estimation on IoT- enabled embedded system.
Ah, and represents the capacity of the battery when fully
charged in Ah. In the SVR method, three types of kernels are commonly used.
(1) The three kernels are the linear kernel, polynomial kernel, and
RBF kernel. The kernel function used in this study is the radial
basis function (RBF) which is the most popular choice of the
Before feed into the modeling phase, the data obtained type of kernel used in SVR, where the mainreason for its use
earlier need to pass through the data preparation phase (pre- is because of its localized and limited response across the real
processing), such as filtering and removing outlier data to x-axis range. Finding the suitable values of ε, C, and γ of the
minimize errors during modeling. In the modeling (model SVR parameters is very complicated so the GridSearchCV
training), the performance of ANN and SVR algorithms to method in python (sci-kit-learn library) is used to determine
estimate the battery system SoC are compares. the optimal SVR parameters.
A. SVR Method B. ANN Method
SVR is part of SVM that specifically design to dealing Unlike the SVR method that only requires few parameters,
with the regression process. It presents the prediction result as the ANN method has many parameters to consider. The
a whole number instead of a group or classification [12]. The parameters include the number of neurons in the input layer,
training data from the previous phase was used to build a the number of hidden layers, the number of neurons in the
battery model using the SVR algorithm. The optimal model hidden layer, the activation method, the optimization methods,
acquires by tuning the SVR parameters includes kernel, the dropout value, batch_size, iteration/epoch [22]. Modify one of
cost function (C), epsilon function (ε), and gamma parameter the parameters can produce different outputs. Hence, the
(γ) [18]. Parameter C is a representation of the range of the tuning process on each parameter is carried out by altering one
boundary field to consider the training data owned. The larger parameter and hold the other parameters at a value.
the value of C, the model will attempt to make adjustments to
more training data. If the value of C is too large, the model After obtaining the optimal model for the ANN method,
will be susceptible to interference (noise) and can cause loss the model then compares with the model obtained using the
of the general nature of the model (overfitting). Parameter SVR method. The model metrics mean absolute error (MAE),
Gamma (γ) is the inversion of variance and the Epsilon mean percentage error (MAPE), and root mean square error
parameter (ε) is the width of the allowed tolerance radius from (RMSE) become the baseline to select the best model
the training data to determine the model boundary area. (validation) [3], [9]. The MAE is the mean of the absolute
values of the difference between the actual and predicted
values. It penalizes the errors due to bias or variance evenly.
Meanwhile, the MAPE is the MAE of the prediction against
the mean of actual value. The model metric that differentiates
the error due to bias and variance is the RMSE. It is the root

78

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
of squares of the difference between the actual and the protocol. Each cell board utilized a DS2438 chip to convert
predicted values. analog to digital data with a typical error of only 10 milliVolt.
To verify the model accuracy, the MCU is put on the test to
The last step is deploying the better model to IoT-enabled estimate the battery system SoC in real-time, and the digital
MCU, in this case, ESP32 acting as the BMS - local module data is converted and can be sent to another device by using a
(deployment). The experimental setup of BESS consists of 15 WiFi/ Ethernet connection based on JSON data format.
cells of high capacity of Lithium Ferro Phosphate (LFP)
prismatic batteries with nominal voltage 3.2VDC and 100Ah III. RESULTS AND DISCUSSION
with each cell characteristic as shown in Table I. The battery
cells are connected in series as one LIB module to achieve a The battery SoC modeling begins by collecting battery
BESS nominal voltage of 48VDC as shown in Fig. 2. The data from discharging and charging the battery. The process
BMS local module then predicts the SoC of each cell based on of discharging and charging the battery is carried out using a
the deployed model. Fig. 3. shows the actual experimental 5A current for 20hours with 1 second sampling time. The
setup to test the SoC estimation model. total measurement data obtained is 74,935 data.
Fig. 4 illustrates the discharging characteristics of the
BMS-Local Module battery module voltage as a function of its capacity. We
JSON TCP/IP Microcontroller
WiFi / Eth. Connection
observe that the battery module has a nearly linear voltage to
Interface (Monitoring and SOC Estimation ) discharge capacity when the voltage of the battery module is
Communication Input Actuator Output Actuator Output Input
one-wire bus Charger Load Current Sensor from 3.2 V to 2.87 V, while the charging period is from 3.31
V to 3.4 V. Outside those ranges, the relationship is highly
non-linear. To estimate SoC in this study, will use
discharging data.
Battery Arrangement

Before the data is used for training, the data goes through
Cell
Board 1
Cell
Board 2
Cell
Board 3
Cell
Board 4
Cell
Board ..
Cell
Board ..
Cell
Board ..
Cell
Board
data preparation. Such as filtering and removing outlier, so
15 that there are no significant errors when creating the model
Battery Battery Battery Battery Battery Battery Battery Battery and calculation of SoC on battery data using (1).
cell 1 cell 2 cell 3 cell 4 cell … cell … cell … cell 15
The next step is to estimate the battery SoC using the SVR
method. At this step, training data is used in the form of
Fig. 2. BMS based on IoT- Enabled Embedded System.
voltage data and SoC calculations when discharging. It is
TABLE I. HIGH CAPACITY LFP BATTERY CHARACTERISTICS known that three SVR parameters will determine the accuracy
of the resulting model, namely parameters C, gamma, and
Parameter Value epsilon. Incorrect parameter values will result in a poor
Model Type NS36130290-100PF
estimation model. To be able to determine the best parameters,
Nominal capacity 100 Ah
Nominal voltage 3.2 V use the Gridsearch CV module in Python. The result of the
Cut-off voltage (charging; discharging) 3.65 V; 2.2 V grid search is in the form of an average test score which can
Recommend charging; discharging current 0.5 C; Under 2C then be referred to as a parameter score, with a range of 0 to
Max pulse charging; discharging current 2C @ 10 s; 5C @ 5 s 1. The higher the score indicating more suitable SoC model.
Life Cycles (at 100% DoD, 0.5 C) 2500
Operating temp. (charging; discharging) 0 ~ 50 oC ; (-20) ~ 50 oC The best combination obtained from Gridsearch are C =
Weight 3.0 kg 5.125, gamma = 10, and epsilon = 0.8, with a score of 0.9978.

Fig. 4. Battery module voltage as a function of battery capacity.

Fig. 3. BESS SoC experimental setup.
To validate that the gamma value obtained is optimum, the
The voltage of each cell measure by the cell board sends values of C and epsilon are kept constant, while the gamma
to the local module using the onewire communication values are increased and decreased until the highest score is

79

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
obtained. The results of the validation of the optimum epsilon reference can be seen in Fig. 5 and Fig. 6. The results are quite
parameters are presented in Table II. good, but when there is a turn at a voltage of 3.28 V to 3.4 V
the model cannot adjust. The sign of the SVR model is quite
TABLE II. SVR MODEL PARAMETER TUNING difficult to fit non-linear graphics. It can also be seen in Fig.
Parameter 5 that the error value when the SoC estimation of the voltage
Tunning Score value is quite volatile. It can be seen that the SoC estimate
C ε γ
1 5.215 0.8 8 0.837 based on SVR is a much better match with its reference until
2 5.215 0.8 10 0.928 90% with the linear condition.
3 5.215 0.8 15 0.902
Meanwhile, modeling using the ANN method requires
Based on Table II, it appears that the gamma value is several parameters, and these parameters are different from
directly proportional to the score. The greater the gamma SVR. The parameters needed include the number of neurons
value, the greater the parameter score value. In tunning 3 it in the input layer, the number of hidden layers, activation
can be seen that the score dropped from 0.929 to 0.902. This functions, and optimizer. To determine the number of hidden
indicates that the optimum model is the previous model, layers and the optimal activation function, you can use
namely during the second tuning, with an epsilon gamma Gridsearch CV as well as SVR. However, to determine the
value of 10. Hence, the results of the second tuning parameter number of neurons in each layer and the optimal optimizer it
are used to estimate the battery SoC using the SVR method. must be trial and error.
After tuning using Gridsearch, the optimal parameter is the
activation function using RELU with the number of hidden
layers 8. In addition, parameters such as the number of
neurons and the optimizer are carried out by trial and error.
For the number of neurons, the optimal result is 8 neurons at
the input layer. And for the hidden layer, the number of
neurons in a row for each layer is 12, 10, 4, 6, 8, 4, 6, 8 with
500 epochs and 500 batch sizes.
Another parameter that is also important to obtain an
accurate ANN model is the optimizer. The ANN optimizer
consists of 8 optimizers. In this study, we tried all optimizers
but briefly summarized 4 optimizer results with good results.
These results can be seen in Table III, and graphically can be
seen in Fig. 7.
TABLE III. ANN OPTIMIZER
Metrics rmsprop Adam Adamax Nadam
MAE 0.02 0.01 0.01 0.04
MAPE 4.2% 1.3 % 3.58% 3.45%
Fig. 5. Battery SoC Model obtain using the SVR method. RMSE 0.15 0.02 0.12 0.18

Fig. 6. Battery SoC error model obtain using the SVR method.
Fig. 7. Optimizer ANN method.
The SVR model is used to estimate battery data when
discharging, to see the accuracy of the model when it is
implemented with the results of the SoC reference. The result
of the SoC estimate by using SVR that is compared to its

80

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
 by ANN on the voltage value less volatile, tends to be
constant, and has a low value compared to the SVR modeling
error.
For quantitative assessment, the ANN and SVR model
metrics are compared side by side in Table III. It verified the
Fig. 8 result that the ANN model produces more accurate SoC
estimation. From all the metrics models, the ANN model
outperforms the model obtain using the SVR method. SVR
model fails outside the 3.0V - 3.1V voltage range.

TABLE IV. COMPARISON OF ANN AND SVR SOC MODEL

Model Metrics ANN SVR
MAE 0.01 0.06
MAPE 1.3 % 4.3%
RMSE 0.02 0.07

If this work is used for real applications in the field, such

Fig. 8. Battery SoC Model obtain using ANN method. as to estimate the SoC on BMS where the SoC working range
is usually set on the range of 20% to 80% [1], this SVR result
couldn’t cover it. Because SVR just covers the range of 20%
to 60%. If we compared with ANN, ANN results could cover
it. By using the SoC working range, the shorter battery life due
to the depth of discharge and battery overcharged risk can be
avoided.
Based on the above result, the ANN model was selected as
the better model compared with SVR. The model then deploys
to the MCU unit to be further tested. To deploy the python
model to the MCU, the model was converted to C code by
utilizing the tinymlgen library.
The MCU uses to estimates the 15-serial battery LFP 3.2V
100Ah in operation (online estimation). The difference
between the SoC estimation (est. SoC) from the SoC
references data prepared for the modeling phase (ref. SoC)
consider an estimation error as percentage of SoC. The
maximum estimation error is 0.4%, with an average
estimation error of approximately 0.2%, as summarized in
Fig. 9. Battery SoC error model obtain using ANN method. Table V.

The results of the four optimizers graphically and model TABLE V. BATTERY SOC ESTIMATION ERROR
metrics produce a fairly good model and the values of the four est. SoC
ID Voltage (V) ref. SoC (%) Error (%)
optimizers are quite close. However, if look more clearly, the (%)
rmsprop optimizer has a fairly large error compared to other 1 3.26 99.4 99.6 0.2
optimizers, and on the graph, the rmsprop optimizer does not 2 3.27 99.5 99.7 0.2
3 3.28 99.7 99.8 0.1
fit the reference. For the adamax and nadam optimizers, even
4 3.23 99.0 99.3 0.3
when the voltage is 2.8 V – 3.1 V, it does not fit the reference 5 3.33 100 99.9 0.1
value. So that the best optimizer among the others is the adam 6 3.31 100 99.8 0.2
optimizer with MAE 0.01, MAPE 5.26%, and RMSE 0.03. 7 3.32 100 99.9 0.1
When compared to the SVR model, the four optimizers 8 3.24 99.1 99.5 0.4
produce smaller errors. 9 3.28 99.6 99.7 0.1
10 3.35 100 99.9 0.1
The results of the ANN modeling with the adam optimizer 11 3.32 100 99.9 0.1
are shown in Fig. 8. It can be seen that the estimation and 12 3.24 99.1 99.5 0.4
reference results are very fit, although at a voltage of 3.55 – 13 3.27 99.5 99.7 0.2
3.6 V there is overfitting, this is not a big problem. The error, 14 3.31 100 99.8 0.2
which is located in Fig. 8 is not fluctuating, dominant is 15 3.30 99.9 99.8 0.1
constant, and has a fairly low value.
IV. CONCLUSIONS
Fig. 5. and Fig. 8. show the ANN model can fit with most
of the data points. However, the model and the estimated value In this study, the battery SoC models were developed
generated by the SVR and ANN methods obtained good using two machine learning algorithms, SVR and ANN. The
performance. The results of the two models show that the SoC result of SoC estimation shows the ANN model provides
estimate based on ANN is much better than SVR, matches better accuracy than SVR. Both qualitative perspective (curve
with its reference starting from 2.9 V is the start of the voltage plot) and quantitative perspective (model metrics) justify the
to the end of the voltage. In addition, the SoC error generated accuracy of the ANN model. The ANN model also

81

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
successfully deploys on MCU as an embedded system. The [10] Rivera-Barrera JP, Muñoz-Galeano N, Sarmiento-Maldonado HO. Soc
embedded system was then put on the test for online SOC estimation for lithium-ion batteries: Review and future challenges.
2017. Epub ahead of print 2017. DOI: 10.3390/electronics6040102.
estimation of a 15-serial LFP battery. It demonstrates the
[11] Ma Y, Duan P, Sun Y, et al. Equalization of Lithium-Ion Battery Pack
online estimation accuracy with the maximum SoC estimation Based on Fuzzy Logic Control in Electric Vehicle. IEEE Trans Ind
error is 0.4% and an average SoC estimation error of Electron 2018; 65: 6762–6771.
approximately 0.2% as percentage of SoC. This result still [12] He Z, Wen X, Liu H, et al. A comparative study of artificial neural
does not consider the error produced by the uncertainty of network, adaptive neuro fuzzy inference system and support vector
current and temperature measurement. Hence, it can be a machine for forecasting river flow in the semiarid mountain region. J
future research challenge. Hydrol 2014; 509: 379–386.
[13] Zhang L, Li K, Du D, et al. A Sparse Learning Machine for Real-Time
ACKNOWLEDGMENT SOC Estimation of Li-ion Batteries. IEEE Access 2020; 8: 156165–
156176.
This research was partially supported by the ITB Research [14] Yang F, Li W, Li C, et al. State-of-charge estimation of lithium-ion
Grant, Lembaga Pengelola Dana Pendidikan, the Ministry of batteries based on gated recurrent neural network. Energy 2019; 175:
Research and Technology/National Research and Innovation 66–75.
Agency of Indonesia and PT Pertamina INV. [15] Rivera-Barrera JP, Muñoz-Galeano N, Sarmiento-Maldonado HO.
SoC Estimation for Lithium-ion Batteries: Review and Future
REFERENCES Challenges. Electronics ; 6. Epub ahead of print 2017. DOI:
10.3390/electronics6040102.
[1] Salas-Puente R, Marzal S, Gonzalez-Medina R, et al. Practical
Analysis and Design of a Battery Management System for a Grid- [16] Cui X, He Z, Li E, et al. State-of-charge estimation of power lithium-
Connected DC Microgrid for the Reduction of the Tariff Cost and ion batteries based on an embedded micro control unit using a square
Battery Life Maximization. Energies ; 11. Epub ahead of print 2018. root cubature Kalman filter at various ambient temperatures. Int J
DOI: 10.3390/en11071889. Energy Res 2019; 43: 3561–3577.
[2] Zhang S, Zhai B, Guo X, et al. Synchronous estimation of state of [17] Ipek E, Eren MK, Yilmaz M. State-of-Charge Estimation of Li-ion
health and remaining useful lifetime for lithium-ion battery using the Battery Cell using Support Vector Regression and Gradient Boosting
incremental capacity and artificial neural networks. J Energy Storage Techniques. In: 2019 International Aegean Conference on Electrical
2019; 26: 100951. Machines and Power Electronics (ACEMP) & 2019 International
Conference on Optimization of Electrical and Electronic Equipment
[3] Shen M, Gao Q. A review on battery management system from the (OPTIM). 2019, pp. 604–609.
modeling efforts to its multiapplication and integration. Int J Energy
Res 2019; 43: 5042–5075. [18] Haq IN, Saputra RH, Edison F, et al. State of charge (SoC) estimation
of LiFePO4 battery module using support vector regression. In:
[4] Hannan MA, Lipu MSH, Hussain A, et al. Toward Enhanced State of Proceedings of the Joint International Conference on Electric
Charge Estimation of Lithium-ion Batteries Using Optimized Machine Vehicular Technology and Industrial, Mechanical, Electrical and
Learning Techniques. Sci Rep 2020; 10: 4687. Chemical Engineering (ICEVT & IMECE). 2015, pp. 16–21.
[5] Tran N-T, Khan AB, Nguyen T-T, et al. SOC Estimation of Multiple [19] Haq IN, Saputra RH, Edison F, et al. State of charge (SoC) estimation
Lithium-Ion Battery Cells in a Module Using a Nonlinear State of LiFePO4 battery module using support vector regression. Proc - Jt
Observer and Online Parameter Estimation. Energies ; 11. Epub ahead Int Conf Electr Veh Technol Ind Mech Electr Chem Eng ICEVT 2015
of print 2018. DOI: 10.3390/en11071620. IMECE 2015 2016; 16–21.
[6] Friansa K, Haq IN, Santi BM, et al. Development of Battery [20] Álvarez Antón JC, García Nieto PJ, de Cos Juez FJ, et al. Battery state-
Monitoring System in Smart Microgrid Based on Internet of Things of-charge estimator using the SVM technique. Appl Math Model 2013;
(IoT). Procedia Eng 2017; 170: 482–487. 37: 6244–6253.
[7] Nashirul Haq I, Kurniadi D, Leksono E, et al. Performance Analysis of [21] Xing Y, He W, Pecht M, et al. State of charge estimation of lithium-
Energy Storage in Smart Microgrid Based on Historical Data of ion batteries using the open-circuit voltage at various ambient
Individual Battery Temperature and Voltage Changes. J Eng Technol temperatures. Appl Energy 2014; 113: 106–115.
Sci 2019; 51: 149.
[22] Lin Q, Wang J, Xiong R, et al. Towards a smarter battery management
[8] Vidal C, Malysz P, Kollmeyer P, et al. Machine Learning Applied to system: A critical review on optimal charging methods of lithium ion
Electrified Vehicle Battery State of Charge and State of Health batteries. Energy 2019; 183: 220–234.
Estimation: State-of-the-Art. IEEE Access 2020; 8: 52796–52814.
[9] yu Z, Huai R, Xiao L. State-of-Charge Estimation for Lithium-Ion
Batteries Using a Kalman Filter Based on Local Linearization.
Energies 2015; 8: 7854–7873.

82

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on January 09,2022 at 16:31:04 UTC from IEEE Xplore. Restrictions apply.
View publication stats