Decision Analytics Journal 8 (2023) 100255

Contents lists available at ScienceDirect

Decision Analytics Journal

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

A feedforward deep neural network for predicting the state-of-charge of

lithium-ion battery in electric vehicles
Bukola Peter Adedeji, Golam Kabir ∗
Industrial Systems Engineering, University of Regina, Regina, SK, Canada, S4S 0A2

ARTICLE INFO ABSTRACT

Keywords: This study proposes a feedforward deep neural network to predict the parameters of the lithium-ion battery in
Artificial neural network electric vehicles. Correlation analysis is used to select the candidate parameters for the proposed model with
Feedforward deep neural network no categorical variable. A direct feedforward deep artificial neural network is developed to predict a battery’s
Lithium-ion battery
charge state and develop the proposed inverse model. The predicted state-of-charge of the direct model is
State-of-charge
combined with four virtual functions to form the input variables for the proposed inverse model. Furthermore,
Electric vehicles
virtual functions are incorporated to enhance the predicting capability of the proposed inverse function model.
The predicted multi-output variables of the proposed inverse model are speed, mileage, voltage, velocity, and
state-of-charge. The proposed inverse model is superior to the feedforward deep neural network previously
proposed in the literature because of its multiple output capabilities. Also, the proposed model makes decision-
making easier when used for the design simulation than the single-output deep neural networks, which predict
the state-of-charge of a battery only. The mean square error is used as the metric for accurate measurement.
During the simulation by the proposed inverse model (with virtual functions), accuracy was 44.43 times higher
than the traditional inverse deep neural network model. Redefined parameters were used to verify the findings
of the model. This result suggests that incorporating virtual functions into a deep neural network model’s
inverse model can improve the accuracy of battery and electric vehicle parameter predictions.

1. Introduction duties by monitoring the temperature of each cell in a battery [3,5].

Consequently, it is incorporated as safety equipment [7]. Thus, battery
The battery is one of the essential components of electrified vehicles. management system functions may help extend the life of a battery by
It is a source of tractive energy for all electric vehicles (AEVs) or pure protecting it from damage. Other functions of the battery management
electric vehicles (PEVs), plug-in hybrid electric vehicles (PHEVs), and system are collection of datasets (like datasets on temperature, current,
hybrid electric vehicles (HEVs). The battery determines how far an and voltage), estimation of states (state of charge, state of the health,
electric vehicle (EV) can travel in pure electric or all-electric mode. and state of life), and cell balancing [3].
That is why a battery electric vehicle with a large battery capacity has
A Li-ion battery can be merged with another storage device, such as
a higher driving range or mileage than one with a low battery capacity
an ultracapacitor, to form a hybrid energy storage system (HESS). Other
[1]. Lithium-ion and lead–acid batteries are commonly employed in
storage devices that can be combined with Li-ion batteries to form HESS
automobiles. The batteries of electric vehicles consist of cells that are
are the fuel cell, superconductive magnetic device, and flywheel [3].
connected in parallel or series. Lithium-ion (Li-ion) batteries have been
The batteries of electric vehicles have certain parameters that can
potentially useful in electric vehicles due to certain qualities like a high
level of energy density [2], extraordinary efficiency, a large or lengthy be estimated using a data-driven approach. Some parameters cannot
cycle life [3], and a low rate of self-discharging [4]. be easily measured directly, according to [8]. For example, a battery
Li-ion batteries have been found to be sensitive to high tempera- management system cannot directly measure the energy stored in the
tures, over-discharging, and over-charging [5]. One of the vital func- battery. Other battery parameters that are difficult to calculate include
tions of the battery management system (BMS) is to ensure that the the aging factor, ambient temperature, and battery chemistry [9]. A
batteries of electric vehicles operate under harmless or favorable con- data-driven method like an artificial neural network (ANN) is needed
ditions [6]. The BMS essentially controls the current and voltage during to estimate or calculate such non-measurable parameters. Measurable
charging and discharging to prevent the battery from being damaged. parameters can also be predicted or forecasted by artificial neural
The battery management system can also provide thermal management networks. This is one of the benefits of artificial neural networks.

∗ Corresponding author.
E-mail addresses: [email protected] (B.P. Adedeji), [email protected] (G. Kabir).

https://doi.org/10.1016/j.dajour.2023.100255
Received 8 April 2023; Received in revised form 14 May 2023; Accepted 27 May 2023
Available online 1 June 2023
2772-6622/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

and annoyance. The result of the study was compared with that of the
Nomenclature linear regression model that was based on psychoacoustic parameters.
The result of the study shows that most of the parameters did not have
Acronyms a linear impact on annoyance perception.
AEVs All Electric Vehicle Hussein [13] used a neural network to predict capacity fade in
lithium-ion (Li-ion) batteries used in electric vehicles. The results of
PEVs Pure Electric Vehicles
the research indicated that artificial neural networks could perform
HEVs Hybrid Electric Vehicles
effectively in situations where there are variations in the internal
PHEVs Plug-in Hybrid Electric Vehicles
parameters. It demonstrated that it is possible to improve a battery’s
EV Electric Vehicle state of charge estimation and remaining life. Tian et al. [14] applied
HESS Hybrid Energy Storage System artificial neural networks to investigate the electric vehicle air condi-
BMS Battery Management System tioning system’s (EVACS) performance. The study developed a model
Li-ion Lithium-ion based on the data collected from the experiment for electric vehicle
SOC State of Charge air conditioning system performances (such as refrigerant mass flow
SOH State of Health rate, condenser heat rejection, refrigeration capacity, and compressor
ANNs Artificial Neural Networks power consumption). The result of the study shows that artificial
MAE Mean Absolute Error neural network is suitable for the estimation of air conditioning system
performances of electrified vehicles.
MSE Mean Square Error
Ansari et al. [15] employed a neural network model to predict
RMSE Root Mean Square Error
the remaining useful life (RUL) of lithium-ion batteries under various
CNNs Convolutional Neural Networks training datasets. The study compared a multi-channel input (MCI)
DNN Deep Neural Network profile to a single-channel input (SCI) or single input (SI) profile using
ANOVA Analysis of Variance different samples from a National Aeronautics and Space Administra-
NASA National Aeronautics and Space Administra- tion (NASA) dataset. To reconstitute the input training profile, ten (10)
tion sample values of current, voltage, and temperature were collected at
LSTM Long Short-Term Memory equal intervals from each charging cycle. The result shows that the root
RNN Recursive Neural Network mean square error was lower in the multi-channel input profile than the
OCV Open-Circuit-Voltage single-channel input. Ma and Faye [16] used hybrid long short-term
memory (LSTM) neural networks to predict multistep electric vehicle
charging station occupancy. The results show that the proposed method
has a high degree of accuracy. Zhao et al. [17] used a recursive neural
The neural network approach can predict or estimate parameters like network (RNN) to improve the data representation. Also, the study
current, state of charge, voltage, mileage, and current. This study used multi-channel extended convolutional neural networks to predict
introduces a multiple input and multiple output model for predicting lithium-ion battery (LIB) state of charge. The approach increased pre-
the five selected parameters of a Li-ion battery and an electric vehicle. diction performance when compared with recurrent neural networks
The motivation behind this study is that previously introduced and the current counting method, respectively.
models for Li-ion batteries could only predict one output variable. Nie and Farzaneh [18] introduces a real-time dynamic predictive
This output variable could be current, voltage, state of charge, state cruise control (PCC) system to minimize the energy consumption of
of health, mileage, voltage, etc. In a situation where the battery manu- electric vehicles. The study employs machine learning to build a model
facturer needs to estimate more than one battery parameter, decision- for estimating instantaneous energy consumption. The results of the
making becomes challenging. For example, if a battery manufacturer study show that predictive cruise control improves energy savings.
wishes to estimate the state of charge (SOC) and mileage of a battery Narasimhulu et al. [19] proposed an artificial neural network and
at a particular speed, the previously introduced model would not be Aquila optimizer algorithm (AOA) for energy management control. The
able to do this. This is a case of the estimation of multiple output hybrid energy storage system (HESS) architecture was developed for all
variables. Most models in the literature were developed for only one electric vehicles to store photovoltaic energy. The approach would aid
output variable. This is the primary motivation behind this study. in designing, maintaining, and evaluating a variety of energy storage
The remaining components of this study have been subdivided into systems.
different sections. Section 2 of the study contains the literature review, Hong et al. [20] proposed a joint-prediction model that imple-
while Section 3 of the paper contains the methodology used in the mented long short-term memory neural networks and multiple linear
study. Section 4 contains the results of all the predictions in the study. regression (LR) algorithms to predict the state of charge of batteries on
Section 5 is the discussion and comparative analysis section, while real-world operational electric vehicles. The accuracy of the model was
Section 6 is the conclusion. within an acceptable range. Charkhgard and Farrokhi [21] proposed a
model using ANN and an extended Kalman filter (EKF) to predict the
2. Literature review state of charge of lithium-ion batteries. Experimental results for state of
charge were reported to be good. Zheng et al. [22] implemented long
Many studies have used artificial neural networks (ANN) for predic- short-term memory to predict the performance of a proton-exchange
tion and estimation of battery parameters. Alvarez et al. [10] applied membrane fuel cell (PEMFC). The accuracy of long short-term memory
ANN to predict the energy consumption of electric vehicles by taking model was found to be higher than that of the recurrent neural network
into consideration the driving behavior of the driver. The variables and the back-propagation neural network. Wang et al. [23] applied
for the model were acceleration, velocity, and jerk, while the output a multilayer perceptron and convolutional neural network to predict
variable was the remaining battery charge. Modi et al. [11] employed the life of a lead–acid battery. Premkumar et al. [24] employed a
convolutional neural networks (CNNs) to estimate the energy consump- deep neural network (DNN) approach to predict state of charge. The
tion of electric vehicles considering velocity, tractive effort, and road values of the mean absolute error and root mean square error were
elevation. The study found that the model with seven layers performed significantly small.
better when images converted from time-series data were implemented. Vidal et al. [25] estimated the state of charge of batteries using
Steinbach and Altinsoy [12] employed ANN as a tool for prediction transfer learning (TL) and deep neural networks. The result shows an

2
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

increase in accuracy. Chandran et al. [26] used six machine learn- 3.1. State-of-charge (SOC) of a battery
ing algorithms to predict the state of charge of lithium-ion battery
systems for electric vehicles. The employed algorithms are artificial The state of charge of the battery of an electric vehicle refers to
neural networks, support vector machines (SVM), linear regression the percentage value of electrical power that is stored or available on
(LR), Gaussian process regression (GPR), ensemble bagging (EBa), and board [8]. This is given as the ratio of current remaining capacity of
ensemble boosting (EBo). The accuracy of Gaussian process regression the battery to the current rated (or maximum current) capacity of the
(GPR) and ANN were found to be higher than the remaining algorithms. battery.
How et al. [9] estimated the state of charge of a lithium-ion battery
𝑄𝑎
using an improved deep neural network approach in electric vehicle 𝑆𝑂𝐶 = × 100% (1)
applications. The results of the study show that increasing the number 𝑄𝑏
of hidden layers will minimize error and support the estimation of state where Qa is the available current (or remaining capacity) of a battery
of charge. and Qb is the current rated or maximum current capacity of the battery.
Adedeji [27,28] applied inverse function models of an artificial The current rated (Qb ) is the maximum or highest electric charge that a
neural network to estimate the parameters of a battery electric vehicle battery can store. The value of state of charge in a cell range between
and the fuel economy label of PHEVs. These models did not estimate 100% and 0%. At 100%, the cell is considered fully charged, and at
battery parameters like current, voltage, state of charge, and speed. 0%, it is considered fully discharged. However, the threshold for a cell
Therefore, the inverse models previously introduced are significantly is practically 50% for it to be recharged [43]. It should be noted that
different from the proposed model in this paper. The model introduced a fully charged, aged battery may have a state of charge equivalent to
in this paper is for the simulation of Li-ion battery parameters. The
80%–70% of the state of charge of a new battery. The battery’s state of
summary of the literature review is shown in Table 1.
charge may be affected as the battery becomes older [43].
Most of the previously proposed models above for battery param-
Different kinds of methods have been proposed for determining the
eters are simple models with a single output variable. This makes
state of charge of a battery. Examples of these methods are Coulomb
the estimating capability of the proposed model limited. This study
counting method and voltage-based State-of-Charge–Open-Circuit-
introduces an approach for the prediction of multiple outputs of battery
and electric vehicle parameters for operational and design simulation Voltage (SOC–OCV) mapping [44]. In the Coulomb counting approach,
purposes. ANNs were used for the simulation of critical parameters of a energy stored in the battery is calculated by integrating the discharged
lithium-ion battery and that of an electric vehicle. These parameters are current in the battery over time. This approach is current-dependent
the state-of-charge of the battery, voltage, current, speed, and mileage. [44]. In order to use this approach, the initial battery capacity, or
The study implemented an inverse function model of an artificial neural initial SOC before the trip, must be known or determined. This method
network to predict multiple outputs of Li-ion battery parameters for determines the average amount of current that is discharged over time
electric vehicles. Electric vehicle parameters were equally predicted. [45]. When the integrated current is used for estimation, state of charge
The accuracy was found to be higher during the prediction of the can be given by [46].
inverse function of the state of charge than that of the other four case 𝑡
∫ 𝐼𝑏 (𝑡)𝑑𝑡
studies. 𝑆𝑂𝐶 𝐹 = SOC0 − 0 (2)
𝐶𝑏
√
2.1. Research gap and contributions 𝐸 − 𝐸 2 − 4𝑅𝑃𝑏
𝐼𝑏 = (3)
2𝑅
Previous studies mentioned above predicted the state of charge as
where Ib = battery current, E = open circuit voltage, R = internal
a single output in artificial neural network models. This limits the
resistance of the battery, SOCF = present or final state of charge of
capability of the models, as initially stated. This current study intro-
the battery, SOC0 = initial state of charge of the battery before the
duces an artificial neural network model that simulates many battery
trip, Cb = battery capacity, Pb = battery power.
parameters simultaneously. Also, previous multiple output models of
artificial neural networks were used for the design simulation of electric The Coulomb counting is the most popular method for determining
vehicle [27] and simulation of fuel economy label parameters [28]. the state of charge of a battery. This is because it is easy, effective, and
These previous multiple output models of ANN did not predict battery gives appreciable results [45]. However, errors are likely to aggregate
parameters like state of charge, voltage, and current. The following are over time due to the integration process or term [44].
the contributions made by this study Another approach for estimating the state of charge level is by
the open circuit voltage or a regressed voltage-based State-of-Charge–
• A feedforward deep artificial neural network-based inverse model Open-Circuit-Voltage (SOC–OCV) mapping. This approach is voltage-
is developed for battery parameter simulations. dependent [44]. In this method, the open circuit voltage (OCV) is
• A multi-output variable model of an artificial neural network is used to estimate the state of charge level. This method estimates the
introduced to facilitate decision-making in the design simulations state of charge level using open circuit voltage (OCV) by mapping
of lithium-ion batteries for electric vehicles. via a nonlinear relationship [44]. The approach maps state of charge
• A superior output model is introduced for the prediction of with open circuit voltage using mathematical relationships between
lithium-ion battery parameters the battery capacity and the remaining or available voltage [45]. This
• The feedforward deep artificial neural network-based inverse method is said to be unpracticable because it cannot estimate state
model can simultaneously predict parameters such as state of of charge while batteries are charging and discharging. It means that
charge, voltage, current, mileage, and speed. the state of charge cannot be measured in a closed circuit using this
approach. This method is commonly employed in the laboratory [45].
3. Material and methods

As one of the objectives of this study is to estimate the state of 3.2. Datasets collection and exploration
charge of the battery using artificial neural networks, the state of
charge was chosen as the target variable of the direct function model. Some battery performance parameters in the real world can be
On the other hand, the proposed inverse function model has multiple simulated using industrial battery testers. Battery testers are capa-
target variables because it is a multi-output model. The selection pro- ble of charging and discharging batteries at different rates. They can
cess is discussed in Section 3.2. The proposed inverse function model equally simulate current profiles for various drive cycles. Thus, in-
implements feedforward-back propagation algorithms. Fig. 1 shows the dustrial battery testers are used for generating datasets like current
flow process for this study. for experimental batteries. Battery Management System cannot directly

3
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 1
Summary of the literature on application of ANNs.
Research objective Research approach Contribution/Achievement Reference
To determine the annoyance of different Artificial neural network (ANN) was employed The study found that most parameters have Steinbach and
electric vehicle sounds for a constant speed, for annoyance estimations. 150 ANNs were no linear influence on the annoyance Altinsoy [12]
single car pass-by situation implemented in the research and the best perception.
models were compared with linear regression
models based on psychoacoustic parameters
To estimate the fading capacity of electric The study employed neural network for The result of the study shows an Hussein [13]
vehicle batteries estimation of fading capacity of two identical improvement in the state-of-charge (SOC)
3.6-V/16.5Ah lithium-ion (Li-on) batteries of accuracy for the battery lifespan prediction.
electric vehicles The result of the work also shows accuracy
in the prediction of the remaining service
time of the battery.
To predict the current of a short circuited The work implemented only the details of the The result shows that maximum estimation Yang et al. [29]
cell using only the voltage information voltage to predict or estimate the current of error increased in value
the short-circuited cell using ANN.
To investigate the electric vehicle air An artificial neural network (ANN) model for Result shows that ANN is suitable for Tian et al. [14]
conditioning system (EVACS) performances EVACS performances (such as refrigerant mass prediction of electric vehicle air
flow rate, condenser heat rejection, conditioning system (EVACS) performances.
refrigeration capacity and compressor power
consumption) prediction was developed based
on experimental data.
To predict the remaining useful life (RUL) An ANN based multi-channel input (MCI) Accuracy is higher for the proposed Ansari et al.
of lithium-ion batteries under various profile is implemented and compared with ANN-based MCI than SCI using B0005 [15]
training datasets. single-channel input (SCI) or single input (SI) dataset in terms of RMSE.
with diverse datasets from NASA.
The research proposes a predictive control Back propagation neural network (BPNN)-based The result of the study shows that the Lin et al. [30]
strategy based on velocity prediction method was used to predict velocity and then proposed velocity prediction approach
incorporated into the predictive control strategy estimated vehicle velocity with high
for a plug-in fuel cell electric vehicle (FCEV) precision. Also, the proposed energy
management strategy (EMS) will reduce the
hydrogen consumption of the plug-in FCEV
The study proposed a new mixed long the proposed model separates different types of The results show that the proposed method Ma and Faye
short-term memory neural network features and handles them differently with produces very accurate and performs better [16]
incorporating both historical charging state mixed neural network architecture than the benchmark approaches.
sequences and time-related features for
multistep discrete charging occupancy state
prediction.
A deep learning-based SOC prediction A recursive neural networks (RNNs)-based The proposed method improved prediction Zhao et al. [17]
model is proposed to ensure reliable vector method was used to improve the data performance by 4.3% and 11.3% compared
representation and sufficient feature representation and a multi-channel extended with recurrent neural networks and Ah
extraction convolutional neural networks (CNNs)-based counting method, respectively
method was used to predict lithium-ion battery
(LIB) SOC.
The study proposes a real-time dynamic This research used machine learning to develop The results of the study showed that the Nie and
predictive cruise control (PCC) system to a new instantaneous energy consumption proposed PCC system can Farzaneh [18]
minimize the energy consumption for estimation model based on the strong mapping save energy between the rates of 8.5% to
electric vehicles (EVs) relationship between energy consumption and 15.6%.
EV’s state variables and dynamic
characteristics’’
To develop a hybrid energy storage system The energy management control algorithms The outcomes of this study could aid Narasimhulu
(HESS) architecture for an all-electric called Artificial Neural Network (ANN) and researchers and professionals in the et al. [19]
vehicle that stores photovoltaic (PV) energy Aquila Optimizer Algorithm (AOA) are automotive industry. Also, professionals and
utilizing batteries and ultracapacitor system proposed. third parties involved in designing,
(UCs). maintaining, and evaluating different energy
sources and storage systems will benefit
from it.
The study proposed a This paper demonstrates a novel The prediction steps of LSTM are within Hong et al. [20]
machine-learning-enabled method to joint-prediction model employing the LSTM acceptable accuracy.
perform real-time multi-forward-step SOC networks and multiple linear regression (LR) The smooth implementation of
prediction for battery systems using a algorithms on real-world operational EVs multi-forward-step SOC prediction on
recurrent neural network with long real-world vehicles can eliminate drivers’
short-term memory (LSTM) mileage anxiety and safeguard vehicle
operation in a big way
To model and estimate the state of charge AN NN is trained offline using the data The results show a good estimation of the Charkhgard and
(SOC) of lithium-ion (Li-Ion) batteries using collected from the battery-charging process. A SOC and fast convergence of the extended Farrokhi [21]
neural networks (NNs) and the extended state-space model of the SOC is proposed that Kalman filter (EKF) state variables
Kalman filter (EKF) is approximated using a neural network (NN).
Then, by using this model and the extended KF
(EKF), the battery SOC is estimated

(continued on next page)

4
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 1 (continued).
Research objective Research approach Contribution/Achievement Reference
The study proposes a performance ‘‘LSTM network is compared with traditional The prediction accuracy of the LSTM Zheng et al. [22]
prediction method of proton-exchange recurrent neural network (RNN) and network is found to be higher than that of
membrane fuel cell (PEMFC) using long back-propagation neural network (BPNN) to traditional recurrent neural network and
short-term memory (LSTM) recurrent neural prove its superiority’’. back-propagation neural network
network (RNN).
To predict the voltage and lifetime of a The MLP(multilayer perceptron) and CNN obtained higher accuracy than MLP Wang et al. [23]
lead–acid battery CNN(convolutional neural network) methods
To estimate the SOC of Lithium-Ion Battery Deep Neural Network (DNN) DNN-based SOC estimate approach produces Premkumar
for Electric Vehicles a low mean absolute error and et al. [24].
root-mean-square-error values less than 5%
error
To estimate SOH and SOC of battery of The study employed two identical dynamically The method estimated battery Chaoui and
electric vehicle recurrent networks (DDRNs) with Nonlinear state-of-charge(SOC) and state-of -health Ibe-Ekeocha
Auto-regressive Exogenous (NARX) architecture (SOH) with minimal errors October [31]
for battery analysis of electric vehicle
SOC estimation using a deep neural Transfer learning is a method that uses the Results show up to 64% better accuracy Vidal et al.
network (DNN) with Transfer Learning (TL). learnable parameters from a trained DNN to and similar or better accuracy with a [6,32].
help train another DNN. Transfer learning has reduced amount of training data.
the potential to improve SOC estimation as well
as reduce DNN training time and data required
To estimate the SOC of lithium-ion battery The employed algorithms are artificial neural ANN and GPR are found to be the best Chandran et al.
systems using six machine learning network (ANN), support vector machine (SVM), methods based on MSE and RMSE. [26]
algorithms for electric vehicles application linear regression (LR), Gaussian process
regression (GPR), ensemble bagging (EBa), and
ensemble boosting (EBo).
To estimate battery state of charge using The study implemented recurrent neural The non-recurrent feedforward neural Vidal et al. [25]
recurrent and non-recurrent neural networks networks, which have inherent memory of the network with filtered inputs is found to be
past, and non-recurrent feedforward neural 23% more accurate and require less training
networks, which can effectively have memory time when compared with a recurrent
added through exogenous filtered inputs. neural network.
To estimate SOC of Li-ion Battery in This paper develops a state of charge (SOC) This study also shows that the 7-layer of How et al [9]
Electric Vehicles. estimation model for a lithium-ion battery DNN training on dynamic stress test (DST)
using an improved deep neural network (DNN) drive cycle can predict SOC values
approach in electric vehicle applications unexpectedly well.
Addition of hidden layers in DNN decreases
the error rate and improves the SOC
estimation.
To propose a convolutional neural network A deep convolutional neural network-based CNN model with seven layers was Modi et al. [11]
for estimation of energy consumption of solution was developed for the estimation of performed well using the image converted
EVs. energy consumption of EVs, considering three from time-series data. This model features
external parameters, including road elevation, high estimation accuracy, real-time
tractive effort, and driving velocity. acquisition of the road elevation data is
problematic for the model application in
eco-driving system development.
To predict the parameters of pure electric Pure electric vehicle parameters such as battery The results show a high degree of accuracy Adedeji [1]
vehicles recharge time, electrical energy consumption in for each of the predictions. The results also
cities, energy consumption on highways, show that a single model can be used to
combined electrical energy consumption, and predict more than one parameter.
range were predicted for design purposes using Government agencies can use the model to
backpropagation ANN. form standards for electric vehicles and
battery manufacturers.
Predicted SOC Temporary CNN on real world data Obtained a good level of accuracy at any Yang et al. [33]
temperature
Predicted SOC DNN on laboratory data at different DNN has higher accuracy than recurrent El Fallah et al.
temperatures unit recurrent neural network (GRU-RNN) [34]
Predicted the state of charge of lithium-ion Hybrid of radial basis function-neural network, ‘‘Maximum error of the proposed model is Fu et al. [35]
battery golden section method(GSM) and Sparrow less than 2%’’.
Search Algorithm (SSA) were used
To extract the parameters of proton Multiple learning neural network(MLNN) and Multiple learning neural network(MLNN) [36]
exchange fuel cell (PEMFC) Neural network algorithm (NNA) outperform Neural network algorithm
(NNA)
To predict the trajectory of aging of a Deep learning, transfer learning technique Accurate life cycle prediction was achieved Zhou et al. [37]
Li-ion battery strategy and Bayesian approach by deep learning
To predict football player’s health Recurrent neural networks and wearable The findings illustrates the practicality and Alghamdi [38]
technology reliability of the approach
To diagnose gearbox fault using vibration One-dimensional deep convolution neural deep convolution neural networks performs Ong et al. [39]
signals networks better than decision tree, random forest and
support vector machine

(continued on next page)

5
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 1 (continued).
Research objective Research approach Contribution/Achievement Reference
To improve prediction accuracy in order to ANN The result can be used as a standard for Agbemenou
save time and cost of design and research et al. [40]
construction
To estimate the lateral velocity in long short-term memory neural networks High level of accuracy was achieved Kong et al. [41]
autonomous vehicle
To provide warning in order to avoid Convolutional Long–Short Term Memory Convolutional Long–Short Term Memory Zheng et al. [42]
overheating of battery during charging (ConvLSTM) (ConvLSTM) has better accuracy than CNN
and LSTM

Fig. 1. Workflow for the predictions.

measure the energy stored in the batteries of electrified vehicles as of The entire datasets were explored by knowing their quantities and
now. types. The datasets are tabulated into columns. The variables are also
Previous work on state of charge and voltage prediction has used classified into quantitative (or continuous) variables and qualitative (or
a variety of datasets for predictions and estimations. Trinandana et al. categorical) variables based on their uniqueness values. The list of the
[8] used voltage and current to predict the state of charge of a battery explored variables is in Table 2. Columns that do not affect the targeted
using ANN in MATLAB. Ismail et al. [44] equally predicted the state variables are deleted. Rows and columns with more than 20% missing
of charge of a lithium-ion (Li-ion) battery using current and voltage. datasets were deleted. Missing values in the continuous variables were
Ali et al. [45] predicted the state of charge for three batteries using replaced by the median values, and missing values in the categorical
current, voltage, and temperature with a feed-forward neural network. variables were replaced by modal values.
The datasets used for the study were collected by He [47] using a
data acquisition system that can record mileage, instantaneous speed,
state of charge in battery and an electric vehicle. The datasets contain 3.3. Variable selection and correlation analysis
parameters like current, voltage, speed, state of charge, and driving
range (mileage) for a specified battery electric vehicle. The BEVs incor- High level of correlation between the continuous variables and
porated a lithium-ion battery with a capacity of 41 kWh. The collected target variable could help in obtaining good training process and
data were then preprocessed in this current study using TensorFlow, a high degree of accuracy. In this study, Pearson’s correlation was used
Python software package to determine the statistical correlation values. The correction matrix

6
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 2
Correlation matrix.
Correlation matrix
Variables Speed Mileage Voltage Current SOC Unique values Variable types
Speed 1 0.142531 −0.17558 0.451764 −0.13259 577 Continuous
Mileage 0.142531 1 −0.96287 −0.00736 −0.99799 159 Continuous
Voltage −0.17558 −0.96287 1 −0.12039 0.975032 445 Continuous
Current 0.451764 −0.00736 −0.12039 1 0.006591 518 Continuous
Soc −0.13259 −0.99799 0.975032 0.006591 1 81 Continuous

( )
Table 3 𝛥wij = 𝜂 𝑇𝑖 − 𝐴𝑖 𝑥𝑖 (7)
Sample of selected variables.
Speed Mileage Voltage Current SOC where 𝜂 = learning rate, 𝑥𝑖 = input vector or signal, Ti = target output
for a specific node or pattern recognition, Ai = actual output for a
1 6.5 120 562 370.2 3.2 100
2 5.4 120 562 370.2 3.1 100 specific node and 𝛥wji is the weight or parameter adjustment.
3 3.7 120 562 370.1 3.5 100
4 9.9 120 562 369.9 3.9 100 3.6. Development of single output model (or direct model)
5 9.2 120 562 369.6 6.3 100
6 9.6 120 562 369.8 4.3 100 A single output model was developed using the preprocessed
7 9.2 120 562 369.7 5.4 100
datasets as shown in Fig. 3. There were five number of selected
8 13.4 120 562 369.4 7.1 100
9 9.4 120 562 369.6 4.9 100 variables. Four of the selected variables were used as the input variables
10 18 120 562 369 8.3 99 while one was used as the target variable. The state of charge was
11 13.3 120 562 369.5 4.3 99 selected as the target variable. Of course, the state of charge would
12 10 120 562 369 10.1 99 be the output variable.
13 17.2 120 562 368.9 8.9 99
Given the set of input variables as
14 24.8 120 562 368.4 12.3 99
{𝑋} = {𝑥1 , … , 𝑥𝑛 } (8)
{𝑋} = {𝑥1 , … , 𝑥𝑛 } = {𝑥1 }𝑈 … 𝑈 {𝑥𝑛 } (9)
between the five continuous variables is shown in Table 2 above.
{𝑋} = {𝑥1 , … , 𝑥𝑛 } = {𝑥1 }𝑈 … 𝑈 {𝑥4 } (10)
There was no categorical variable in this study. In a situation where
categorical variables exist among the candidate variables; an ANOVA where {X} is a set of the input variables
test can be conducted to determine the suitability of the variables. Similarly, the set of output variables can be given as
The final sets of variables selected for the artificial neural network are
{𝑌 } = {𝑦1 , … , 𝑦𝑛 } (11)
current, voltage, state of charge, mileage, and speed. Table 3 above
shows the sample for the five selected variables. {𝑌 } = {𝑦1 , … , 𝑦𝑛 } = {𝑦1 }𝑈 … 𝑈 {𝑦𝑛 } (12)
Since there is only one output variable in this situation as shown in
3.4. Datasets preprocessing Fig. 3, then

Standard Scaler was used for the normalization of the values of {𝑌 } = {𝑦1 } (13)
the selected datasets. Normalization transforms datasets into a range where {Y} is a set of the output variables
between 1 and 0. The selected datasets in the columns are transformed
into one scale by normalizing them. 80% of the datasets were employed 𝑋 ∈ 𝑅𝑛 𝑎𝑛𝑑 𝑌 ∈ 𝑅𝑛
for training the model, while 20% were used for testing the model. The R are real numbers
testing datasets were employed in evaluating the trained model. Fig. 2
(a, b, c, d, & e) show the distribution of each of the selected variables. 3.7. Development of the inverse function models

3.5. Components of an artificial neural network To simulate the output of the single output model (or direct function
model), another model is needed. This is known as the ‘‘inverse function
model’’. The direct model’s output in Fig. 3 would be used as an input
A feedforward-back propagation algorithm is employed in this
for the inverse function model. The inverse function model has multiple
study. An artificial neural network is made up of three layers, namely:
output variables. In this study, two types of inverse function models
the input layer, the hidden layer, and the outer layer [1]. Also, an ANN
were developed. The first inverse function model shown in Fig. 4 did
consists of neurons or nodes that are interconnected. The three nodes
not incorporate virtual functions. In this model, only the output of the
can be represented by the 𝑖th, 𝑗th, and 𝑘th nodes, respectively. direct function model was used as the input for the inverse function
At the output layer, during the forward pass, we have the following: model. This model has one input and five outputs.
∑
𝑛 In the second type of the inverse function model shown in Fig. 5,
𝑛𝑒𝑡𝑘 = (𝑤𝑗𝑘 𝑦𝑗 + b𝑘 ) (4) the output of the direct function is merged with four virtual functions
𝑗=1
to become the input variables for the direct function model. Therefore,
𝑦𝑘 = 𝑔(𝑛𝑒𝑡𝑘 ) (5) there are five input variables that comprise the output of the direct
function model and virtual functions. In this second type of inverse
where i = ith node in the input layer, j = jth node in the hidden function model, the number of the output variables must be equal to the
layer, k = kth node in the output layer, and y = output. b = bias, number of the input variables to have the exact solution of the neural
wij = connecting weight or strength from node i to j. g = activation network [48]. Thus, the output variables of the model are five.
function or non-linear activation function and wjk = connecting weight The matrices for the inputs and outputs variables can be given as:
or strength from node j to k. ( )
⎡y1 ⎤ ⎡ g1 x1,1 , x1,2 , x1,3 , … , x1,m ⎤
The backward pass or back propagation in multilayer neural net- ⎢ ⎥ ⎢ ⎥
work is based on the principle of gradient descent principle below: ⎢⋮⎥ = ⎢ ⋮ ⎥ (14)
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ( ) ⎥
𝑤𝑖𝑗 (𝑡 + 1) = 𝑤𝑖𝑗 (𝑡) + 𝛥𝑤𝑖𝑗 (gradient descent) (6) ⎣yn ⎦ ⎣ gn xn,1 , x𝑛,2 , x𝑛,3 , … , xn,m ⎦

7
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 2. Dataset Virtualization (a) speed;(b) mileage;(c) current; (d) SOC; (e) voltage.

Fig. 3. Direct function model.

8
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 4. Inverse function model without dummy function.

Fig. 5. Inverse function with dummy functions.

where n = number of input variables and m = the number of output Table 4

variables. Variables and their units.

Using Axion of set, we can also determine the union between the Parameters Units

variables in the inverse function case. Given that a set of input variables Virtual function from voltage volts
is Virtual function from speed km/h
Virtual function from mileage km
{𝑌 } = {𝑦0 , … , 𝑦4 } = {𝑦0 }𝑈 … 𝑈 {𝑦4 } (15) Virtual function from current Amperes
Virtual function from soc %
Similarly, the set of output variables can be given as Speed km/h
Mileage km
{𝑋} = {𝑥1 , … , 𝑥5 } = {𝑥1 }𝑈 … 𝑈 {𝑥5 } (16) Voltage volts
Current Amperes
where {X} is a set of the output variables for the inverse function model SOC %

X ∈ Rn and y ∈ Rn
R are real numbers
(D) is selected in such that values of the virtual function is greater than
3.8. Calculation of the virtual functions 0.
𝑋𝑚𝑎𝑥(1) 𝑋𝑚𝑎𝑥(2) 𝑋𝑚𝑎𝑥(4)
𝑦1 = ( − 𝑥1 )2 , 𝑦2 = ( − 𝑥2 )2 , … , 𝑦4 = ( − 𝑥4 )2 (18)
The virtual functions, or dummy variables, were needed for aug- 2 2 2
mentation so that the number of target variables would be equal to the
number of input variables for the inverse function model. Otherwise, 𝑋𝑚𝑎𝑥(1) 𝑋𝑚𝑎𝑥(2) 𝑋𝑚𝑎𝑥(𝑛−1)
𝑦1 = ( − 𝑥1 )2 , 𝑦2 = ( − 𝑥2 )2 , … , 𝑦𝑛−1 = ( − 𝑥𝑛−1 )2
the inverse neural network model would not predict properly, as ini- 2 2 2
tially stated. The virtual functions were calculated using the equations (19)
below [28].
𝑋𝑚𝑎𝑥(𝑖) Table 4 Contains the details of the virtual functions and other
𝑦𝑖 = ( − 𝑥𝑖 )2 𝑦𝑖 > 0 (17) variables
𝐷
where i = 1,2. . . ., n − 1.
𝑥𝑖 = variable of the BEVs, 𝑦i = virtual function value for a particular 3.9. Model formation for the inverse function model
variable,
n = number of input variables and 𝑋max = The maximum data value The components of the proposed model are shown in Table 5 and
for a particular variable. D = Virtual function denominator factor. D Fig. 6. There are 128 neurons, 64 neurons, 32 neurons, and 5 neurons
is assigned the value of 2 in this study. Eq. (17) is used to generate in the first hidden layer, second hidden layer, third hidden layer, and
Eqs. (18) and (19). The value of the virtual function denominator factor outer layer of the proposed model, respectively. The model was built

9
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 6. Deep neural network model.

with the aid of Python libraries. The dimension of the input layer is Table 5
Components of the proposed model.
five because there are five input variables. The metric for assessment
Model parameters No of neurons Activation function Initializer
of the model was mean square error. Adam was used to optimize the
1st hidden layer 128 neurons Relu Normal
weight. Adam was equally used for tuning the hyperparameters. Relu
2nd hidden layer 64 neurons Tanh Normal
was the activation function in the first hidden layer, while Tanh was 3rd hidden layer 32 neurons Tanh Normal
the activation function for the second and third hidden layers. Outer Layers 5 neurons Normal
Input layer dimension 5
Rectified linear unit (ReLu) activation function is suitable for deep
Regression loss function Mean square error
neural networks in the hidden layer. ReLu is superior to sigmoid Optimizer Adam
and other previously introduced activation functions in training deep
neural networks [49]. It approximates faster than all other activation
functions. That is why it is commonly used for training deep neural
In this study, different case studies were for the simulation of the
networks. That is the reason for choosing it for the first hidden layer.
inverse function. The case studies are listed below.
ReLu approximates values from 0 to infinity. The virtual functions have
non-negative values. This makes ReLu activation function suitable for Case I: Combination of SOC (the predicted output of the single output
mapping any values of the virtual functions. The only problem with model) and four virtual functions as the inputs of the proposed model
ReLu is that it considers negative values to be zero. This will make the Case II: Combination of current (the predicted output of the single
gradient zero, and the neurons will be unable to respond to variations in output model) and four virtual functions as the inputs of the proposed
error. It is, however, suitable for the hidden layer. The ReLu activation model
function can be expressed as
Case III: Combination of voltage (the predicted output of the single
𝑔(𝑥) = 𝑚𝑎𝑥(0, 𝑥) (20) output model) and four virtual functions as the inputs of the proposed
{ model
𝑥 𝑥≥0
𝑔(𝑥) = Case IV: Combination of speed (predicted output of the single output
0 𝑥<0 model) and four virtual functions as the inputs of the proposed model
where g = activation function. Case V: Combination of mileage (the predicted output of the single
The Tanh activation function was chosen for the second and third output model) and four virtual functions as the inputs of the proposed
hidden layers due to negative values. The hyperbolic tangent activation model
function approximates values from −1 to +1. The activation function
Case VI: SOC (predicted output of the single output model) with no
will map the negative inputs as strongly negative, and the zero inputs virtual functions
will be mapped close to zero. One of the selected parameters (current)
for training the proposed model comprises negative and positive values. Case VII: Current (predicted output of the single output model) with
The Tanh activation is suitable for mapping the negative values of no virtual functions
such a parameter. The hyperbolic tangent activation function is also Case VIII: Voltage (predicted output of the single output model) with
characterized by high accuracy. This is another reason for choosing it no virtual functions
in the second and third hidden layers. Both ReLu and Tanh activation
Case IX: Speed (predicted output of the single output model) with no
functions are superior to sigmoid. They can be used for training mul- virtual functions
tilayer neural networks. The hyperbolic tangent function is given as:
Case X: Mileage (predicted output of the single output model) with no
virtual functions
𝑥 1 − 𝑒−𝑥
𝑡𝑎𝑛ℎ( ) = (21) The input and the output variables are listed below.
2 1 + 𝑒−𝑥
Input Variables for the inverse Model
The proposed inverse function model implemented feedforward 𝑦0 = Output variable from the direct model
back propagation algorithms. 𝑦1 = Calculated virtual function from state of charge

10
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 6
Summary of the inputs and output variables for the case studies.
Case studies Input variables Predicted outputs Predicted variables
Case I Direct model output (SOC) + (𝑦2 , 𝑦3 , 𝑦4 , 𝑦5 ) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case II Direct model output (Current) + (𝑦1 , 𝑦3 , 𝑦4 , 𝑦5 ) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case III Direct model output (Voltage) + (𝑦1 , 𝑦2 , 𝑦4 , 𝑦5 ) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case IV Direct model output (Mileage) + (𝑦1 , 𝑦2 , 𝑦3 , 𝑦5 ) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case V Direct model output (Speed) + (𝑦1 , 𝑦2 , 𝑦3 , 𝑦4 ) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case VI Direct model output (SOC) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case VII Direct model output (Current) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case VIII Direct model output (Voltage) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case IX Direct model output (Mileage) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5
Case X Direct model output (Speed) SOC, Current, Voltage, Mileage & Speed 𝑋1 , 𝑋2 , 𝑋3 , 𝑋4 𝑋5

𝑦2 = Calculated virtual function from current By integrating both sides

𝑦3 = Calculated virtual function from voltage 𝑡=0
1
𝑦4 = Calculated virtual function from mileage 𝑉 = [𝐹 − (𝐹𝑎𝑑 + 𝐹𝑟𝑜𝑙𝑙 + 𝐹ℎ𝑖𝑙𝑙 + 𝐹𝑎𝑛𝑔 )]dt (26)
𝑚 ∫𝑡
𝑦5 = Calculated virtual function from speed
where V is the linear velocity, m is the mass of the car, F is the tractive
Output Variables for the inverse Model force, Fad is the aerodynamic force, Froll is the rolling force and Fhill is
𝑋1 = State of charge as the output variable of the inverse model the force to climb the hill, Fang is the angular acceleration force. The
𝑋2 = Current as the output variable of the inverse model initial velocity = 0 at time 𝑡 = 0 and the final velocity attain a certain
𝑋3 = Voltage as the output variable of the inverse model value at time t = t
𝑋4 = Mileage as the output variable of the inverse model
𝑋5 = Speed as the output variable of the inverse model 3.10.4. Modeling of energy and efficiencies
The summary of the inputs and outputs for case studies are dis- Energy management is crucial in electric vehicles. The battery is
played in Table 6 below. the source of tractive energy for battery electric vehicles, as shown
in Fig. 7. To determine this tractive energy, the efficiencies of some
3.10. Modeling of parameters critical components are required. The three efficiencies associated with
BEV power are battery efficiency, electric motor efficiency, and gear
3.10.1. Mileage of a battery efficiency. These are shown in Fig. 7.
Driving range(mileage) is a critical parameter in finding the change Electric Motor Efficiency (𝜂𝒎 )
trajectory of the state of charge of a battery. If the value of the driving The efficiency of the DC electric motor and its controller varies with
range is known, the trajectory of the state of charge can be determined the motor size, torque, and power. The efficiency of the motor power
linearly [50]. If the driving range is known (by the historical datasets and the controller can be considered together and can be modeled as
or the Global Positioning System), the state of charge reference value [51].
at any position ‘s’ can be calculated by the equation [50]: Power output from the motor
𝜂𝑚 = (27)
(𝐷 − 𝐷𝐼 (𝑠)) Power input into the motor
𝑆𝑂𝐶 𝑅𝑒𝑓 (𝑠) =[ 𝑇 .(𝑆𝑂𝐶0 − 𝑆𝑂𝐶𝐹 )] + 𝑆𝑂𝐶𝐹 (22)
𝐷𝑇 𝑃𝑚−𝑜𝑢𝑡
𝜂𝑚 = (28)
𝑃𝑚−𝑖𝑛
3.10.2. Work done by the electric current or charges
where 𝜂𝑚 𝑖𝑠 the motor efficiency, 𝑃𝑚−𝑖𝑛 is 𝑡ℎ𝑒 motor power input and
Energy is used whenever work is done. If the potential difference
𝑃𝑚−𝑜𝑢𝑡 is the motor power output.
applied is Vb (in volts) and the quantity of electricity that passes from
the battery to the electric motor is Q (coulombs), the work done (W) Power output from the motor
𝜂𝑚 = (29)
would be given as (Power output from the motor) + Losses
The power losses in the electric motor are divided into four cate-
𝑊 = 𝑉𝑏 𝑄 (in joules) (23) gories which are discussed below [51].
But quantity of electricity Q = It (where I is the current and t is the ∙ Copper losses is due to the resistance of the wire and carbon
time) brushes which causes heat energy. Power = I2 R and because current
is proportion to the torque, the copper losses is given by
𝑊 = 𝐼𝑉𝑏 𝑡 (24)
𝐿c = 𝑘1 𝑇 2 (30)
From ohms law, 𝑉b = IR or I = 𝑉b /R. Thus, the expressions for work
where 𝑘1 is constant and it depends on the resistance of the wire and
done (or electrical energy) can be given in the following forms:
the carbon brushes. 𝑇 is the torque
𝑊 = 𝐼𝑉b 𝑡 (24a) ∙ Iron losses is caused by the magnetic force or effects in the iron of
2 the rotor of electric motor, and it is given as
𝑊 = (𝑉b ) 𝑡∕𝑅 (24b)
𝐿i = 𝐾2 𝜔 (31)
𝑊 = 𝐼 2 𝑅𝑡 (24c)
where 𝑘2 is a constant and it depends on the frequency of the changes
where I is the current, 𝑉b is the voltage, R is the resistance and t is the
in the magnetic flux. 𝜔 is the angular speed
time in seconds.
∙ Losses due to friction and windage resistance. There are frictional
losses in the bearing and brushes of the motor. The windage loss is due
3.10.3. Modeling the velocity (or speed) of the vehicle to the resistance caused by the fan attached to the rotor for cooling.
The linear velocity of a battery electric vehicle is determined from Friction loss is proportional to the torque, and it is constant, while the
the linear acceleration force by integrating the linear acceleration. It windage resistance is proportional to the square of the angular speed
can be given as 𝜔.
𝑑𝑣
𝑚 = 𝐹 − (𝐹𝑎𝑑 + 𝐹𝑟𝑜𝑙𝑙 + 𝐹ℎ𝑖𝑙𝑙 + 𝐹𝑎𝑛𝑔 ) (25) Windage loss L3 = 𝑘3 𝜔3 (32)
𝑑𝑡

11
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 7. Flow of energy consumption in a battery powered BEV with regenerating baking.

∙ The type of loss is the power used by the electrical control unit. The power input of a DC electric motor can also be given as [52]
Power is always needed by the electrical control unit, and it is loss Pm−out 𝐹𝑡 𝑉
attributed to this is given by C. 𝑃𝑚−𝑖𝑛 = = (46)
𝜂m 3600 𝜂𝑇 𝜂𝑚
𝑘1 𝑇 2 + 𝐾2 𝜔 + 𝑘3 𝜔3 + C (33) where 𝜂𝑇 Is the mechanical efficiency of the transmission system, 𝜂𝑚
But is the efficiency of the electric motor, 𝐹𝑡 is the tractive force, 𝑉 is the
Power output from the motor speed of the car.
𝜂𝑚 = (34) 𝐹𝑉
(Power output from the motor) + Losses Substituting for 𝑃𝑚−𝑖𝑛 = 3600𝜂𝑡 𝜂 in Eq. (46) above
𝑇 𝑚
Torque x angular speed 𝑡 t 𝐹𝑡 𝑉
𝜂𝑚 = (35)
(Torque x angular speed) + ( Losses) 𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = 𝑑𝑡 (47)
∫0 ∫0 3600 𝜂𝑇 𝜂𝑚 𝜂b
T𝜔 𝑡 t
𝜂𝑚 = (36) 1
T𝜔 + (𝑘1 𝑇 2 + 𝐾2 𝜔 + 𝑘3 𝜔3 + C) 𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = 𝐹 𝑉 𝑑𝑡 (48)
∫0 3600 𝜂𝑇 𝜂𝑚 𝜂b ∫0 𝑡
Gear Efficiency (𝜂𝒈 ) But tractive force is given as.
Assuming there is only one gear in the gear system of BEVs, the
1 𝐺2 𝑑𝑣
efficiency of the gear system (𝜂𝑔 ) is constant. It can be given as [51]. 𝐹t = [μmg + 𝜌 CD AF V2 + (m + I ) ] (49)
2 𝜂𝑔 𝑟2 𝑑𝑡
Power output from the motor
𝜂𝑔 = (37) V = velocity of the car, G = gear ratio, m = mass of the vehicle, g
Tractive power at the wheel
𝑃𝑚−𝑜𝑢𝑡 = acceleration due to gravity, 𝜇 = Rolling resistance coefficient, CD =
𝜂𝑔 = (38) aerodynamic coefficient, AF = frontal area in m2 , a = acceleration of
𝑃𝑡𝑟𝑎𝑐𝑡𝑖𝑣𝑒
the vehicle, I = moment of inertial, r = radius of the tyre, 𝜂𝑔 = gear
Efficiency of the Battery (𝜂𝒃 )
system efficiency, and 𝜌 = air density.
The battery power = 𝑃𝑚−𝑖𝑛 + 𝑃𝑎𝑐𝑐𝑒𝑠𝑠 (39) Substituting for tractive force in Eq. (48) above.
𝑡 t
where 𝑃𝑎𝑐𝑐𝑒𝑠𝑠 is the power used by the accessories such as horn and 1 1
𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = [μmg + 𝜌 𝐶𝐷 𝐴𝐹 𝑉 2
lights, and 𝑃𝑚−𝑖𝑛 is the power input to the electric motor. ∫0 3600 𝜂𝑇 𝜂𝑚 𝜂b ∫0 2
Power output from the battery 𝐺2 𝑑𝑣
𝜂𝑏 = (40) + (m + I ) ]𝑉 𝑑𝑡 (50)
Power input into the battery 𝜂𝑟2 𝑑𝑡
𝑡 t
𝑃𝑚−𝑖𝑛 + 𝑃𝑎𝑐𝑐𝑒𝑠𝑠 1 1
𝜂𝑏 = (41) E= 𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = [μmg + 𝜌 𝐶𝐷 𝐴𝐹 𝑉 2
𝑃𝑏−𝑜𝑢𝑡 ∫0 3600 𝜂𝑇 𝜂𝑚 𝜂b ∫0 2
and 𝑃𝑏−𝑜𝑢𝑡 is the output power from the battery 𝐺2 𝑑𝑣
+ (m + I ) ]𝑉 𝑑𝑡 (51)
𝑃 + 𝑃𝑎𝑐𝑐𝑒𝑠𝑠 𝜂𝑔 𝑟2 𝑑𝑡
𝑃𝑏−𝑜𝑢𝑡 = 𝑚−𝑖𝑛 (42) t
𝜂𝑏 1 1
Energy E = [μmg + 𝜌𝐶𝐷 𝐴𝐹 𝑉 2
3600 𝜂𝑇 𝜂𝑚 𝜂b ∫0 2
Modeling of Energy Consumption
𝐺2 𝑑𝑣
𝑃𝑚−𝑖𝑛 + 𝑃𝑎𝑐𝑐𝑒𝑠𝑠 + (m + I ) ]𝑉 𝑑𝑡 (52)
𝑃𝑏−𝑜𝑢𝑡 = (43) 𝜂𝑔 𝑟2 𝑑𝑡
𝜂𝑏
By integrating both sides Substituting for
𝑡 t T𝜔
Pm−in + Paccess 𝜂𝑚 = in Eq. (52) above
𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = 𝑑𝑡 (44) T𝜔 + (𝑘1 𝑇 2 + 𝐾2 𝜔 + 𝑘3 𝜔3 + C)
∫0 ∫0 𝜂b
T.𝜔 + 𝑘1 𝑇 2 + 𝐾2 𝜔 + 𝑘3 𝜔3 + C t
1
If Paccess is assumed to be zero for experimental purposes, Equation Energy E = [μmg + 𝜌 𝐶𝐷 𝐴𝐹 𝑉 2
3600 𝜂𝑇 𝜂b T.𝜔 ∫0 2
above becomes [52]
𝑡 t 𝐺2 𝑑𝑣
Pm−in + (m + I ) ]𝑉 𝑑𝑡 (53)
𝑃𝑏−𝑜𝑢𝑡 𝑑𝑡 = 𝑑𝑡 (45) 𝜂𝑔 𝑟2 𝑑𝑡
∫0 ∫0 𝜂b

12
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 7
Testing and training errors.
Input variables Testing errors Training errors
MAE MSE RMSE MAE MSE RMSE
Case I (SOC + dummy functions) 0.08286 0.02184 0.1478 0.04654 0.0048311 0.06951
Case II (current + dummy functions) 0.1049 0.05098 0.2258 0.05216 0.006912 0.08314
Case II (voltage + dummy functions) 0.1982 0.1725 0.4153 0.0991 0.03257 0.1805
Case IV (speed + dummy functions) 0.1193 0.06093 0.2468 0.08176 0.02311 0.15203
Case V ( mileage + dummy functions) 0.09222 0.03479 0.1865 0.04473 0.004735 0.06927

It is assumed that the grade angle is zero. i.e., the vehicle is moving
on a flat road without any angle of inclination.
where T = torque, 𝜔 = angular speed, 𝑘1 = copper loss constant, 𝐾2
= Iron losses constant, 𝑘3 = windage loss constant, C = power used by
the control unit, 𝜂𝑇 = mechanical efficiency of the transmission system,
and 𝜂b = Efficiency of the battery.

4. Results

4.1. Results of the predictions of the inverse function model with virtual
functions

This section contains the details of the predictions for the case
studies where dummy or virtual functions were employed. In each
of the case studies, five variables were predicted. These variables are
voltage, mileage, SOC, current, and speed. The results of the case
studies are discussed below. Table 7 contains the results of Cases I to
V where virtual functions were employed. Fig. 8. Model losses behaviors for current (the predicted output of direct model) and
virtual functions as inputs variables.
The proposed model was evaluated using the following error func-
tions :
1∑
𝑛
𝑀𝐴𝐸 = |𝑋 − 𝑋𝑇 𝑎𝑟𝑔𝑒𝑡𝑒𝑑(𝑖) | (54) behaviors associated with the losses are shown in Fig. 8 above. Table 8
𝑛 𝑖=1 𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑(𝑖) illustrates the details of the predicted and targeted variables.
Fig. 9(a, b, c, d, and e) display estimated and targeted variables
1∑
𝑛
for case 1. The degree of overlap between the predicted and target
𝑀𝑆𝐸 = (𝑋 − 𝑋𝑇 𝑎𝑟𝑔𝑒𝑡𝑒𝑑(𝑖) )2 (55)
𝑛 𝑖=1 𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑(𝑖) variables shows the level of accuracy. Based on the level of overlap,
√ accuracy is higher in Fig. 9 (b, c, d, and e) than in Fig. 9(a).
√
√ 1∑ 𝑛
𝑅𝑀𝑆𝐸 = √ (𝑋 − 𝑋𝑇 𝑎𝑟𝑔𝑒𝑡𝑒𝑑(𝑖) )2 (56)
𝑛 𝑖=1 𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑(𝑖) 4.1.2. Prediction of the inverse function model for case of current (the
output of direct model) and virtual functions.
4.1.1. Prediction of the inverse function model for the case study of SOC Fig. 10 displays the model loss behaviors when SOC (the predicted
(the predicted output of the direct model) and virtual functions output of the direct model) and virtual functions were input variables
Fig. 8 shows the graph of losses against the epoch. One epoch for the proposed inverse model. The figure shows the changing at-
implies that all the datasets for the training have been completely tributes of the model’s testing and training losses. The reducing pattern
used to train the neural network for one cycle [53]. It will take a is noticed from 0 to about 80 epochs in both testing and training
lot of epochs to train a neural network. An epoch is made up of one errors. This indicates a good level of accuracy in the prediction of
or more batches of data points. Training one batch of data points is the output variables. The reduction in losses with increasing epoch
a process known as iteration. In this study, the batch size used is 5. is directly proportional to the model’s accuracy and precision. The
Under normal conditions, training and testing losses should decrease reducing pattern is more pronounced in the training errors than the
until the neural network converges. The training must stop at this point testing errors.
of convergence, otherwise, the ANN will begin to overfit [53]. This is
known as early stopping, a regularization method. Additional training 4.1.3. Prediction of the inverse function model for case of voltage (the
after convergence may result in an increase in testing errors. In this output of direct model) and virtual functions
study, 100 epochs were found suitable for early stopping. For the case study of voltage (the output of the direct model) and
For the case study of current (output of the direct function model) virtual functions, the behavior of the training and testing losses can be
and virtual functions as inputs, the behavior of the training loss can seen to have a reducing pattern from 0 to 100 epochs. However, there
be seen to be reducing as the number of epochs increases from 0 to were occasional rises in the error values. For instance, at about epoch
80. It then rises at 90 and falls again as it approaches 100 epochs. 62, the testing error rises significantly. Notwithstanding, the error
Also, the testing loss is reducing as the number of epochs increases values kept on reducing after this point. The pattern of the learning
from 0 to 70. It begins to rise at 75 epochs and continues to fall from curve implies a reduction in losses. A reduction in losses signifies an
80 to 100 epochs. The overall decreasing pattern of the learning curve increase in the accuracy and precision of the predictions. It should
losses demonstrates a high degree of accuracy in prediction based on be noted that the reduction in losses with increasing epoch is directly
the behavior of the training and testing errors. Whenever the learning proportional to the model’s accuracy and precision. The behavior of
curve shows a decrease in losses as the epoch increases, it indicates the training and testing errors shows a high level of accuracy in the
an increase in accuracy. Reduction in losses with increase in epoch prediction of the output variables. The behaviors associated with the
is directly proportional to the model’s accuracy and precision. The losses are shown in Fig. 11 below.

13
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Table 8
Predicted and targeted variables.
Estimated speed Speed Estimated mileage Mileage Estimated soc SOC Estimated voltage Voltage Estimated current Current
1 40.272 49.6 120 581.37 120 581 90.628 91 360.078 360.2 8.244 8.4
2 28.250 25.2 120 576.05 120 576 92.998 93 359.583 359.3 37.683 38.1
3 44.155 37.6 120 585.11 120 585 88.986 89 354.330 354.1 51.540 52.3
4 15.872 18 120 562.14 120 562 98.823 99 368.942 369 7.093 8.3
5 80.648 79.2 120 569.77 120 572 95.560 94 357.121 356.1 70.598 65.5
6 35.448 34.7 120 575.08 120 575 93.322 94 363.367 363.3 2.442 2.1
7 37.481 50.1 120 591.53 120 592 86.171 86 356.619 356.4 4.369 2.8
8 38.586 27.2 120 577.94 120 578 92.209 92 359.167 358.9 31.356 33.6
9 30.584 32.8 120 589.42 120 590 87.280 88 357.921 358.2 1.013 1
10 12.008 12.5 120 599.16 120 599 83.114 83 354.507 354.5 2.534 1.2

Fig. 9. (a, b, c, d, and e) Predictions for Case 1.

4.1.4. Prediction of the inverse function model for case of speed(the output as input variables of the proposed inverse model. The training loss was
of direct model) and virtual functions noticed to be decreasing as the epoch increased from 0 to 100 epochs,
Fig. 12 represents the model loss behaviors when voltage (the while the testing loss was fluctuating in nature. The value of the testing
predicted output of the direct model) and virtual functions were used error was found to be higher at epoch 30 and dropped significantly

14
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 10. Model loss behaviors for SOC (the predicted output of the direct model) and
virtual functions as input variables.
Fig. 12. Model losses behaviors for speed (predicted output of the direct function
model) and virtual functions as inputs variables.

Fig. 11. Model losses behaviors for voltage (output of direct function model) and Fig. 13. Model losses behaviors for mileage (output of direct function model) and
virtual functions as inputs variables. virtual functions as inputs variables.

high degree of accuracy. This is due to the linear relationship between

between 30 and 100 epochs. Between the 30 and 100 epochs, there reduction in model losses and precision as epoch increases.
is a significant reducing pattern. This also signifies good precision and
accuracy in the estimation of output variables. Whenever the learning
curve shows a decrease in losses as the epoch increases, it indicates 4.1.6. Predictions of the inverse function model (without virtual function)
a high level of accuracy. There is a linear relationship between the This section contains the details of the results of the inverse model
reduction in losses and accuracy as the number of epochs increases. The
that did not implement virtual functions. Five variables were predicted
reduction in model losses with increasing epoch is directly proportional
in each of the case studies. These variables are voltage, mileage, SOC,
to the accuracy and precision of the model. As the number of epochs
current, and speed.
increases, the artificial neural network is moving towards convergence.
The testing and training losses behaviors for the inverse function
4.1.5. Prediction of the inverse function model for case of mileage (the model without virtual functions are depicted in Fig. 14(a, b, c, d, e). In
output of direct model) and virtual functions each of the figures, the training errors are found to be decreasing from
Fig. 13 displays the model loss behaviors when mileage (the output 0 to 100 epochs. In Fig. 14(d & e), the testing errors are noticed to be
of the direct model) and virtual functions were input variables of the decreasing as the number of epochs increases from 0 to 100. In Fig. 14
inverse function model. The training and testing losses increase after 10 (a, b, & c), the testing errors were found to be fluctuating. Table 9
epochs and then decrease from 15 to 100 epochs. This illustrates accu- illustrates that the errors found in each of the cases are significantly
racy in the estimation of the output variables. Whenever the learning high. It means that the accuracy of the inverse function model would
curve shows a decrease in losses as the epoch increases, it indicates a not be as high as that of the model that implemented virtual functions.

15
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 14. Model loss behaviors for the inverse function model without virtual functions.

Table 9 MSE, and RMSE are found to be highest for Case III (voltage + dummy
Results of the predictions.
function). Among all the case studies in the inverse function model,
Inputs variables Testing errors Training errors
Case I (SOC + dummy function) has the lowest MAE, MSE, and RMSE.
MAE MSE RMSE MAE MSE RMSE This implies that Case I (SOC + dummy function) has the highest
(Case VI) SOC only 0.8606 0.9703 0.985 0.8265 0.9745 0.9872 accuracy value for predicting the inverse functions, while the former
(Case VII) Current only 0.886 1.0378 1.0187 0.8213 0.9751 0.9874
has the lowest accuracy.
(Case VIII) Voltage only 0.8743 1.0032 1.0016 0.8361 0.9914 0.9957
(Case XI) Speed only 0.8664 1.0026 1.0013 0.8237 0.9634 0.9815 The accuracy in terms of MSE for Case I (SOC + dummy functions)
(Case X) Mileage only 0.8596 0.9701 0.9849 0.8292 0.9778 0.9889 is 2.33 times higher than for Case II (current + dummy functions) and
7.90 times higher than for Case III (voltage + dummy function). More
so, accuracy for Case I (SOC + dummy functions) is 2.79 times higher
5. Discussions than for Case IV (speed + dummy functions), and it is 1.59 times higher
than that of Case V (mileage + dummy function). Fig. 15(b) shows the
5.1. Comparative analysis for cases(I to V) with virtual functions comparative analysis in terms of MSE. Again, Case I (SOC + dummy
functions) has the highest accuracy when compared with all other case
This section contains the details of the analysis of the results. studies in terms of MSE. It has the lowest value of MSE as shown in
Fig. 15(a) compares the testing errors for all the case studies. MAE, Fig. 15(b)

16
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 15. Comparative analysis for the testing errors.

Table 10 compares the error functions for the case studies(I to Table 10
V) when the target variables were redefined. The results show that Testing errors for the redefined parameters.

the highest values of error functions were found when voltage (the Input Variables Testing errors

predicted output of the direct model) and virtual functions were input MAE MSE RMSE
variables. Also, the lowest values of errors were found when SOC (the Case I 0.09068 0.02723 0.165
predicted output of the direct model) and virtual functions were the Case II 0.1294 0.07936 0.2817
input variables. This denotes that SOC (the predicted output of the Case III 0.2151 0.2032 0.4507
Case IV 0.1285 0.07067 0.2658
direct model) and virtual functions as input variables in the case study Case V 0.096697 0.039809 0.19952
have the highest accuracy value for predicting the inverse functions,
while the lowest accuracy was found when voltage (the predicted
output of the direct model) and virtual functions were input variables.
was used as an input variable, the accuracy of Case III (voltage +
5.2. Comparativeness between case studies with virtual functions and case dummy functions) is 5.812 times higher than that of Case VIII (without
studies without virtual functions virtual functions). Case IV (speed + virtual functions) is 16.46 times
more accurate than Case IX (without virtual functions) when speed is
For cases where SOC was used as an input, the accuracy of Case I used as an input. For cases of mileage, the accuracy of Case V (mileage
(SOC + dummy function) is 44.43 times higher than Case VI (without + dummy functions) is 27.88 times higher than that of Case X (without
virtual functions). For cases with current as input, the accuracy of Case virtual functions). The highest accuracy is found in Case I. Accuracy is
II (current + dummy functions) is 20.357 times higher than the accu- measured in terms of MSE. Fig. 16 shows the details of the comparative
racy of Case VII (without virtual functions). For cases where voltage analysis of the 10 case studies.

17
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

Fig. 16. Comparative analysis for the case studies with and without virtual functions.

Fig. 17. Functionality of the model.

5.3. Implication of the proposed feedforward deep inverse function model 5.5. Benefits of the proposed model to manufacturers of batteries and
electric vehicles.
The proposed feedforward deep inverse function model has several
applications both in academia and industry. In the automobile industry, In this study, current, voltage, and SOC are classified as battery pa-
the feedforward inverse function model can be used for the design rameters. Mileage and speed are classified as electric vehicle operation
simulation of lithium-ion batteries for electric vehicles. The ability of parameters. On the basis of these classifications, the functionality of the
the model to simulate parameters like state of charge, voltage, current, model can be divided into two functions: battery parameter estimations
and mileage makes it an indispensable simulating machine for battery and electric vehicle parameter estimations. This functionality is shown
and automobile manufacturers. In academia, it is an indispensable in Fig. 17.
breakthrough in the area of artificial neural networks. It makes artificial
The proposed model would be beneficial to battery and electric
neural networks useful as multifunctional simulation tools. Also, the
vehicle manufacturers. This is because the proposed model can predict
proposed model is useful in the laboratory for design simulations in
vital battery and electric vehicle parameters. Parameters like state
automobile research.
of charge, voltage, speed, mileage, and current can be predicted or
5.4. Advantages of the proposed feedforward deep inverse function model estimated using the proposed model. Another major benefit is that
over others the five parameters can be simulated at once. This makes decision-
making easier than when only one parameter is simulated by most of
One of the benefits of the proposed feedforward deep neural net- the previously proposed models in the literature.
work model is its multi-functionality. For instance, the proposed DNN One of the key factors in electric vehicle design is the estimation
by Premkumar et al. [24] can only predict the state of charge of a of mileage. With this proposed model, the manufacturers of batteries
lithium battery, while the proposed model can predict the state of and electric vehicles would be able to estimate the mileage that a
charge, voltage, current, and mileage of a lithium-ion battery con- battery can cover during the design simulation stage. Another critical
currently. This makes the proposed model superior to the previously parameter is the state of charge. The state of charge determines the
introduced model. mileage that an electric vehicle can cover. This capability makes this

18
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

proposed model so useful and beneficial for design simulations and the [5] Battery Solution for a Long and a Safe Battery Life, Battery Solution
operations of EVs. for a Long and a Safe Battery Life, Technical Article, 2022, www.
powerelectronicsnews.com/bms-solutions-for-a-long-and-safe-batterylife/#:~:
Comparing this proposed model with the previous models in the
text=To%20measure%20the%20temperature%20in%20and%20around%20the,
literature, it is evident that the former is more superior. It is superior to%20protect%20the%20battery%20from%20overheating%20during%20work.
in simultaneously predicting mileage, voltage, SOC, current, and speed. Accessed on 04 July 2022.
This is not obtainable for most simple models that have been proposed [6] C. Vidal, O. Gross, R. Gu, P. Kollmeyer, A. Emadi, xEV Li-ion battery
low-temperature effects, IEEE Trans. Veh. Technol. 68 (5) (2019) 4560–4572.
in the literature for battery parameter simulations.
[7] Environment Health and Safety, Lithium Battery Safety, University of
Washington, 2021.
6. Conclusion [8] G.A. Trinandana, A.W. Pratama, E. Prasetyono, D.O. Anggriawan, Real time state
of charge estimation for lead acid battery using artificial neural network, in: 2020
International Seminar on Intelligent Technology and its Applications, ISITIA,
This study introduces an approach for the prediction of multiple
IEEE, 2020, pp. 363–368.
outputs of battery parameters for operational and design simulation [9] D.N. How, M.A. Hannan, M.S.H. Lipu, K.S. Sahari, P.J. Ker, K.M. Muttaqi, State-
purposes. Artificial neural networks were used for the simulation of of-charge estimation of li-ion battery in electric vehicles: A deep neural network
five critical parameters selected for an electric vehicle and a lithium approach, IEEE Trans. Ind. Appl. 56 (5) (2020) 5565–5574.
[10] A.D. Alvarez, F.S. Garcia, J.E. Naranjo, J.J. Anaya, F. Jimenez, Modeling the
battery. These parameters are the state-of-charge of the battery, volt-
driving behavior of electric vehicles using smartphones and neural networks,
age, current, speed, and mileage. The study implemented an inverse IEEE Intell. Transp. Syst. Mag. 6 (3) (2014) 44–53.
function model of an artificial neural network to predict multiple out- [11] S. Modi, J. Bhattacharya, P. Basak, Estimation of energy consumption of electric
puts of battery parameters for electric vehicles. Case I (SOC + dummy vehicles using deep convolutional neural network to reduce driver’s range
functions) has the lowest MAE, MSE, and RMSE. This implies that Case anxiety, ISA Trans. 98 (2020) 454–470.
[12] L. Steinbach, M.E. Altinsoy, Prediction of annoyance evaluations of electric
I (with SOC and dummy functions as inputs) has the highest accuracy vehicle noise by using artificial neural networks, Appl. Acoust. 145 (2019)
value for predicting the inverse functions. The accuracy was found to be 149–158.
higher during the prediction of the inverse function of the SOC than for [13] A.A. Hussein, Capacity fade estimation in electric vehicle li-ion batteries using
the other nine cases (Case II to X). The proposed model is suitable for artificial neural networks, IEEE Trans. Ind. Appl. 51 (3) (2014) 2321–2330.
[14] Z. Tian, C. Qian, B. Gu, L. Yang, F. Liu, Electric vehicle air conditioning system
lithium-ion battery parameter estimations. Some of the contributions of
performance prediction based on artificial neural network, Appl. Therm. Eng. 89
this study are the following: (2015) 101–114.
[15] S. Ansari, A. Ayob, M.S. Hossain Lipu, A. Hussain, M.H.M. Saad, Multi-
• A multi-output feedforward deep inverse function model of an ar- channel profile based artificial neural network approach for remaining useful
tificial neural network is introduced for predicting the parameters life prediction of electric vehicle lithium-ion batteries, Energies 14 (22) (2021)
7521.
of lithium-ion batteries.
[16] T.Y. Ma, S. Faye, Multistep electric vehicle charging station occupancy prediction
• A novel model that facilitates decision-making in the design using hybrid LSTM neural networks, Energy 244 (2022) 123217.
simulation of lithium-ion batteries is proposed. [17] F. Zhao, Y. Li, X. Wang, L. Bai, T. Liu, Lithium-ion batteries state of charge
• The proposed model is superior to the previously introduced prediction of electric vehicles using RNNs-CNNs neural networks, IEEE Access 8
(2020) 98168–98180.
feedforward deep neural network model for predicting the state
[18] Z. Nie, H. Farzaneh, Real-time dynamic predictive cruise control for enhancing
of charge of a battery. It can predict additional parameters like eco-driving of electric vehicles, considering traffic constraints and signal phase
voltage, current, mileage and speed. and timing (SPaT) information, using artificial-neural-network-based energy
• The proposed feedforward model(with virtual functions) is supe- consumption model, Energy 241 (2022) 122888.
rior to the traditional inverse function model. [19] N. Narasimhulu, R.S.R. Krishnam Naidu, P. Falkowski-Gilski, P.B. Divakarachari,
U. Roy, Energy management for PV powered hybrid storage system in electric
vehicles using artificial neural network and aquila optimizer algorithm, Energies
The proposed model is primarily for a lithium-ion battery. It cannot 15 (22) (2022) 8540.
be used for a lead–acid battery, Sodium sulfide battery and Nickel [20] J. Hong, Z. Wang, W. Chen, L.Y. Wang, C. Qu, Online joint-prediction of
Metal Hydride battery. Future studies may develop similar models for multi-forward-step battery SOC using LSTM neural networks and multiple linear
lead–acid batteries, sodium sulfide battery and nickel metal hydride regression for real-world electric vehicles, J. Energy Storage 30 (2020) 101459.
[21] M. Charkhgard, M. Farrokhi, State-of-charge estimation for lithium-ion batteries
batteries.
using neural networks and EKF, IEEE Trans. Ind. Electron. 57 (12) (2010)
4178–4187.
Declaration of competing interest [22] L. Zheng, Y. Hou, T. Zhang, X. Pan, Performance prediction of fuel cells using
long short-term memory recurrent neural network, Int. J. Energy Res. 45 (6)
(2021) 9141–9161.
The authors declare that they have no known competing finan- [23] Z.H. Wang, Hendrick, G.J. Horng, H.T. Wu, G.J. Jong, A prediction method
cial interests or personal relationships that could have appeared to for voltage and lifetime of lead–acid battery by using machine learning, Energy
influence the work reported in this paper. Explor. Exploit. 38 (1) (2020) 310–329.
[24] M. Premkumar, R. Sowmya, S. Sridhar, C. Kumar, M. Abbas, M.S. Alqahtani, K.S.
Nisar, State-of-charge estimation of lithium-ion battery for electric vehicles using
Data availability deep neural network, CMC-Comput. Mater. Contin. 73 (3) (2022) 6289–6306.
[25] C. Vidal, P. Malysz, M. Naguib, A. Emadi, P.J. Kollmeyer, Estimating battery
Data will be made available on request state of charge using recurrent and non-recurrent neural networks, J. Energy
Storage 47 (2022) 103660.
[26] V. Chandran, C.K. Patil, A. Karthick, D. Ganeshaperumal, R. Rahim, A. Ghosh,
References State of charge estimation of lithium-ion battery for electric vehicles using
machine learning algorithms, World Electr. Veh. J. 12 (1) (2021) 38.
[1] B.P. Adedeji, Parametric predictions for pure electric vehicles, World Electr. Veh. [27] B.P. Adedeji, A novel method for estimating parameters of battery electric
J. 12 (2021) 257, http://dx.doi.org/10.3390/wevj12040257. vehicles, Intell. Syst. Appl. 15 (2022) 200089.
[2] H. Shi, L. Wang, S. Wang, C. Fernandez, X. Xiong, B.E. Dablu, W. Xu, A [28] B.P. Adedeji, A multivariable output neural network approach for simulation of
novel lumped thermal characteristic modeling strategy for the online adaptive plug-in hybrid electric vehicle fuel consumption, Green Energy Intell. Transp. 2
temperature and parameter co-estimation of vehicle lithium-ion batteries, J. (2) (2023) 100070.
Energy Storage 50 (2022) 104309. [29] R. Yang, R. Xiong, S. Ma, X. Lin, Characterization of external short circuit faults
[3] Xiaosong Hu, Changfu Zou, Caiping Zhang, Yang Li, Technological developments in electric vehicle Li-ion battery packs and prediction using artificial neural
in batteries, IEEE Power Energy Mag. (2017) 42–44. networks, Appl. Energy 260 (2020) http://dx.doi.org/10.1016/j.apenergy.2019.
[4] R. Pu, S. Wang, X. Chen, J. Huang, M. He, W. Cao, A novel cuckoo search 114253.
particle filtering strategy for the remaining useful life prediction of the lithium- [30] X. Lin, Z. Wang, J. Wu, Energy management strategy based on velocity prediction
ion batteries in hybrid electric vehicle, Int. J. Energy Res. 46 (15) (2022) using back propagation neural network for a plug-in fuel cell electric vehicle,
21703–21715. Int. J. Energy Res. 45 (2021) 2629–2643, http://dx.doi.org/10.1002/er.5956.

19
B.P. Adedeji and G. Kabir Decision Analytics Journal 8 (2023) 100255

[31] H. Chaoui, C.C. Ibe-Ekeocha, State of charge and state of health estimation for [42] X. Zheng, D. Gao, Z. Zhu, Q. Yang, An early warning protection method for
lithium batteries using recurrent neural networks, IEEE Trans. Veh. Technol. 66 electric vehicle charging based on the hybrid neural network model, World
(10) (2017) 8773–8783, http://dx.doi.org/10.1109/TVT.2017.2715333. Electr. Veh. J. 13 (7) (2022) 128.
[32] C. Vidal, P. Kollmeyer, E. Chemali, A. Emadi, Li-ion battery state of charge [43] H. Abdi, B. Mohammadi-ivatloo, S. Javadi, A.R. Khodaei, E. Dehnavi, Energy
estimation using long short-term memory recurrent neural network with transfer storage systems, Distrib. Gener. Syst. 7 (2017) 333–368.
learning, in: 2019 IEEE Transportation Electrification Conference and Expo, ITEC, [44] M. Ismail, R. Dlyma, A. Elrakaybi, R. Ahmed, S. Habibi, Battery state of charge
IEEE, 2019, pp. 1–6. estimation using an Artificial Neural Network, in: 2017 IEEE Transportation
[33] X. Yang, J. Hu, G. Hu, X. Guo, Battery state of charge estimation using temporal Electrification Conference and Expo, ITEC, IEEE, 2017, pp. 342–349.
convolutional network based on electric vehicles operating data, J. Energy [45] O. Ali, M.K. Ishak, F. Memon, M.S.M. Asaari, Estimation of battery state-
Storage 55 (2022) 105820. of-charge using feedforward neural networks, in: 2022 19th International
[34] S. El Fallah, J. Kharbach, Z. Hammouch, A. Rezzouk, M.O. Jamil, State of Conference on Electrical Engineering/Electronics, Computer, Telecommunications
charge estimation of an electric vehicle’s battery using Deep Neural Networks: and Information Technology (ECTI-CON), IEEE, 2022, pp. 1–4.
Simulation and experimental results, J. Energy Storage 62 (2023) 106904. [46] Z. Chen, N. Guo, J. Shen, R. Xiao, P. Dong, A hierarchical energy management
[35] B. Fu, W. Wang, Y. Li, Q. Peng, An improved neural network model for battery strategy for power-split plug-in hybrid electric vehicles considering velocity
smarter state-of-charge estimation of energy-transportation system, Green Energy prediction, IEEE Access 6 (2018) 33261–33274.
Intell. Transp. (2023) 100067. [47] Chao He, Real-World Data Collected from a BEV Operating on Actual Roads in
[36] Y. Zhang, C. Huang, H. Huang, J. Wu, Multiple learning neural network 2018 and 2019, Mendeley Data, 2021, http://dx.doi.org/10.17632/mm9rvfvmrb.
algorithm for parameter estimation of proton exchange membrane fuel cell 2, V2.
models, Green Energy Intell. Transp. 2 (1) (2023) 100040. [48] H. Shafiee, R.V. Mayorga, A novel method for estimating the inverse function
[37] Z. Zhou, Y. Liu, M. You, R. Xiong, X. Zhou, Two-stage aging trajectory prediction of the black–scholes option pricing model using artificial neural networks, Call
of LFP lithium-ion battery based on transfer learning with the cycle life Pap./Case Stud. 5 (7) (2014) 849.
prediction, Green Energy Intell. Transp. 1 (1) (2022) 100008. [49] J. Loy, Neural Network Projects with Python: The Ultimate Guide to using
[38] W.Y. Alghamdi, A novel deep learning method for predicting athletes’ health Python to Explore the True Power of Neural Networks Through Six Projects,
using wearable sensors and recurrent neural networks, Decis. Anal. J. 7 (2023) Packt Publishing Ltd, 2019.
100213. [50] Y. Ma, P. Wang, T. Sun, Research on energy management method of plug-in
[39] P. Ong, Y.K. Tan, K.H. Lai, C.K. Sia, A deep convolutional neural network for hybrid electric vehicle based on travel characteristic prediction, Energies 14 (19)
vibration-based health-monitoring of rotating machinery, Decis. Anal. J. 7 (2023) (2021) 6134.
100219. [51] J. Larminie, J. Lowry, Electric Vehicle Technology Explained, John Wiley & Sons,
[40] A.K.H. Agbemenou, R. Motamed, A. Talaei-Khoei, A predictive analytics model UK, 2012.
for designing deep underground foundations using artificial neural networks, [52] H. Wei, C. He, J. Li, L. Zhao, Online estimation of driving range for battery
Decis. Anal. J. 7 (2023) 100220. electric vehicles based on SOC-segmented actual driving cycle, J. Energy Storage
[41] D. Kong, W. Wen, R. Zhao, Z. Lv, K. Liu, Y. Liu, Z. Gao, Vehicle lateral velocity 49 (2022) 104091.
estimation based on long short-term memory network, World Electr. Veh. J. 13 [53] Baeldung, Epoch in Neural Network, Baeldung on Computer Science, 2023,
(1) (2022) 1. www.baeldung.com/cs/epoch-neural-networks, 23 March, 2023.

20