Journal of Power Sources 400 (2018) 242–255

Contents lists available at ScienceDirect

Journal of Power Sources

journal homepage: www.elsevier.com/locate/jpowsour

State-of-charge estimation of Li-ion batteries using deep neural networks: A T

machine learning approach
Ephrem Chemalia,∗, Phillip J. Kollmeyera, Matthias Preindlb, Ali Emadia
a
Department of Electrical and Computer Engineering, McMaster Institute for Automotive Research and Technology, McMaster University, Hamilton, ON, Canada
b
Department of Electrical Engineering, Columbia University in the City of New York, New York, NY, USA

H I GH L IG H T S

• Deep neural network used to map battery signals directly to SOC.

• Deep neural network self-learns network weights.
• Neural network SOC estimator is shown to be computationally efficient.
• Increased SOC estimation accuracy and robustness by adding noise to training data.
• One deep neural network learns to estimate SOC over many ambient temperatures.

A R T I C LE I N FO A B S T R A C T

Keywords: Accurate State of Charge (SOC) estimation is crucial to ensure the safe and reliable operation of Li-ion batteries,
Battery management systems which are increasingly being used in Electric Vehicles (EV), grid-tied load-leveling applications as well as
Deep neural networks manned and unmanned aerial vehicles to name a few applications. In this paper, a novel approach using Deep
Energy storage system Feedforward Neural Networks (DNN) is used for battery SOC estimation where battery measurements are di-
Li-ion batteries
rectly mapped to SOC. Training data is generated in the lab by applying drive cycle loads at various ambient
Machine learning
State of charge estimation
temperatures to a Li-ion battery so that the battery is exposed to variable dynamics. The DNN's ability to encode
the dependencies in time into the network weights and in the process provide accurate estimates of SOC is
presented. Moreover, data recorded at ambient temperatures lying between −20 °C and 25 °C are fed into the
DNN during training. Once trained, this single DNN is able to estimate SOC at various ambient temperature
conditions. The DNN is validated over many different datasets and achieves a Mean Absolute Error (MAE) of
1.10% over a 25 °C dataset as well as an MAE of 2.17% over a −20 °C dataset.

1. Introduction transportation sector. Furthermore, half of the overall health-related

economic cost, estimated to be $865 billion in 2010, is credited to air
Li-ion batteries are not only heavily used in most portable electro- pollution [4]. Nowadays, some countries are taking proportionate ac-
nics and Electric Vehicles (EV) but are also used in smart-grid tech- tion to counteract these negative effects by banning new petrol and
nology for load levelling as well as in newer technologies like diesel powered vehicles by 2030 or as early as 2025, in the case of
Unmanned Aerial Vehicles and passenger drones aimed for medium to Norway. Due to the advantages of the Li-ion batteries outlined above,
short range distances [1]. This can be attributed to many advantages electrified vehicles powered by Li-ion batteries are currently one of the
that Li-ion batteries offer over other batteries. These include a high best ways to mitigate these issues.
specific energy and energy density which allows electrified vehicles A reliable state of charge estimation is required to ensure an accu-
longer electric-only driving range, high cycle life, high Coulombic ef- rate gauge of a vehicle's remaining driving range as well as proper
ficiency (up to 98%) and low self-discharge [2,3]. In 2015, 50% of all balancing of the battery pack [3,5,6]. Due to unpredictable driving
nitrogen oxide air pollutants in the world, corresponding to 53 million habits and the repeated acceleration and deceleration of a vehicle, the
tonnes of airborne nitrogen oxide emissions, can be attributed to the battery can be exposed to highly dynamic load demands. As a result of

∗
Corresponding author. Department of Electrical and Computer Engineering, McMaster Institute for Automotive Research and Technology, McMaster University,
Hamilton, ON, Canada.
E-mail address: [email protected] (E. Chemali).

https://doi.org/10.1016/j.jpowsour.2018.06.104
Received 16 December 2017; Received in revised form 21 June 2018; Accepted 30 June 2018
Available online 16 August 2018
0378-7753/ © 2018 Published by Elsevier B.V.
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

these dynamic load demands, SOC estimation is a tedious task. SOC is Advancements in modern machine learning techniques are accel-
not an observable quantity, therefore its accurate estimation becomes erating faster than ever before due to constantly improving computing
essential for reliable and safe operation of the vehicle [5,7]. power and increased access to vast pools of data. Nowadays, machine
SOC is defined as the remaining charge within the battery and is defined learning algorithms have become deeply entrenched in our lives. They
as the ratio of the residual capacity of the battery to its nominal capacity are now the dominant algorithms used for object recognition in images
[3]. The relationship between the battery's observable signals to the esti- and video sequences, natural language processing on smartphones and
mated SOC is a highly non-linear one, varying with temperature and dis- predictive analytics in many industries, to name a few [26].
charge/charge currents [8,9]. Traditionally, the two main estimation This work showcases how a machine learning technique like
methods have used open circuit voltage based techniques and coulomb Feedforward Neural Networks (FNN) as well as Deep Feedforward
counting [7,10]. These methods are known to have their limitations and Neural Networks (DNN), can accurately estimate SOC without the help
have been generally displaced by more sophisticated methods. They include of Kalman filters or any other inference methods. Specifically, this work
Luenberger observer [7,11], adaptive observer [7,12], sliding mode ob- contributes the following novelties. (1)A DNN can map observable
server [7,13,14], and Kalman filters [15–17]. Typically, in observer signals from the battery like voltage, current and temperature directly
methods, the parameters of an equivalent circuit model like resistances and to the battery SOC, avoiding additional filters and estimation algo-
open circuit voltage are fit to observed battery current and voltage data. An rithms like Kalman filters used in traditional systems. (2)The DNN can
estimate is issued by mapping these parameters to SOC. In Kalman filter- self-learn its own weights by using learning algorithms like gradient
based algorithms, it is typically required to linearize around an operating descent. This is markedly different than incumbent techniques like
point which can significantly increase computational load. The measured lumped parameter models, equivalent circuit or electrochemical models
current, voltage and the previously estimated SOC are provided to the al- which require a great deal of time to hand-engineer and parameterize.
gorithm and the filter issues an estimate of SOC at the next time step. These (3)It will be shown that one DNN can learn to estimate SOC at different
techniques are often tied to some battery model, like a lumped parameter ambient temperature conditions. This is beneficial since traditional
model or an equivalent circuit model which require arduous model iden- estimation techniques must use different models or different look-up
tification to adequately represent the non-linear behavior of a battery. In tables for estimation at different ambient temperatures.
addition, they often require large numbers of parameters or different ver- After a brief introduction, the second section will discuss the deep
sions of the model to perform SOC estimation at varying ambient condi- neural networks constructed in this work. In the third section, the ex-
tions. perimental apparatus for the battery testing and data logging is de-
Strategies involving classic machine learning algorithms have also been scribed. In the fourth section, the performance of the DNN is tested with
used in the past. The benefit of these types of techniques is that they can be many validation datasets recorded at constant and at varying ambient
trained with real world data and self-learn SOC estimation without the need temperatures.
for hand-engineered models. However, when neural network were solely
used, the results were typically not accurate enough, and therefore required 2. Deep neural networks for SOC estimation
the additional use of Kalman filters or other inference mechanisms to
achieve sufficient estimation accuracy. Although some works have used There are many examples where deep learning architectures have
Kalman filters in conjunction with combined battery models or equivalent made significant improvements over conventional algorithms. In 2012,
circuit battery models [18], many other works have also used them in AlexNet, a deep convolutional neural network won the ImageNet
conjunction with NN battery models. In Ref. [19], a trained 2-layer Neural competition where teams are tasked with classifying over 1 million high
Network (NN) with 30 neurons in the hidden layer estimates terminal resolution images in 1000 different categories. AlexNet achieved a top-
voltage within a 4% Root Mean Square (RMS) error. However, to estimate 5 error rate of 15.3% compared to a more traditional model taking
SOC and to further reduce the RMS error to 2%, the NN is used as a battery second place with a top-5 error of 26.2% [27]. Recently, Microsoft
model in an Extended Kalman Filter (EKF). In Ref. [20], an Extreme Research's deep learning algorithm, called a deep residual network,
Learning Machine is used at a constant ambient temperature of 25 °C. An won the 2015 ImageNet challenge with an error rate of 3.57% which
SOC estimation error of under 1.5% is claimed however this is only even surpasses human level accuracy valued at 5.1% [28].
achieved in conjunction with a Kalman filter as well. Furthermore, the ex- Traditional machine learning techniques contain no more than one
treme learning machine is trained on constant discharge pulses hence their or two layers of non-linear and linear transformations [29]. With the
performance in transient load demand, experienced in real world scenarios, advent of faster computational power and an abundance of available
is unknown. In Ref. [21], a SVM is used with a moving window to increase real world data, deeper architectures were investigated which, in many
computational efficiency when modeling the battery; a Mean Absolute Error cases, allowed researchers to make striking improvements in many
(MAE) of less than 2% is achieved. However, as is the case for the above applications [27,30–33].
works, it achieves this MAE in conjunction with an EKF. In Ref. [22], a load Feedforward neural networks, whose 2-layer and multi-layer DNN ar-
classifying neural network is trained on 12 US06 drive cycles however chitectures are shown in Fig. 1, can, in principle, model any non-linear
different neural networks are used for idling, charging and discharging system by mapping the observables to a desired output. Once trained off-
operation. The method achieves an average estimation error of 3.8% or line, FNN and DNN offer fast computational speeds online since they are
2.6% when additional filtering is performed. Furthermore, validation is composed of a series of matrix multiplications, as opposed to other strate-
performed on pulse discharge tests hence the method's performance in real gies which can contain computationally intensive calculations like partial
world applications is unknown. differentials equations. When FNN and DNN are applied for SOC estimation,
More recently, additional works have utilized model-based and a typical dataset that is used to train the networks is defined by D =
machine learning-based approaches for battery SOC estimation. One {(ψ (1), SOC (1)*), (ψ (2), SOC (2)*), …, (ψ (τ ), SOC (τ )*)} , where SOC (t )*
such approach uses a moving average estimation with a reduced elec- and ψ (t ) are the ideal state-of-charge value and the vector of inputs at time
trochemical model which is able to perform estimation without line- step t, respectively. The current measurement used to determine the ideal
arization error and allows for constraints on states like the internal SOC (t )* is described in more detail in the next section of this paper. The
resistance state and Li-ion concentration [23]. In Ref. [24], a fuzzy C- vector of inputs is defined as ψ (t ) = [V (t ), T (t ), Iavg (t ), Vavg (t )] where
means and subtractive clustering method is used along with a SVM for V (t ) , T (t ) , Iavg (t ) and Vavg (t ) represent the voltage, temperature, average
SOC estimation. The work performed in Ref. [25] builds on the latter current and average voltage of the battery at time step t. The average cur-
fuzzy-SVM approach by using a genetic algorithm-based fuzzy C-means rent and voltage are both calculated over ξ precedent time steps, which
clustering technique with a backpropagation algorithm to estimate SOC ranged from 50 to 400 time steps. This is not to be confused with the total
and is claimed to outperform classical fuzzy modeling techniques. dataset time span defined by τ, where ξ < τ . Many different types of inputs

243
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 1. Architecture of 2-layer neural network (top) and architecture of Deep Neural Network (DNN) (bottom). The input data is given by
ψ (t ) = [V (t ), T (t ), Iavg (t ), Vavg (t )] where V (t ) , T (t ) , Iavg (t ) and Vavg (t ) represent the voltage, temperature, average current and average voltage of the battery at time
step t. The output of the DNN is the estimated SOC at every time step.

were examined and it was found that Iavg , Vavg are better values to feed into take advantage of parallel computing and to accelerate the training
the network rather than feeding many antecedent values of current and process, when performing real time estimation after training is com-
voltage. The number of weights in the first layer of the network increases plete, the algorithms discussed in his paper can be flashed on a standard
proportionally to the number of additional inputs. For example, in a net- microprocessor. In this work, typical training times required to train a
work which has 10 neurons in the first hidden layer, each additional input DNN can be anywhere from a few hours to 40–50 hours, depending on
will require 10 additional weights for a fully-connected layer. Therefore, whether the DNN is trained over a single ambient temperature or over 5
better computational efficiency and lower memory requirements are different ambient temperatures.
achieved by capturing the system dynamics in time through an averaging of Feedforward neural networks are matrix-based and can be re-
current and voltage. presented by the below composite function. A few important variables
In this paper, TensorFlow [34], a machine learning framework, is are first defined. Let w jl, k denote the weight connection between neuron
used in conjunction with two NVIDIA Graphical Processing Units j in layer l − 1 and neuron k in layer l. Let bkl and hkl be the bias and the
(GPU); a TITAN X and a GeFORCE GTX 1080 TI GPU. The TensorFlow activation function, respectively, of neuron k in layer l. The activation
framework provides the ability to quickly prototype and test different function is computed as follows;
network architectures as a result of its ability to automatically compute
backpropagation. The latter describes the process by which the network
⎛ ⎞
weights can be updated at the end of every training epoch and will be hkl (t ) = η ⎜∑ (w jl, k hkl− 1 (t ) + bkl ) ⎟
further discussed below. Although training is performed on a GPU to ⎝ k ⎠ (1)

244
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

average error as well as the maximum error value and does not prior-
itize the minimization of one value over another. A forward pass starts
when the training data is fed into the network and is complete when the
SOC estimate for each time step is calculated as well as when the overall
loss function is determined. A full training epoch, ∈, includes one for-
ward pass and one backward pass; describing the process of sending the
overall loss signal backward through the network to update the
weights. To perform this backward pass, an optimization method called
Adam [35] is used, which updates the network weights and biases
based on the gradient of the loss function. This is defined by the fol-
lowing composite function;
mε = β1 mε − 1 ∇ (wε − 1)

Fig. 2. Experimentally verified relationships between computational time and

rε = β2 rε − 1 ∇ (wε − 1)2
∼ = m /(1 − β ε )
m
number of neurons as well as number of layers. (right) Semi-log plot of com- ε ε 1
putational time vs. the number of neurons. The number of layers in the DNN is r͠ ε = rε /(1 − β2ε )
fixed to 4 layers. (left) Computational time as a function of the number of ∼
m ε
layers. The number of neurons in each layer of the DNN is fixed to 50 neurons. wε = wε − 1 − α r͠ , (6)
ε−κ

where β1 and β2 are decay rates set to 0.9 and 0.999, respectively,
Table 1 α = 10−4 is the learning rate and κ is a constant term set to 10−8 . The
Computational time of a DNN and of an extended Kalman filter. network parameters at the current training epoch are denoted by wε .
SOC Estimation Technique Computational Speed (seconds) The network weights as well as the biases are initialized with a nor-
mally distributed random number generator having mean 0 and stan-
DNN (2 layers, 256 neurons) 0.07
dard deviation of 0.05. It is only during training that a forward and
Combined Model + EKF 0.66
backward pass are performed to iteratively update the network weights
Results obtained in MATLAB using Intel Xeon E5630, 2.53 GHz and 32 Gb of until a convergence criteria is met. With the backward pass, the net-
RAM. work self-learns its weights and biases, offering significant improve-
ments over traditional SOC estimation strategies which require the
Table 2 time-consuming construction and parameterization of hand-engineered
Tested drive cycles and corresponding battery cell power characteristics. battery models.
Test Use Mean RMS Peak
Training of the networks is performed offline and only when the
convergence critera is met and the overall loss function is minimized
Power(W) Power(W) Power(W) can the networks be applied online. During validation, which simulates
online operation, we need only perform a forward pass to calculate the
Cycle 1 Training 3 7 35
Cycle 2 Training 3 7 35
estimated SOC values at each time step and no backward pass is re-
Cycle 3 Training 3 7 35 quired since the network weights and biases have already been learned
Cycle 4 Training 3 7 35 offline during training. FNNs and DNNs offer an advantage of lower
US06 Validation 6 13 34 computational overhead, once trained, since a forward pass is com-
HWFET Validation 5 6 19
prised mainly of consecutive matrix multiplications. This, in general,
UDDS Training 2 4 19
LA92 Training 2 7 35 has higher computational efficiency than other algorithms, which
NN Training 3 10 39 might contain differential equations. This point is further quantified
below through a comparison between the computational time of a DNN
and that of an EKF model. In addition, a DNN, as will be shown in the
where, results section of this paper, has the ability to encode the inherent re-
sponse of a battery at various ambient temperatures thus reducing the
hkl (t ) = SOC (t ) for l = L (2)
memory required to store different parameters for different ambient
SOC (t ) is the estimated state-of-charge at time step t calculated by the temperatures as is typically done in traditional battery models. This
network. Due to its simplicity during training as well as during vali- offers great benefits when performing modelling and state estimation
dation, the nonlinearity used in these networks is called Rectified for large quantities of cells found in vehicle battery packs.
Linear Units (ReLU) and is given by; To evaluate the SOC estimation performance of our networks, a few
different performance metrics are used. These include the Mean
η = max (0, h) (3)
Absolute Error (MAE), the standard deviation of the errors (STDDEV)
To understand how accurate the SOC estimate is compared to the ideal and finally the maximum error (MAX).
SOC value, an error signal is generated at the output of the network for To gain a better understanding of the computational overhead dis-
each time step and is given by cussed above, an analytical representation of the computational effi-
ciency of these DNNs is given. Since most of the computational cost is
e (t ) = SOC (t ) − SOC * (t ) (4) due to floating-point operations, then it follows that the computational
To gain a good understanding of the overall loss of the network, the time of a forward pass is proportional to the number of multiplication
following loss function is computed at the end of a forward pass; and addition operations. Referring to Fig. 1, for each neuron of the
DNN, there exists d multiplications and d − 1 additions resulting, ide-
τ
1 ally, in 2d − 1 operations per neuron. Let N be the number of neurons in
L = [max (e )]2 + ∑ (e (t ))2
τ t=0 (5) each layer, τ be the total number of time steps in the sequential dataset
and L be the number of layers in the DNN, then the number of floating-
where τ is the length of the sequence, as mentioned above. Both the point operations is given by;
average error and the maximum error are included in the overall error
function so that the network expends its energy on minimizing both the F (τ , L, N ) = τ [N 2 (2L − 4) + N (2d − L + 3) − 1] (7)

245
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 3. (a) US06 drive cycle applied on the same battery cell at an ambient temperature of 25 °C (orange) and −20 °C (blue). Recorded voltage (top right), current
(top left), capacity (bottom right) and cell temperature (bottom left). (b) Cell temperature vs. Ah count for the different drive cycles applied on the same battery cell
recorded at 5 different ambient temperatures. (c) Schematic of the battery test bench and data logging system. (For interpretation of the references to colour in this
figure legend, the reader is referred to the Web version of this article.)

246
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Table 3 characterize vehicle fuel consumption [37]. A power profile is gener-

Panasonic 18650 PF cell parameters. ated based on a large light duty passenger vehicle with a 35 kWh
Nominal Open Circuit Voltage 3.6 V battery pack for each of these drive cycles. These are then scaled for a
single battery cell in order to apply these drive cycles in the
Capacity Min. 2.75 Ah/Typ. 2.9 Ah laboratory.
Min/Max Voltage 2.5 V/4.2 V
Four standard drive cycles are used in addition to a custom drive
Mass/Energy Storage 48 g/9.9 W h
Minimum Charging Temperature 10 °C
cycle which was built for this work. The standard drive cycles used are
Cycles to 80% Capacity 500 (100% DOD, 25 °C) the Urban Dynamometer Driving Schedule (UDDS), the Highway Fuel
Economy Driving Schedule (HWFET), the Unified Driving Schedule
(LA92) and the Supplemental Federal Test Procedures or US06 [37].
Table 4 The custom drive cycle, called the NN, was designed for this work and
Test equipment specifications. has additional dynamics useful for training the neural network. As an
Cycler Manufacturer Digatron Firing Circuits example, the voltage, current, capacity and cell temperature recorded
from the same battery cell throughout the US06 drive cycle are shown
Test Channel Used 25 A, 0–18 V channel in Fig. 3a). Fig. 3b) shows the cell temperature as a function of the amp-
Voltage/Current Accuracy ± 0.1% Full Scale
hour count for some of the different drive cycles applied to the same
Data Acquisition Rate Used 10 Hz
Thermal Chamber Cincinatti Sub Zero ZP-8
battery cell at the five different ambient temperatures. In compliance
Size 8 cu. Ft. with the battery manufacturer's recommendations, regenerative
Accuracy ± 0.5 °C braking below 10 °C is not performed which is the reason for the lack of
charging current at −20 °C in Fig. 3a). The other four cycles, cycles 1
through 4, are composed of a mixture of the four standard drive cycles
Therefore, when letting τ and L remain constant, it is immediately performed in a random, repeating order.
clear that the number of floating-point operations grows as the square The DNNs are validated using the US06 cycle, an aggressive and
of the number of neurons or in Big-O notation as O (N 2) . On the other highly dynamic profile, and the HWFET cycle, a profile with moderate
hand, if τ and N remain constant, and the number of layers is allowed to dynamics. The data for the remaining seven tests are used to train the
vary, then it follows that the number of floating-point operations grows neural network, as is indicated in Table 2.
linearly with the number of layers, L, or in Big-O notation as O (L) .
Hence, it is typically preferred that networks grow in depth rather than
in width to reduce computational time. This relationship is tested by 3.2. Test procedure, battery cell, and cycling equipment
timing the forward pass for varying numbers of neurons as well as for
varying numbers of layers and are shown in Fig. 2. The test procedure used is as follows: (1) set the thermal chamber
Although a good understanding of a DNN's computational com- temperature to 25 °C, (2) wait 10 min or until the measured battery
plexity is important, it is also essential to compare the computational temperature is greater than 10 °C, (3) perform a full charge, (4) set the
efficiency of the DNN with respect to other competing algorithms. To chamber to the current test temperature, (5) wait for the battery tem-
this end, the computational time of the DNN is compared to that of an perature to stabilize, and (6) perform the drive cycle. The drive cycle is
EKF approach used in Ref. [18]. The goal of this test was to measure the repeated until the battery voltage reaches 2.5 V for the 25 and 10 °C
time required by these two techniques in order to compute through tests. For the lower temperature tests, the battery resistance is much
13000 data points. The results are shown in Table 1 and are based on an higher, causing the terminal voltage to hit 2.5 V periodically
average of 20 separate executions of the two algorithms over the same throughout the test. To accommodate the higher resistance, an amp-
dataset. Therefore, on average, the DNN is almost an order of magni- hour cut off 2.32Ah at 0 °C, 2.03Ah at −10 °C, and 1.74Ah at −20 °C is
tude faster than the EKF. This can be attributed to the fact that the EKF used instead. The battery tester still has a lower voltage limit of 2.5 V,
is required to linearize the non-linear system at every time step while but the test is not terminated when this voltage is hit. The tester reduces
the DNN mainly performs a series of matrix multiplications to achieve current to prevent voltage from falling below the limit. The battery may
the desired output which is an inherently much simpler series of op- hit the lower voltage limit several times when at higher currents and
erations. The DNN outlined in Table 1 requires less than 3000 floating low SOC. The nine drive cycle tests were performed in the order given
point operations per second (FLOPs), including the data preprocessing in Table 2, first at 25 °C, and then at 10, 0, −10, and −20 °C. Following
phase, which is within the realm of many conventional embedded those tests, the nine drive cycles were also performed at a variable
processors able to handle up to 109 FLOPs. ambient temperature where the temperature was increased from 10 °C
to about 25 °C. These varied tests are used to evaluate the SOC esti-
3. Experimental data collection for training and validation of mation algorithm for cases when the battery is warming up throughout
neural network SOC algorithm the drive.
The tested battery cell is a Panasonic NCR18650PF, nickel cobalt
3.1. Drive cycles aluminum (NCA) chemistry Li-ion battery, with a nominal capacity of
2.9Ah and other specifications as described in Table 3 [38,39]. The
A series of vehicle drive cycle and charging profiles were applied on battery cell is tested with a 25 amp, 0–18 V rated channel of a Digatron
a battery cell, and the resulting measured data [36] is used to train and Firing Circuits Universal Battery Tester, and placed in a Cincinnati Sub
validate the neural network SOC algorithm. The nine different drive Zero thermal chamber, as is described in Table 4. The battery tester is
cycles used are given in Table 2. The cycles were chosen to cover a very accurate; rated for less than 0.1% error, which is important be-
range of mean, RMS, and peak power values for the battery, as would cause the measured battery amp-hours is used to calculate the reference
be experienced in a vehicle application, and to provide enough data to or ideal SOC value. The maximum error of 0.1% translates to a 25 mA
train the neural network. The tests were also performed at a range of current error, which would result in a maximum of 25 mAh of error for
temperatures, down to −20 °C where nonlinear resistance, battery a 1 h drive cycle test, a value small enough to not have a major effect on
self-heating, and diffusion effects make SOC estimation more tedious. the results. The battery system and instrumentation is illustrated in
Procedurally, the dynamometer drive schedules obtained from the Fig. 3c), which shows that the voltage sensing leads are connected di-
United States Environmental Protection Agency (EPA) are provided to rectly to the battery terminals and that a thermocouple is connected to
an electric vehicle model in the form of a speed profile and is used to the center of the case.

247
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 4. (a) DNN estimation accuracy, estimation error over the HWFET discharge validation dataset recorded at 25 °C and the mean absolute error as a function of
training epochs. (b) Estimation accuracy and estimation error over validation charging profile also recorded at 25 °C. DNN is composed of 3 layers and 4 neurons in
each layer.

3.3. Training data augmentation augmented by injecting Gaussian noise into the battery measurement
signals. Specifically, a normally distributed random number generator
To robustify the deep neural network, the training data is is used to generate noise for each of the battery's signals. We specify the

248
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Table 5 recommended charge rate and selecting an even higher rate is excessive
SOC estimation accuracy of deep neural network trained on fixed ambient and potentially dangerous.
temperature data of 25 °C during validation. Good performance is observed both in slow and fast dynamics as
Validation Dataset MAE (%) STDDEV (%) MAX (%) well as in the charging validation dataset, as can be seen in Fig. 4 and in
Table 5, where an MAE of 1.35%, 1.85% and 0.39% is achieved over
HWFET Dataset 1.35 0.94 3.80 the HWFET, US06 and charging datasets, respectively. The DNN used to
US06 Dataset 1.85 1.20 5.14
obtain these results is composed of 3 layers where the first two layers
Charging Dataset 0.39 0.26 1.21
each contains 4 neurons and the last layer contains 1 neuron. The
network is trained for up to 85000 epochs which culminates to 5 h of
mean and variance of this normally distributed random number gen- training time on the GPU. The MAE calculated over the HWFET vali-
erator based on typical noise levels observed in off-the-shelf measure- dation dataset is plotted as a function of the training epochs and is also
ment devices. In particular, Gaussian noise with 0 mean and a standard shown in Fig. 4a.
deviation of 2–4% is injected into the voltage, current and temperature Different tests are performed to examine effects of various structural
measurements. In addition, to robustify against offsets and gains in- aspects of the DNN on SOC estimation accuracy. The first of such tests
herent in battery measurement devices, an offset is applied to all evaluates the influence of the ξ parameter, which is used when com-
measurement signals and a gain is applied to the current measurement. puting the average voltage and current of the input data, on estimation
Specifically, an offset of up to ± 150 mA and a gain of up to ± 3% is accuracy. In effect, this would examine how exposure to increased
applied to the current measurements, an offset of up to ± 5 mV is ap- amounts of historical data could increase estimation accuracy of the
plied to the voltage measurement and an offset of up to ± 5 °C is applied DNN.
to the temperature measurement. New versions of the initial recorded A DNN is trained on input data with ξ = 100 and another is trained
training data is created with different noise, offset and gain levels. We on input data with ξ = 400. The results from these two networks are
create up to 20 new versions of the training data. shown in Fig. 5a) and summarized in Table 6. Exposing the DNN to 400
All the results showcased in this work are based on DNNs trained on historical data points as opposed to 100 offers good performance gains
data obtained from Panasonic 18650 PF cells. Nevertheless, the DNNs with a reduction in the MAE and MAX of approximately 30% and
discussed in this work can be trained on any other type of battery cell. 30–40%, respectively. To ensure an unbiased comparison, training is
The network architectures will remain unchanged if used to perform stopped at 160,000 epochs.
SOC estimation on another battery however, the network might need to The battery measurements performed in the lab are obtained in
be retrained or better yet transfer learning can be used to simply retrain isolated and controlled environments to ensure good quality data.
the last one or last few layers of the network. This would be less time- However, in the real world, battery measurements, like voltage, current
consuming than re-parameterizing a model for an alternative battery and temperature measurements, can be of variable quality. A DNN can
cell as would be typically performed for traditional models. be taught to handle such noisy environments at training time. As a
result, a comparison showing the estimation performance achieved on
the augmented training dataset versus the unaugmented training da-
4. State-of-charge estimation results taset is shown in Table 7 and in Fig. 5b). From the latter table and
figure, it becomes evident that using the augmented dataset for training
As mentioned earlier, the vector of inputs fed into the DNNs is given achieves a significant reduction in MAE and MAX of 21–41% and
by ψ (t ) = [V (t ), T (t ), Iavg (t ), Vavg (t )] where V (t ) , T (t ) , Iavg (t ) and 32–43%, respectively. By injecting Gaussian noise, offsets and gains on
Vavg (t ) represent the voltage, temperature, average current and average
the measurement devices, not only is the accuracy of the DNN improved
voltage of the battery at time step t. The output of the DNN is the es-
but the network is robustified for real world application.
timated SOC at t. The drive cycles used for training and validation are
Another important test is to verify the DNN's performance when it is
recorded at a sampling frequency of 1 Hz and are between 4000 and
initialized incorrectly. To that end, instead of providing the correct
13000 s long. The following two subsections examine the SOC estima-
measured initial battery voltage of 4.2 V to the network, an incorrect
tion accuracy of the DNN at fixed ambient temperature and, thereafter,
initial value of 3.6 V is given. This is shown in Fig. 5c), where the DNN
at variable ambient temperatures.
struggles to output a good estimate of SOC at the beginning of the
dataset, where the error is over 50%, but quickly converges to good
4.1. SOC estimation at constant ambient temperature SOC estimation within the first 10 s of the dataset.
A large fraction of the performance results discussed above are
In this section, the DNN is trained on up to seven complete dis- obtained on DNNs with 3 layers; where the first two layers contain 4
charge datasets which are augmented as described in subsection 3.3. To neurons and the last fully-connected layer contains 1 neuron.
verify the DNN's performance in both fast and slow dynamics, valida- Interestingly, this corresponds to 36 network weights which is rela-
tion is conducted on the US06 and HWFET datasets, respectively. These tively small in size and can be flashed on any embedded processor when
latter datasets as well as all other datasets used to obtain performance operating in real time. To determine an optimal network architecture,
results are only used during validation and never during training. Re- the estimation accuracy is evaluated on networks with different num-
generative breaking is also allowed for the discharge datasets recorded bers of layers and neurons. This is shown in Fig. 6a) and in Fig. 6b),
at the higher temperatures however not at lower ambient temperatures respectively. For each test case shown in Fig. 6, ξ = 400 and training is
since the battery is not rated for charging at temperatures lower than stopped at 100,000 epochs to ensure an unbiased comparison. It is clear
10 °C. When regenerative braking occurs at the higher temperatures, that a deeper network can achieve better estimation accuracy as going
the charging current is seen to spike up to 8 A. This is useful in order to from 2 layers to 4 offers a 10% and a 24% reduction in MAE and MAX
ascertain the DNN's performance at higher charging currents even if error, respectively. In the same vein, going from 4 to 6 layers reduces
momentary in nature. These higher charging spikes cannot be main- the MAE and MAX error by 9% and 15%, respectively, however, in-
tained for longer than a few seconds since this would also exceed the creasing the depth up to 8 layers achieves no reduction in MAE or MAX.
peak charge current rating of the battery. To evaluate the DNN's per- This could very well be a result of the gradients, defined in eq. (6),
formance in charging scenarios for longer periods of time, the DNN is becoming much smaller as more layers are added to the network which
applied on a charging validation dataset. leads to vanishing gradients and the inability to update the network
The charge rate used in this dataset is 1C. There is interest in using a weights during training. With respect to the number of neurons in the
charge rate that is higher than this however 1C is already twice the network, Fig. 6b) indicates that going from 1 neuron to 4 per layer

249
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 5. (a) Performance of DNN trained over data with ξ = 100 and over data with ξ = 400. (b) Performance of DNN trained over augmented training data and DNN
trained over unaugmented training data. (c) Performance over incorrectly initialized network. DNNs have 3–8 layers and 4–8 neurons per layer. Validation is
performed over 25 °C HWFET dataset.

250
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Table 6 DNN can estimate SOC at different ambient temperatures. The seven
SOC estimation accuracy using different averaging parameter ξ in training data. training datasets recorded at the 5 different ambient temperatures
Validation Dataset MAE (%) STDDEV (%) MAX (%) (25 °C, 10 °C, 0 °C, −10 °C and −20 °C), mentioned in Table 2, are used
for training. These datasets are augmented, as described in subsection
HWFET with ξ = 100 0.96 0.76 3.91 3.3, to increase the robustness of the DNN which raises the training
HWFET with ξ = 400 0.61 0.49 2.38
dataset size to 20 times the initial size. Two validation datasets from
US06 with ξ = 100 1.16 0.86 4.54
US06 with ξ = 400 0.084 0.61 3.14
each of the 5 ambient temperatures are then used to evaluate the DNN's
performance. The second objective is to investigate the DNN's cap-
*All validation results are obtained on 25 °C datasets. ability to interpolate it's ability to estimate SOC at ambient tempera-
tures other than the ones one which it is trained. The DNN used in this
Table 7 section is 4 layers deep and has 8, 16, 32 and 1 neurons in the re-
SOC estimation accuracy of DNN trained over augmented and unaugmented spective layers. The training data is composed of the battery voltage,
data. temperature, average voltage and average current, as is used in the
Training Data Validation MAE (%) STDDEV (%) MAX (%) previous section, and the averaging parameter, ξ is set to 400. The time
Augmentation Dataset required to train this DNN is about 40 h using one of the GPUs discussed
in Section 2.
No HWFET 1.35 0.94 5.04
Fig. 7 showcases the DNN's estimation performance over one vali-
Yes HWFET 1.06 0.49 3.41
No US06 2.67 2.08 15.14
dation dataset from 3 different ambient temperatures; 25 °C, 0 °C and
Yes US06 1.59 1.23 7.14 −20 °C. In addition, Fig. 8a shows a bar graph outlining the DNN's
performance over the two validation datasets for every one of the five
*All validation results are obtained on 25 °C datasets. ambient temperatures. Typically, higher estimation error is expected
for the lower ambient temperatures since the disparity between the
measured surface temperature of the battery and its internal tempera-
ture increases for lower ambient temperatures. Though, interestingly,
the lowest MAE is achieved over a 0 °C dataset at 0.91%. This can be
attributed to the added regenerative braking at the higher ambient
temperatures which can add uncertainty. However, the MAE achieved
at 25 °C is still quite low at 1.44%.
In most cases, the MAE over the HWFET datasets which are less
dynamic in nature are lower than those obtained over the US06 datasets
which are much more dynamic in nature. However, this phenomenon is
not exhibited for the case of −20 °C. For the −20 °C ambient tem-
perature, the MAE over the HWFET dataset is higher. This is likely
because the average battery temperature for the HWFET test is −14 °C,
which is much lower than the average temperature of −6.7 °C for the
US06 test. While the US06 temperature is greater for all of the test
cases, the temperature difference has more significance at these lower
temperatures where battery resistance increases dramatically. The
maximum error is not entirely representative of the performance of the
DNN since a few outliers over thousands of estimated values do not
represent the overall performance of the network, however MAX is
provided in the results for completeness.
The DNN has good estimation performance when validated over
constant ambient temperatures. However, depending on the geo-
graphical location within which a battery-powered vehicle may op-
erate, a variation of 5 to 10 °C in ambient temperature is possible over
the course of one day. Hence, a worthwhile exercise would be to vali-
date the DNN over a dataset which has an ambient temperature that
changes over time. Therefore, the DNN's estimation performance is
tested over a validation cycle which increases from 10 °C to 25 °C over
the course of the dataset, which is shown in Fig. 8b. As can be seen, the
Fig. 6. (a) Estimation accuracy measured during validation versus number of DNN performs well over this validation dataset, maintaining good es-
layers in DNN. The number of neurons per layer is fixed to 4 neurons. (b) timation accuracy even at ambient temperatures lying between 10 °C
Estimation accuracy measured during validation versus number of neurons. The and 25 °C, on which the network is not trained. This interpolating
number of layers in the DNN is fixed to 4 layers. All tests are performed over ability can be a great advantage when training data is scarce.
validation datasets recorded at 25 °C.
5. Conclusion
offers a reduction in MAE and MAX of 32–35% and 16–45%, respec-
tively. However, 8 or more neurons offer no additional reduction in In the final analysis, this paper offers three different unique con-
MAE or MAX error. tributions. Firstly, on the modelling and estimation front, the DNNs
used in this paper map the measured battery signals like voltage, cur-
rent and temperature directly to SOC and achieves competitive esti-
4.2. SOC estimation at variable ambient temperature mation performance with MAEs below 1%. Table 8 compares the DNN's
estimation performance to other algorithms mentioned in literature for
In this section, two objectives are pursued. The first is to train the a more direct comparison. Secondly, on the parametrization front, the
DNN on data taken at various ambient temperatures such that a single DNN self-learns all its weights, eliminating the need to hand-engineer

251
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 7. Estimation performance of one DNN at 3 different ambient temperatures. DNN is composed of 4 layer, where 8, 16, 32 and 1 neurons are used in the
respective layers.

252
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Fig. 8. (a) SOC estimation accuracy of DNN trained on varying ambient temperature data during validation. (b) Estimation performance of DNN on validation set
recorded at an increasing ambient temperature. From top to bottom, the plots shown are of the estimated SOC compared to the ideal SOC, the resulting errors at each
time step and the ambient temperature. The validation dataset is composed of a mixture of HWFET, US06, LA92 and the NN drive cycle.

and parametrize traditional models, which is a very time-consuming the DNNs presented in this work are robustified against measurement
and costly process. Thirdly, on the inference front, once trained, a DNN offsets, gains and noise such that they can retain great estimation
can operate online with relatively low computational time. In addition, performance regardless of the imperfections found in a vehicle's

253
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

Table 8
Comparison of SOC estimation error for six studies.
Method Lowest Error Temp. Test Case Li-ion Battery

Model Adaptive-Improved EKF [17] < 1.5% MAE 25 °C Two drive cycles, +2 to −1 A 1.1Ah A123 APR18650m1
AEKF w/ANNbattery model [19] ˂3% RMS 20 °C (ambient) ± 1 A charging pulse profile 1.2Ah
AUKF w/extreme learning machine ˂1.5% MAX 25 °C (ambient) 0.52 A 50% duty cycle pulse discharge 2.6Ah Samsung
battery model [20]
Fuzzy NN w/genetic algorithm [40] ˂0.9% APE 25 °C (ambient) Constant resistance discharge (13 A) 10Ah MRL/ITRI
Radial Bias Function NN [41] 0.02% AAPE Unspecified Constant discharge rate (0.3C, 1C, 0.7C) 10Ah Lyno Power
LYS347094S
DNN (estimator in this paper) 0.61% MAE 0.78% RMS −20 to 25 °C Dynamic drive cycles, ± 18 A, Range of 2.9Ah Panasonic
2.38% MAX (25 °C) (ambient) ambient temperatures NCR18650PF

measurement devices. Furthermore, the beauty of using this machine [9] Z. Li, J. Huang, B.Y. Liaw, J. Zhang, On state-of-charge determination for lithium-
learning algorithm is that the intrinsic behavior of the battery at dif- ion batteries, J. Power Sources 348 (Supplement C) (2017) 281–301.
[10] M.A. Roscher, D.U. Sauer, Dynamic electric behavior and open-circuit-voltage
ferent ambient temperatures can be encoded into the weights of the modeling of LiFePO4-based lithium ion secondary batteries, J. Power Sources 196
DNN such that one single network can offer an accurate and robust (1) (2011) 331–336.
estimation strategy at different ambient conditions. Finally, the results [11] J. Li, J.K. Barillas, C. Guenther, M.A. Danzer, A comparative study of state of charge
estimation algorithms for lifepo4 batteries used in electric vehicles, J. Power
from extensive validation tests shown in this paper, illustrates that the Sources 230 (2013) 244–250.
DNN offers competitive estimation performance. As a result, it can be [12] L. Liu, L.Y. Wang, Z. Chen, C. Wang, F. Lin, H. Wang, Integrated system identifi-
concluded that machine learning techniques are powerful tools when cation and state-of-charge estimation of battery systems, IEEE Trans. Energy
Convers. 28 (1) (2013) 12–23.
applied to Li-ion battery SOC estimation and potentially to other bat- [13] M. Gholizadeh, F. Salmasi, Estimation of state of charge, unknown nonlinearities,
tery diagnostics. Specifically, for future work, applying these machine and state of health of a lithium-ion battery based on a comprehensive unobservable
learning algorithms on State-of-Health estimation or State-of-Power model, IEEE Trans. Ind. Electron. 61 (3) (2014) 1335–1344.
[14] X. Chen, W. Shen, Z. Cao, A. Kapoor, A novel approach for state of charge esti-
estimation with different time horizons can be very beneficial. Other
mation based on adaptive switching gain sliding mode observer in electric vehicles,
future works can also include the development of a more generalized J. Power Sources 246 (2014) 667–678.
machine learning algorithm not only able to estimate SOC at various [15] C. Hu, B.D. Youn, J. Chung, A multiscale framework with extended kalman filter for
ambient temperatures but also for different battery cells. Given the lithium-ion battery {SOC} and capacity estimation, Appl. Energy 92 (2012)
694–704.
amount of data that is generated by a Li-ion battery pack is vast, it [16] F. Sun, X. Hu, Y. Zou, S. Li, Adaptive unscented kalman filtering for state of charge
becomes instinctive to utilize data-driven approaches like machine estimation of a lithium-ion battery for electric vehicles, Inside Energy 36 (5) (2011)
learning models to perform state and parameter estimation. This work 3531–3540.
[17] S. Sepasi, R. Ghorbani, B.Y. Liaw, Improved extended kalman filter for state of
illustrates how these algorithms can self-learn their own weights and charge estimation of battery pack, J. Power Sources 255 (Supplement C) (2014)
achieve competitive estimation performance over a large range of 368–376.
ambient temperatures. [18] W. Wang, D. Wang, X. Wang, T. Li, R. Ahmed, S. Habibi, A. Emadi, Comparison of
kalman filter-based state of charge estimation strategies for li-ion batteries,
Transportation Electrification Conference and Expo (ITEC), 2016 IEEE, IEEE, 2016,
Acknowledgments pp. 1–6.
[19] M. Charkhgard, M. Farrokhi, State-of-charge estimation for lithium-ion batteries
using neural networks and ekf, IEEE Transactions on Industrial Electronics 57 (12)
This work was supported in part by the Canada Excellence Research (2010) 4178–4187.
Chairs, Government of Canada in Hybrid Powertrain Program. Battery [20] J. Du, Z. Liu, Y. Wang, State of charge estimation for li-ion battery based on model
testing was performed at the Wisconsin Energy Institute at the from extreme learning machine, Contr. Eng. Pract. 26 (2014) 11–19.
[21] J. Meng, G. Luo, F. Gao, Lithium polymer battery state-of-charge estimation based
University of Wisconsin-Madison. The authors gratefully acknowledge
on adaptive unscented kalman filter and support vector machine, IEEE Trans. Power
the support of NVIDIA Corporation with the donation of the Titan X Electron. 31 (3) (2016) 2226–2238.
Pascal GPU used for this research. [22] S. Tong, J.H. Lacap, J.W. Park, Battery state of charge estimation using a load-
classifying neural network, Journal of Energy Storage 7 (Supplement C) (2016)
236–243.
References [23] X. Hu, D. Cao, B. Egardt, Condition monitoring in advanced battery management
systems: moving horizon estimation using a reduced electrochemical model, IEEE
[1] X. Hu, C. Zou, C. Zhang, Y. Li, Technological developments in batteries: a survey of ASME Trans. Mechatron. 23 (1) (2018) 167–178.
principal roles, types, and management needs, IEEE Power Energy Mag. 15 (5) [24] X. Hu, F. Sun, Fuzzy clustering based multi-model support vector regression state of
(2017) 20–31. charge estimator for lithium-ion battery of electric vehicle, Intelligent Human-
[2] A. Emadi, Advanced Electric Drive Vehicles, CRC Press, New York, 2015. machine Systems and Cybernetics, 2009. IHMSC’09. International Conference on,
[3] E. Chemali, M. Preindl, P. Malysz, A. Emadi, Electrochemical and electrostatic vol. 1, IEEE, 2009, pp. 392–396.
energy storage and management systems for electric drive vehicles: state-of-the-art [25] X. Hu, S.E. Li, Y. Yang, Advanced machine learning approach for lithium-ion battery
review and future trends, IEEE Journal of Emerging and Selected Topics in Power state estimation in electric vehicles, IEEE Transactions on Transportation elec-
Electronics 4 (3) (2016) 1117–1134. trification 2 (2) (2016) 140–149.
[4] World Energy Outlook, Special Report Energy and Air Pollution, International [26] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444.
Energy Agency, 2016. [27] A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep con-
[5] P. Malysz, R. Gu, J. Ye, H. Yang, A. Emadi, State-of-charge and state-of-health es- volutional neural networks, Adv. Neural Inf. Process. Syst. (2012) 1–9.
timation with state constraints and current sensor bias correction for electrified [28] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition,
powertrain vehicle batteries, IET Electr. Syst. Transp. 6 (2) (2016) 136–144. Arxiv.Org 7 (3) (2015) 171–180.
[6] L. McCurlie, M. Preindl, A. Emadi, Fast model predictive control for redistributive [29] L. Deng, Deep Learning: Methods and Applications, Found. Trends® in Signal
lithium-ion battery balancing, IEEE Trans. Ind. Electron. 64 (2) (2017) Process. 7 (3–4) (2014) 197–387.
1350–1357. [30] D. CireşAn, U. Meier, J. Masci, J. Schmidhuber, Multi-column deep neural network
[7] W. Waag, C. Fleischer, D.U. Sauer, Critical review of the methods for monitoring of for traffic sign classification, Neural Network. 32 (2012) 333–338.
lithium-ion batteries in electric and hybrid vehicles, J. Power Sources 258 (2014) [31] G. Hinton, L. Deng, D. Yu, G.E. Dahl, A. r. Mohamed, N. Jaitly, A. Senior,
321–339. V. Vanhoucke, P. Nguyen, T.N. Sainath, B. Kingsbury, Deep neural networks for
[8] R. Ahmed, M. El Sayed, I. Arasaratnam, J. Tjong, S. Habibi, Reduced-order elec- acoustic modeling in speech recognition: the shared views of four research groups,
trochemical model parameters identification and SOC estimation for healthy and IEEE Signal Process. Mag. 29 (6) (2012) 82–97.
aged Li-Ion batteries. Part I: parameterization model development for healthy [32] J. Ma, R.P. Sheridan, A. Liaw, G.E. Dahl, V. Svetnik, Deep neural nets as a method
batteryies, IEEE Journal of Emerging and Selected Topics in Power Electronics 2 (3) for quantitative structureactivity relationships, J. Chem. Inf. Model. 55 (2) (2015)
(2014) 659–677. 263–274.

254
E. Chemali et al. Journal of Power Sources 400 (2018) 242–255

[33] D. Wang, A. Khosla, R. Gargeya, H. Irshad, A.H. Beck, Deep Learning for Identifying https://www.epa.gov/vehicle-and-fuel-emissions-testing/dynamometer-drive-
Metastatic Breast Cancer, ArXiv Preprint, (2016) 1–6arXiv:1591239. schedules#vehicleDDS, (2017).
[34] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado, [38] Panasonic, Panasonic NCR18650PF Lithium-ion Battery Datasheet (June 2016),
A. Davis, J. Dean, M. Devin, et al., Tensorflow: Large-scale Machine Learning on (2016).
Heterogeneous Systems, (2015) url h ttp, Software available from: tensorflow. [39] Panasonic, Introduction of NCR18650PF, Panasonic, (2013).
org. [40] Y.S. Lee, W.Y. Wang, T.Y. Kuo, Soft computing for battery state-of-charge (bsoc)
[35] D.P. Kingma, J. Ba, Adam: a method for stochastic optimization, CoRR abs/ estimation in battery string systems, IEEE Trans. Ind. Electron. 55 (1) (2008)
1412.6980, http://arxiv.org/abs/1412.6980. 229–239.
[36] Phillip Kollmeyer, Panasonic 18650PF Li-ion Battery Data, Mendeley Data, v1, [41] W.-Y. Chang, Estimation of the state of charge for a lfp battery using a hybrid
(2018) https://doi.org/10.17632/wykht8y7tg.1. method that combines a rbf neural network, an ols algorithm and aga, Int. J. Electr.
[37] United States environmental protection agency, vehicle and fuel emissions testing, Power Energy Syst. 53 (2013) 603–611.

255