Content uploaded by Huseyin Ayhan Yavasoglu
Author content
All content in this area was uploaded by Huseyin Ayhan Yavasoglu on Jan 09, 2023
Content may be subject to copyright.
SPECIAL ISSUE RESEARCH ARTICLE
Neural network-based energy management of multi-source
(battery/UC/FC) powered electric vehicle
Huseyin A. Yavasoglu
1
| Yusuf E. Tetik
1
| Huseyin Gunhan Ozcan
2,3
1
Marmara Research Center, Energy
Institute, Robotics and Automation
Technologies Group, TUBITAK, Kocaeli,
Turkey
2
Department of Energy Systems
Engineering, Yasar University, Izmir,
Turkey
3
Department of Mechanical Engineering,
Porto University, Porto, Portugal
Correspondence
Huseyin A. Yavasoglu, Marmara Research
Center, Energy Institute, Robotics and
Automation Technologies Group,
TUBITAK, Kocaeli, Turkey.
Email: huseyin.yavasoglu@tubitak.gov.tr
Summary
Due to increased environmental pollution and global warming concerns, the use
of energy storage units that can be supported by renewable energy resources in
transportation becomes more of an issue and plays a vital role in terms of clean
energy solutions. However, utilization of multiple energy storage units together in
an electric vehicle makes the powertrain system more complex and difficult to
control. For this reason, the present study proposes an advanced energy manage-
ment strategy (EMS) for range extended battery electric vehicles(BEVs) with com-
plex powertrain structure. Hybrid energy storage system (HESS) consists of
battery, ultra-capacitor (UC), fuel cell (FC) and the vehicle is propelled with two
complementary propulsion machines. To increase powertrain efficiency, traction
power is simultaneously shared at different rates by propulsion machines. Propul-
sion powers are shared by HESS units according to following objectives:extending
battery lifetime, utilizing UC and FC effectively. Primarily, to optimize the power
split in HESS, a convex optimization problem is formulated to meet given objec-
tives that results 5 years prolonged battery lifetime. However, convex optimiza-
tion of complex systems can be arduous due to the excessive number of
parameters that has to be taken into consideration and not all systems are suitable
for linearization. Therefore, a neural network (NN)-based machine learning
(ML) algorithm is proposed to solve multi-objective energy management problem.
Proposed NN model is trained with convex optimization outputs and according to
the simulation results the trained NN model solves the optimization problem
within 92.5% of the convex optimization one.
KEYWORDS
artificial neural network, convex optimization, electric vehicle, energy management strategy,
fuel cell, hybrid energy storage system, machine learning, ultra-capacitor
Abbreviations: AC, alternating current; ANN, artificial neural network; BEV, battery electric vehicle; BMS, battery management system; DC, direct
current; DOD, depth of discharge; EMS, energy management strategy; EOL, end-of-life; ESS, energy storage system; EV, electric vehicle; FC, fuel cell;
FCV, fuel cell vehicle; HESS, hybrid energy storage system; HEV, hybrid electric vehicle; ICE, internal combustion engine; IDC, Istanbul driving
cycle; LA92, the California unified cycle; MIMO, multi input multi output; min, minimum; ML, machine learning; MLP, multi-layer perceptron;
MSE, mean squared error; NN, neural network; NYCC, New York city cycle; PEM, proton exchange membrane; ReLU, rectified linear unit; SOC,
state of charge; SOCP, Second-order cone programming; UC, ultra-capacitor; UDDS, urban dynamometer driving schedule; USA, United States of
America; WLTC, the worldwide harmonized light vehicles test cycles; ZEV, zero emission vehicle.
Received: 1 October 2019 Revised: 9 March 2020 Accepted: 20 March 2020
DOI: 10.1002/er.5429
Int J Energy Res. 2020;1–14. wileyonlinelibrary.com/journal/er © 2020 John Wiley & Sons Ltd 1
1|INTRODUCTION
The USA (United States of America) Energy Information
Administration 2017 data indicates that almost 36.2% of
the USA's energy source is petroleum, and almost 92% of
the petroleum is used in the transportation sector, generat-
ing 45% of the USA's energy-related carbon dioxide emis-
sions.
1
In 2016, transport contributed by 27% to total
greenhouse gas emissions in Europe and 72.1% of this
emission is due to road transport.
2
Furthermore, vehicle
operating life is mostly exceeded, therefore emission rates
become more than set point defined by European anti-
pollution standards.
3
Electric vehicles (EVs) are regarded
as a potential solution to environmental pollution, energy
source shortages and global climate issues, highly con-
cerned worldwide caused by transportation with fast
development of automobile industry and increasing car
ownership.
4,5
Electric vehicles with internal combustion
engine (ICE) are called as hybrid electric vehicle (HEV)
6-8
and result low-emission
9,10
compared to conventional
ones. However, the goal is to have zero-emission vehicles
(ZEVs)
11
which has no ICE in its powertrain.
EV technology is not fully mature yet and has been
under dynamic development for producing more efficient
powertrain with an extended driving range. Researches
are mainly focusing on the optimization of both the
powertrain structure and energy storage systems (ESS).
12
As an example, the dual motor coupled drivetrains
allow the motors to operate in efficient areas by reasonably
synthesizing and decomposing the output power.
13
By uti-
lizing two propulsion machines instead of a large central
one, the powertrain efficiency could be improved up to
10%.
14
This approach is adopted by the industry and many
carmakers such as Audi, BYD, Jaguar, Mercedes-Benz,
Tesla, GM, Porsche and Genovation Cars. Typically, they
use multiple electric propulsion machines in their
powertrain to obtain improved efficiency results and bet-
ter arrangement under the hood.
Despite a significant increase in driving range of EVs
with the introduction of lithium-ion battery technology,
still one of the biggest challenges is the storage of electric
energy. HESS can be utilized by combining different energy
sources to explore the benefits of each of them within the
EVs for getting an extended driving range and longer bat-
tery lifetime.
15
Because of large impact on acceleration and
drive range on a single charge, battery is considered one of
the most important components in HESS.
16
In some vehicle
applications,
14,17
FCs are considered as an alternative
energy converter to battery. If hydrogen is produced by a
renewable energy source, the FC propulsion system pro-
duces no carbon emissions.
18
Hydrogen can be safely stored
onboard in tanks with high pressure.
19
Hybridization of FC
with ESS (batteries and/or UCs) is essential to reduce FC
size, hydrogen consumption and cost of the powertrain.
20
An UC, also termed as supercapacitor or electric double-
layer capacitor, offers a high value of power density but low
energy density. It can be charged/discharged millions of
times and at a faster rate than battery.
21
Hence, UC is nearly
maintenance free and offers a very long lifetime. Thus, UC
is considered as a buffer unit in the HESS. In a HESS, FC
and/or UC could help to reduce the stress on the battery,
resulting in battery lifetime extension with an effective
energy management.
22,23
However, the number of input and output of the sys-
tem rises with each added unit. The use of more than one
propulsion machines in the powertrain and utilizing
HESS increases the complexity of the system architec-
ture, what needs a reliable energy management. There-
fore, many EMS shave been reported in the literature for
different types of power transmission systems.
The EMS for EVs with HESS are mainly divided into
rule-based and optimization-based strategies.
24
Rule-based
EMS is set in advance according to human experience and
easy to utilize, however does not guarantee an optimal solu-
tion. Optimization-based strategies, as dynamic program-
ming, can lead to a global optimum solution, however it is
usually difficult to be implemented as real time control due
to high computation times.
25
Convex optimization used in
studies
24,26,27
seems to have an increased practical signifi-
cance, due to its high computational efficiency and global
optimization capability. Nevertheless, not all the systems
are suitable for linearization and convex optimization is not
applicable to non-linear systems. Thus, ML algorithms such
as artificial neural network (ANN) can be used as an alter-
native solution for such complex optimization problems.
The various HESS configurations require different energy
management control strategies special to each individual
HESS structure.
28
One of the most common HESS in the lit-
erature consist of battery and UC
29
where its sizing and
power splitting optimization is studied in Ref.
30
The paper
in Ref.
31
reviews and discusses the structures and the EMSs
of HESSs comprised of battery and UC. FC/battery
32
and
FC/UC
33
are the other well-known HESS types for hydro-
gen powered EVs. A more complex HESS type is battery/
UC/FC and its different types of power conversion configu-
rations are discussed in Ref.
20
Sizing optimization is dis-
cussed for battery/UC/FC HESS considering battery
degradation and operating cost.
34
In another paper, EMS
based on fuzzy logic is proposed for a fuel cell hybrid ship,
combining battery/UC/FC.
35
Nevertheless, fuzzy rules
based on expert knowledge are obtained on the basis of
observations and does not guarantee global optimum. In a
study,
36
a convex optimization problem for power splitting
is formulated to extend the battery lifetime for battery/UC/
FC HESS. However, due to the excessive number of param-
eters that has to be taken into consideration, convex
2YAVASOGLU ET AL.
optimization of complex systems may not provide online
solution. On the other hand, not all systems are suitable for
linearization. Therefore, a NN-based ML algorithm could
be utilized as proposed in this study to solve multi-objective
energy management problem.
The most important contribution of this study is that
NN-based optimization has demonstrated its usefulness in
implementing energy management system strategies for
complex power transmission systems. As far as the authors
know, an ML model to solve EMS problem of a multi-input
(battery/UC/FC HESS), multi-output (two-propulsion
machines) EV powertrain is not proposed in the literature
yet. Especially in online control applications like EVs, such
approaches come to the forefront. Methods such as convex
optimization can be used for HESS problems. However,
it may have difficulty in performing online when the
powertrain complexity increases and objectives multiplied.
This study also shows comparison of NN-based optimization
with convex one in terms of accuracy and computational
complexity.
In this study, convex optimization method was devel-
oped for the energy management of two propulsion
machines and three energy storage units aiming to explore
benefits of HESS effectively by considering each of compo-
nents, namely battery, UC and FC separately. Moreover, an
NN-based ML model was suggested for the proposed topol-
ogy study given in Ref. 36 where EV has two propulsion
machines and the results of the suggested optimization
method are discussed. Therefore, the present paper is orga-
nized as follows. In Section II, proposed HESS topology is
introduced and discussed. In Section III, the multi-objective
power split problem that aims to efficiently utilize HESS
units for extend battery life, is given. In Section IV, how
convex optimization is used to solve presented optimization
problem is explained. In Section V, proposed NN model, its
train and test procedures are presented. In Section VI,
suggested convex optimization and NN-based approach
results are discussed. Finally, Section VII presents the
conclusions.
2|SYSTEM TOPOLOGY
Powertrain configuration is one of the most important parts
of the design of EVs with HESS. The power flow path and
the coupling method of the power sources are determined
by the configuration, which in turn determine the fuel
economy and dynamic performance. To design highly per-
formance HEVs, a number of configurations were pro-
posed.
37
Focusing on increased efficiency, robustness, long
driving range, etc., propulsion machines and energy storage
units, in other words, inputs and outputs of the powertrain,
can be arranged differently in an EV. In such systems,
power distribution canbe performed with multi input multi
output (MIMO) DC/DC convertor.
14,35,38
Having fast re-fueling time, utilizing FC in EVs is a
beneficial option. Hybridization of FC combined with
auxiliary energy storage devices like batteries and UC can
offer FC reliability and improved efficiency.
In HESS with FC, when the battery is the main
energy storage unit, FC is a component that supports the
battery, thus the vehicle is classified as BEV with FC
range extender. In contrast, when the FC is the main
HESS unit, the battery is utilized as a buffer unit and the
vehicle is classified as plug-in hybrid FC vehicle as shown
in Equation (1).
18
PEl,Bat
PEl,FC
> 1, BEV with range extender
< 1, plug−in hybrid FCV
:ð1Þ
In Equation (1), P
El,Bat
and P
El,FC
are maximum elec-
tric output power of battery and FC, respectively.
In this study, BEV with range extender is considered
where FC stack has 8 kW output power for a range
extender purpose and the main and detailed power
demands are supplied by the battery and UC, respec-
tively. Additionally, the considered hydrogen tank in this
study has a capacity of 44 kg/60 L, and it can store up to
2.5 kg of H
2
@700 bar corresponding to ~42 kWh of useful
energy. FC and battery show high useful energy content
but with smaller power density. Especially for sudden
power demands, another energy storage unit as UC can
deliver high power densities for a short duration of time.
For achieving high fuel economy and long battery
lifetime, the topology given in Ref.
36
is recommended. It
composed of two complementary electric propulsion
machines and battery/UC/FC HESS as represented in
Figure 1.
As shown in Figure 1, FC system consists of stack,
hydrogen tank, controller, cooler fan, hydrogen valve
(S
1
) and hydrogen purge (S
2
) valve. Hydrogen is pro-
vided through S
1
. FC controller detects an operating
state of a FC stack and determines an opening time of a
hydrogen purge valve (S
2
) based on information regard-
ing the stack operating state and also controls stack
cooling fan. Battery management system (BMS) is uti-
lized for both safety and monitoring issues of battery
pack. MIMO DC/DC converterenablespowerflowin
desired direction to meet inverter powers. A supervisory
controller communicates with FC controller, BMS and
MIMO DC/DC convertor to overall power management.
Control and voltage supply lines of electronic devices
are shown in dashed and red lines, respectively. High
voltage lines of HESS units and propulsion machines
shown with bold black lines.
YAVASOGLU ET AL.3
In given topology, while propulsion machines can
operate in either propulsion or regenerative breaking mode,
battery and UC power could be negative or positive
depending on their discharging or charging period. The only
unidirectional unit in the studied system is the proton
exchange membrane (PEM) FC,
39
since it has no regenera-
tive operation mode. The HESS and vehicle parameters are
given in Table 1 and the power flows of the studied topology
are illustrated in Figure 2. Vehicle speed provided by driving
cycles and required for demand calculation were considered
to be as in Refs.
40,41
P
Bat
,P
UC
and P
FC
are powers of battery, UC and FC,
respectively. P
M1
and P
M2
are powers of propulsion
machines 1 and 2, respectively. Power flow directions in
Figure 2 shown in red arrows.
Structure and operation principle of the powertrain can
be summarized as follows. The studied powertrain system
has two complimentary propulsion machines. These elec-
tric motors have different efficiency characteristic due to
different torque and speed values. Propulsion machines
cover demand power of the vehicle that occurs during driv-
ing. As the proposed vehicle has two propulsion machines,
demand power is shared between electric motors simulta-
neously. Depending on the current speed and torque
requirement, one of the motors works alone or together to
meet the demand power while keeping efficiency high. Out-
puts of the HESS are inputs of propulsion machines. In the
hybrid powertrain system, battery/UC/FC have been used
as power sources. Among the available power sources, the
FC generates low-grade DC voltage, and it is converted as
useful constant voltage by boost DC/DC converter. The bat-
tery is the main power source. FC and ultra-capacitor assist
the power for transient load conditions. Each propulsion
machines can be powered by any HESS unit via a bi-
directional DC/DC convertor.
3|THE OPTIMIZATION
PROBLEM
In an EV with multi propulsion machines and a multi
ESS, optimization was performed in two separate per-
spectives, namely in propulsion machines and in HESS,
respectively.
3.1 |Power split on propulsion
machines
Primarily, the demand load power(P
Load
)was calculated
based on the force acting on a moving vehicle, which can
be determined by the traction effort and the resistance, to
conduct the power split optimization.P
Load
is calcu-
lated as
42
:
FIGURE 1 System architecture of the plant [Colour figure can be viewed at wileyonlinelibrary.com]
TABLE 1 Vehicle and HESS parameters
Specification Value Unit
Total vehicle weight 1752.30 kg
Battery pack nominal capacity 19.25 kWh
Stored hydrogen in the tank 2.50 kg
FC stack maximum output power 8.00 kW
SOC of battery 100~20 %
SOC of UC 100~50 %
UC capacity 9.00 F
4YAVASOGLU ET AL.
PLoad tðÞ
=vtðÞ δMacc tðÞ+1
2ρCDAfvtðÞ
2+MgCrcos αðÞ+Mgsin αðÞ
,
ð2Þ
where the mass factor δrepresents the equivalent mass
increase due to the angular moments of the rotating com-
ponents. The other quantities in Equation (2) are total
vehicle mass M, the air density ρ, the vehicle frontal area
A
f
, the standard gravity g, angle defining the slope of the
road α, the drag coefficient C
D
and the rolling coefficient
C
r
. The vehicle speed v(t) and acceleration acc(t) are con-
sidered to be a function of time. The traction torque τ(t)
and rotational speed of the propulsion machines ω
m
(t)
can be determined as follows:
τtðÞ=rwh PLoad tðÞ
vtðÞGr
ð3Þ
ωmtðÞ=vtðÞ
rwh
Gr,ð4Þ
where G
r
is the gear ratio and r
wh
is the wheel radius.
Optimal load power sharing by propulsion machines
is obtained by calculating power loss on each electric
motor. The power loss P
loss
in M1 (P
loss,1
) and M2 (P
loss,2
)
can be calculated as in Equation (5) for propulsion mode
(with positive torques) and regenerative braking mode
(with negative torques):
Ploss,itðÞ=
ωtðÞτitðÞ 1−ηitðÞðÞif τitðÞ≥0
ωtðÞτitðÞ 1−ηitðÞðÞ
ηitðÞ if τitðÞ≤0
8
<
:9
=
;
:ð5Þ
Here {η
i
(t),i = 1,2} is the propulsion machine effi-
ciency at each time step. The efficiencies of M
1
and M
2
are denoted as η
1
and η
2
.ω(t) is the rotation speed of the
two motors. τ(t),τ
1
(t) and τ
2
(t) are the total demand tor-
que, the torque on M
1
and the torque on M
2
, respectively.
To minimize the total power loss on propulsion
machines, the power split optimization which is elabo-
rately given in Ref. 24 is used.
3.2 |Power splitting in HESS
After calculating propulsion machine powers corresponding
to load power, P
M1
(t) and P
M2
(t) should be shared by the
HESS units optimally. Objectives of the supervisory control
problem are listed as below:
•Extend battery lifetime
•Minimize battery power magnitude
•Minimize battery power variations
•Utilize UC effectively
•Control state of charge of UC (SOC
UC
) corres-
ponding to vehicle speed
•Utilize FC effectively
•Minimize power loss on FC
•Control FC output power (P
FC
) corresponding to
state of charge of the battery (SOC
Bat
)
The life of battery can be diminished through the
high current rate (C-rate) operation or the large varia-
tions of current flow that generate excessive heat and
increase the internal resistance of the battery.
43
To extend
battery lifetime, the objective function f
1
is formulated to
minimize battery power magnitude and variations as:
P
Bat
,P
FC
,P
UC
,P
Conv
.
f1=aP
Bat tðÞjj+bP
Bat tðÞ−PBat t−ΔtðÞðÞ
2
:ð6Þ
In Equation (6), P
Bat
is the battery power.
To efficiently use UC as a buffer unit during accelera-
tion and deceleration periods, it is required to control
SOC
UC
wisely, as it has very low energy density. At this
point vehicle speed could be useful on active control of UC
as described next. If the vehicle speed is high, vehicle is
more likely slowdown in the upcoming time steps, thus UC
should be ready to capture regenerative breaking energy.
Likewise, if the vehicle speed is low, vehicle is more likely
accelerate in the upcoming time steps and UC should be
ready to provide enough energy. Based on this logic, a refer-
ence SOC of UC parameter (SOCref
UC)is formulated as:
FIGURE 2 Inputs and outputs of the battery/UC/FC HESS
topology where energy flow direction is shown with red arrows
[Colour figure can be viewed at wileyonlinelibrary.com]
YAVASOGLU ET AL.5
SOCref
UC tðÞ=SOCmax
UC +SOCmin
UC −SOCmax
UC
vmax
vtðÞ,ð7Þ
where SOCmin
UC ,SOCmax
UC are the minimum and the maxi-
mum SOC values of the UC, respectively, and v(t) is the
vehicle speed.
Thus, objective function f
2
is formulated as in Equa-
tion (8) to minimize error between actual and reference
SOC values of UC.
f2=c SOCref
UC tðÞ−SOCUC tðÞ
:ð8Þ
Due to physical limitations of the FCs, such as low
power density and slow dynamics, EVs solely powered by
FCs are often criticized for their long start-up time, and slow
power response. On the other hand, FC efficiency is highly
related with its operation power level. Thus, it requires
active control for utilizing efficiently. In this study, as given
in Figure 3 generic FC efficiency model
44
is utilized.
Moreover, the power loss on the fuel cell is formu-
lated as in Equation (9),
27
PFC,loss tðÞ=PFC tðÞ 1−β:
PFC tðÞ
PFC,nom
+ξ
,ð9Þ
where βand ξare trend-line coefficients. According to study
given in Ref.
22
FC operates in maximum efficiency around
its 20% nominal power and its efficiency is decreasing
almost linearly through its maximum power. The mini-
mum FC operation power is limited to 20% of its nominal
power due to reduce losses. A FC reference power function
Pref
FC tðÞ is formulated in Equation (10), which keeps the
FC in its high efficient operation area and increases the
FC power inversely proportional corresponding to
SOC
Bat
(t) to support battery when it is necessary.
Pref
FC tðÞ=Pmax
Fc −Pmin
Fc −Pmax
Fc
SOCmax
Bat 0:8ðÞ
SOCBat tðÞ−SOCmax
Bat 0:2
,
ð10Þ
where Pmin
Fc and Pmax
Fc are the minimum and maximum
FC power and SoCmin
Bat and SoCmax
Bat are the minimum
and the maximum battery state of charge values,
respectively.
Both minimizing power loss in FC and supporting
battery, objective function f
3
is formulated with Equa-
tion (11) to utilize FC effectively.
f3=dP
ref
FC tðÞ−PFC tðÞ
+ePFC,loss tðÞ:ð11Þ
Coefficients a, b, c, d, e used in f
1
,f
2
,f
3
are calculated
as in Ref.
25
Finally, total objective function is formulated
as in Equation (12).
minðf1+f2+f3:ð12Þ
As shown in Figure 2, FC and battery are connected
to the same DC-link and UC is interfaced with this link
through a bidirectional DC/DC converter. Therefore, the
power flows between ESS unites and propulsion
machines are given with the equalities below:
PBat tðÞ+PFC tðÞ−PConv tðÞ=PM1tðÞ ð13Þ
PUC tðÞ+PConv tðÞ=PM2tðÞ,ð14Þ
where P
Conv
is DC/DC converter power. On the other
hand, ESS units should work in their operations limits as
given with the equations below:
Pmin
Bat ≤PBat tðÞ≤Pmax
Bat ð15Þ
Pmin
UC ≤PUC tðÞ≤Pmax
UC ð16Þ
Pmin
Conv ≤PConv tðÞ≤Pmax
Conv ð17Þ
0≤PFC tðÞ≤Pmax
FC ð18Þ
Emin
UC ≤EUC t−ΔtðÞ−PUC tðÞΔt≤Emax
UC ð19Þ
Emin
Bat ≤EBat t−ΔtðÞ−PBat tðÞΔt≤Emax
Bat ,ð20Þ
where E
UC
and E
Bat
are the energy amounts in UC and
battery, respectively. Here, min and max indices indicate
FIGURE 3 Generic FC efficiency model [Colour figure can be
viewed at wileyonlinelibrary.com]
6YAVASOGLU ET AL.
minimum and maximum value of used variable. Each
time step is Δt.
4|CONVEX OPTIMIZATIONS
Convex optimization method given in Ref.,
45
was utilized
for the optimization problem given in Section 3.2. The
convex optimization toolbox
46
in MATLAB software was
utilized to efficiently solve the optimization problem.
The norm and huber functions were considered for
convex optimization. The Huber penalty function is con-
vex and has been provided in the CVX function library.
So, solving the Huber penalty minimization problem in
CVX toolbox is simple as details given in Ref. 46 Here,
CVX automatically transforms this problem into an
SOCP, which the core solver then solves. Constraints are
given in matrix form with Equation (21).
;x=H
Ψx≤J
θx≥K
,ð21Þ
where
x=u1u2u3u4
½
Tð22Þ
u1=PBat tðÞ
PBat thresh u2=PBat tðÞ−PBat t−ΔtðÞ
PBat var thresh u3=Pconv tðÞ
Pconv thresh
u4=PFC tðÞ−Pref
FC tðÞ
PFC thresh
ð23Þ
H=
PBat t−ΔtðÞ
PM1tðÞ−Pref
FC tðÞ
"#
;=PBat_thresh −PBat_var_thresh 0
PBat_thresh 0−PConv_thresh
0
PFC_thresh
"#
ð24Þ
J=
Pmax
Conv
Pmax
Bat var
Pmax
Bat
Pmax
FC
2
6
6
6
6
6
6
4
3
7
7
7
7
7
7
5
Ψ=
00PConv thresh 0
0PBat var thresh 00
PBat thresh 000
00 0PFC thresh
2
6
6
6
4
3
7
7
7
5
ð25Þ
θ=
PBat thresh 000
0PBat var thresh 00
00PConv thresh 0
00 0PFC thresh
2
6
6
6
4
3
7
7
7
5
K=
−Pmax
Bat
−Pmax
Bat var
−Pmax
Conv
−Pmax
FC
2
6
6
6
4
3
7
7
7
5
ð26Þ
Here, Pmax
Bat is the maximum power of battery,
P
Bat_thresh
is the threshold value of battery power, Pmax
Bat var
is the maximum variation power of battery, P
Bat_var_thresh
is the threshold variation value of power of battery, Pmax
Conv is
the maximum power of DC/DC converter, P
conv_thresh
is the threshold value of DC/DC converter power. P
UC
is
calculated with a rule-based post-processing control that
is why it is called quasi-optimum.
5|NEURAL NETWORK
NN is one of the mostly used techniques in ML.
Asindicated in Ref. 47 it has a long history with many
contributors such as the first artificial neuron model by
McCulloch & Pitts, perceptron by Rosenblatt and multi-
layer perceptron by Rumelhart et al. Deep learning also
utilizes deep neural networks, that is, NNs with many
hidden layers and novel transfer functions such as ReLU
(Rectified Linear Unit).
48
The framework for the training and testing of the pro-
posed energy management system is given in Figure 4. NN
was trained by using supervised learning method. Therefore,
FIGURE 4 ANN training and testing flowchart
YAVASOGLU ET AL.7
a training dataset with target values are needed. The datasets
used in training and testing are introduced in Section 5.1.
Then in order to generate input values for the NN model a
feature extraction scheme was applied as given in Section 5.2.
Target values were found by convex optimization method
discussed in previous section. By using these target values
and input features an NN model was trained as shown in
Figure 4A. Architecture and configuration parameters of NN
model are explained in Section 5.3. Then with a test dataset,
that is, a dataset not used while training, trained model was
tested as shown in Figure 4B. Training and testing of the
model discussed in Section 5.4.
5.1 |Training and testing datasets
For the training and testing of the proposed EMS, reliable
datasets are needed. The driving cycles given in Figure 5
were used to produce target and input values for the pro-
posed model.
While UDDS (Urban Dynamometer Driving Schedule),
NYCC (New York City Cycle) and IDC (Istanbul driving
cycle) have urban driving characteristic, LA92 (The Califor-
nia Unified Cycle) is more aggressive one with fewer stops
and idle time. It also has higher speed and acceleration
compared the other train drive cycles. WLTC (The
Worldwide harmonized Light vehicles Test Cycles) has
both urban and highway driving characteristic. Driving
cycle specificationsare summarized in Table 2.
19,49,50
Here, UDDS, IST, NYCC, LA92 driving cycles were
utilized during training and WLTC driving cycle was
used for test procedure.
5.2 |Extracting features
Input values for the model were found by extracting rele-
vant features from the velocity time data as seen in flow-
chart given in Figure 4. For the training, four driving
cycles with urban and highway driving characteristics
were chosen to generate features as seen in the flowchart
given in Figure 4A.
Each cycle is a velocity-time graph that shows the
instantaneous velocity of the vehicle. During training at
every second, long-term and short-term features such as
maximum and average velocity and acceleration values
for the last 50 and 10 seconds were computed, respec-
tively. Since hard brakes and rapid accelerations affected
system differently, positive and negative accelerations
separately considered. Brakes might produce power via
regenerative braking whereas rapid acceleration demands
high power for short durations from battery and ultra-
capacitor.
At each time step, that is, Δt= 1, the parameters
given below were utilized as input features to our model.
Input variables of ANN
PM1tðÞ
PM2tðÞ
SOCBat t−ΔtðÞ
SOCUC t−ΔtðÞ
SOCref
UC tðÞ
nFC tðÞ
PFC t−ΔτðÞ
Pref
FC tðÞ
VtðÞ
acc tðÞ
9
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
;
Instantfeatures tðÞ
Vmax t−10,tðÞ
Vavg t−10,tðÞ
posaccmax t−10, tðÞ
posaccavg t−10,tðÞ
negaccmax t−10, tðÞ
negaccavg t−10, tðÞ
9
>
>
>
>
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
>
>
>
>
;
Short term features tðÞ
FIGURE 5 Drive cycles, (IST, LA92, NYCC, UDDS, WLTC)
[Colour figure can be viewed at wileyonlinelibrary.com]
8YAVASOGLU ET AL.
Vmax t−50,tðÞ
Vavg t−50,tðÞ
posaccmax t−50, tðÞ
posaccavg t−50,tðÞ
negaccmax t−50,tðÞ
negaccavg t−50,t
ðÞ
9
>
>
>
>
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
>
>
>
>
;
Long term features tðÞ:ð27Þ
Here, acc is acceleration and avg, pos and neg indices
are used for average, positive and negative of the utilized
parameter, respectively.
By using these 22 features, the proposed model
predicted 4 real time values such as P
Bat
,P
UC
,P
FC
and
P
Conv
. Target values to be predicted were generated by
using quasi-optimum convex optimization.
5.3 |Architecture and configuration
parameters of NN
In this presented study, a feed forward NN was used, that is,
Multi-Layer Perceptron (MLP), which is capable of approxi-
mating any arbitrary continuous non-linear function.
51
Number of inputs and outputs to the NN model determines
thenumberofnodesintheinput and output layers, respec-
tively. However, there is no general rule for determining nei-
ther the number of hidden layers nor the number of nodes in
a hidden layer. Actually, these are hyper-parameters and
they are generally found by hyper-parameter tuning tech-
niques such as grid search, that is, trying a set of parameters
and choosing the one that gives the best validation results.
52
Theproposedcasehas22inputvalues, that is, features and
4 output values. Thus, input layer has 22 nodes and output
layer has 4 nodes. The number of hidden layers was chosen
as one with 25 nodes by using hyper-parameter tuning. Uti-
lized ANN model is given in Figure 6.
5.4 |Training and testing of NN
By using these 22 features, the proposed model predicted
4 real time values such as P
Bat
,P
UC
,P
FC
and P
Conv
. Target
values to be predicted were generated by using quasi-
optimum convex optimization. The optimization
problem discussed in Section 3 solved with the global
optimization method explained in Section 4. Gathered
results then used to train NN model. While training,
mean squared error (MSE) was used as the cost function
to compute the error between target and predicted values.
Node weights were updated by using scaled conjugate
gradient backpropagation in order to decrease the error
found by the cost function. Input and output data were
normalized before training in order to make gradient
descent work more efficiently. Training was stopped after
100 iterations. Randomly selected 30% of training data
was used for the validation purpose, that is, not included
in the training set.
From computational aspect, the computation time of an
ANN depends on the number of layers and node counts at
each layer. For a three-layered feed-forward NN, computa-
tional complexity is given as the following
53,54:
TABLE 2 Driving cycle specifications
Driving cycle Duration, s Distance, km Average speed, km/h Maximum speed, km/h
UDDS 1369 12.07 31.50 91.25
NYCC 598 01.89 11.40 44.60
IDC 1003 31.00 30.90 78.00
LA 92 1435 15.80 24.40 108.10
WLTC 1800 23.20 46.50 131.00
FIGURE 6 ANN model [Colour figure can be viewed at
wileyonlinelibrary.com]
YAVASOGLU ET AL.9
Oij+jkðÞ,ð28Þ
i,jand kare the number of neurons in input, hidden
and output layers, respectively. (ij) denotes the number
of multiplications when inputs are fed to the hidden
layer, and (jk) denotes the number of multiplications
when outputs of hidden neurons are fed to the output
layer.
In other words, for a single prediction, computation
time only depends on the number of input features, output
predictions and hidden nodes for a single hidden layer feed-
forward ANN.Thus, in this presented study, a single predic-
tion took O(ij+jk)=O(22 25 + 25 4) = 650 flops
(floating point operations).
Convex optimization also requires the multiplication
of matrices given in Section 4. There are three matrix
multiplications were computed at each iteration. The
computational complexity of matrix multiplication of
matrices X
ij
.Y
jk
is O(ijk), where i,jand kare dimen-
sions of these matrices. Also computation of xrequires
4 divisions.
Convex optimization computes the global minima in t
iterations.
55
Therefore, computational complexity of con-
vex optimization can be given as O(t(O(;x)+O
(Ψx)+O(θx)+O(x))). In this presented case, these
three multiplications (;x,Ψx, and θx) plus compu-
tational of vector xrequired 44 flops. Convex optimiza-
tion found the output values at most 25 iterations, that is,
t= 25. Therefore, convex optimization found output
values for a single prediction in 1100 flops. When com-
pared to NN, complexity increased 69.2% with convex
optimization.
When computation times were measured on the same
computer (Intel i7 6500U @ 2.5 GHz CPU 8GB RAM) for
convex optimization and NN respectively, it has found
that NN computes EMS parameters about 6 times faster
than convex optimization. (NN = 0.16 seconds per pre-
diction, convex = 0.98 seconds per prediction.) Most of
this speed increase was due to efficient implementa-
tion of NN.
6|SIMULATION RESULTS
Prolonging the battery life-time and having efficient
HESS system are top priority criteria in EVs. The focus
designs are determined to be assisting batteries during
transient hard states especially with other HESS units.
23
Parallel to these targets, Battery, UC and FC powers
corresponding to WLTC driving cycle were predicted by
the NN model as given in the Figure 7.
These results showed that trained NN model success-
fully follows the optimization objectives. Here, the main
load power was supplied by the battery and FC, while the
power fluctuations at the load power were being met by
the UC. Additionally, FC operated at its efficient power
rate unless high power was demanded and SOC of bat-
tery is low. The FC was operating at a power level that
corresponds to around 20% of its nominal power (where
it has maximum efficiency) in the first parts of WLTC,
which reflected the urban driving style. The last part of
the WLTC reflected highway driving, where the capacity
of the battery was reduced due to increased power
requirements, so the power drawn from the FC was also
slightly increasing proportionally to support the battery.
As an example, in the 1200th second of the test cycle
SOC
Bat
= 92% and P
FC
= 2.2 kW, in the 1750th second
SOC
Bat
=85% and P
FC
=2.79 kW. These results complied
with our energy management objectives stated in previ-
ous sections.
The objective of the optimization problem in this
study is to prolong the life-time of the battery, as well as
the efficient and effective use of the FC and UC. C-rate is
one of the key effective parameters on battery life-time
and refers a measure of the rate at which a battery is dis-
charged or charged relative to its maximum capacity.
Here convex optimization and NN based EMSC-rate
results are compared as shown in the Figure 8.
C-rate mean value of the convex and NN-based opti-
mization under WLTC driving cycle are 0.22 and 0.24,
respectively. If the powertrain is not optimized by a
supervisory control, mean value of this C-rate goes up to
0.36, where FC power is set to constant 20% of its nomi-
nal power all the time and no active control is applied to
UC operation. Although it is still a subject of research
that how peak current impacts on battery degradation
and cycle life, there are established models that account
for the effect of C-rate on battery cycle life, which is
incorporated in the battery cycle life estimation model.
56
Battery degradation model given in Ref.
44
was utilized in
this study for the calculations. This generalized battery
FIGURE 7 ANN power results corresponding to WLTC
driving cycle [Colour figure can be viewed at
wileyonlinelibrary.com]
10 YAVASOGLU ET AL.
cycle-lifetime model was constructed based on the experi-
ment data in Ref.
57
The model estimates the battery life-
time extension with considering the temperature, the
depth of discharge (DOD) and the C-rate effects on
capacity loss. For battery end of life (EOL) estimation, a
daily commute was assumed to be 72 km long which was
equal to three consecutive WLTC driving cycles. Battery
EOL estimations are compiled in the Table 3.
It was estimated that the battery EOL is 8.1 years if
no optimization control was applied and 13.3 years for
quasi-optimum convex optimization and 12.3 years for
suggested NN model under the same drive cycle and
daily commute assumptions. When the EMS results were
evaluated in terms of battery life-time predictions, NN
achieved 92.48% close result to convex optimization one
and the lifetime was estimated to have a 51.85% exten-
sion compared to one without optimization.
7|CONCLUSION
The design of vehicle is nowadays arranged around ESS
and one or more propulsion machines rather than an
internal combustion engine. To have more efficient long
lasting and fast refillable green vehicles, utilized number of
propulsion machines and ESS units are increasing. Due to
the multiple-power-source, multiple-propulsion machines'
nature and the complex configuration, the control strategy
of an EV is more complicated than that of an engine-only
vehicle. This kind of supervisory control strategy of a HESS
electric vehicle coordinates the operation of vehicle sub-
systems to achieve targets such as maximizing fuel econ-
omy, prolonging ESS end of life and improving vehicle per-
formance. In this paper, supervisory control of an EV
powertrain with HESS (battery, UC and FC) and two pro-
pulsion machines that determines proper power split
between ESS units, while satisfying constraints such as driv-
ability, component reliability is studied. Proposed ANN
algorithm successfully accomplished given objectives such
as extending battery life-time, utilizing UC effectively and
minimizing power loss on FC while meeting propulsion
machines powers. Although there are other ML algorithms
that can be utilized to attack the problem in this work,
ANN is a good alternative due to its capability in handling
complex non-linear relationships between input and output
parameters. This makes it adaptable to solving a diverse set
of problems. With proper datasets, ANN models with high
prediction abilities can be generated.
Utilizing proposed supervisory control, 51.85% better bat-
tery life was achieved compared to the passively controlled
one. NN model was trained with quasi optimum convex opti-
mization results and NN achieved 92.48% close result to con-
vex optimization one. Naturally, an NN model trained with
globally optimal results would give even better results.
This study demonstrated that ML can be used in place
of control solutions that require a rule-based structure or
mathematical model, and that a well-trained NN model can
be successfully used in energy management systems for
complex powertrain of EVs. In our future work, second or
higher order models for HESS units will be conducted and
we will investigate the enabling EMS based on a fix system
size, where experimental or HIL validation will be necessar-
ily conducted. Moreover, integrating an online learning
algorithm instead of a pre-trained one will be investigated.
By this way, we will aim to develop a dynamic supervising
EMS that will adopt itself according to dynamic parameters.
NOMENCLATURE
A
f
frontal area, m
2
a, b weights of f
1
acc acceleration, m/s
2
cweight of f
2
C
D
,C
r
drag, rolling coefficients
d,e weights of f
3
Eenergy, kWh
FIGURE 8 C-rate results of convex and NN-based
optimization under WLTC driving cycle [Colour figure can be
viewed at wileyonlinelibrary.com]
TABLE 3 Battery EOL estimation
EMS Battery EOL, y
Battery EOL
extension, %
No control 8.10 N/A
Convex 13.30 64.20
ANN 12.30 51.90
YAVASOGLU ET AL.11
f
1
,f
2
,f
3
objective functions
ggravity, m/s
2
G
r
gear ratio
H, J, K constraint matrices
i, j, k number of neurons in input, hidden and out-
put layers
Mtotal vehicle mass, kg
neg_acc negative acceleration, m/s
2
Ocomputational complexity
Ppower, W
pos_acc positive acceleration, m/s
2
rradius, m
SOC state of charge, %
ttime interval, s
V,vvelocity, m/s
SUBSCRIPTS
avg average
Bat battery
conv converter
El electric
FC fuel cell
Load load
loss loss
M1,
M2
propulsion machine one, propulsion
machine two
max maximum
nom nominal
thresh threshold
UC ultra-capacitor
var variation
wh wheel
8|SUPERSCRIPTS
max maximum
min minimum
ref reference
GREEK LETTERS
;,Ψ,θconstraint matrices
αslope angle,
Δdifference
δmass factor
ηefficiency
ξ,βtrend line coefficients
ρdensity, kg/m
3
τtorque, Nm
ωrotational speed, rad/s
ACKNOWLEDGEMENTS
The authors would like to gratefully acknowledge the
support of Dr. Szabolcs Varga from the Department of
Mechanical Engineering at the Faculty of Engineering of
the University of Porto, Portugal.
ORCID
Huseyin A. Yavasoglu https://orcid.org/0000-0001-
8145-719X
REFERENCES
1. Esmaeeli R, Aliniagerdroudbari H, Hashemi SR, et al. Design,
modeling, and analysis of a high performance piezoelectric
energy harvester for intelligent tires. Int J Energy Res. 2019;43:
5199-5212. https://doi.org/10.1002/er.4441.
2. European Environment Agency Web Portal. Greenhouse gas
emissions from transport in Europe in 2016. https://www.eea.
europa.eu/data-and-maps/indicators/transport-emissions-of-
greenhouse-gases/transport-emissions-of-greenhouse-gases-11.
Accessed April 12, 2020.
3. Cabata R. Urban and extra-urban hybrid vehicles: a technologi-
cal review. Energies. 2018;11(11):2924. https://doi.org/10.3390/
en11112924.
4. Zhou X, Zou L, Ma Y, Gao Z. Research on impacts of the elec-
tric vehicles charging and discharging on power grid. Paper
presented at: 2017 29th Chinese Control And Decision Confer-
ence (CCDC); May 28-30, 2017; Chongqing, China. https://doi.
org/10.1109/CCDC.2017.7978736.
5. Torcat AB, Poddar T, Almansoori A. A realistic framework to a
greener supply chain for electric vehicles. Int J Energy Res.
2019;43:3346-3364. https://doi.org/10.1002/er.4373.
6. ZhangS,HuX,XieS,SongZ,HuL,HouC.Adaptively
coordinated optimization of battery aging and energy man-
agement in plug-in hybrid electric buses. Appl Energy.
2019;256:113891. https://doi.org/10.1016/j.apenergy.2019.
113891.
7. Hu X, Zou C, Tang X, Liu T, Hu L. Cost-optimal energy manage-
ment of hybrid electric vehicles using fuel cell/battery health-
aware predictive control. IEEE Trans Power Electron. 2020;35(1):
382-392. https://doi.org/10.1109/TPEL.2019.2915675.
8. Kong Z, Shi G, Wang B, Wang H. Development of the electric
control unit for a full hybrid power-train structure. World
Electr Veh J. 2011;4(1):550-557. https://doi.org/10.3390/
wevj4030550.
9. Martinez CM, Hu X, Cao D, Velenis E, Gao B, Wellers M.
Energy management in plug-in hybrid electric vehicles: recent
Progress and a connected vehicles perspective. IEEE Trans Veh
Technol. 2017;66(6):4534-4549. https://doi.org/10.1109/TVT.
2016.2582721.
10. Bianchi D, Rolando L, Serrao L, et al. Layered control strategies
for hybrid electric vehicles based on optimal control. Int J
Electr Hybrids Veh. 2011;3(2):191-217. https://doi.org/10.1504/
IJEHV.2011.042147.
11. Rath R, Kumar P, Mohanty S, Nayak SK. Recent advances,
unsolved deficiencies, and future perspectives of hydrogen fuel
cells in transportation and portable sectors. Int J Energy Res.
2019;43:8931-8955. https://doi.org/10.1002/er.4795.
12 YAVASOGLU ET AL.
12. Yazici MS, Yavasoglu HA, Eroglu M. A mobile off-grid plat-
form powered with photovoltaic/wind/battery/fuel cell hybrid
power systems. Int J Hydrogen Energy. 2013;38:11639-11645.
https://doi.org/10.1016/j.ijhydene.2013.04.025.
13. Hu J, Niu X, Jiang X, Zu G. Energy management strategy based
on driving pattern recognition for a dual-motor battery electric
vehicle. Int J Energy Res. 2019;43:3346-3364. https://doi.org/10.
1002/er.4474.
14. Yavasoglu HA, Shen J, Shi C, Khaligh A. A supervisory con-
troller for a hybrid energy storage system with two propulsion
machines in electric vehicles. Paper presented at: 2015 Intl
Aegean Conference on Electrical Machines & Power Electron-
ics (ACEMP), 2015 Intl Conference on Optimization of Electri-
cal & Electronic Equipment (OPTIM) & 2015 Intl Symposium
on Advanced Electromechanical Motion Systems (ELECTRO-
MOTION); September 2-4, 2015; Side, Turkey. https://doi.org/
10.1109/OPTIM.2015.7426995.
15. Song Z, Li J, Hou J, Hofmann H, Ouyang M, Du J. The battery-
supercapacitor hybrid energy storage system in electric vehicle
applications: a case study. Energy. 2018;154:433-441. https://
doi.org/10.1016/j.energy.2018.04.148.
16. Eckert JJ, Silva LCA, Santiciolli FM, Costa ES, Corrêa FC,
Dedini FG. Energy storage and control optimization for an
electric vehicle. Int J Energy Res. 2018;42:3506-3523. https://
doi.org/10.1002/er.4089.
17. Oldenbroek V, Hamoen V, Alva S, Robledo CB, Verhoef LA,
Wijk AJ. Fuel cell electric vehicle to grid: experimental feasibil-
ity and operational performance as balancing power plant. Fuel
Cells. 2018;18:649-662. https://doi.org/10.1002/fuce.201700192.
18. Walters M, Kuhlmann A, Ogrzewalla J. Fuel cell range
extender for battery electric vehicles. Paper presented at: 2015
International Conference on Electrical Systems for Aircraft,
Railway, Ship Propulsion and Road Vehicles (ESARS); March
3-5, 2015; Aachen, Germany. https://doi.org/10.1109/ESARS.
2015.7101509.
19. Salahuddin U, Ejaz H, Iqbal N. Grid to wheel energy efficiency
analysis of battery- and fuel cell–powered vehicles. Int J Energy
Res. 2018;42:2021-2028. https://doi.org/10.1002/er.3994.
20. Kasimalla VK, Sirinivasulu N, Velisala V. A review on energy
allocation of fuel cell/battery/ultracapacitor for hybrid electric
vehicles. Int J Energy Res. 2018;42:4263-4283. https://doi.org/
10.1002/er.4166.
21. Arya Y. Impact of ultra-capacitor on automatic generation con-
trol of electric energy systems using an optimal FFOID control-
ler. Int J Energy Res. 2019;43:8765-8788. https://doi.org/10.
1002/er.4767.
22. Yavasoglu HA, Shen J, Shi C, Gokasan M, Khaligh A. Power
split control strategy for an EV powertrain with two propulsion
machines. IEEE Trans Transport Electrific. 2015;1(4):382-390.
https://doi.org/10.1109/TTE.2015.2504406.
23. Carter R, Cruden A, Hall PJ. Optimizing for efficiency or battery
life in a battery/supercapacitor electric vehicle. IEEE Trans Veh
Technol. 2012;61(4):1526-1533. https://doi.org/10.1109/TVT.2012.
2188551.
24. Kong Z, Cui N, Li P. Energy management strategy coordinating
lithium-ion battery and ultra-capacitor for electric vehicle.
Paper presented at: 2017 36th Chinese Control Conference
(CCC); July 26-28, 2017; Dalian, China. https://doi.org/10.
23919/ChiCC.2017.8028837.
25. WangW,ZhangZ,ShiJ,LinC,GaoY.Optimizationofadual-motor
coupled powertrain energy management strategy for a battery elec-
tric bus based on dynamic programming method. IEEE Access. 2018;
6:32899-32909. https://doi.org/10.1109/ACCESS.2018.2847323.
26. Keyser AD, Vandeputte M, Crevecoeur G. Convex mapping for-
mulations enabling optimal power split and design of the electric
drivetrain in all-electric vehicles. IEEE Trans Veh Technol. 2017;
66(1):9702-9711. https://doi.org/10.1109/TVT.2017.2745101.
27. Gao M, Du J. Design method of energy management strategy
for range-extended electric buses based on convex optimiza-
tion. Paper presented at: 2016 11th International Forum on
Strategic Technology (IFOST); June 1-3, 2016; Novosibirsk,
Russia. https://doi.org/10.1109/IFOST.2016.7884249.
28. Geetha A, Subramani C. A comprehensive review on energy
management strategies of hybrid energy storage system for
electric vehicles. Int J Energy Res. 2017;41:1817-1833. https://
doi.org/10.1002/er.3730.
29. Sellali M, Abdeddaim S, Betka A, Djerdir A, Drid S, Tiar M.
Fuzzy-Super twisting control implementation of battery/super
capacitor for electric vehicles. ISA Trans. 2019;95:243-253.
https://doi.org/10.1016/j.isatra.2019.04.029.
30. Eldeeb HH, Elsayed AT, Lashway CR, Mohammed O. Hybrid
energy storage sizing and power splitting optimization for plug-
in electric vehicles. IEEE Trans Ind Appl. 2019;55(3):2252-2262.
https://doi.org/10.1109/TIA.2019.2898839.
31. Xiong R, Chen H, Wang C, Sun F. Towards a smarter hybrid
energy storage system basedon battery and ultracapacitor - A criti-
cal review on topology and energy management. J Clean Prod.
2018;202:1228-1240. https://doi.org/10.1016/j.jclepro.2018.08.134.
32. MokraniZ,RekiouaD,MebarkiN,RekiouaT,BachaS.Proposed
energy management strategy in electric vehicle for recovering
power excess produced by fuel cells. Int J Hydrogen Energy. 2017;42
(30):19556-19575. https://doi.org/10.1016/j.ijhydene.2017.06.106.
33. Majeed MA, Khan MG, Asghar F. Nonlinear control of hybrid
energy storage system for hybrid electric vehicles. Int. Trans. Electr.
Energy Syst. 2020;30:e12268. https://doi.org/10.1002/2050-7038.12268.
34. Jiang H, Xu L, Li J, Hu Z, Ouyang Mc. Energy management
and component sizing for a fuel cell/battery/supercapacitor
hybrid powertrain based on two-dimensional optimization
algorithms. Energy. 2019;177:386–396. https://doi.org/10.1016/
j.energy.2019.04.110
35. Zhu L, Han J, Peng D, Wang T, Charpentier JF. Fuzzy logic based
energy management strategy for a fuel cell/battery/ultra-
capacitor hybrid ship. Paper presented at: 2014 First International
Conference on Green Energy ICGE 2014; March 25-27, 2014;
Sfax, Tunisia. https://doi.org/10.1109/ICGE.2014.6835406.
36. Yavasoglu HA, Shi C, Gokce K. Energy storage systems for EVs with
two propulsion machines. Paper presented at: 2017 IEEE Transpor-
tation Electrification Conference and Expo (ITEC); June 22-24,
2017; Chicago, IL. https://doi.org/10.1109/ITEC.2017.7993354.
37. Hu J, Jiang X, Zheng L. Design and analysis of hybrid electric
vehicle powertrain configurations considering energy transfor-
mation. Int J Energy Res. 2018;42:4719-4729. https://doi.org/10.
1002/er.4225.
38. Ling Z, Wang H, Yan K, Sun Z. A new three-port bidirectional
DC/DC converter for hybrid energy storage. Paper presented at:
2015 IEEE 2nd International Future Energy Electronics Confer-
ence (IFEEC); November 1-4, 2015; Taipei, Taiwan. https://doi.
org/10.1109/IFEEC.2015.7361598.
YAVASOGLU ET AL.13
39. Karanfil G. Importance and applications of DOE/optimization
methods in PEM fuel cells: a review. Int J Energy Res. 2020;44:
4-25. https://doi.org/10.1002/er.4815.
40. Sciarretta A, Back M, Guzzella L. Optimal control of parallel
hybrid electric vehicles. IEEE Trans Control Syst Technol. 2004;
12(3):352-363. https://doi.org/10.1109/TCST.2004.824312.
41. Shen C, Khaligh A. An energy management strategy for an EV
with two propulsion machines and a hybrid energy storage sys-
tem. Paper presented at: 2015 IEEE Transportation Electrifica-
tion Conference and Expo (ITEC); June 14-17, 2015; Dearborn,
MI. https://doi.org/10.1109/ITEC.2015.7165791.
42. Hofman T, Ebbesen S, Guzzella L. Topology optimization for
hybrid electric vehicles with automated transmissions. IEEE
Trans Veh Technol. 2012;61(6):2442-2451. https://doi.org/10.
1109/TVT.2012.2196299.
43. Choi M, Kim S, Seo S. Energy management optimization in a
battery/supercapacitor hybrid energy storage system. IEEE
Trans Smart Grid. 2011;3(1):463-472. https://doi.org/10.1109/
TSG.2011.2164816.
44. Barbir F. PEM Fuel Cells, Theory and Practice. 1st ed. Waltham,
MA: Elsevier; 2013:65-163.
45. Choi M, Seo S. Robust energy management of a
battery/supercapacitor Hybrid Energy Storage System in an
electric vehicle. Paper presented at: 2012 IEEE International
Electric Vehicle Conference; March 4-8, 2012; Greenville, SC.
https://doi.org/10.1109/IEVC.2012.6183166.
46. Grant M, Boyd S. CVX: Matlab software for disciplined convex
programming, version 2.0 beta. September 2013. http://cvxr.
com/cvx. Accessed April 12, 2020.
47. Masson E, Wang YJ. Introduction to computation and learning
in artificial neural networks. Eur J Oper Res. 1990;47(1):1-28.
https://doi.org/10.1016/0377-2217(90)90085-P.
48. Schmidhuber J. Deep learning in neural networks: an over-
view. Neural Netw. 2015;61:85-117. https://doi.org/10.1016/j.
neunet.2014.09.003.
49. Xun Q, Liu Y. Evaluation of fluctuating voltage topology with
fuel cells and supercapacitors for automotive applications. Int J
Energy Res. 2019;43:4807-4819. https://doi.org/10.1002/er.4622.
50. Shen J, Khaligh A. Design and real-time controller implemen-
tation for a battery-ultracapacitor hybrid energy storage system.
IEEE Trans Ind Inf. 2016;12(5):1910-1918. https://doi.org/10.
1109/TII.2016.2575798.
51. Hush DR, Horne BG. Progress in supervised neural networks.
IEEE Signal Process Mag. 1993;10(1):8-39. https://doi.org/10.
1109/79.180705.
52. Bergstra J, Bengio Y. Random search for hyper-parameter opti-
mization. J Mach Learn Res. 2012;13(1):281-305. https://doi.
org/10.5555/2503308.2188395.
53. Alpaydin E, Kaynak C, Alimo
glu F. Cascading multiple classi-
fiers and representations for optical and pen-based hardwritten
digit recognition. Paper presented at: Seventh International
Workshop on Frontiers in Handwriting Recognition; 2000;
Amsterdam, Holland. In EPRINTS-BOOK-TITLE. s.n. 2004.
54. Otsuka Y, Fukumoto Y, Owaki S, Nakamura M. Computational-
complexity comparison of artificial neural network and volterra
series transfer function for optical nonlinearity compensation.
Paper presented at: 23rd Opto-Electronics and Communications
Conference (OECC); July 2-6, 2018; Jeju Island, South Korea.
https://doi.org/10.1109/OECC.2018.8730100.
55. McGowern LK. Computational Analysis of Real-Time Convex
Optimization for Control Systems [PhD thesis]. Department
ofAeronautics and Astronautics, Massachusetts Institute of
Technology; 2000.
56. Shen J, Dusmez S, Khaligh A. Optimization of sizing and battery
cycle life in battery/ultracapacitor hybrid energy storage systems
for electric vehicle applications. IEEE Trans Ind Inf. 2014;10(4):
2112-2121. https://doi.org/10.1109/TII.2014.2334233.
57. Wang J, Liu P, Garner JH, et al. Cycle-life model for graphite-
LiFePO
4
cells. J Power Sources. 2011;196(8):3942-3948. https://
doi.org/10.1016/j.jpowsour.2010.11.134.
How to cite this article: Yavasoglu HA,
Tetik YE, Ozcan HG. Neural network-based energy
management of multi-source (battery/UC/FC)
powered electric vehicle. Int J Energy Res. 2020;
1–14. https://doi.org/10.1002/er.5429
14 YAVASOGLU ET AL.
