BATTERY STATE OF CHARGE ESTIMATION

USING AN ARTIFICIAL NEURAL NETWORK

Mahmoud Ismail, Rioch Dlyma, Ahmed Elrakaybi, Ryan Ahmed, and Saeid Habibi

Mechanical Engineering Department

McMaster University
Hamilton, Canada
E-mail: [email protected], [email protected], [email protected],
[email protected], and [email protected]

The automotive industry is currently experiencing a accurately estimating, in real-time, the stored energy in the
paradigm shift from conventional, diesel and gasoline- vehicle batteries. This is a relatively complex task due to many
propelled vehicles into the second generation hybrid and aspects. First, there is no direct measurement method that can
electric vehicles. Since the battery pack represents the most directly measure the stored energy in the battery pack. Two
important and expensive component in the electric vehicle methods currently exist for directly evaluating the battery pack
powertrain, extensive monitoring and control is required. SOC from measurable variables such as current and voltage;
Therefore, extensive research is being conducted in the field namely: Coulomb Counting and regressed-voltage based State
of electric vehicle battery condition monitoring and control. of Charge – Open Circuit Voltage (SOC-OCV) mapping. The
In this paper, an Artificial Neural Network (ANN) is used first method is current dependent, in which the stored energy
for Lithium-Ion (Li-Ion) battery state-of-charge (SOC) in the battery is calculated based on integrating the
estimation. When properly trained using the random current charge/discharge current values over time, as shown in
profile described in this paper, a single-layered Neural Equation (1) [5]:
Network is capable of capturing the non-linear
SOC = 1 -
ò i ×h × dt (1)
characteristics of a battery. The ANN is able to estimate a
Cn
non-measurable parameter such as battery SOC level based
Where 𝑡 is time, 𝑖 is the current, 𝐶$ is the nominal capacity,
on battery measurable parameters such as voltage and
and 𝜂 is the coulomb efficiency. This method is prone to errors
current. The ANN in this paper is trained using
as it is an open loop estimation system and any small error
experimental data generated from an experimental battery
will accumulate with time due to the integration term. The
using a R-RC model with SOC/OCV relationship. The
SOC/OCV relationship was derived from a commercial 3.6V second method depends on voltage measurements. It uses the
OCV to estimate the SOC level. OCV measurements can be
3.4Ah Li-Ion battery cell. The network is trained using
mapped to SOC through a nonlinear relationship similar to the
current, and voltage as inputs and SOC as the output. The
trained network is tested using benchmark driving cycles to one shown in Figure 1.
be capable of estimating the battery SOC with a relatively
high degree of accuracy.

I. INTRODUCTION
Electric Vehicles (EVs) and Hybrid Electric Vehicles
(HEVs) demand is growing due to the global concerns of the
traditional fuel supply and emissions [1, 2].In order to
successfully replace conventional internal combustion engine
vehicles with electric vehicles, the latter should be as reliable
and cost efficient as the former. At present, the main challenge
for the wide acceptance of electric vehicles lies mainly in their
initial cost due to using battery systems, which represent the
most expensive components. Manufacturers tend to over Figure 1: OCV-SOC-OCV relationship for both charging and discharging [1].
design the battery system in EVs to compensate for any
uncertainties in battery system modeling. Therefore improving This estimation method is also not accurate due to the flat
battery system modeling accuracy results in smaller pack size, region that exists in the mid SOC region (25% and 85%.)
which in turn directly lowers the system’s cost and especially for Lithium-Iron Phosphate (𝐿𝑖𝐹𝑒𝑃𝑂, ) chemistry.
consequently, the vehicles’ cost [3, 4]. Another issue in this method is its impracticality due to the
The Battery Management System (BMS) is responsible for requirement of measuring OCV which means SOC estimation
is not available while the battery is charging or discharging.

978-1-5090-3953-1/$31.00 ©2017 IEEE 342

 These reasons are the motive to build robust estimators to is the OCV-R-RC model. The OCV-R-RC circuit is as shown
accurately predict the SOC levels of EV batteries. In the in Figure 2.
literature, various battery models have been presented, such as
Equivalent Circuit-based Models, behavioral (empirical)
models and electrochemical models, [6]. These models are
either estimating SOC levels directly or they are coupled with
estimator algorithms such as Kalman Filter (KF).
In this paper, a battery model is built using Artificial Neural
Network. SOC estimation error is the main performance
criterion along with other criteria such as computational
efficiency, the required amount of training data, and
computational complexity.
The paper is organized as follows: Section I provides a Figure 2: OCV-R-RC Equivalent-circuit-based Model
background and overview of the paper. Section II goes over
In Figure 2, Voc is the open circuit voltage, R0 is the
the experimental battery used in this paper. This is then
series resistance representing the battery internal resistance, Rd
followed by the electric vehicle model and data generation in
and Cd are the resistance and the capacitance that represent the
Section III. Section IV involves a literature review of the
nonlinear behavior of the battery. In discrete form the model is
Neural Network applications in Battery systems in EVs and represented by the following equations:
followed by a brief Neural Network anatomy scientific basis 𝑖- 𝑘 + 1 = exp −
6
𝛥𝑡 𝑖7 6 𝑘 + 1 − exp −
6
𝛥𝑡 𝑖 𝑘 (2)
and types of Neural Network. Section V provides an overview 7 9
8 8 7 9 8 8

of ANNs as found in Matlab. Section VI outlines the

𝑣 𝑘 = 𝑂𝐶𝑉 𝑧 𝑘 − 𝑅6 𝑖7 6 𝑘 − 𝑅? 𝑖[𝑘] (3)
experiment, and the unique current profile used in order to
obtain our results. Lastly, section VII includes the conclusion BC
𝑧 𝑘+1 =𝑧 𝑘 − 𝜂 𝑘 𝑖[𝑘] (4)
and recommendations for future work. D

II. EXPERIMENTAL BATTERY Where:

• ∆𝑡 is the sampling time
Training of a neural network to represent a battery model
• 𝑖78 is the current flowing through the R1 branch
requires a large amount of experimental data, which is
attainable through extended battery test. Industrial battery • 𝑧[𝑘] is the state of charge of the battery
testers are usually used for the purpose of collecting this data. • 𝑂𝐶𝑉(𝑧 𝑘 )Is the open circuit voltage corresponding
Battery testers are able to charge/discharge the batteries at to the state of charge
different charge/discharge current rates. They are also able to • 𝑄 is the overall battery nominal capacity
use current profiles that represent driving cycles so as to • 𝜂 𝑘 is the Coulomb efficiency
witness battery performance under real world conditions. The
driving cycles are discussed more in details later. The open circuit voltage is dependent on the SOC level; this
relationship is dependent on the battery chemistry. In this
Collecting the required charge/discharge cycle data from a paper, this relationship is taken from a commercial Li-Ion
real battery would have been very time consuming, instead we battery, 3.6V 3.4Ah cell. This relationship, as shown in Figure
choose to generate battery approximated data using an 3. The R-RC model used the following parameters:
experimental battery. This is a common methodology for • R0 = 0.003 Ohms
generating experimental data to test different battery condition • R1 = 0.0029 Ohms
monitoring techniques. The experimental battery is a battery • C1 = 3760 Farad
approximated model that is able to generate a similar behavior
to actual batteries. In literature there are a number of battery
models, and they were developed mainly for battery SOC
estimation techniques. Examples of battery modelling
techniques are:

• Electrochemical Models
• Stochastic models
• Analytical models
• Electrical circuit models

Electrical circuit models are commonly used due to their

simplicity, fast implementation, and low running Figure 3: OCV-SOC Relationship of Experimental Battery
computational cost requirements. One of the simplest models

343
 III. ELECTRIC VEHICLE MODELING AND CURRENT Batte ry Pac k Curre nts fo r UDDS, US06, HWFET Cyc le s

Curre nt [Amps ]
GENERATION 150
Pack Current - UDDS
0
Common driving cycles are usually used in order to test the -150
performance of battery models. These include: Urban
2 4 6 8 10 12 14 16 18 20 22
Dynamometer Driving Schedule (UDDS), a light duty drive Time [Mins ]
cycle for high speed and high load (US06), and a Highway

Curre nt [Amps ]
250 Pack Current - US06
Fuel Economy Test (HWFET), [7]. Figure 4 displays the 0
velocity vs. time of the mentioned driving cycles. For this -250

experiment, a combination of these cycles was created called a 1 2 3 4 5 6 7 8 9

‘mixed driving cycle’. Time [Mins ]

Curre nt [Amps ]
150
Pack Current - HWFET
0
Ve lo c ity Pro file s fo r UDDS, US06, and HWFET Cyc le s
Ve locity [Kph]

100 -150
80 Velocity Profile - UDDS
60 2 4 6 8 10 12
40
20 Time [Mins ]
0
0 5 10 15 20 25
Time [Mins ] Figure 6: Pack Current Profiles for the UDDS , US06, and HWFET Cycles
Ve locity [Kph]

120
90 Velocity Profile - US06
60
30 IV. NEURAL NETWORKS
0
0 1 2 3 4 5 6 7 8 9 10 A. Neural Network Applications in Battery systems
Time [Mins ]
Artificial Neural Networks (ANN) are intelligent machine-
Ve locity [Kph]

100
80
60
Velocity Profile - HWFET learning tools that have been used in various applications such
40
20
as system modeling [9], classification [10] and control [11].
0
ANNs are adaptive, self-learning and can fit any non-linear
0 2 4 6 8 10 12 14
Time [Mins ] system. ANNs do not require any physical knowledge of the
modeled system which makes them a good choice for complex
Figure 4: Velocity profiles for the UDDS, US06, and HWFET Cycles, [7] systems. However, as a self-learning tool, ANNs need system
data to learn from. The amount of learning data required is
The UDDS cycle is used to simulate driving habits of an often dependent on the complexity of the system modeled.
average North American driver in the city. The US06 is used With the development of ANN, new battery estimation
to describe an aggressive, high acceleration driving habits, and techniques have been developed. For State of Charge (SOC)
HWFET is used to simulate highway driving conditions, [7]. estimation, different types of ANN techniques were studied
In order to generate the current profile 𝑖I from the velocity for example; in [12] B. Sun et al. divided the SOC working
profile from these driving cycles, a mid-size Battery Electric range into three parts and by using a back propagation neural
Vehicle (BEV) is modeled in MATLAB/Simulink/SimScape network (BPNN) for each part, they can estimate the SOC
as shown in Figure 5, with a sampling rate of 0.1 sec. The with acceptable precision for NiMH batteries.
electric vehicle model has been modified from a hybrid Wen-Yeau et al. [13] used Back-Propagation Neural
electric vehicle model presented in [8]. Networks (BPNN) and Radial Basis Function Neural
Networks (RBFNN) to estimate the SOC for LiFePO4
batteries.
In [14], Adaptive Wavelet Neural Network (AWNN) was
implemented for Lithium batteries SOC estimation, improved
convergence speed and accuracy over traditional techniques
were presented.
In [15], Elman Neural Network (ENN) demonstrated good
dynamic properties and more accurate results for SOC
Figure 5: All-electric mid-size sedan simulation model in Simscape
(Adopted from [8])
estimation compared to BPNN. Liu, Zhao and Huang [16]
used RBFNN on a Lead-Acid battery system taking time
The electric vehicle model consists of a lithium-Ion battery based characteristics into consideration by training the
pack, DC-DC convertor, vehicle dynamics model, vehicle network with time varying curve.
controller, and electric motor. The BEV driving range is Hybrid techniques were also investigated where ANN
approximately 200 kms per full charge. The cell-level current combined with other algorithms such as Extended Kalman
profile 𝑖I is obtained by scaling down the pack current profile Filters (EKF) [17] were studied to achieve effective dynamic
assuming cell balancing is maintained across the pack. The operation of the battery system. This study showed accurate
current profile generated form the model is as shown below in SOC estimations with 1% estimation error. In [18], Xu, Wang
Figure 6 and Cheng estimated the SOC using EKF and Stochastic

344
Fuzzy Neural Network (SFNN) on a NiMH battery to capture
the dynamics of the system instead of a regular ANN. The
technique demonstrated enhanced accuracy with maximum
estimation error of 0.6%.
Another study using Unscented Kalman Filters (UKF) with
ANN was done by W. He, N. Williard, C. Chen and M. Pecht
[19] to improve the estimation accuracy and reduce the
estimation variance for the Li-Ion battery system, their model
showed an RMS error of 2.5% and maximum error of 3.5% at Figure 8: Node (n+1, i) representation, [20]
different temperatures.
Where, bS` signifies the node offset (bias) for node(n, i),
B. Neural Network Scientific Basis x`S t indicates the output of node(n, j), w`,d
S
represents the
A multi-layer artificial neural network (ANN) link weight from node(n, j) to the node(n + 1, i). The
typically consists of an input layer, one or more hidden layers function φ(. ) is a nonlinear sigmoid activation function
and an output layer, [20]. A feed-forward ANN, as shown in defined by, [20]:
Figure 7, consists of numerous sensory units in which a 6
nonlinear activation function is applied. These nodes are 𝜑 𝑤 =
6Vg hij
𝑎 > 0 and − ∞ < 𝑤 < ∞ (6)
connected to adjacent layers by links (weights), [20]. In feed-
forward networks, a static mapping exists between the input For simplicity, the node bias is considered as a link weight
and output, in which the input signal spreads from the input by setting the last input 𝑁$ to 𝑛𝑜𝑑𝑒 𝑛 + 1, 𝑖 to the value of
layer through the hidden layers and then to the output in a one as follows:
forward direction and on a layer-by-layer fashion, [20]. 𝑥Z$[ 𝑡 = 1, 1≤𝑛≤𝑘 (7)
Let k indicates the total number of hidden, input and output
layers. Node(n, i) Denotes the iQR node in the nQR layer,
$
𝑤U,Z[
= 𝑏U$V6 , 1≤𝑛 ≤𝑘−1 (8)
and NS − 1 is the total number of nodes in the nQR layer. As Therefore, Equation (2) can be rewritten in the following
shown in Figure 8, node(n + 1, i) evaluates the following form:
Z[
Equation, [20]: 𝑥U$V6 𝑡 = 𝜑( Y]6 $
𝑤U,Y 𝑥Y$ 𝑡 ) (9)
C. Types of Neural Networks
Z[ \6 $ $
𝑥U$V6 𝑡 = 𝜑( Y]6 𝑤U,Y 𝑥Y 𝑡 + 𝑏U$V6 ) (5) One of the most common ways to categorize Neural
Networks is by the type of learning algorithm as this is one of
the main characteristics that decides the application and
usefulness of the Network for a specific task. For example, a
Neural Network’s ability to automatically learn and draw
conclusions from input data without human interference can
be very beneficial for data mining applications.
D. Supervised
Supervised learning is a type of machine learning algorithm
where the Neural Network is given a collection of input and
output data usually called a dataset. From this the Neural
Network tries to make a model that predicts the output values
in a way that agrees with it. And usually the larger the dataset
used in training, the more accurate the model is going to be.
Below are some examples of common supervised ANN types.
Some examples of supervised networks are Feedforward
neural network, Radial basis function (RBF) network,
Learning Vector Quantization (LVQ) & Dynamic neural
networks.
Figure 7: Schematic of feed-forward multilayer perceptron network, [20]:

Figure 9: Supervised learning block diagram [22]

345
1) Feedforward neural network
In this type of networks, the connection is one way from
input to output and does not form a cycle. It could consist of a
single or multilayers by increasing the number of hidden
layers. This type of network is thoroughly explained in
Section IV. B.
1) Radial basis function (RBF) network
A RBF Neural Network is a type that uses radial basis
function as an activation function. This Network training is
made in two steps the first step is within the hidden layer
where the value of the center vector is determined statistically
in an unsupervised manner. And secondly, based on the
distance from this center vector the weights of each neuron is
determined.
The Output function of this network is a scalar function of
the input. For example, for a single hidden layer; Figure 11: A reduced dataset using Principal component analysis.
𝑦 𝑥 = Z U]6 𝑤U 𝜑(
𝑥 − 𝑥U ) (10)
Where N is the number of neurons, w is the weight of each
neuron, 𝜑( 𝑥 − 𝑥U ) is a radial basis function from a center
point.
E. Unsupervised
Unsupervised learning is a type of machine learning
algorithm where the Neural Network tries to represent a
dataset which consists only of inputs in a particular pattern or
relation depending on a statistical structure of the given
dataset. Some of the common examples of unsupervised
Neural Networks are Competitive layers & Self-organizing
maps.

Figure 12: .Clustered data using self-organizing maps neural network

V. ARTIFICIAL NEURAL NETWORK. MATLAB

In this paper the commercial software, MathWorks
Matlab®, was used. Matlab provides a library of different
types of neural networks in which different methods were
applied and tested. This section provides a brief overview of
the tested types. The libraries for Neural Networks are based
on functional categorization and is broken down to:
• NN Fitting tools
• NN Pattern Recognition tools
• NN Clustering tools
Figure 10: Output of a competitive layers NN. Graph represents input
vectors versus weights. And it appears to be clustered • NN Time series tools
For battery SOC estimation, Pattern Recognition tools and
Clustering tools are out of the scope. Neural network fitting,
and Time series tools were tested with different inputs were
used as part of this paper. Following is a brief overview of
how they work:
A. Time Series Tools
Time series tools are used when the output is a time series
parameter; that is, its value depends on time history as well as
other inputs. In Matlab there are three main subcategories in
this category, (1) Nonlinear Autoregressive with External

346
input (NARX), (2) Nonlinear Autoregressive (NAR), and (3) C. Artificial Neural Network Notes
Nonlinear Input-Output. Their architectures are shown in In both of the used tools (Fitting tools, NARX), the user can
Figure 13. The second option is not applicable to batteries choose the back propagation training algorithm from a library
SOC estimation as the output is only dependent on its own or define a new custom training algorithm such as:
time history, however battery’s SOC level depends on many • Bayesian regularization backpropagation
parameters such as battery voltage. The third option is • BFGS quasi-Newton backpropagation
applicable to batteries SOC estimation; however it can be • Conjugate gradient backpropagation with Powell-Beale
equivalently deployed in NN Fitting tools with proper input restarts
handling. The first type (NARX) is very unique as the output • Conjugate gradient backpropagation with Fletcher-
depends on its own history and on other inputs and their Reeves updates
history as well. • Conjugate gradient backpropagation with Polak-Ribiére
updates
• Gradient descent with adaptive learning rate
backpropagation
• Gradient descent with momentum backpropagation
• Gradient descent with momentum and adaptive learning
rate backpropagation
• Levenberg-Marquardt backpropagation
• Resilient backpropagation
• Scaled conjugate gradient backpropagation
Each method has its advantages and disadvantages and
training algorithm selection is not an easy task. The selection
Figure 13: Time Series tools found in Matlab
is highly dependent on the fitting problem and trial and error is
required. According to Matlab documentation, Levenberg-
In Matlab NARX are trained in open loop fashion and then Marquardt (LM) backpropagation is one of the fastest,
can be tested in closed loop configuration. The open loop however it is a memory demanding algorithm. Each training
architecture is as shown in Figure 14. function has a different set of properties to define for training
the network. For example, the properties of LM as shown in
Table 1. Other neural network architectures available in
MATLab are
• Radial Basis Neural Networks
• Learning vector quantization(LVQ) networks

Figure 14: Open Loop NARX

After the model is well trained, it can be deployed in closed

loop configuration as shown in Figure 15. This type of neural
network was tested with using battery current and voltage as
external inputs and SOC level as an output.

Table 1: Properties of Levenberg-Marquardt backpropagation

However these networks were not tested in the scope of this
Figure 15: Closed Loop NARX
paper and it is recommended in future work section to test
B. Fitting Tools these different architectures.
The purpose of NN Feedforward Fitting tools in Matlab is
VI. THE EXPERIMENT
mainly to fit a nonlinear function between the input and the
output. The architecture could be customized, however a A. Data Generation & Network Training
common architecture uses 3 layers, input, hidden, and output Using the experimental model described above a set of
layers. This architecture is shown in Figure 16 training data was generated in order to train the neural
network. The type of data generated for training purposes is an
important factor when training the neural network to estimate
state of charge. The data generated for training should be able
to describe all the characteristics of the battery for the network
Figure 16: Fitting Tool Architecture to accurately estimate the state of charge.

347
 In order to accurately calculate the dynamics of a battery, • Max Epochs: 2000
the training profile needed to consist of four inputs: • Performance Goal: 1e-08
• Current at time [𝑘] • Minimum Gradient: 1e-08
• Voltage at time [𝑘] • Mu: 1e+10
• Current at time [𝑘 − 1] • Validity Checks: 50
• Voltage at time [𝑘 − 1]
The inclusion of previous time points is what allows the neural
network to learn the dynamics of the battery by allowing it to
better understand the relation of current and voltage drops to
state of charge. This is a very important inclusion as it
prevents the neural network from simply associating the
voltage to the state of charge.

Using the inputs above constant current profiles, charging

and discharging, were the first to be considered as they
correctly account for the entire battery state of charge range.
The current being supplied, voltage and state of charge at each
step were recorded. However, this was not sufficient in
teaching the neural network the dynamics of the battery’s state Figure 18: Mean Squared Error vs Epochs
of charge. While the neural network accurately estimated the
state of charge in areas of constant current, it failed to estimate B. Final Results
the state of charge correctly under very dynamic conditions. At 2000 epochs, the network was able to achieve a mean
squared error as low as 1.6148e-06 where SOC values vary
In order to better train the network to understand these from 0 to 1. As seen in Figure 18, the largest drops in error
conditions a new set of profiles were generated to include a were achieved at approximately 180, 460 and 1400 epochs.
normally distributed random current profile with a mean shift. Although, it is likely that a longer training time would result in
This was an important addition to the neural network as it now higher accuracy, the value gained from it would not justify the
allowed for the network to better learn how to perform under time taken.
dynamic conditions such as those found in drive cycle tests.
Figure 17, shows the current profile used for training the The final result was tested against a UDDS drive cycle
neural network. profile as well as a mixed drive cycle and the output state of
charge from the network was compared to the true state of
charge, obtained through Coulomb counting. The results, as
seen in Figure 19 and Figure 20, are accurate enough to
replace a traditional battery model in this situation. The state
of charge estimation by the neural network was able to
achieve a mean-squared error of approximately 3.853e-09
when tested on the UDDS cycle and 4.91e-08 on the mixed
driving cycle.

Figure 17: Constant & Random Current Profile

By training the neural network on this unique profile, it is
now able to better estimate the state of charge during high
current draws or dramatic voltage drops. This is what results
in the accurate state of charge estimation obtained in the
results.

With the profile displayed in Figure 17, the network trained

up to 2000 epochs as it had reached the max Epochs set by
ourselves and the error achieved was sufficiently low. The
network chosen for this purpose was FitNet using the Figure 19: State of Charge Estimation of a UDDS Cycle over time
Levenberg-Marquardt training function and the following
network properties:

348
 REFERENCES
[1] K. Parks, P. Denholm, and A. J. Markel, Costs and emissions
associated with plug-in hybrid electric vehicle charging in the Xcel
Energy Colorado service territory: National Renewable Energy
Laboratory Golden, CO, 2007.
[2] A. Nordelöf, M. Messagie, A.-M. Tillman, M. L. Söderman, and J.
Van Mierlo, "Environmental impacts of hybrid, plug-in hybrid, and
battery electric vehicles—what can we learn from life cycle
assessment?," The International Journal of Life Cycle Assessment,
vol. 19, pp. 1866-1890, 2014.
[3] N. Xue, W. Du, T. A. Greszler, W. Shyy, and J. R. Martins,
"Design of a lithium-ion battery pack for PHEV using a hybrid
optimization method," Applied Energy, vol. 115, pp. 591-602,
2014.
[4] A. M. Joshi, H. Ezzat, N. Bucknor, and M. Verbrugge,
"Optimizing battery sizing and vehicle lightweighting for an
extended range electric vehicle," SAE Technical Paper2011.
[5] J. Zhang and J. Lee, "A review on prognostics and health
monitoring of Li-ion battery," Journal of Power Sources, vol. 196,
pp. 6007-6014, 8/1/ 2011.
Figure 20: State of Charge Estimation of a Mixed Cycle over time [6] A. Seaman, T.-S. Dao, and J. McPhee, "A survey of mathematics-
based equivalent-circuit and electrochemical battery models for
VII. CONCLUSION hybrid and electric vehicle simulation," Journal of Power Sources,
vol. 256, pp. 410-423, 6/15/ 2014.
The artificial neural network outlined in this paper [7] U. EPA. Dynamometer Drive Schedules [Online]. Available:
provided accurate results compared to other forms of state http://www.epa.gov/nvfel/testing/dynamometer.htm
[8] S. Miller. Hybrid-Electric Vehicle Model in Simulink [Online].
of charge estimation such as battery models. The ANN was Available:
able to self-learn the battery dynamics, something which is http://www.mathworks.com/matlabcentral/fileexchange/28441-
very favorable. It will allow application on different battery hybrid-electric-vehicle-
chemistries without intensive engineering effort to create [9] C.-W. Chen, "Modeling and control for nonlinear structural
systems via a NN-based approach," Expert Systems with
new battery models. These accurate results were achieved Applications, vol. 36, pp. 4765-4772, 2009.
with a relatively simple Neural Network architecture. Once [10] G. A. Carpenter, S. Grossberg, and J. H. Reynolds, "ARTMAP:
the network is trained offline, it can be loaded on a neural Supervised real-time learning and classification of nonstationary
data by a self-organizing neural network," Neural networks, vol. 4,
network chip where the SOC can be estimated very fast pp. 565-588, 1991.
using the four easy to measure inputs. The performance of [11] K. S. Narendra and K. Parthasarathy, "Identification and control of
the ANN was tested using standard driving cycles and it dynamical systems using neural networks," Neural Networks,
IEEE Transactions on, vol. 1, pp. 4-27, 1990.
achieved a mean squared error of 3.853e-09 on UDDS [12] B. Sun and L. Wang, "The SOC estimation of NIMH battery pack
driving cycle, which allows it to compete with the for HEV based on BP neural network," in Intelligent Systems and
traditional SOC estimation techniques. Moreover the inputs Applications, 2009. ISA 2009. International Workshop on, 2009,
pp. 1-4.
of the ANN do not include the previous SOC level which [13] W.-Y. Chang, "State of charge estimation for LiFePO4 battery
makes the SOC estimation more robust as the system will using artificial neural network," INTERNATIONAL REVIEW OF
not drift with time which is the case in Coulomb counting. ELECTRICAL ENGINEERING-IREE, vol. 7, pp. 5874-5880, 2012.
[14] F. Zhou, L. Wang, H. Lin, and Z. Lv, "High accuracy state-of-
charge online estimation of EV/HEV lithium batteries based on
VIII. FUTURE WORK
Adaptive Wavelet Neural Network," in ECCE Asia Downunder
This paper showcases a proof of concept that an artificial (ECCE Asia), 2013 IEEE, 2013, pp. 513-517.
[15] S. Qingsheng, Z. Chenghui, C. Naxin, and Z. Xiaoping, "Battery
neural network is capable of performing accurate state of state-of-charge estimation in electric vehicle using elman neural
charge estimation. However, to further the use of artificial network method," in Control Conference (CCC), 2010 29th
neural networks in this field we aim to testing this concept Chinese, 2010, pp. 5999-6003.
[16] L. Yanwei, Z. Kegang, H. Xiangdong, and P. Feng, "A new
on a commercial battery currently on the market.
method based on RBFNN in SOC estimation of HEV battery," in
Control Conference (CCC), 2010 29th Chinese, 2010, pp. 4923-
4927.
[17] Z. Chen, S. Qiu, M. A. Masrur, and Y. L. Murphey, "Battery state
of charge estimation based on a combined model of extended
Kalman filter and neural networks," in Neural Networks (IJCNN),
The 2011 International Joint Conference on, 2011, pp. 2156-2163.
[18] L. Xu, J. Wang, and Q. Chen, "Kalman filtering state of charge
estimation for battery management system based on a stochastic
fuzzy neural network battery model," Energy Conversion and
Management, vol. 53, pp. 33-39, 2012.
[19] W. He, N. Williard, C. Chen, and M. Pecht, "State of charge
estimation for Li-ion batteries using neural network modeling and
unscented Kalman filter-based error cancellation," International
Journal of Electrical Power & Energy Systems, vol. 62, pp. 783-
791, 2014.
[20] S. Haykin, Kalman filtering and neural networks vol. 47: John
Wiley & Sons, 2004.

349