You are on page 1of 12

Received: 29 May 2019 Revised: 27 July 2019 Accepted: 30 July 2019

DOI: 10.1002/er.4809

RESEARCH ARTICLE

Effective methodology based on neural network optimizer


for extracting model parameters of PEM fuel cells

Mohamed Fawzi1 | Attia A. El‐Fergany2 | Hany M. Hasanien3

1
Department of Electrical
Summary
EngineeringFaculty of Engineering,
Mansoura University, Mansoura, Egypt I/V polarization curves of proton‐exchange membrane fuel cells (PEMFCs) are
2
Electrical Power and Machines used to characterize the performance of single cells and stacks. Numerous
DepartmentFaculty of Engineering, semi‐empirical models are presented to predict such polarization curves by
Zagazig University, Zagazig 44519, Egypt
3 determining the unknown parameters of mathematical model of the PEMFCs
Electrical Power and Machines
DepartmentFaculty of Engineering, Ain stack. In this paper, a novel optimization approach, namely neural network
Shams University, Cairo 11517, Egypt algorithm (NNA) is applied for an estimation of the unknown PEMFC model

Correspondence
parameters. The NNA is employed to minimize adopted objective function,
Attia A. El‐Fergany, Electrical Power and which is formulated as the sum of squared deviations (SSD) between the actual
Machines Department, Faculty of data and estimated voltage points subjects to set of inequality constraints are
Engineering, Zagazig University, Zagazig
44519, Egypt.
satisfied. Three commercial types of PEMFCs stack namely Ballard Mark V,
Email: el_fergany@ieee.org. BCS‐500 W, and Nedstack PS6 are numerically simulated to show the effective-
ness of the proposed NNA‐based tool for parameter identification. The mini-
mum values of SSD are 0.8536 V2 for Ballard Mark V, 0.011698 V2 for BCS‐
500 W stack, and 2.14487 V2 for Nedstack PS6, respectively. The obtained
results of the NNA are compared with other optimizers recently published in
the literature such as flower pollination algorithm, slap swarm optimizer, grey
wolf algorithm, grasshopper optimization algorithm, and shark smell algo-
rithm under the same conditions. The comparisons and other performance
tests indicate the robustness and the competition of the adopted NNA‐based
method for producing accurate I/V polarization curves under different operat-
ing scenarios.

KEYWORDS
optimization algorithms, parameter extraction, polymer electrolyte membrane fuel cells, steady‐state
modeling

1 | INTRODUCTION properties to achieve high performance and resilience


required for FCs. The FCs can be categorized into a num-
The contribution of renewable energy resources became ber of major types as presented in Liu et al and Kim and
mandatory because of the of scarcity of fossil fuel, price Lee.7,8 Among them, proton‐exchange membrane fuel
inflation, and environmental pollution.1-3 The fuel cell cell (PEMFC) has low operating pressure, low operating
(FC) is one of the important recently developed clean temperature, low relative humidity, high efficiency, green
energy technologies with high efficiency and fuel flexibil- source, no contamination, and waste less other than
ity.4-6 The ‐electrolyte membranes (PEM) have sufficient other types.7,9,10 Current can directly be produced during
Int J Energy Res. 2019;1–12. wileyonlinelibrary.com/journal/er © 2019 John Wiley & Sons, Ltd. 1
2 FAWZI ET AL.

the normal operation between hydrogen and oxygen in Therefore, recent heuristic‐based optimization methods
the presence of platinum as reagent.9 It can be utilized have been used to overcome these shortfalls and easily
for different power applications. The steadiness of electro- resolve these problems. One of the well‐known heuristic
chemical, mechanical, and thermal process of PEM using techniques is the genetic algorithm (GA). Simple GA has
a thinner membranes leads to produce effective been employed to define PEMFC parameters.16 An
generation.7 adaptive RNA‐GA has been used to evaluate PEMFC
The identification of the best utilized properties of the performance along with improving the convergence char-
FC is a big challenge for the simple and accurate model- acteristics and reducing the absolute voltage error.17 A
ing6 . Many factors affect PEMFC selection and perfor- mechanistic model, describing all internal phenomena in
mance such as stack life, startup speed, efficiency, a single PEMFC with artificial neural network (ANN)
temperature, and modularity.11 One of the most signifi- has been adopted.18 To reduce unpredictability in evaluat-
cant FC issues is the improvement of the used com- ing control variables and FC parameters, neural network
pounds leading to reliable operation with high (NN) is mixed with a particle‐swarm optimizer (PSO).19
performance. To achieve optimum operation and control In addition, a metaheuristic optimization technique
for FCs, many researches have been exerted to accurately stipulated based on an effective modified PSO addressed
identify its parameters, modeling, simulation, and con- more PEMFC complicated optimization problem.20 A dif-
trol. The model procedures can be classified as mechani- ferential evolution (DE) approach has been applied to
cal or experimental modeling.9 The mechanical modeling accurately identify the PEMFC parameters model.4 Tradi-
deals with the electrochemical, heat, and mass parame- tional DE is updated to increase its quality by using rank‐
ters. The experimental modeling is stimulated partially based mutation and rank‐based different strategies.21,22
from the mechanical one and formulated as empirical In addition, a Markovian switching system is used to reg-
formulas. ulate DE population size and distribution.23 A DE conver-
Many researchers studied the PEMFC as fitting prob- gence trend is enhanced by using two neighborhood
lem using the most extensively used model which is con- search, dynamic crossover probability, and integrated
cerned with the reversible voltage, ohmic voltage drop, scaling factor.24 Moreover, flower pollination algorithm
the activation, and diffusion overvoltage. They are has been adopted to estimate unknown PEMFC
searching for a set of unknown parameters minimizing parameters.25
error between real and estimated data. The most fre- A fast and precise parameter identification for differ-
quently addressed issues with PEMFC parameters identi- ent types of FCs with decomposition and recent optimiza-
fication are (a) difficulty of the applying conventional tion method is addressed.26 Another grouping‐based
methods, (b) degree of correspondence between PEMFC harmony‐search procedure is introduced to define a
volt and current, (c) algorithm speed, (d) type of objective PEMFC model and to evaluate PEMFC parameters.27
function, and (e) the identified parameters precision.12 DE has been updated with global optimization and
Formerly, many deterministic methods have been applied to identify seven parameters using the empirical
introduced to model the PEMFC and identify its parame- polarization properties of known PEMFC stack.28 The
ters. The PEMFC identification problem is reformulated biogeography‐based optimization algorithm with muta-
by using an emulator based on a dynamic model to obtain tion of DE including chaos theory has been utilized to
fast and high accuracy solutions as reported in identify PEMFC parameters.5
Voottipruex et al.13 The PEMFC transient response model PEMFC has been modeled and represented by
is proposed for automotive implementations.14 The metaheuristic optimization algorithms such as multiverse
changes of the magnetic field caused by the pinholes in optimizer,2 adaptive DE,4 three metaheuristic optimiza-
the PEM membrane are simulated using a current dipole tion algorithms such as shuffled frog‐leaping algorithm
flaw model utilizing the linear superposition theory. The (SFLA), imperialist competitive algorithm (ICA), and fire-
proton current density can be identified due to magnetic fly optimization algorithm (FOA) in Kandidayeni et al10
contrary and field decline ramp.15 The deterministic and grey wolf optimizer (GWO)29 to preciously obtain
methods require continuous, convexity, and differentia- its characteristics at various states. The simplified
bility conditions.5 It cannot perfectly deal with nonlinear, teaching‐learning based optimizer (STLBO) is applied
nonconvex, multivariable, and powerfully joined PEMFC with the purpose of solving the PEMFC's problem. Using
problem. These methods have problems in finding the STLBO search algorithm means population quality is pro-
unknown values of the unidentified parameters because moted, and local search with the aid of a chaotic map is
of restrictions of the data which leads to large deviations conducted to improve the global best solution.30 TLBO
between mathematical representations and the real and DE have been combined to improve and accelerate
PEMFC characteristics.4 the search process to solve the PEMFC problem.31 PSO
FAWZI ET AL. 3

has been applied to generate a detailed PEMFC modeling based on NNA to simulate their real behaviors under var-
to assess system behavior regardless temperature value.32 ious operating conditions, and (c) generating the various
A backtracking search optimizer to define PEMFC's polarization characteristics for further investigations by
polarization curves is presented.33 other researchers.
Many other heuristic methods have been utilized for In this current research, the NNA is employed to iden-
solving the PEMFC problems such as salp swarm opti- tify (optimally) the seven unknown parameters of
mizer (SSO),34 grasshopper optimization algorithm PEMFCs for further modeling and simulations. The pro-
(GOA),35 aging and exciting P‐systems based optimiza- cess of parameters extraction is adopted as an optimiza-
tion procedures using semi‐empirical models,9 bird mat- tion problem pending that predefined inequality
ing optimizer,36 atom search optimizer,37 and shark constraints are satisfied. Three typical commercial
smell algorithm (SSA).38 PEMFC systems are used to demonstrate and emphasize
Some researchers have been concerned with the the viability of the proposed NNA‐based tool. Necessary
PEMFC controller design in addition to parameters to discussions and detailed comparisons are in order to eval-
enhance its performance and characteristics.39-42 Differ- uate the results of the NNA. In addition, simple statistical
ent nonlinear controllers methods were designed using measures are made as well.
a linearization approach to extend the PEMFC stack The rest of this article is structured as follows: Section
lifespan. A sliding‐mode based on methodology to control 2 announces the mathematical model of PEMFC based
compressor motor to feed FC with air has been on the generalized steady‐state model along with problem
reported.40 The researchers were interested with control- definition and parameters identifications. The procedures
ler optimization such as stable and optimal controls,39 of the NNA are presented in Section 3. Section 4 reveals
PEMFC fuel‐cell reduced‐order model and a sliding‐mode the numerical cropped results with necessary subsequent
control,40 stable optimal control,41 and PEMFC control- discussions and further validations. At last, the conclu-
lers stability and H‐disturbance attenuation.42 sion and future work are drawn in Section 5.
Over and above the aforesaid literature survey, a
recent comprehensive review in this regard can be found
2 | M A T H E M A T I C A L M O D E L OF
in Priya et al.12 As per the aforementioned survey, the
PEM F UEL CELLS A N D
reader can note that many heuristic‐based optimization
PARAMETERS' IDENTIFICATIONS
approaches have been attempted, and some of them
achieve good results. Following no‐free‐launch theorem,
An electrochemical model of PEMFC, which is developed
still there is a room of improvement in terms of the con-
by Mann et al,45 widely used in the literature to charac-
vergence trend, steady pattern ahead toward the mini-
terize the steady‐state performance of the PEMFC is
mum cost function, and precision of the generated
employed in this current work. In this model, three volt-
results.
age drops per cell are indicated, specifically called (a) cell
The NN algorithm (NNA) is considered a novel
activation overpotential (vact), (b) cell ohmic voltage drop
metaheuristic algorithm, which is motivated by biological
(vΩ), and (c) concentration overpotential (vcon).
nervous systems and the behavior of the ANNs. It is
The voltage of PEMFCs stack (VStack) of many cells
invented by Ali Sadollah and other researchers in
connected in series assuming the same behavior for all
2018.43 It was applied to solve several complex optimiza-
the cells is computed using (1).34,45
tion problems, and it is examined by many well‐known
benchmark functions. A comparison of its results with V Stack ¼ N cells :ðE Nernst − vact − vΩ − vcon Þ (1)
other optimization methods has proven its high efficacy
and performance.43 It is mainly based on the ANN struc- where Ncells is the number of cells connected in series,
ture and its coherent concepts to produce new candidate and ENernst is the cell reversible voltage.
solutions. It has several merits such as its high speed of A closer look to a typical I‐V polarization characteris-
convergence and its lower number of parameters to be tics of the PEMFCs, initially the cell voltage drops rapidly
designed. This helps the designer in selecting a proper due the activation overpotential and then drops linearly
design of the NNA when dealing with a nonlinear optimi- and gradually because of the Ohmic resistance. Once
zation problem. Recently, NNA has been successfully again, at higher current values, the cell voltage comes
applied to tune a fractional order fuzzy PID control of down faster due to the convection overpotential.
an automotive PEMFC air feed system.44 The ENernst is calculated using Equation (2) which
This current research study aims at (a) identifying the assumes that the reference temperature equals 25°C.
best values of the unknown PEMFC parameters, (b) The aforementioned three types of voltage drops can be
developing an effective model of commercial PEMFCs determined spending Equations (3) to (5), respectively.45
4 FAWZI ET AL.

 
E Nernst ¼ 1:229 − 0:85 × 10−3 T fc − 298:15 is adopted to minimize the sum of squared deviations
 pffiffiffiffiffiffiffi
þ 4:3085 × 10−5 T fc ln PH 2 PO2 (2) (SSD), commonly used in the literature for sake of fair
comparisons, between experimental measured cell volt-
age and estimated voltage values by the proposed NNA‐
vact ¼ − ½ξ1 þ ξ2 Tfc þ ξ3 Tfc lnðCO2 Þ þ ξ4 Tfc lnðIfc Þ based model which is depicted in Equation (6).

where
OF ¼ Minimize
ðSSDÞ
Nsamples 2 
P o2   ¼ Minimize ∑m¼1 V FC;meas ½m−V FC;est ½m (6)
C o2 ¼ 6 exp 498=T fc (3)
5:08:10
where Nsamples denotes the number of measured sam-
ρ l ples, m defines the iteration counter, VFC,meas is the exper-
vΩ ¼ I fc ðRm þ Rc Þ; Rm ¼ m (4)
MA imental measured voltage (V), and VFC,est defines model
calculated voltage (V).
where
The OF subjects to the provisions of number of
   2  2:5
inequality restrictions as follows:
I fc T I fc
181:6 1 þ 0:03 þ 0:062 303fc
MA MA
8 h i
ρm ¼  
>
I fc T −303
4:18: fcT >
> ξ i ∈ ξ i; min ; ξ i; max ∀i ¼ f1; 2; 3; 4g
λ − 0:634 − 3 e >
>
<
fc
MA λ ∈ ½λmin ; λmax 
S:t: (7)
>
> Rc ∈ Rc; min ; Rc; max
  >
>
J >
:
vcon ¼ −β: ln 1 − (5) β ∈ ½βmin ; βmax 
J max
where ξi,min and ξi,max are min/max interval of ξi, λmin and
where Tfc is the operating FC's temperature (K), PH 2 and
λmax are min/max values of λ, Rc,min and Rc,max are
PO2 are the partial pressures of H2 and O2 (atm), respec-
min/max boundaries of Rc, and βmin and βmax are
tively, Ifc is the FC's current (A), MA defines the active
min/max interval of the constant β.
area of the membrane (cm2), CO2 denotes the concentra-
tion of O2 in (mol/cm3), ξi (i ∈ {1 to 4}) are semi‐
empirical coefficients, Rm is the membrane resistance 3 | NEURAL NETWORK
(Ω), Rc is the connections' resistance (Ω), l is the mem- O P T I M I Z A T I O N A L GO R I T HM
brane thickness (cm), ρm is the membrane resistivity
(Ω. cm), λ denotes an adjustable parameter concerning The ANNs are composed of computational models that
the water content and humidity, β is a parametric coeffi- are inspired by the biological nervous systems. The
cient which is treated as a constant, and J and Jmax stand ANN structure consists of several artificial neurons,
for the actual and maximum current densities (A/cm2), which are connected together to form the NN. Based on
respectively. the ANN structure, it can be classified into two types
It is worth noting that the estimation of water con- feed‐forward or recurrent NNs, and it relies on its struc-
tent is difficult and changes during the utilization of ture whatever open loop or feedback, respectively. More-
the FC as well. The value of λ equals 13 under ideal sit- over, the recurrent NNs have either a local or global
uations and is equal to 23 under supersaturated situa- feedback structure based on its type.
tions. λ is being considered as an adjustable parameter As all the metaheuristic algorithms, the NNA starts its
with a possible maximum value of 23 in this research. process with generating an initial population entitled pat-
It is worthy emphasizing that the value of λ depends tern solutions within the search space. The NNA is likely
on relative humidity and other many factors.45 In this to imitate the ANN behavior, where its input data and
current study, the assumption of a proper membrane output/target data are received and then predicting the
water content at all possible operating conditions is relationship between them. These input data can be pro-
valid. duced from experiments or simulation results.43,44 In this
From the above stated formulas, it can be concluded algorithm, the best solution at each iteration is consid-
that there are seven variables in the above model required ered the output data, and the main goal is to decrease
to be extracted carefully to characterize the PEMFCs, the error between this best solution and the other pre-
namely ξ1,ξ2,ξ3,ξ4,λ,Rc & β. dicted solutions. This target solution is updated with the
With the aim of estimating the abovementioned course of iterations. In any D dimensional optimization
unknown seven parameters, an objective function (OF) problem, the candidate solution or the pattern solution
FAWZI ET AL. 5

is an array, as written in Equation (8), and it is the input W update ðt þ 1Þ ¼ W i ðt Þ þ 2 × rand × ðW target ðt Þ − W i ðt ÞÞ;
i
data to the NNA.
i ∈ N POP (16)
Pattern solution ¼ ½x 1 ; x 2 ; x 3 ; ……:; x D  (8)
The bias operator can act as a noise of the system. It
A pattern solution matrix (X) is produced to start the can represent the exploration process within the search
optimization process. Its number of rows represents the area, and it is similar to the mutation property of GA. It
number of population NPOP, and its number of columns affects the new pattern solution (Equations 14 and 15)
is the number of design variables D. It can be written as and the weight updating of Equation (16). The following
follows: Pseudo code indicates the calculation of a bias operator
2 3 for new pattern solutions and update the weight under
x 11 x 12 ⋯ x 1D the condition rand ≤ ψ, where ψ is the modification fac-
6 7
X ¼4 ⋮ ⋱ ⋮ 5: (9) tor that demonstrates the percentage of pattern solutions
x N1 POP x N2 POP ⋯ x ND POP needed to be altered.

The fitness function is evaluated for each pattern solu- For i = 1 to NPOP
tion as follows: %% Bias for new pattern solutions
  Nb = Round (D×ψ)
C i ¼ f x i1 ; x i2 ; ……:; x iD : (10) For j = 1 to Nb
xinput(i, integer rabd [0,D]) = LB+(UB ‐ LB) ×
The target solution is considered. Therefore, the NNA rand .
has now X input matrix data and one target solution. The
weight matrix is a squared matrix with NPOP side. It gener- End For
ates random numbers which lie in the range [0, 1] through %% Bias for update weight matrix
the iterations. It is expressed by the following equation: Nwb = Round (NPOP×ψ)
2 3 For j = 1 to Nwb
W 11 W i1 ⋯ W N1 POP
6 7 Wupdate(j, interg rand [0, NPOP]) = U(0,1).
W ðt Þ ¼ 4 ⋮ ⋱ ⋮ 5 End For
W 1N POP W iN POP ⋯ W NN POP End For
2 POP 3
W 11 ……… W i1 ⋯ W N POP 1
6 7 Under the condition rand > ψ, the transfer function
¼4 ⋮ ⋱ ⋮ 5: (11)
W 1 N POP W i N POP ⋯ W N POP N POP (TF) operator can be applied as follows:

The constraints of the weight values are mathemati- X *i ðt þ 1Þ ¼ X i ðt þ 1Þ þ 2 × rand × ðX target ðtÞ − X i ðt þ 1ÞÞ;
cally modeled by the following equation: i ∈ N POP : (17)

N POP Then, the fitness value for all updated pattern agents is
∑ wij ðtÞ ¼ 1; i ∈ N POP (12)
j¼1 calculated. A complete flowchart of the NNA is
established in Figure 1.

wij ∈ U ð0; 1Þ; i& j ∈ N POP (13)


4 | NUMERICAL SIMULATIONS,
After determining the weight matrix W, a new X CASE STUDIES, VALIDATIONS, AND
matrix is created based on the ANN strategy, and it is DISCUSSIONS
modeled as follows:
Simulations are demonstrated on three case studies under
N POP
various operating conditions, and the results are pre-
j ðt
X new þ 1Þ ¼ ∑ wij ðt Þ × X i ðt Þ; j ∈ N POP (14)
i¼1 sented in comparison to other competing optimizers. At
first, the best values of the unknown seven parameters
X i ðt þ 1Þ ¼ X i ðtÞ þ X new are cropped using the NNA. At a later stage, various oper-
i ðt þ 1Þ; i ∈ N POP : (15)
ating conditions such as temperature and partial pres-
Equations (14) and (15) are used to update the pattern sures variations are analyzed. The min/max margins of
solutions with the course of iterations. Then, the these unknowns are listed in Table 1 which are widely
weighting matrix can be updated by Equation (16). used in literatures.2,4,10,29,34,35,37,38
6 FAWZI ET AL.

FIGURE 1 Flowchart of the proposed NNA [Colour figure can be viewed at wileyonlinelibrary.com]

TABLE 1 Min/max limits of unknown parameters

Parameter ξ1 ξ2 × 10−3 ξ3 × 10−5 ξ4 × 10−5 λ Rc(mΩ) β

Low −1.1997 1.00 3.60 ‐26.00 13.00 0.10 0.0136


High −0.8532 5.00 9.80 ‐9.54 23.00 0.80 0.5000

It is worthy stating that the final agreed controlling convergence, ie, SSD is shown in Figure 2. It can be seen
parameters of the NNA which are reached after many tri- that the convergence occurs in steadily and moves fast to
als and errors are NPOP equals to 50, Max ¯ Iter is equal to the final best value in only 100 iterations. Since, the pro-
100, and ψ is equal to 1. It is well‐known that the meta‐ cess of PEMFC parameter extraction is offline in nature
heuristic optimization procedures such as NNA have a which means that the processing time is not the issue.
high level of randomness, and for that reason the NNA Table 2 indicates the obtained values for each targeted
has been implemented 100 times independently, and the parameter after implementing the NNA for PEMFC
best results are reported among these independent trials. parameter extraction. The corresponding best fitness
Some statistical performance parametric measures such
as best, worst, and standard deviation (SD) of the adapted
SSD are performed to appraise the NNA's robustness for
all presented case studies. In the next subsections, the
subsequent case studies are detailed.

4.1 | Case study (1): Ballard Mark V 5 kW

In this case, the data sheet and operating data of this type
of stack are collected from the literatures.25,29,30,35,38 This
stack has 35 cells connected in series with 70‐A maxi-
mum thermal current. After implementing the NNA to
estimate these unknown parameters in order to describe
this PEMFCs stack, the best values are selected carefully FIGURE 2 The signature of best score of SSD convergence
among several trials. The trend of the fitness function [Colour figure can be viewed at wileyonlinelibrary.com]
FAWZI ET AL. 7

TABLE 2 The extracted best parameters along with the lowest SSD in comparisons to others for case study (1)

Parameter NNA FPA25 GWO29 GOA35 SSA38

ξ1 −0.97997 −1.0257 −1.1827 −0.8532 −1.1827


−3
ξ2 × 10 3.6946 3.4000 3.7080 3.4173 3.7080
ξ3 × 10−5 9.0871 6.7952 9.3600 9.8000 9.3600
−5
ξ4 × 10 −16.2820 −12.8500 −11.9250 −15.9555 −11.9250
λ 23.0000 15.6446 11.7603 *
22.8458 11.7603*
Rc(mΩ) 0.1000 0.52906 0.7877 0.1000 0.78773
β 0.0136 0.0614 0.0136 0.0136 0.0136
# **
SSD (best) 0.85361 307.9250 0.002067 0.8710 0.002067**
SSD (Worst) 0.87068 NR NR 0.9097 NR
SSD (SD) 0.00854 NR NR 0.0113 NR

Abbreviation: NR, Not reported.


#
This value is recalculated using the generated parameters by the FPA. * The value of λ violates the min/max range (infeasible solution**).

value produced by the NNA overall 100 independent runs


along with relevant statistical actions is itemized in
Table 2 as well. Additionally, the subsequent compari-
sons among the other competing optimizers are indicated
there. The I/V characteristics of this PEMFC stack com-
bining the experimental and estimated voltage data by
the NNA‐based method are shown in Figure 3A. Simi-
larly to this, the I/P plots for this stack under study are
shown in Figure 3B. It can be confirmed that good
matchings or one can say even full unison between actual
measured and estimated points (either voltage or power)
are indicated. It may be worth mentioning that the num-
ber of measured data is 13 points.

4.2 | Case study (2): 500 W BCS stack

In this subsection, the behavior of the BCS 500‐W


PEMFC stack which consists of 32 cells connected in
series which is manufactured by the American Company
called BCS Technologies is investigated. The main char-
acteristics of this PEMFCs stack are extracted from the lit-
eratures.22,24,25,29,34,38 As can be seen, the thermal rated
current of this system is 30 A. After the implementations
of NNA to define the unknown parameters overall 100
runs, the best results by the NNA in comparisons to other
optimizers are presented in Table 3. In addition to that, FIGURE 3 Curves of Ballard mark V: A, I/V polarization plots;
the quality statistical measures are reported in the same B, I/P plots [Colour figure can be viewed at wileyonlinelibrary.com]
table. The trend of SSD minimization to the best value
is depicted in Figure 4. The performance curves to char-
acterize the steady‐state behavior of this stack such as Figure 5, the worthy performance of the NNA is indicated
I/V and I/P plots are presented in Figure 5A‐5B, respec- in terms of the lower SSD value and good matching
tively, in consort with the experimental data (number of between measured and estimated data by the NNA‐based
measured data equals 18). A closer look to Table 3 and method.
8 FAWZI ET AL.

TABLE 3 The extracted best parameters along with the lowest SSD in comparisons to others for case study (2)

Parameter NNA SFLA10 ICA10 FOA10 FPA25 GWO29 SSO34 SSA38

ξ1 −1.0596 −0.9657 −0.9086 −0.9928 −0.9851 −1.0180 −0.8532 −1.0180


−3
ξ2 × 10 3.7435 3.0800 2.4798 2.6210 2.8000 2.3151 4.8115 (computed) 2.3151
ξ3 × 10−5 9.6902 7.2236 4.4583 3.7464 4.4600 5.2400 9.4334 5.2400
ξ4 × 10−5 −19.3020 −19.3000 −19.3000 −19.3000 −23.2000 −12.8150 −19.2050 −12.8150
λ 20.8772 20.8862 22.6626 21.1011 17.4598 18.8547 23.0000 18.8547
Rc(mΩ) 0. 1000 0.1000 0.2460 0.1000 0.1660 0.7504 0.3499 0.75036
β 0.0161 0.0161 0.0162 0.0163 0.0697 0.0136 0.01589 0.0136
SSD (best) 0.011698 0.011697 0.011856 0.011819 0.01642 7.18890 0.01219 7.18890
SSD (worst) 0.013670 0.011698 0.034665 0.030233 NR NR 0.01520 NR
−4 −8 −4
SSD (SD) 5.640 × 10 5.039 × 10 0.00586 0.00417 NR NR 8.711 × 10 NR

Abbreviation: NR, Not reported.

FIGURE 4 The signature of the best score of SSD convergence


[Colour figure can be viewed at wileyonlinelibrary.com]

4.3 | Case study (3): 6 kW Nedstack PS6


PEMFCs stack

In this last study, Nedstack PS6 PEMFCs stack (in which,


65 cells are linked together in series) with the rated active
power of 6 kW and its main characteristics and operating
data can be found in the literatures.10,34,46-48 The maxi-
mum thermal current of this system is 225 A. The same FIGURE 5 Curves of BCS‐500 W: A, I/V polarization plots; B, I/P
plots [Colour figure can be viewed at wileyonlinelibrary.com]
procedures performed for the above presented two case
studies are followed. The best trend of SSD minimization
over the 100 iterations is described in Figure 6, in which,
the best value of SSD is 2.14487 V2 which is lesser than
obtained by various algorithms (SFLA, ICA, and FOA) generated by other recent competing algorithms. The rep-
employed in,10 FPA,25 SSO,34 GOA,35 and GA,35 as publi- resentations of the steady‐state behavior of this stack are
cized in Table 4. The identified best values of Nedstack depicted in Figure 7A‐7B, respectively (number of mea-
type PS6 are presented in Table 4 together with those sured data equals 30).
FAWZI ET AL. 9

FIGURE 6 The signature of the best score of SSD convergence


[Colour figure can be viewed at wileyonlinelibrary.com]

4.4 | Various operating scenarios:


temperature and pressure variations

In this subsection, further simulations under various


operating temperatures and pressures are investigated.
At this moment, after the best values of the seven
unknown parameters of the above three typical stacks
in the market are cropped, two scenarios of each stack
under study are planned such as (a) temperature varia-
tions under constant partial pressures and (b) varying
the operating partial pressures under constant tempera-
FIGURE 7 Curves of Nedstack PS6: A, I/V polarization plots; B,
ture. The aforementioned means that these proposed sce-
I/P plots [Colour figure can be viewed at wileyonlinelibrary.com]
narios are repeated three times and avoid length article;
just as a representative study, only these scenarios are
presented for Ballard Mark V stack. voltage and output power can be boosted to a certain
The simulated behavior analysis of this stack overall degree. On the other hand, by keeping the cell tempera-
various operating temperatures (at 25oC, 50oC, 70oC, ture constant at 70oC, the system behaves as depicted in
and 90oC) is depicted in Figure 8, at which the PO2 and Figure 9. Once again, it can be noted that by regulating
PH2 are maintained constant at 1/1 atm. The reader can the partial pressures, more terminal voltage and realized
observe that by increasing the temperature, the stack power can be produced by the stack.

TABLE 4 The extracted best parameters along with the lowest SSD in comparisons to others for case study (3)

Parameter NNA SFLA10 ICA10 FOA10 FPA25 SSO34 GOA35 GA35

ξ1 −0.8535 −1.02307 −1.0343 −1.0357 −1.1605 −0.9719 −1.1997 ‐1.1997


−3
ξ2 × 10 2.4316 3.4760 3.3202 2.9502 4.0000 3.3487 3.5505 3.4172
(Computed)
ξ3 × 10−5 3.7545 7.7883 6.4421 3.7669 8.4565 7.9111 4.6144 3.6000
ξ4 × 10−5 −9.5400 −9.5400 −9.5400 −9.5400 −10.1230 −9.5435 −9.5400 −9.5400
λ 13.0802 15.0323 15.0970 15.0297 15.1264 13.0000 13.0092 13.0000
Rc(mΩ) 0.1000 0.1620 0.1650 0.1622 0.12863 0.1000 0.1005 0.1376
β 0.0136 0.0136 0.0136 0.0136 0.0153 0.0534 0.0579 0.0359
*
SSD (best) 2.14487 2.16706 2.16834 2.16709 13.3791 2.18067 2.18586 2.4089
SSD (worst) 2.1645 2.16760 2.51819 2.61422 NR 2.25060 NR NR
−3 −4
SSD (SD) 5.848 × 10 1.0336 × 10 0.07707 0.1298 NR 0.0203 NR NR
*
Abbreviation: NR, Not reported. This value is recalculated using the generated parameters by the FPA.
10 FAWZI ET AL.

FIGURE 9 Behavior analysis of the Ballard Mark V under


FIGURE 8 Behavior analysis of the Ballard Mark V under
regulating pressures: A, I/V polarization curves; B, I/P plots
changed temperatures: A, I/V polarization curves; B, I/P plots
[Colour figure can be viewed at wileyonlinelibrary.com]
[Colour figure can be viewed at wileyonlinelibrary.com]

• Various operating conditions such as different tem-


5 | C O N C L U S I O N S AN D F UT U R E perature variations and pressure regulations are inves-
WORK tigated, and
• Validation of the PEMFC model results‐based NNA
In this paper, the extraction of PEMFCs parameters to with experimental data has been confirmed and
define their characteristics has been treated as an optimi- well‐matched.
zation problem and solved by NNA. According to the
numerical results, the following points can be concluded: It is worthy stating that the authors are scheduling to
extend the current work by implementing a detailed
• Lesser effort in adjusting the NNA controlling param- dynamic model with an appropriate control strategy to
eters to guarantee good performance is observed with improve the dynamic responses of the PEMFC stack
less calculation burden, and to simulate actual working conditions. This charac-
• The best values of the model seven unknown parame- terizes our near future research works.
ters are cropped for accurate characterization of three
real commercial PEMFCs stack under study (eg, Bal-
lard Mark V, BCS‐500 W, and Nedstack PS6) with
A C KN O WL ED G EME N T S
necessary discussions,
• Statistical performance measures have been made to The authors would like to deeply thank the editor and
validate the competency of the proposed NNA‐based anonymous referees for their valid remarks, and con-
tool, structive comments and kind recommendations, which
• Detailed comparisons in this article along with perfor- have significantly upgraded the technical quality and pre-
mance tests have confirmed that the NNA‐based tech- sentation of this article which make it more interesting to
nique is better than the other competing methods, the readers.
FAWZI ET AL. 11

ORCID Phuket, Thailand: IEEE; June 2017:27‐30 10.1109/


ECTICon.2017.8096291.
Attia A. El‐Fergany https://orcid.org/0000-0003-3476-1361 14. Abid RA, Masoud R. Dynamic simulation of a proton exchange
membrane fuel cell system for automotive applications. In: Pro-
ceedings of SEEP2012. Dublin, Ireland: DCU; 05‐08 June
R EF E RE N C E S
2012:311‐316. ISBN:978‐1‐873769‐11‐9.
1. El‐Hay EA, El‐Hameed MA, El‐Fergany AA. Optimized param- 15. Claycomb JR. Algorithms for the magnetic assessment of proton
eters of SOFC for steady state and transient simulations using exchange membrane (PEM) fuel cells. Res in Nondest Eval
interior search algorithm. Energy. 2018;166:451‐461. https:// (RNDE). 2018;9:167‐182. https://doi.org/10.1080/09349847.2017.
doi.org/10.1016/j.energy.2018.10.038 1304596
2. Fathy A, Rezk H. Multi‐verse optimizer for identifying the opti- 16. Priya K, Babu TS, Balasubramanian K, Kumar KS, Rajasekar N.
mal parameters of PEMFC model. Energy. 2018;43:634‐644. A novel approach for fuel cell parameter estimation using sim-
https://doi.org/10.1016/j.energy.2017.11.014 ple genetic algorithm. Sustain Energy Techn Assess.
3. Saleh IMM, Ali R, Zhang H. Simplified mathematical model of 2015;12:46‐52. https://doi.org/10.1016/j.seta.2015.09.001
proton exchange membrane fuel cell based on horizon fuel cell 17. Zhang L, Wang N. An adaptive RNA genetic algorithm for
stack. J Mod Power Syst Clean Energy. 2016;4:668‐679. https:// modeling of proton exchange membrane fuel cells. Int J Hydro-
doi.org/10.1007/s40565‐016‐0196‐5 gen Energy. 2013;38:219‐228. https://doi.org/10.1016/j.
4. Cheng J, Zhang G. Parameter fitting of PEMFC models based on ijhydene.2012.10.026
adaptive differential evolution. Electr Power Energy Syst. 18. Grondin‐Perez B, Roche S, Lebreton CL, Benne M, Damour C,
2014;62:189‐198. https://doi.org/10.1016/j.ijepes.2014.04.043 Kadjo JA. Mechanistic model versus artificial neural network
5. Niu Q, Zhang L, Li K. A biogeography‐based optimization algo- model of a single‐cell PEMFC. Engineering. 2014;6:418‐426.
rithm with mutation strategies for model parameter estimation https://doi.org/10.4236/eng.2014.68044
of solar and fuel cells. Energ Conver Manage. 2014;86:1173‐1185. 19. Chang W. Equivalent circuit parameters estimation for PEM
https://doi.org/10.1016/j.enconman.2014.06.026 fuel cell using RBF neural network and enhanced particle
6. El‐Hay EA, El‐Hameed MA, El‐Fergany AA. Steady‐state and swarm optimization. Math Probl Eng. 2013;2013:67268. https://
dynamic models of solid oxide fuel cells based on satin bower- doi.org/10.1155/2013/672681
bird optimizer. Int J Hydrogen Energy. 2018;43:14751‐14761. 20. Askarzadeh A, Rezazadeh A. Optimization of PEMFC model
https://doi.org/10.1016/j.ijhydene.2018.06.032 parameters with a modified particle swarm optimization. Int J
7. Liu D, Xia S, Tang H, et al. Parameter optimization of PEMFC Energy Res. 2011;35:1258‐1265. https://doi.org/10.1002/er.1787
stack under steady working condition using orthogonal experi- 21. Gong W, Cai Z. Accelerating parameter identification of proton
mental design. Int J Energy Res. 2019;43:2571‐2582. https://doi. exchange membrane fuel cell model with ranking‐based differ-
org/10.1002/er.4131 ential evolution. Energy. 2013;59:356‐364. https://doi.org/
8. Kim YS, Lee KS. Fuel cell membrane characterizations. Polymer 10.1016/j.energy.2013.07.005
Rev. 2015;55:330‐370. https://doi.org/10.1080/15583724.2015. 22. Gong W, Cai Z. Parameter optimization of PEMFC model with
1011275 improved multi‐strategy adaptive differential evolution. Eng
9. Yang S, Chellali R, Lu X, Li L, Bo C. Modeling and optimization App Artif Intell. 2014;27:28‐40. https://doi.org/10.1016/j.
for proton exchange membrane fuel cell stack using aging and engappai.2013.07.016
challenging P systems based optimization algorithm. Energy. 23. Zhu W, Fang J, Zhang W, Xu Y, Tong L. A hybrid differential evo-
2016;109:569‐577. https://doi.org/10.1016/j.energy.2016.04.093 lution for optimum modeling of PEM fuel cells. Arab J Sci Eng.
10. Kandidayeni M, Macias A, Khalatbarisoltani A, Boulon L, 2014;9:2869‐2885. https://doi.org/10.1007/s13369‐014‐0958‐1
Kelouwani S. Benchmark of proton exchange membrane fuel 24. Sun Z, Wang N, Bi Y, Srinivasan D. Parameter identification of
cell parameters extraction with metaheuristic optimization PEMFC model based on hybrid adaptive differential evolution
algorithms. Energy. 2019;183:912‐925. https://doi.org/10.1016/j. algorithm. Energy. 2015;90:1334‐1341. https://doi.org/10.1016/j.
energy.2019.06.152 energy.2015.06.081
11. Kumar P, Kannaiah SKK, Choudhury SR, Rajasekar N. 25. Priya K, Rajasekar N. Application of flower pollination algo-
Genetic algorithm‐based modeling of PEM fuel cells suitable rithm for enhanced proton exchange membrane fuel cell
for integration in DC microgrids. Electr Power Compono Syst. modelling. Int J Hydrogen Energy. 2019;44:18438‐18449.
2017;45:1152‐1160. https://doi.org/10.1080/15325008.2017.1318 https://doi.org/10.1016/j.ijhydene.2019.05.022
980 26. Gong W, Yan X, Hu C, Wangc L, Gao L. Fast and accurate
12. Priya K, Sathishkumar K, Rajasekar N. A comprehensive review parameter extraction for different types of fuel cells with decom-
on parameter estimation techniques for proton exchange mem- position and nature‐inspired optimization method. Energ Conver
brane fuel cell modelling. Renew Sustain Energy Rev. Manage. 2018;174:913‐921. https://doi.org/10.1016/j.enconman.
2018;93:121‐144. https://doi.org/10.1016/j.rser.2018.05.017 2018.08.082
13. Voottipruex K, Sangswang A, Naetiladdanon S, Mujjalinvimut 27. Askarzadeh A, Rezazadeh A. A grouping‐based global harmony
E, Wongyoa N. PEM fuel cell emulator based on dynamic model search algorithm for modeling of proton exchange membrane
with relative humidity calculation. In: 14th Inter. Conf. on Elect. fuel cell. Int J Hydrogen Energy. 2011;36:5047‐5053. https://doi.
Eng./Electronics, Comp., Telecom. and Info. Techn. (ECTI‐CON). org/10.1016/j.ijhydene.2011.01.070
12 FAWZI ET AL.

28. Chakraborty UK, Abbott TE, Das SK. PEM fuel cell modeling 40. Talj RJ, Hissel D, Ortega R, Becherif M, Hilairet M. Experimen-
using differential evolution. Energy. 2012;40:387‐399. https:// tal validation of a PEM fuel‐cell reduced‐order model and a
doi.org/10.1016/j.energy.2012.01.039 moto‐compressor higher order sliding‐mode control. IEEE Trans
29. Ali M, El‐Hameed MA, Farahat MA. Effective parameters' iden- Indust Electron. 2010;57:1906‐1913. https://doi.org/10.1109/
tification for polymer electrolyte membrane fuel cell models TIE.2009.2029588
using grey wolf optimizer. Renew Energy. 2017;111:455‐462. 41. Torres C, Rubio JJ, Aguilar‐Ibáñez CF, Pérez‐Cruz JH. Stable
https://doi.org/10.1016/j.renene.2017.04.036 optimal control applied to a cylindrical robotic arm. Neural
Comput Applic. 2014;24:937‐944. https://doi.org/10.1007/
30. Niu Q, Zhang H, Li K. An improved TLBO with elite strategy for
s00521‐012‐1294‐6
parameters identification of PEM fuel cell and solar cell models.
Int J Hydrogen Energy. 2014;39:3837‐3854. https://doi.org/ 42. Dabkowski P, Galkowski K, Bachelierd O, Rogers E, Sebek M,
10.1016/j.ijhydene.2013.12.110 Kummert A. Control of differential linear repetitive processes
using strong practical stability and H disturbance attenuation.
31. Turgut OE, Coban MT. Optimal proton exchange membrane
Int J Control. 2013;86:636‐649. https://doi.org/10.1080/
fuel cell modelling based on hybrid teaching learning based
00207179.2012.756148
optimization‐differential evolution algorithm. Ain Shams Eng
J. 2016;7:347‐360. https://doi.org/10.1016/j.asej.2015.05.003 43. Sadollah A, Syyaadi H, Yadav A. A dynamic metaheuristic opti-
mization model inspired by biological nervous systems: neural
32. Salim R, Nabag M, Noura H, Fardoum A. The parameter identi-
network algorithm. Appl Soft Comput. 2018;71:747‐782. https://
fication of the Nexa 1.2 kW PEMFC's model using particle
doi.org/10.1016/j.asoc.2018.07.039
swarm optimization. Renew Energy. 2015;82:26‐34. https://doi.
44. AbouOmar MS, Zhang H‐J, Su Y‐X. Fractional order fuzzy PID
org/10.1016/j.renene.2014.10.012
control of automotive PEM fuel cell air feed system using neural
33. Askarzadeh A, Coelho LDS. A backtracking search algorithm network optimization algorithm. Energies. 2019;12:1435. https://
combined with Burger's chaotic map for parameter estimation doi.org/10.3390/en12081435
of PEMFC electrochemical model. Int J Hydrogen Energy.
45. Mann RF, Amphlett JC, Hooper MAI, Jensen HM, Peppley BA,
2014;39:11165‐11174. https://doi.org/10.1016/j.ijhydene.2014.
Roberge PR. Development and application of a generalised
05.052
steady‐state electrochemical model for a PEM fuel cell. J Power
34. El‐Fergany AA. Extracting optimal parameters of PEM fuel cells Sources. 2000;86:173‐180. https://doi.org/10.1016/S0378‐7753
using salp swarm optimizer. Renew Energy. 2018;119:641‐648. (99)00484‐X
https://doi.org/10.1016/j.renene.2017.12.051 46. NedStack PS6 Product Data—Fuel Cell Markets www.
35. El‐Fergany AA. Electrical characterisation of proton exchange fuelcellmarkets.com/content/images/articles/ps6.pdf (Accessed,
membrane fuel cells stack using grasshopper optimizer. IET 4 August, 2018).
Renew Power Gen. 2018;12:9‐17. https://doi.org/10.1049/iet‐ 47. El‐Hay EA, El‐Hameed MA, El‐Fergany AA. Improved perfor-
rpg.2017.0232 mance of PEM fuel cells stack feeding switched reluctance
36. Askarzadeh A, Rezazadeh A. A new heuristic optimization algo- motor using multi‐objective dragonfly optimizer. Neural Comput
rithm for modeling of proton exchange membrane fuel cell: bird Applic. 2018;(In press. https://doi.org/10.1007/s00521‐018‐3524‐z
mating optimizer. Int J Energy Res. 2013;37:1196‐1204. https:// 48. El‐Hay EA, El‐Hameed MA, El‐Fergany AA. Performance
doi.org/10.1002/er.2915 enhancement of autonomous system comprising proton
37. Agwa AM, El‐Fergany AA, Sarhan GM. Steady‐state modeling exchange membrane fuel cells and switched reluctance motor.
of fuel cells based on atom search optimizer. Energies. Energy. 2018;163:699‐711. https://doi.org/10.1016/j.energy.2018.
2019;12:1884. https://doi.org/10.3390/en12101884 08.104
38. Rao Y, Shao Z, Ahangarnejad AH, Gholamalizadeh E, Sobhani
B. Shark smell optimizer applied to identify the optimal param-
eters of the proton exchange membrane fuel cell model. How to cite this article: Fawzi M, El‐Fergany
Energ Conver Manage. 2019;182:1‐8. https://doi.org/10.1016/j. AA, Hasanien HM. Effective methodology based on
enconman.2018.12.057 neural network optimizer for extracting model
39. Rubio JJ. Stable and optimal controls of a proton exchange parameters of PEM fuel cells. Int J Energy Res.
membrane fuel cell. Int J Control. 2014;87:2338‐2347. https:// 2019;1–12. https://doi.org/10.1002/er.4809
doi.org/10.1080/00207179.2014.913201

You might also like