Received December 2, 2017, accepted January 15, 2018, date of publication January 25, 2018, date of current version

March 15, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2797976

Neural Network Approach for Estimating State

of Charge of Lithium-Ion Battery Using
Backtracking Search Algorithm
MAHAMMAD A. HANNAN 1 , (Senior Member, IEEE), MOLLA S. HOSSAIN LIPU2 ,
AINI HUSSAIN2 , (Member, IEEE), MOHAMAD H. SAAD2 , AND AFIDA AYOB2
1 Universiti Tenaga Nasional, Kajang 43000, Malaysia
2 Universiti Kebangsaan Malaysia Faculty of Engineering and Built Environment, Selangor 43600, Malaysia
Corresponding author: Mahammad A. Hannan ([email protected])
This work was supported in part by the Universiti Tenaga Nasional Bold Multi-Track Incentive under Grant 10289176/B/9/2017/36 and in
part by the Universiti Kebangsaan Malaysia under Grant DIP-2015-012.

ABSTRACT The state of charge (SOC) is a critical evaluation index of battery residual capacity. The
significance of an accurate SOC estimation is great for a lithium-ion battery to ensure its safe operation and
to prevent from over-charging or over-discharging. However, to estimate an accurate capacity of SOC of the
lithium-ion battery has become a major concern for the electric vehicle (EV) industry. Therefore, numerous
researches are being conducted to address the challenges and to enhance the battery performance. The main
objective of this paper is to develop an accurate SOC estimation approach for a lithium-ion battery by
improving back-propagation neural network (BPNN) capability using backtracking search algorithm (BSA).
BSA optimization is utilized to improve the accuracy and robustness of BPNN model by finding the optimal
value of hidden layer neurons and learning rate. In this paper, Dynamic Stress Test and Federal Urban Driving
Schedule drive profiles are applied for testing the model at three different temperatures. The obtained results
of the BPNN based BSA model are compared with the radial basis function neural network, generalized
regression neural network and extreme learning machine model using statistical error values of root mean
square error, mean absolute error, mean absolute percentage error, and SOC error to check and validate the
model performance. The obtained results show that the BPNN based BSA model outperforms other neural
network models in estimating SOC with high accuracy under different EV profiles and temperatures.

INDEX TERMS Lithium-ion battery, the state of charge, back propagation neural network, backtracking
search algorithm, electric vehicle.

I. INTRODUCTION energy density, high thermal voltage, low memory effect, low
Global warming, climate change and greenhouse gas (GHG) pollution and light weight [6], [7]. However, the high initial
emissions caused by diesel, petrol-based vehicles have cost, temperature impact, unstable performance during over-
emerged the urges of the implementation of electric vehi- charging and overdischarging operation, reliability and safety
cle (EV) technologies [1], [2]. EVs are regarded as the clean issues and precise residual energy estimation of the lithium-
and energy efficient mode of transportation in recent decades ion battery have become major concerns for the efficient
and the operation of these vehicles are highly dependent lithium-ion battery operation in a longer lifetime for future
on energy storage performance and management system [3]. EV uses [8]. Therefore, research and development are being
EV technologies are growing along with the development actively performed to address the challenges and improve the
of different categories of energy storage devices including lithium-ion battery performance.
lead-acid battery, lithium-ion battery, Ni-MH battery and The state of charge (SOC) is a critical parameter to signify
Ni-Cd battery [4], [5]. Among these, lithium-ion battery has the current and available charge of a battery. However, it is
achieved huge popularity for EV operation because of its difficult to estimate SOC of the battery due to the battery non-
lucrative features including long lifespan, fast charging, high linear characteristics and complex electrochemical reactions.

2169-3536 2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 10069
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

For instance, ampere-hour method uses current integration particle swarm optimization (PSO) [23], backtracking search
to estimate SOC which is the easiest method and can be algorithm (BSA) [24] and artificial bee colony (ABC) algo-
implemented with low power consumption [9]. However, rithm [25]. Among them, BSA algorithm outperforms the
the method suffers from difficulties in determining the initial other algorithms since BSA is easy to implement with few
value of SOC which causes a cumulative effect. Open Circuit parameters that is fast, efficient and robust to control param-
Voltage (OCV) is another frequently used approach which eters. BSA is an advanced optimization technique to find
achieves high accuracy in SOC estimation [10]. Nevertheless, the optimum solution using historical population and map
OCV needs long duration before it reaches a stable condition. matrix to overcome the local minima trap through exploration
The electrochemical models [11], [12] based on the physi- and exploitation. The effectiveness of BSA over other opti-
cal principles use lots of parameters with partial differential mization algorithms can be found in [26]. The BPNN with
equations to predict the battery SOC. However, the heavy BSA has already been used for weights optimization [27]
computational load increases the implementation complexity and node tuning [28]. However, lithium-ion battery modeling
in practical use. Battery impedance is estimated using the with BPNN based BSA being used for both tuning hidden
concept of Electrochemical Impedance Spectroscopy (EIS). layer neurons and learning rate has not been explored.
EIS utilizes inductances and capacitances over a wide fre- SOC is one of the crucial factors in BMS, which indicates
quency spectrum depending on a suitable electrochemical the driving mileage and specifies the residual energy [29].
model. Nonetheless, impedance based models are sensitive High accuracy of SOC estimation not only assures
to aging and temperature that cause difficulty for the repro- reliable operation but also provides information about
duction of EIS results. Kalman filter method has the ability to remaining useful energy [30], [31]. Also, an effective and
reduce the high extent of noise in current measurements [13]. accurate SOC estimation strategy delivers information about
Nonetheless, the Kalman filter model has high mathemati- the rate of charging, discharging and battery health degrada-
cal computation and is extremely vulnerable to aging, tem- tion which is essential for optimal energy management [32].
perature and highly non-linear systems. Fuzzy logic is an Thus, an appropriate SOC estimation will prevent the batter-
intelligent algorithm that estimates SOC considering aging, ies from occurring abnormalities such as over-charge, over-
temperature, and noises [14]. Nonetheless, the fuzzy method discharge and over-heating which will enhance the reliability
requires a huge amount of training data, lots of computation and increase the lifespan of batteries. Nevertheless, accurate
as well as a costly processing unit. Support vector machine estimation of SOC for a lithium-ion battery is a challenging
(SVM) has satisfactory performance in the modeling of a task due to the nonlinear, time-varying characteristics and
nonlinear and high dimensional battery. Besides, SVM can complex electrochemical reactions. Furthermore, lithium-ion
estimate SOC quickly and accurately [15]. However, highly battery is very sensitive to some internal and external factors.
complex computation makes the process difficult to be exe- The SOC is estimated using current capacity divided by
cuted in the battery energy management system (BMS). nominal capacity [33] which is shown in (1).
To address the above shortcomings, a new advanced neural R
i.ηdt
network (NN) technique for estimating SOC is proposed in SOC = SOC init − (1)
this study. NN has robust algorithm to perform under different Cn
battery dynamics, dynamic loads, and different temperatures. where SOC init is the initial value of SOC, i is the battery
The main benefits of using NN are that they do not need current, η is the columbic efficiency, t is time and Cn is the
the mathematical model and they are capable of handling nominal capacity.
any nonlinear and complex system. Nonetheless, NN has In this paper, an optimal BPNN model based BSA algo-
some challenges in determining the most influential features, rithm is developed to improve the accuracy by finding the
activation function, number of neurons in hidden layer and best value of hidden layer neuron and learning rate. The study
learning rate. Many NN based methods have been investi- was conducted with two different EV drive profiles at three
gated for SOC estimation of the lithium-ion battery [16]–[21]. different temperatures. The performance of the proposed
However, most of them have used trial and error for finding model is assessed based on root mean square error (RMSE),
the correct value of hidden layer neurons, learning rate and mean absolute error (MAE), mean absolute percentage error
other parameters that consumes a lot of time. To solve the (MAPE) and SOC error and then compared with the radial
above issue, BPNN with optimization technique is used in basis function neural network (RBFNN), generalized regres-
this study for finding the appropriate value of hidden layer sion neural network (GRNN) and extreme learning machine
neurons and learning rate. (ELM) algorithms for verification and validation purpose.
The optimization technique is important in SOC
estimation for achieving the best values of various param- II. BPNN MODEL
eters in battery models under different conditions. This A three-layer feedforward BPNN model is used in this
paper uses the feedforward BPNN battery modeling which research for the estimation of SOC. The first layer is the
updates the weights and biases of each hidden layer to input layers to characterize the inputs variables, the second
minimize error. There are several techniques for the opti- layer consists of one or more hidden layers and the third
mization of NN such as genetic algorithm (GA) [22], layer is the output layer to characterize the output variables.

10070 VOLUME 6, 2018

M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

Step 4: In this stage, error and biases are updated. Weights are
updated using the following equations
1wk,j = α∂k Sj (9)
wk,j = wk,j + 1wk,j (10)
1wj,i = α∂j xp (11)
wj,i = wj,i + 1wj,i (12)
where α is the learning rate.
Biases are updated using the following equations
1θ k,j = α∂k (13)
θk,j = θk,j + 1θ k,j (14)
1θ j,i = α∂j (15)
θj,i = θj,i + 1θ j,i (16)
The BPNN algorithm has computation complexity such as
FIGURE 1. Flowchart of BPNN model. slow convergence, data overfitting and can easily be trapped
into local minima. However, the computational complexity
The flowchart of BPNN structure is divided into four steps as of BPNN can be addressed by choosing the appropriate
shown in Fig. 1. training algorithm, activation function, number of hidden
The detail explanation of each step is summarized in the layers, number of neurons in the hidden layer and learning
following steps. rate. In this research, BPNN training is performed using
Step 1: Initialize weight and bias to random variables. Levenberg-Marquardt (LM) algorithm because it has good
Step 2: The hidden layer uses log-sigmoid function as a training speed and maintains very good accuracy [34]. The
transfer function which is defined as, most widely used sigmoid activation is used to approximate
the threshold function [35]. BSA is employed to optimize the
1
S (net) = (2) hidden layer neurons and learning rate which will improve
1 + e(−net) the computation efficiency of training BPNN.
For input pattern p, the i-th input layer node holds xp,i . Net
input to j-th node in the hidden layer is, III. DATA DESCRIPTION
Xn
net j = wj,i xp,i + θj,i (3) In this stage, the input variables and samples are collected.
i=0 18650 LiNiMnCoO2 /NMC cell is chosen for the experi-
where, wj,i is the weight from the input layer to the hidden ments. This type of lithium-ion cell is suitable for EV applica-
layer, θj,i represents the bias from the input layer to the hidden tion due to its high voltage and capacity. The test battery has
layer. nominal voltage and capacity of 3.6 V and 2.0 Ah, respec-
Output of j-th node in the hidden layer is, tively. The range of voltage lies between 2.5 V and 4.2 V.
Xn
The test battery can hold a maximum current of 22 A. The

xp,j = Sj wj,i xp,i + θj,i (4)
i=0 specifications of the test battery are shown in Table 1.
Net input to k-th node in the output layer is,
X TABLE 1. Specifications of test battery [36].
net k = wk,j xp,k + θk,j (5)
j
where, wk,j , θk,j , are the weight and bias from the hidden layer
to the output layer, respectively. Output of k-th node in the
output layer is,
X 
Op,k = Sk wk,j xp,k + θk,j (6)
j
Two EV cycles named Dynamic Stress Test (DST) [37]
Step 3: The error is estimated and propagates backward from and Federal Urban Driving Schedule (FUDS) [38] are used
the output layer to the hidden layer. The error in the output for data training and testing. These drive cycles have a diver-
layer is computed as, sity of current profile in terms of different amplitude and
∂k = Sk (1 − Sk )(Tk − Sk ) (7) time duration with regenerative charging. The duration for
completing one cycle of DST and FUDS is 360 seconds
Tk is the true output. The error in the hidden layer is calculated
and 1372 seconds, respectively [36]. In order to validate
as
the proposed model, data were collected from the center for
∂j = Sj (1 − Sj )∂k wk,j (8) advanced life cycle engineering (CALCE) battery group [39].

VOLUME 6, 2018 10071

 M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

data of BPNN can be made more efficient and robust through

appropriate normalization of data. The normalized data can
enhance the convergence rate and able to remove the negative
influence. In this study, input and output are normalized to a
range between -1 and 1, as shown in (17)
2(x − xmin )
x= −1 (17)
xmax − xmin
Where, xmax and xmin are the maximum and the minimum
values of input vector x of the BPNN model. The validation
FIGURE 2. DST EV drive cycle (a) zoom view of current profile (b) current dataset is scaled using the same range used in the training
profile (c) voltage profile. data.
The selected EV drive cycles were scaled down based
on voltage and capacity. In this study, 20 consecutive DST
cycles and 5 consecutive FUDS cycles have been chosen
for discharging the battery. Two EV drive cycles are divided
into two data sets. The developed model is trained using
70% data set (5040 samples for DST and 4802 samples for
FUDS) and tested using 30% data set (2160 samples for DST
and 2058 samples for FUDS). In this study, the maximum
numbers of epochs are 1000 and the performance goal is set as
0.000001. The SOC estimation results of BPNN based BSA
model is compared with the reference value. Statistical error
FIGURE 3. FUDS EV drive cycle (a) zoom view of current profile (b) current terms are used to evaluate the results of the proposed BPNN
profile (c) voltage profile. based BSA model including RMSE, MAE, and MAPE. The
mathematical equations of this statistical errors are expressed
CALCE obtained the data from a custom built battery test as follows:
bench which included NMC cells, an Arbin BT2000 battery r
1 XN
test system for controlling battery charging and discharg- RMSE = (Ies − Ia )2 (18)
ing, a thermal chamber to control temperature and a host N i=1
N
computer with Arbin software to observe and control data 1 X
information [36]. The battery cell was charged by constant MAE = (Ies − Ia ) (19)
N
i=1
current-constant voltage (CCCV) method. The measurements
N
were recorded in a 1-second interval. Data were collected 1 X Ia − Ies
MAPE = (20)
at a low temperature of 0◦ C, room temperature of 25◦ , and N Ia
i=1
a high temperature of 45◦ C. Three most influencing factors
such as voltage, current, and temperature are selected for Where, Ies represents the estimated value, Ia is actual value
data pre-processing and normalization. The selection of three and n is number of observations.
features is logical since current, voltage and temperature have
a substantial impact on battery performance. SOC is varied V. PROPOSED BPNN BASED BSA MODEL
at different charging and discharging current rates. A high The estimation of the SOC begins with the collection of
discharge current rate causes a rapid decrease in SOC. More- three input variables of DST and FUDS drive cycles at three
over, battery SOC has a strong relationship with voltage since different temperatures. Then, data is pre-processed and nor-
increment of SOC causes a rise in voltage. However, open malized. A low pass finite impulse response (FIR) named
circuit voltage as well as charging and discharging currents moving average filter is used for smoothing the sampled data.
of the battery change in different temperature conditions. After that, the hidden layer neurons and learning rate are
The reason is that battery holds a small capacity at low optimized using BSA algorithm based on the minimum value
temperatures due to the rise of internal resistance.The zoom of RMSE. Finally, the data are proceeded for training and
view of current profile, consecutive cycles of current and testing using BPNN activation function. Fig. 4 shows the
voltage profile for DST and FUDS are shown in Fig. 2 and 3, schematic diagram for estimating SOC using BPNN based
respectively. It shows that the current profile of DST is more BSA. The overall process is categorized into four stages
stable and has less oscillation than that of FUDS. which are described as follows:
Stage I. In this stage, the input variables and samples are
IV. TRAINING AND TESTING DATA collected. Three most influencing factors such as
Three important significant factors such as current, voltage voltage, current, and temperature are selected for
and temperature are selected for BPNN training. The training SOC estimation. Then, the sample data is moved

10072 VOLUME 6, 2018

M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

FIGURE 4. Schematic diagram of proposed SOC estimation model based on BPNN with BSA.

through the average filter to reduce the noise. BSA is newly developed a meta-heuristic population-based
After that, the sample data is normalized. effective algorithm to solve the complex and non-linear opti-
Stage II. In this stage, after data pre-processing and nor- mization problems. The goal of BSA optimization is to
malization, the BPNN model is used for SOC improve the performance efficiency of the proposed BPNN
estimation. method as well as give reliable SOC estimation with less vari-
Stage III. The optimal number of hidden layer neurons and ance, less error and the best fitting for the prediction function.
learning rate of BPNN model are computed based BSA is capable to operate large dimensional problem for the
on BSA algorithm to enhance the accuracy of optimum solutions using the historical population and map
SOC. matrix concepts. With the help of historical population, BSA
Stage IV. The proposed model is trained and validated using explores and exploits the better solution to address the local
training and testing data with the optimal value minima trap. On the contrary, map matrix does the necessary
of hidden layer neurons and learning rate. Finally, correction to change the search direction in order to confirm
SOC and SOC error is calculated. the accurate movement during the exploitation search [26].
The BSA algorithm initiates by addressing three quantities,
iteration (T), population size (N), and dimension (D). The
A. BSA ALGORITHM steps of the BSA algorithm are explained as follows [40].
The parameters of BPNN are quite significant for the algo- Step 1 (Initialization): In this stage, the primary and his-
rithm performance. The learning rate parameter can be used torical populations are generated using uniform distribution
to identify if the values are stuck in local minima. In con- within the boundary constraints. Then, the objective function
trast, the hidden layer neurons control the time complexity of each population is calculated. Primary population,
of the algorithm. In general, if the hidden layer neurons are
higher than the optimal value, overfitting may take place, Popn,d ∼ ∪ (lowd , upd ) (21)
and this will lead to high deviation in training performance,
estimation error rate and slow convergence rate. On the other Objective function of primary population,
hand, the underfitting and high variance may occur when the yPop = f (Pop) (22)
numbers of hidden neurons are less than the optimal number.
The initial values of these parameters are set randomly or by Historical population,
experience or by trial and error method, and it is not possible
to ensure that the best solution is achieved. Therefore, the HisPopn,d ∼ ∪ (lowd , upd ) (23)
parameter estimation based on BPNN with trial and error
Objective function of historical population,
process may not be appropriate and takes long duration for
training. To improve the accuracy as well as to increase the yHiisPop = f (HisPop) (24)
computational speed of BPNN, BSA algorithm is employed
for obtaining the best value of hidden layer neurons and Where n ∈ {1, 2, 3, . . . , N } and d ∈ {1, 2, 3, . . . , D}. ∪ is the
learning rate. uniform distribution. Popn,d and HisPopn,d are the individual

VOLUME 6, 2018 10073

 M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

of the main and historical population, respectively. yPop rep-

resents the objective function of the total population size.
lowd and upd are the lower and upper limit of the dimension,
respectively.
Step 2 (Selection-I): BSA has the decision to update the
historical population at the starting of each iteration. The
historical population is revised based on ‘if-then rules’,
as presented by (25).
if < b, then HisPop := Pop|a, b ∼ ∪(0, 1) (25)
where, := is the update operation. BSA defines a population
which is randomly chosen earlier generation as the historical
population and memorizes this historical population until it
is updated. After that, a random shuffling function is applied
to revise the order of the individuals.
HisPop := permuting (HisPop) (26)
Step 3 (Mutation): Mutation is the procedures for obtaining
the new population of the primary and historical population.
The new population is known as trial population which can
be expressed by (27),
Mutant = Pop + F. (HisPop − Pop) (27)
where, F represents the control parameter which changes the
amplitude of the search direction matrix and can be mathe-
matically represented by,
F = 3.randn (28)
FIGURE 5. Pseudo code of BSA algorithm execution.
Where, randn is a function which generates normal distribu-
tion number (0∼1).
Step 4 (Crossover): In this stage, the final version of the Step 5 (Selection-II): In this stage, the trial population gen-
trial population is formed within the boundary condition. This erated in the crossover section is used to verify the objective
section can be divided into two parts: function. Thereafter, the objective function is recalculated for
Step 4(a) (Part-I): The elements of individual are con- each population using (33) and is compared with the previous
trolled by introducing map matrix which have the same size of population (yPop ). Consequently, the objective function is
Pop. Map matrix finds the mutant by governing the individu- updated until the maximum iteration number is achieved.
als to be manipulated and updated through mutation process. Finally, the optimal value of hidden neurons and learning
The development of the map matrix can be expressed using rate are determined using the minimum value of the objective
(29-30). Map matrix initialization, function.
map(1:N ,1:D) = 1 (29) Trial population objective function,

If a < b|a, b ∼ ∪(0, 1), then for n from 1 to N do yTrialPop = f (TrialPop) (33)
Popn = TrialPopn if yn,TrialPop > yn,Pop (34)
mapn,u(1:[mixrate.rand.D]) = 0|u = permuting(h1, 2, 3, . . . , Di)
(30) Detail pseudo code of BSA algorithm is shown in Fig. 5.

where for n from 1 to N do. B. PARAMETER SETTING AND MODEL EXECUTION

Thus, the final trial population, In BSA algorithm, the training process of BPNN algorithm is
(
Popn,d if mapn,d = 1 used to update the hidden layer neurons and learning rate as
TrialPopn,d := (31) well as to assess the objective function of population. First,
Mutantn,d if mapn,d = 0
the initial values of BSA parameters are set. For this study,
Step 4(a) (Part-II): The final trial population is checked the size of the population is 100 and iteration number is 250.
based on boundary range. Also, the value of mix rate is taken as ‘0.9’. The number
of dimensions is assigned as 2. In addition, the boundary
TrialPopn,d = lowd + rand.(upd − lowd ) (32)
limit of two dimensions; hidden layer and learning rate is
If TrialPop TrialPopn,d < lowd or(TrialPopn,d > upd ).


assigned. For hidden layer, the limit is between ‘1’ and ‘50’

10074 VOLUME 6, 2018

M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

and for learning rate, the limit lies between ‘0’ and ‘1’. The electrolyte activity which may support the migration impact
population of hidden layer neurons and learning rate is gener- and ion diffusion [41].
ated randomly using the boundary range. The BPNN training The regression coefficient (R) is a good indicator for eval-
process is executed using the primary and old population of uating the SOC estimation performance. Fig. 7 demonstrates
hidden neurons and learning rate. Then, the objective function that the value of regression coefficient is 0.99852 for DST
(i.e., RMSE) of the primary and old population is assessed. cycle and 0.99831 for FUDS cycle in BPNN based BSA
The equations from (21) to (24) are applied to perform the model. It is seen that the regression coefficient results are in
process. close agreement to unity which validates the accuracy of the
The iteration process initiates to upgrade the historical BPNN-BSA model.
population based on the ‘if-then rule’ using (25). The orders
of the individuals are revised based on random shuffling
function. The trial population is generated through the exe-
cution of mutation process with the primary population and
the updated historical population achieved from the previous
step and control parameter using (27). The binary map matrix
is generated to control the number of the individual to be
updated and manipulated using (30). The final form of the
trial population is generated using (31) and the boundary limit
of the trial population is checked using (32). Accordingly,
the objective function of trial population is evaluated and
compared with the previous value. If the evaluated value of
the objective function is less than that of the previous value, FIGURE 7. Regression plot of the BPNN based BSA algorithm (a) DST
(b) FUDS.
the objective function is updated until the maximum iteration
is achieved. The minimum value of RMSE for hidden layer
and learning rate are determined by the best population value In BPNN based BSA model, the optimal values of hidden
found through all iterations. Thus, the BPNN is trained with layer neurons of 26, 14, 13 and learning rate of 0.958, 0.757,
the optimal number of hidden neurons and learning rate to 0.591 are achieved after 192, 218 and 134 iterations in DST
achieve the best parameters. Finally, the SOC is calculated cycle at 0◦ C, 25◦ C, and 45◦ C, respectively. Similarly, for
and the performance of the proposed model is verified using FUDS cycles, 46, 111, 230 iterations deliver the best opti-
different statistical error terms. mization results presenting hidden layer neurons of 23, 24,
19 and learning rate of 0.745, 0.259, 0.374 at 0◦ C, 25◦ C and
45◦ C, respectively. The optimal values in DST and FUDS
profiles are shown Table 2.

TABLE 2. Optimal parameter of BPNN based BSA model.

FIGURE 6. Convergence response curves for (a) DST (b) FUDS.

The SOC estimation results using BPNN-BSA algorithm

VI. RESULTS AND DISCUSSION are compared with the conventional BPNN and actual SOC
A. SOC ESTIMATION RESULTS for DST and FUDS cycles, as shown in Fig. 8 and Fig. 9,
Fig. 6 presents the convergence characteristics curves of DST respectively.
and FUDS profiles. In DST cycle, the minimum values of The conventional BPNN model estimates hidden layer
the objective function (RMSE) of 0.0183, 0.0148, 0.0119 are neurons and learning rate by trial and error method. Before
attained after 192, 218 and 134 iterations at 0◦ C, 25◦ C, and conducting the experiment, the battery was fully charged
45◦ C, respectively. On the contrary, the FUDS cycle achieves which means that the SOC was 100%. The current sensor was
the lowest objective function of 0.0219, 0.0172, 0.0114 after carefully adjusted which resulted in negligible integration
46, 111 and 230 iterations at 0◦ C, 25◦ C, and 45◦ C, respec- error. Hence, SOC computed by ampere-hour method is the
tively. It is clearly visible from the two cycles that the value actual SOC which is considered as a reference. The reference
of RMSE decreases as the temperature rises.Ambient temper- SOC and SOC estimated by conventional BPNN, BPNN-
ature has a significant effect on battery SOC. The increment BSA are standardized to a range from 100% to 0%. It is
of ambient temperature results in an increase capacity of the observed from Fig. 8 and Fig. 9 that the SOC calculated by
battery. This is due to the fact that the temperature accel- BPNN-BSA is nearly aligned with actual SOC which proves
eration results in a reduction of viscosity and a rise of the that the proposed model achieves satisfactory performance in

VOLUME 6, 2018 10075

 M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

However, in the BPNN-BSA model, SOC error is found at

between −2.4% and 3.5% in FUDS cycle while in DST cycle,
the SOC error is obtained at between −2.1% and 3.2%. The
SOC error achieves high at 0◦ C reaching from −4.8% to 9.8%
in FUDS cycle and −8.5% to 8.8% in DST cycle.

B. PERFORMANCE COMPARISON
To assess the superiority of the proposed method, a compar-
ative study is performed between the proposed BPNN-BSA
model and the RBFNN-BSA, GRNN-BSA and ELM-BSA
methods based on evaluation parameters such as RMSE,
MAE, MAPE and SOC error for DST and FUDS cycles
at 0◦ C, 25◦ C and 45◦ C, respectively. Similar to the pro-
posed BPNN-BSA model, the hyper parameters of RBFNN,
GRNN, and ELM models are tuned using BSA as presented
in Table 3. All the models take current (I ), voltage (V ),
temperature (T ) as inputs.

TABLE 3. Optimal parameters setting used in RBFNN, GRNN and ELM.

FIGURE 8. SOC estimation results and SOC error for DST cycle (a-b) 0◦ C,
(c-d) 25◦ C, and (e-f) 45 ◦ C.

TABLE 4. Comparison of SOC estimation in DST cycle.

FIGURE 9. SOC estimation results and SOC error for FUDS cycle (a-b) 0◦ C,
(c-d) 25◦ C, and (e-f) 45 ◦ C.

estimating SOC with high accuracy. On the contrary, conven-

tional SOC diverges from the actual SOC and has high SOC A detailed comparative analysis of four optimal mod-
error. SOC estimation for DST cycle is more stabilized than els based on different evaluation indicators are illustrated
that of FUDS cycle because current profile in DST is more in Table 4 and Table 5. It is evident that the temperature has
stable and has less fluctuation than that of FUDS. a significant influence on battery performance as the high
Fig. 8 and Fig. 9 also illustrate the SOC error which is temperatures have a low error and high accuracy.
defined as the difference between actual SOC and estimated It is observed that, for DST cycle, the BPNN-BSA model
SOC. It is observed that SOC error reduces as the temperature computes RMSE as 1.47%, 0.81%, and 0.48%, respectively
increases. At 45◦ C, the range of SOC error determined by whereas for FUDS cycle, RMSE is calculated as 1.74%,
BPNN lies at between −3.5% and 4.3% for FUDS cycle and 0.91% and 0.57% at 0◦ C, 25◦ C and 45◦ C, respectively. The
−2.4% and 3.3% for DST with some sudden high peaks. value of RMSE for DST cycle for the proposed method

10076 VOLUME 6, 2018

M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

TABLE 5. Comparison of SOC estimation in FUDS cycle. • The performance of the BPNN method is significantly
improved using BSA by finding the appropriate value of
hidden layer neurons and learning rate.
• The conventional BPNN model based on trial and error
method can estimate SOC; however, the performance
is not satisfactory due to under-fitting and over-fitting
problems. Thus, BSA optimization algorithm is utilized
in BPNN model to enhance the accuracy, robustness and
increase the computational speed of SOC estimation.
The proposed model can be considered a good trade-
off between desired accuracy and complexity of the
SOC estimation algorithm for lithium-ion battery while
avoiding the time-consuming procedures implemented
is improved and reduced by 34%, 62%, and 56%, respec- by other methods.
tively, whereas for FUDS cycle, RMSE is decreased by 59%, • Temperature impact is considered to evaluate the model
58%, and 61%, respectively in comparison to RBFNN-BSA, robustness, adaptability and efficiency of SOC estima-
GRNN-BSA and ELM-BSA methods at 25◦ C. The proposed tion under different dynamic loads for several EV drive
model also has the lower RMSE in comparison to other profiles. In this study, DST and FUDS EV drive profiles
methods at 0◦ C and 45◦ C. Moreover, the value of MAE for are used for testing the BPNN based BSA model at 0◦ C,
DST cycle reaches at 0.32% at 45◦ C, however, the value is 25◦ C, and 45◦ C. The results confirm the superiority of
increased by 138% and 50% at 0◦ C and 25◦ C respectively. BPNN based BSA model with high accuracy.
There is also a decrease in MAE of the proposed method for • The evaluation results of BPNN-BSA model are com-
DST cycle where the error is reduced by 43%, 68%, and 63%, pared with other commonly used neural network models
respectively compared to RBFNN-BSA, GRNN-BSA and such as RBFNN, GRNN, and ELM using the statistical
ELM-BSA methods at 25◦ C. In contrast, the value of MAE at errors of RMSE, MAE, MAPE and SOC error. It is
0◦ C for FUDS cycle is achieved at 0.87% which is decreased observed that the BPNN based BSA model reduces
by 32% and 56% compared to values obtained at 25◦ C and RMSE by 34%, 62%, and 56%, respectively for DST
45◦ C, respectively. Also, there is a reduction in MAE of the cycle, whereas for FUDS cycle, RMSE is reduced
proposed method for FUDS cycle which is dropped by 65%, by 59%, 58%, and 61%, respectively compared to
54%, and 67%, respectively compared to RBFNN-BSA, RBFNN-BSA, GRNN-BSA and ELM-BSA models at
GRNN-BSA and ELM-BSA methods at 25◦ C. 25◦ C. Likewise, the proposed model also minimizes
The performance of the proposed method is further eval- MAE, MAPE and SOC error values in all temperature
uated based on MAPE and SOC error. The MAPE in DST conditions. Thus, the developed model is very effective
cycles is estimated as 7.15% at 25◦ C for the proposed method in reducing errors which in turns increase the SOC
whereas the RBFNN-BSA, GRNN-BSA, and ELM-BSA estimating capability and works efficiently at different
methods have MAPE of 14.58%, 23.39%, and 18.47%, discharge load profiles and temperatures which demon-
respectively which is a 104%, 227% and 158% increase from strate the model robustness.
the proposed method at 25◦ C. Besides, there is a reduction
Comprehensive experimental tests would confirm the adapt-
in MAPE of the proposed method for FUDS cycle which is
ability of the developed SOC estimation model and algorithm
dropped by 32%, 23%, and 39%, respectively compared to
to a wide number of lithium-ion battery storage systems in
other three methods at 25◦ C. Moreover, the proposed method
real time EV and sustainable energy applications.
achieves very narrow SOC error compared to RBFNN-BSA,
GRNN-BSA and ELM-BSA methods, respectively.
REFERENCES
It can be concluded that BPNN with BSA optimization
[1] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, ‘‘A review of
technique has the best performance in terms of accuracy and lithium-ion battery state of charge estimation and management system in
robustness under different EV profiles and temperatures. electric vehicle applications: Challenges and recommendations,’’ Renew.
Sustain. Energy Rev., vol. 78, pp. 834–854, Oct. 2017.
VII. CONCLUSION [2] Y. Xiang, H. Zhou, W. Yang, J. Liu, Y. Niu, and J. Guo, ‘‘Scale evolution
of electric vehicles: A system dynamics approach,’’ IEEE Access, vol. 5,
In this paper, an accurate SOC estimation method is con- pp. 8859–8868, 2017.
structed using improved BPNN based BSA optimization [3] N. Kularatna, ‘‘Rechargeable batteries and their management,’’ IEEE
algorithm. The main contributions of this paper are high- Instrum. Meas. Mag., vol. 14, no. 2, pp. 20–33, Apr. 2011.
[4] A. Dubey and S. Santoso, ‘‘Electric vehicle charging on residential
lighted as follows, distribution systems: Impacts and mitigations,’’ IEEE Access, vol. 3,
• A robust BPNN-BSA estimation strategy is devel- pp. 1871–1893, 2015.
oped which has worked efficiently without considering [5] H. Farzin, M. Fotuhi-Firuzabad, and M. Moeini-Aghtaie, ‘‘A prac-
tical scheme to involve degradation cost of lithium-ion batteries in
battery model, parameters, and complex mathematical vehicle-to-grid applications,’’ IEEE Trans. Sustain. Energy, vol. 7, no. 4,
equations. pp. 1730–1738, Oct. 2016.

VOLUME 6, 2018 10077

 M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

[6] D. Ansean, M. González, J. C. Viera, V. M. García, J. C. Álvarez, [27] J. C. Yang, ‘‘A novel wind power capacity combined forecasting method
and C. Blanco, ‘‘Evaluation of LiFePO4 batteries for electric vehicle based on backtracking search algorithm,’’ in Proc. Int. Ind. Inform. Com-
applications,’’ IEEE Trans. Ind. Appl., vol. 51, no. 2, pp. 1855–1863, put. Eng. Conf., 2015.
Mar./Apr. 2015. [28] B. Wang, L. Wang, Y. Yin, Y. Xu, W. Zhao, and Y. Tang, ‘‘An improved neu-
[7] M. A. Hannan, M. Hoque, S. E. Peng, and M. N. Uddin, ‘‘Lithium-ion ral network with random weights using backtracking search algorithm,’’
battery charge equalization algorithm for electric vehicle applications,’’ Neural Process. Lett., vol. 44, no. 1, pp. 37–52, Aug. 2016.
IEEE Trans. Ind. Appl., vol. 53, no. 3, pp. 2541–2549, May/Jun. 2017. [29] H. He, R. Xiong, and H. Guo, ‘‘Online estimation of model parameters and
[8] S. Dey, S. Mohon, P. Pisu, and B. Ayalew, ‘‘Sensor fault detection, isola- state-of-charge of LiFePO4 batteries in electric vehicles,’’ Appl. Energy,
tion, and estimation in lithium-ion batteries,’’ IEEE Trans. Control Syst. vol. 89, no. 1, pp. 413–420, 2012.
Technol., vol. 24, no. 6, pp. 2141–2149, Nov. 2016. [30] M. Cacciato, G. Nobile, G. Scarcella, and G. Scelba, ‘‘Real-time model-
[9] K. S. Ng, C.-S. Moo, Y.-P. Chen, and Y.-C. Hsieh, ‘‘Enhanced coulomb based estimation of SOC and SOH for energy storage systems,’’ IEEE
counting method for estimating state-of-charge and state-of-health of Trans. Power Electron., vol. 32, no. 1, pp. 794–803, Jan. 2017.
lithium-ion batteries,’’ Appl. Energy, vol. 86, no. 9, pp. 1506–1511, 2009. [31] P. Y. Kong and G. K. Karagiannidis, ‘‘Charging schemes for plug-in
[10] M. Coleman, C. K. Lee, C. Zhu, and W. G. Hurley, ‘‘State-of-charge hybrid electric vehicles in smart grid: A survey,’’ IEEE Access, vol. 4,
determination from EMF voltage estimation: Using impedance, terminal pp. 6846–6875, Nov. 2016.
voltage, and current for lead-acid and lithium-ion batteries,’’ IEEE Trans. [32] M. M. Hoque, M. A. Hannan, and A. Mohamed, ‘‘Voltage equaliza-
Ind. Electron., vol. 54, no. 5, pp. 2550–2557, Oct. 2007. tion control algorithm for monitoring and balancing of series connected
[11] K. A. Smith, C. D. Rahn, and C.-Y. Wang, ‘‘Model-based electrochemical lithium-ion battery,’’ J. Renew. Sustain. Energy, vol. 8, no. 2, p. 25703,
estimation and constraint management for pulse operation of lithium ion Mar. 2016.
batteries,’’ IEEE Trans. Control Syst. Technol., vol. 18, no. 3, pp. 654–663, [33] H. Dai, X. Wei, Z. Sun, J. Wang, and W. Gu, ‘‘Online cell SOC estimation
May 2010. of Li-ion battery packs using a dual time-scale Kalman filtering for EV
[12] T. R. Ashwin, A. McGordon, W. D. Widanage, and P. A. Jennings, ‘‘Modi- applications,’’ Appl. Energy, vol. 95, pp. 227–237, Jul. 2012.
fied electrochemical parameter estimation of NCR18650BD battery using [34] M. T. Hagan and M. B. Menhaj, ‘‘Training feedforward networks with
implicit finite volume method,’’ J. Power Sources, vol. 341, pp. 387–395, the Marquardt algorithm,’’ IEEE Trans. Neural Netw., vol. 5, no. 6,
Feb. 2017. pp. 989–993, Nov. 1994.
[13] Z. Chen, Y. Fu, and C. C. Mi, ‘‘State of charge estimation of lithium-ion [35] J. Han and C. Moraga, ‘‘The influence of the sigmoid function parameters
batteries in electric drive vehicles using extended Kalman filtering,’’ IEEE on the speed of backpropagation learning,’’ in From Natural to Artificial
Trans. Veh. Technol., vol. 62, no. 3, pp. 1020–1030, Mar. 2013. Neural Computation. Berlin, Germany: Springer, 1995, pp. 195–201.
[14] I. H. Li, W. Y. Wang, S. F. Su, and Y. S. Lee, ‘‘A merged fuzzy neural [36] F. Zheng, Y. Xing, J. Jiang, B. Sun, J. Kim, and M. Pecht, ‘‘Influence of
network and its applications in battery state-of-charge estimation,’’ IEEE different open circuit voltage tests on state of charge online estimation for
Trans. Energy Convers., vol. 22, no. 3, pp. 697–708, Sep. 2007. lithium-ion batteries,’’ Appl. Energy, vol. 183, pp. 513–525, Dec. 2016.
[15] J. C. A. Antón, P. J. G. Nieto, C. B. Viejo, and J. A. V. Vilán, ‘‘Support [37] T. Q. Duong, ‘‘USABC and PNGV test procedures,’’ J. Power Sources,
vector machines used to estimate the battery state of charge,’’ IEEE Trans. vol. 89, no. 2, pp. 244–248, Aug. 2000.
Power Electron., vol. 28, no. 12, pp. 5919–5926, Dec. 2013. [38] Y. Xing, W. He, M. Pecht, and K. L. Tsui, ‘‘State of charge estimation
[16] W. He, N. Williard, C. Chen, and M. Pecht, ‘‘State of charge estimation of lithium-ion batteries using the open-circuit voltage at various ambient
for Li-ion batteries using neural network modeling and unscented Kalman temperatures,’’ Appl. Energy, vol. 113, pp. 106–115, Jan. 2014.
filter-based error cancellation,’’ Int. J. Elect. Power Energy Syst., vol. 62, [39] CALCE. Lithium-Ion Battery Experimental Data. Accessed: Jan. 5, 2017.
pp. 783–791, Nov. 2014. [Online]. Available: http://www.calce.umd.edu/batteries/data.htm
[17] M. Charkhgard and M. Farrokhi, ‘‘State-of-charge estimation for lithium- [40] N. N. Islam, M. A. Hannan, H. Shareef, and A. Mohamed, ‘‘An application
ion batteries using neural networks and EKF,’’ IEEE Trans. Ind. Electron., of backtracking search algorithm in designing power system stabilizers
vol. 57, no. 12, pp. 4178–4187, Dec. 2010. for large multi-machine system,’’ Neurocomputing, vol. 237, pp. 175–184,
[18] S. Tong, J. H. Lacap, and J. W. Park, ‘‘Battery state of charge estima- May 2017.
tion using a load-classifying neural network,’’ J. Energy Storage, vol. 7, [41] D. H. Doughty, E. P. Roth, C. C. Crafts, G. Nagasubramanian,
pp. 236–243, Aug. 2016. G. Henriksen, and K. Amine, ‘‘Effects of additives on thermal stability
[19] L. Kang, X. Zhao, and J. Ma, ‘‘A new neural network model for the state- of Li ion cells,’’ J. Power Sources, vol. 146, nos. 1–2, pp. 116–120, 2005.
of-charge estimation in the battery degradation process,’’ Appl. Energy,
vol. 121, pp. 20–27, May 2014.
[20] Y. Shen, ‘‘Adaptive online state-of-charge determination based on neuro-
controller and neural network,’’ Energy Convers. Manage., vol. 51, no. 5,
pp. 1093–1098, 2010.
[21] H. Chaoui and C. C. Ibe-Ekeocha, ‘‘State of charge and state of health
estimation for lithium batteries using recurrent neural networks,’’ IEEE
Trans. Veh. Technol., vol. 66, no. 10, pp. 8773–8783, Oct. 2017.
[22] K. Thirugnanam, T. P. Ezhil Reena Joy, M. Singh, and P. Kumar, ‘‘Math-
ematical modeling of Li-ion battery using genetic algorithm approach MAHAMMAD A. HANNAN (M’10–SM’17)
for V2G applications,’’ IEEE Trans. Energy Convers., vol. 29, no. 2, received the B.Sc. degree in electrical and elec-
pp. 332–343, Jun. 2014. tronic engineering from the Chittagong Univer-
[23] J. Zhao, F. Wen, Z. Y. Dong, Y. Xue, and K. P. Wong, ‘‘Optimal dispatch of sity of Engineering and Technology, Chittagong,
electric vehicles and wind power using enhanced particle swarm optimiza- Bangladesh, in 1990, and the M.Sc. and Ph.D.
tion,’’ IEEE Trans Ind. Informat., vol. 8, no. 4, pp. 889–899, Nov. 2012.
degrees in electrical, electronic, and systems engi-
[24] H. Duan and Q. Luo, ‘‘Adaptive backtracking search algorithm for induc-
neering from Universiti Kebangsaan Malaysia,
tion magnetometer optimization,’’ IEEE Trans. Magn., vol. 50, no. 12,
Dec. 2014, Art. no. 6001206. Bangi, Malaysia, in 2003 and 2007, respectively.
[25] N. K. Paliwal, N. K. Singh, and A. K. Singh, ‘‘Optimal power flow in grid He is currently a Professor of Intelligent Systems
connected microgrid using artificial bee colony algorithm,’’ in Proc. IEEE with the Department of Electrical Power Engineer-
Region 10 Conf. (TENCON), Nov. 2016, pp. 671–675. ing, College of Engineering, Universiti Tenaga Nasional, Malaysia. His
[26] P. Civicioglu, ‘‘Backtracking search optimization algorithm for numer- research interests include intelligent controllers, power electronics, hybrid
ical optimization problems,’’ Appl. Math. Comput., vol. 219, no. 15, vehicles, energy storage systems, image and signal processing, and artificial
pp. 8121–8144, 2013. intelligence. He is an Associate Editor of IEEE ACCESS.

10078 VOLUME 6, 2018

M. A. Hannan et al.: Neural Network Approach for Estimating SOC of Lithium-Ion Battery Using BSA

MOLLA S. HOSSAIN LIPU received the B.S. MOHAMAD H. SAAD received the B.Sc. degree
degree in electrical and electronic engineer- in mechanical and materials engineering and
ing from the Islamic University of Technology, the Ph.D. degree from Universiti Kebangsaan
Gazipur, Bangladesh, in 2008, and the M.S. degree Malaysia in 2017. He is currently a Senior Lecturer
in energy from the Asian Institute of Technology, with the Department of Mechanical and Materi-
Bangkok, Thailand, in 2013. He is currently pursu- als Engineering, Faculty of Engineering and Built
ing the Ph.D. degree with the Centre for Integrated Environment, Universiti Kebangsaan Malaysia.
Systems Engineering and Advanced Technologies, His research interests include artificial intelli-
FKAB, Universiti Kebangsaan Malaysia, Bangi, gence, intelligent systems, and complex event
Malaysia. He is currently an Assistant Professor processing.
of Electrical and Electronic Engineering with the University of Asia Pacific,
Dhaka, Bangladesh. His research interest includes energy storage systems,
artificial intelligence, model optimization, and hybrid renewable energy
system.

AINI HUSSAIN (M’98) received the B.Sc. degree

in electrical engineering from Louisiana State Uni- AFIDA AYOB received the B.Sc. degree in elec-
versity, USA, the M.Sc. degree from the University trical and electronic engineering from the Univer-
of Manchester Institute of Science and Technol- sity of Manchester in 2000 and the Ph.D. degree
ogy, U.K., and the Ph.D. degree from Universiti from the Newcastle University in 2012. She is
Kebangsaan Malaysia. She is a Professor with the a Senior Lecturer with the Centre for Integrated
Centre for Integrated Systems Engineering and Systems Engineering and Advanced Technologies,
Advanced Technologies, Faculty of Engineering Faculty of Engineering and Built Environment,
and Built Environment, Universiti Kebangsaan Universiti Kebangsaan Malaysia. Her research
Malaysia. Her research interests include decision interests include power electronic, energy storage,
support systems, machine learning, pattern precognition, and signal and and artificial intelligence.
image processing.

VOLUME 6, 2018 10079