2016 review (in-sun)
TRANSCRIPT
-
8/19/2019 2016 Review (in-Sun)
1/14
Modeling and operation optimization of a proton exchange membrane
fuel cell system for maximum efficiency
In-Su Han a, Sang-Kyun Park b,⇑, Chang-Bock Chung c
a R&D Center, GS Caltex Corp., 359 Expo-ro, Yusung-gu, Daejeon 341222, Republic of Koreab Division of Marine Information Technology, Korea Maritime and Ocean University, Busan 49112, Republic of Koreac School of Applied Chemical Engineering, Chonnam National University, Gwangju 61186, Republic of Korea
a r t i c l e i n f o
Article history:
Received 26 November 2015
Accepted 17 January 2016
Keywords:
Proton exchange membrane fuel cell
Modeling
Simulation
Operation optimization
Artificial neural network
a b s t r a c t
This paper presents an operation optimization method and demonstrates its application to a proton
exchange membrane fuel cell system. A constrained optimization problem was formulated to maximize
the efficiency of a fuel cell system by incorporating practical models derived from actual operations of the
system. Empirical and semi-empirical models for most of the system components were developed based
on artificial neural networks and semi-empirical equations. Prior to system optimizations, the developed
models were validated by comparing simulation results with the measured ones. Moreover, sensitivity
analyses were performed to elucidate the effects of major operating variables on the system efficiency
under practical operating constraints. Then, the optimal operating conditions were sought at various sys-
tem power loads. The optimization results revealed that the efficiency gaps between the worst and best
operation conditions of the system could reach 1.2–5.5% depending on the power output range. To verify
the optimization results, the optimal operating conditions were applied to the fuel cell system, and the
measured results were compared with the expected optimal values. The discrepancies between the mea-
sured and expected values were found to be trivial, indicating that the proposed operation optimization
method was quite successful for a substantial increase in the efficiency of the fuel cell system.
2016 Elsevier Ltd. All rights reserved.
1. Introduction
Fuel cells have been actively studied for the last several decades
because they have been regarded as the most promising alterna-
tives to conventional power generation systems such as internal
combustion engines and gas turbines [1,2]. Several types of fuel
cells, including solid oxide fuel cells (SOFCs), phosphoric acid fuel
cells (PAFCs), molten carbonate fuel cells (MCFCs), direct methanol
fuel cells (DMFCs), alkaline fuel cells (AFCs), and proton exchange
membrane (PEM) fuel cells, have been commercialized for various
applications [3]. Their working principles, advantages and disad-
vantages have been well explained in various references including
a textbook [4]. Among these, PEM fuel cells are suitable for both
stationary and transportation applications such as residential
power generators, cars, buses, forklifts, bicycles, and watercraft
because they offer many advantages, including high efficiencies,
high power densities, short startup times, and low emissions of
pollutants [5].
As fuel cell systems have spread, the need for their operational
optimization to heighten performance or reduce operating costs
has gained increased attention. To maximize the efficiency of a fuel
cell system, and thereby minimize its operating cost, it is essential
that it operates near its optimal operating conditions. This can be
usually achieved by performing operation optimization techniques
based on mathematical models [6]. However, the model-based
optimization of a fuel cell system is a challenging task because
accurate models for all its components must be available in order
to find real optimal operating conditions that will deliver a sub-
stantial improvement in performance. A number of papers dealing
with the operation optimizations of fuel cells have been published
in the open literature. However, most have focused on the opti-
mization of single components [7–12] or sub-systems [13–15]
rather than complete systems [16–22].
A considerable number of papers on the operation optimization
of single fuel cells or sub-systems have been published. Mawardi
et al. [7] proposed a model-based optimization to maximize the
power density of a single PEM fuel cell. Meidanshahi and Karimi
[8] performed an optimization study using a one-dimensional
dynamic model for a singlePEMfuel cell. Zhanget al.[9] determined
the optimal operating temperature of a high-temperature PEM fuel
http://dx.doi.org/10.1016/j.enconman.2016.01.045
0196-8904/ 2016 Elsevier Ltd. All rights reserved.
⇑ Corresponding author. Tel.: +82 514104579; fax: +82 514043985.
E-mail addresses: [email protected] (I.-S. Han), [email protected]
(S.-K. Park), [email protected] (C.-B. Chung).
Energy Conversion and Management 113 (2016) 52–65
Contents lists available at ScienceDirect
Energy Conversion and Management
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e n c o n m a n
-
8/19/2019 2016 Review (in-Sun)
2/14
cell by considering its performance, CO tolerance, and durability.
Kanani et al. [10] used a response surface method to maximize the
power output of a single PEM fuel cell. Ni et al. [11] carried out aparametric study using an electrochemical model to elucidate the
effects of operating variables on the performance of a single SOFC.
Tafaoli-Masoule et al. [12] employed a genetic algorithmand a quasi
two-dimensional, isothermal model to determine the optimal oper-
ating temperature and pressure of a single DMFC. Subramanyan
et al. [13] performed a multi-objective optimization for a hypothet-
ical SOFC–PEM hybrid sub-system both to minimize the CO2 emis-
sion and to maximize the performance. Caliandro et al. [14]
presented a multi-objective optimization for a SOFC–GT (gas tur-
bine) hybrid sub-system both to maximize the efficiency and to
minimize the capital costs. Ranjbar et al. [15] analyzed the effects
of operating variables on the energy and exergy efficiencies of a
hybrid SOFC sub-system, using a zero-dimensional mathematical
model.Several papers on system-level operation optimizations of PEM
fuel cells have appeared in the open literature. Godat and Marechal
[16] performed a simulation study to find the optimal process
structure and operating conditions for a stationary fuel cell system
consisting of a PEM fuel cell stack and fuel processing units. They
analyzed the sensitivity of the major decision parameters (the
steam-to-carbon ratio, reforming and cell temperatures, and fuel
utilization) on the overall efficiency of the fuel cell system. Bao
et al. [17] carried out an optimization study for a hypothetical
PEM fuel cell system, using a hybrid model that combined a
neural-network model with a first-principles model, to find the
optimal operating conditions that maximized net power genera-
tion. The optimal values of two operating variables (the air stoi-
chiometry and cathode outlet pressure) were sought using agenetic algorithm under three different configurations of the
air-supply system. Wu et al. [18] presented an optimization
approach to find the optimal operating conditions for a 25-cm2 sin-
gle PEM fuel cell coupled with a hypothetical compressor and ahumidifier. They employed a meta-modeling approach in which
the input-output relations were approximated with radial basis
functions (RBFs) using the data obtained from a simulator, to
reduce the computational burden in locating an optimal solution.
Four decision variables—the cell temperature, cathode stoichiome-
try, cathode gas pressure, and cathode relative humidity—were
sought under ideal and realistic system assumptions after accom-
plishing a model validation for the fuel cell. Hasikos et al. [19]
adopted a dynamic first-principles model, which was originally
proposed by Pukrushpan et al. [23], as a hypothetical PEM fuel cell
system composed of a stack and auxiliary units to generate opera-
tional data for optimizations. A meta-modeling approach was
employed to build the optimization models from the operational
data using an RBF neural network. They formulated an optimiza-tion problem to minimize the stack current at a given power
demand, and then the optimal operating conditions were used as
set-points for the dynamic matrix controls (DMCs) of the hypothet-
ical system. Wishart et al. [20] performed a system-level optimiza-
tion for an experimental system comprising a Ballard Mark IV fuel
cell stack, a compressor, and pumps. They demonstrated two dif-
ferent optimization cases to find the optimal operating conditions
for vehicular and stationary applications. Mert et al. [21] presented
an optimization of a PEM fuel cell system for vehicular applica-
tions. They carried out a multi-objective optimization of the vehic-
ular fuel cell system both to maximize the power output, energy,
and exergy efficiencies and to minimize the cost of the produced
work. A simple electrochemical model for a Ballard Xcellsis TM HY-
80 fuel cell engine was employed for the optimization. Fran-gopoulos and Nakos [22] performed optimization simulations for
Nomenclature
b bias vector in an artificial neural network modelC pw heat capacity of the cooling water (4.186 kJ kg
1 K1).F flow rate (SLPM)F stoicH2 stoichiometric flow rate of hydrogen entering the stack
(SLPM)
F purge purge gas flow rate from the stack (SLPM) f transfer function of an artificial neural network modeleF Faraday constant (96,485 C mol1) g transfer function of an artificial neural network modelI current (A) J objective function (%)MWa molecular weight of air (28.97 kg kg-mol
1)N s number of cells in the stackP pressure (gauge pressure in kPa)P a discharge pressure of the air from the air blower (gauge
pressure in kPa)P e ambient pressure (gauge pressure in kPa)R universal gas constant (8.314 J mol1 K1)
RMSE root mean squared error defined by
ffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffiffi ffiffiffiPni¼1 ^ yi yið Þ
2=n
q where n is the number of measurements, yi the mea-sured variable, and ^ yi the predicted variable
R2 coefficient of determinationT temperature (C)V cell average cell voltage of the stack (V)W power (kW)W demand power demand of the electricity users (kW)
Greek lettersca mean adiabatic exponent of air (1.402)DT temperature difference (C)
DT g–wa temperature difference between the exhaust gas andthe humidified air (C)
g efficiency (%)gfuel fuel utilization efficiency of the stack (%)h vector of the decision variablesqw density of the cooling water (0.981 kg L 1)x weight matrix in an artificial neural network model
SubscriptsA air blowera airB pump and other balance of plantsc cooling waterg exhaust gas exiting the cathode of the stackP power converterS stackT fuel cell systemwa humidified (wet) air to the stack
SuperscriptsHL hidden layer of an artificial neural network modelin power inputlb lower boundmax maximummin minimum powerOL output layer of an artificial neural network modelout power outputub upper bound
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 53
-
8/19/2019 2016 Review (in-Sun)
3/14
a PEM fuel cell system designed for a marine application. They
used a simplified semi-empirical model to investigate the effects
of the operating temperature and current density on the perfor-
mance and capital cost of the fuel cell system.
However, most of the papers mentioned above considered
hypothetical systems [16,17,19,22] based on simulation models
rather than actual fuel cell systems, and omitted validation steps
[16,17,19–22] for the modeling and optimization. In this work,
an operation optimization method is proposed and applied to an
actual PEM fuel cell system. For the practical application of the
operation optimization method to an actual fuel cell system, this
paper deals with all the necessary steps for the optimization,
including model development based on empirical and semi-
empirical modeling for most of the fuel cell components, the for-
mulation of an optimization problem, and the verification of the
models and optimization results.
2. System description
Fig. 1 shows the PEM fuel cell system for which the operating
conditions are optimized in this study. This system was originally
designed as a stationary power supplier for various electricityusers. As shown in Fig. 1, the fuel cell system consists of two cab-
inets containing the main module and the power converter. The
main module contains most of the fuel cell components, including
the stack, humidifier, air blower, pump, heat exchanger, control
and cell voltage monitoring (CVM) boards, and other balance of
plants (BOPs). The power converter, which elevates the voltage of
the direct current (DC) power to a desired level, is configured in
a separate cabinet.
Fig. 2 illustrates a schematic process diagram of the PEM fuel
cell system. Pure hydrogen is supplied as a fuel to the stack from
high-pressure hydrogen storage, and air is supplied as an oxidant
from an air blower to the stack. The air blower has a single-stage
centrifugal configuration, and its pressure ratio can reach as high
as 1.42 at a flow rate of 2000 SLPM (standard liter per minute).The supplied air is humidified using a membrane-type gas-to-gas
humidifier attached to the cathode side of the stack. In the humid-
ifier, both the humidity and temperature of the dry air from the air
blower are elevated by absorbing both moisture and heat from the
damp, hot exhaust gas exiting the cathode outlet of the stack. A
stack comprising 146 cells was fabricated for the fuel cell system;
it is capable of generating as much as25.5 kW electric power. The
stack has a unique design on the anode side; the cells are divided
into four stages (or blocks) by inserting barriers between the cells,
to maximize the fuel utilization efficiency without a hydrogen
recirculation pump. The hydrogen travels the four stages in a
cascaded manner, whereas the air passes through the undivided
cathode side, as in a conventional PEM fuel cell stack. Han et al.
[24] proposed this stack design and demonstrated a fuel utilization
efficiency as high as 99.6%. The pressure of the hydrogen entering
the stack is regulated at a specified value using a pressure regula-
tor. Since the anode side of the stack operates in dead-end mode, a
purge valve at the anode outlet must be periodically and briefly
opened to prevent the cell voltage from dropping below a certain
limit. Cooling water is used to remove the heat generated from
the stack using the heat exchanger through which the heat is trans-
ferred from the cooling water to the chilling water supplied from a
utility facility. The cooling water temperature at the stack inlet is
controlled by adjusting the flow rate of the chilling water, and that
at the stack outlet is regulated by adjusting the circulation rate of
the cooling water through the stack. The power converter increases
the voltage of the power generated fromthe stack to a desired level
before being supplied to various electricity users. An electronic
load (Model PLW36K-400-1200, manufactured by AMREL Power)
simulates the power consumption of the electricity users. A small
amount of the power from the stack is supplied via the BOP power
inverter, which lowers the voltage of the stack to around 24 V, for
the operation of the BOPs, including the water pump, sensors, and
control boards. The air flow rate is measured using a mass flow
meter (MFM), and the purge gas flow rate is measured using an
MFM (Model M-20SLPM, manufactured by Alicat Scientific) tem-
porarily installed at the anode outlet of the stack. This flow meter
automatically integrates the purge-gas flow rate for a certain per-
iod of time. The flow rate of the cooling water is measured using a
rotary flow meter. In Fig. 2, T and P indicate temperature and pres-
sure measurements, respectively.
3. Formulation of the optimization problem
The fuel cell system was designed to deliver electric power to
various users with variable demands. The objective of the opti-
mization is to maximize the efficiency of the fuel cell system at a
given power demand. Therefore, the objective function of the opti-
mization problem can be described as follows:
max J ðhÞ ¼ gT ¼W T gSW S
¼W TV cellgfuel
1:482W Sð1Þ
where h is the vector of the decision variables and gT is the system
(electrical) efficiency. In this study, the following four operating
variables are the decision variables to be determined by solving
the optimization problem: (1) the stack current (I S), (2) the air flow
rate to the stack (F a), (3) the cooling water temperature (T c), and (4)
the temperature rise of the cooling water through the stack (DT c).
Main module Power converter
Control boards
Air blower
Humidifier Stack Water pump
CVM boards
Fig. 1. Photo of the PEM fuel cell system for which the operating conditions are optimized.
54 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
4/14
Fig. 3 presents the power generation and supply diagram of the
system components and electricity users. The stack is the onlypower generation source and all others are consumers. In Fig. 3,
the points captioned with (I , V ) denote the measurements of cur-
rent and voltage, which enable us to directly measure or calculate
the electric power at the required points for modeling and opti-
mization. Based on the diagram, the objective function (1) must
be subject to the following power balances among the system com-
ponents and electricity users:
W T W demand ¼ 0 ð2Þ
W T W outP þ W A ¼ 0 ð3Þ
W S W inP W B ¼ 0 ð4Þ
W S V cellI SN S=1000 ¼ 0 ð5Þ
In the equations above, the power output of the fuel cell system
(W T) must be equal to the power delivered to the electricity users
(W demand), as described in Eq. (2). A portion of the total power from
the power converter (W outP ) is consumed by the air blower (W A)
andthe remainingpower is delivered to the electricity users, satisfy-
ing the power balance described in Eq. (3). The power generated
from the stack (W S) is delivered to both the power converter (W inP )
and the BOP power inverter (W B), satisfying the power balancedescribed in Eq. (4). DC power of 24 V is supplied through the BOP
power inverter to the water pump and other BOPs. The stack is the
only power generator in the fuel cell system and its power genera-
tion can be calculated according to Eq. (5).
The optimal values for all the decision variables and two other
operating variables must be found within the following bounds:
I lbS 6 I S 6 I
ubS ð6Þ
F lba 6 F a 6 F
uba ð7Þ
T lbc 6 T c 6 T
ubc ð8Þ
DT lbc 6 DT c 6 DT
ubc ð9Þ
F lbc 6 F c 6 F
ubc ð10Þ
T g T wa 6 DT g—wa ð11Þ
The lower and upper bounds are imposed on the decision vari-
ables as described in Eqs. (6)–(9). The flow rate of the cooling water
to the stack (F c) has upper and lower bounds because the cooling
water must flow rapidly enough to fill the cooling channels of
the stack and the cooling water pump has a limited capacity, which
is defined by Eq. (10). As described in Eq. (11), the temperature dif-
ference between the stack exhaust gas (T g) and the wet (humidi-
fied) air from the humidifier (T wa) should be less than a specifiedvalue (DT g–wa) to prevent the humidified air from condensing.
Pressureregulator
PC
Air blower
Hydrogen
TC
P
Powerconverter
HX
Humidifier
Stack
Water vessel
Electricity users(Electronic load)
BOP powerinverter
OtherBOPs
MFM
Purge valve control
VC
Chillingwater
Purge gas
Chillingwater return
Air
T
PT
T
Flow meter
Water pump
P
T
T
T
P
TCCooling water
Air exhaust
MFM(for temporary use)
Fig. 2. Schematic process diagram of the PEM fuel cell system. T and P indicate temperature and pressure measurements, respectively.
Air blower
Power
converter
Stack
Electricity users(Electronic load)
BOP power
inverter
OtherBOPs
Water pump
(I, V)
(I, V)
(I, V)
(I, V)
Fig. 3. Power generation and supply diagram for the PEM fuel cell system.
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 55
-
8/19/2019 2016 Review (in-Sun)
5/14
The optimization problem defined in Eqs. (1)–(11) is a nonlinear
programming (NLP) problem with equality and inequality con-
straints. A sequential quadratic programming (SQP) method is suit-
able for solving this NLP problem [25] efficiently, and therefore,
was employed in this study. The function, fmincon, implementing
an SQP algorithm included in the MATLAB optimization toolbox
[26] was used to solve the optimization problem and perform sen-
sitivity analyses. Because the SQP algorithm does not always guar-
antee a global optimum, a number of starting points for the
decision variables, which were randomly generated, were tested
to locate near-global optimum points. The partial derivatives of
the objective function and nonlinear constraints were approxi-
mated using finite-difference gradients.
4. Experimental and data collection
Mathematical models for the fuel cell system are necessary to
solve the optimization problem. The first step is to collect experi-
mental data to build the models. In this study, test operations of
the fuel cell system were performed to collect operational data
for modeling. The operational data were divided into a modeling
data set and a testing data set for model building and validation,respectively. To collect the modeling data set, the four decision
variables were changed in turn according to the combination of
the decision variables as described in Table 1 during a test opera-
tion. Then, to collect the testing data set, an additional test opera-
tion was carried out by altering the decision variables arbitrarily.
As described in Table 1, three different testing cases that adjusted
the air flow rate to the stack were applied in the test operations.
For each testing case, the flow rate was regulated as a function of
the stack current. As a result, the stoichiometry of air was set to
decrease with increasing stack current, and ranged from 1.59 to
3.66 depending on the stack current and testing case. The cooling
water temperature was controlled at two different temperatures
(54 and 65 C), and the maximum temperature rise of the cooling
water through the stack was also regulated at two different tem-perature gaps (7 and 10 C). Note that the cooling water tempera-
ture rise may not reach its set-point in the lower current range
because the amount of heat generated is insufficient to maintain
the temperature in this range. The stack current was raised from
0 to 250 A in 10- or 20-A steps, and was then reduced to 0 A while
fixing the other three decision variables at a given combination, as
listed in Table 1. The fuel cell system logged the operational data in
a database at 2 s intervals. The first test operation was performed
for about 180 min to collect the modeling data set, and then an
additional test operation was carried out for about 55 min to col-
lect the testing data set. The raw data stored in the database were
averaged every 30 s after removing the data gathered during the
startup and shutdown, to finally afford 311 and 104 data points
for the modeling and testing data sets, respectively.
5. Model development
Empirical and semi-empirical models for most of the system
components, including the stack, humidifier, air blower, power
converter, pump, and other BOPs as shown in Figs. 1 and 2, are pre-
sented in this section.
5.1. Stack–humidifier
The stack and humidifier are treated as a single unit for model-
ing. The unknown variables in modeling the stack–humidifier are
the average cell voltage of the stack (V cell), the temperature of the
stack exhaust gas (T g), the temperature of the humidified air from
the humidifier (T wa), and the fuel utilization efficiency of the stack
(gfuel). To predict the three output variables (V cell, T g, and T wa)
among these unknown variables, an empirical model for the
stack–humidifier was developed by employing artificial neural net-
works (ANNs) [27]. The remaining variable, the fuel utilization effi-
ciency of the stack, was predicted using a semi-empirical model.
Fig. 4 depicts the structure of the neural network, which con-
sists of the five input variables (I S, F a, T c, DT c, and T a), a single
hidden-layer, and the three output variables. ANNs have been
widely used for nonlinear modeling in various fields and also been
applied to the field of fuel cells because they allow greater flexibil-
ity in determining model structures and typically give good predic-
tive performance [28,29]. The basic structure of an ANN consists of
a number of interconnected computing processors, called neurons
or nodes, grouped into input, hidden, and output layers. The
Table 1
Combination of the decision variables altered during the test operations for obtaining the modeling data set.
Decision variables F a T c (C) DT c (C) I S
Combination of the decision variables 358–1054SLPM (corresponding to the stack current 0–250 A) 54 Max 7 0–250 A (10–20A step)
Max 10 ”
65 Max 7 ”
Max 10 ”
421–1240 SLPM (corresponding to the stack current 0–250 A) 54 Max 7 ”
Max 10 ”
65 Max 7 ”
Max 10 ”
484–1426 SLPM (corresponding to the stack current 0–250 A) 54 Max 7 ”
Max 10 ”
65 Max 7 ”
Max 10 ”
Fig. 4. Neural network structure used for modeling of the stack–humidifier.
56 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
6/14
strengths of the connections, called weights, among the nodes are
adjusted to obtain a desired output behavior using given informa-
tion and a learning algorithm. Various types of ANNs have been
proposed according to their structures and learning algorithms
for determining the weights [27]. In this study, a feed-forward net-
work with one hidden layer was employed to model the stack–hu-
midifier. It has been shown that this kind of network can
approximate virtually any function of interest to any degree of
accuracy, as long as enough hidden units are available [30]. In pre-
dicting the unknown variables, the final form of the neural net-
work model can be expressed by a function of weight matrices
(xHL and xOL ) and bias vectors (bHL and bOL ):
^ y ¼ bV cell ; bT g ; bT wah i ¼ g OL xOL f HL xHL x þ bHL þ bOL ð12Þwhere x stands for the vector of the input variables and ^ y for the
vector of the predicted output variables. The log-sigmoid transfer
function ( f HL ( x ) = 1/(1 + exp ( x )) and the pure linear transfer
function ( g OL ( x ) = x ) were used for the hidden and output nodes,
respectively. The weight matrix (xHL ) and bias vector (bHL ) for the
hidden layer and those (xOL and bOL ) for the output layer are the
major model training parameters. These parameters were deter-mined using a back-propagation training algorithm [27].
The fuel utilization of the stack was predicted using Eqs. (13)–
(15). The stack operates in dead-end mode and vents purge gas
periodically to avoid flooding. If the flow rate of the purge gas
from the stack is zero, the amount of hydrogen fed into the stack
will be the same as the stoichiometric flow rate of hydrogen
required for the electrochemical reactions. Then, the fuel utiliza-
tion efficiency of the stack will be 100% according to the follow-
ing equation:
gfuel ¼ 1 F purge
F stoicH2
! 100% ð13Þ
In this equation, the flow rate of the purge gas is linearly pro-
portional to the stack output current [24]. A linear regressionwas carried out using the modeling data set to obtain the following
relationship between the purge gas flow rate and the stack current:
F purge ¼ 0:00341 I S þ 0:00534 ð14Þ
The above regression equation explains 99.1% (R2 (coefficient of
determination) = 0.9908) of the total variability of the measured
data for the purge gas flow rate. In Eq. (13), the stoichiometric flow
rate of hydrogen entering the stack can be calculated by
F stoicH2 ¼N SI S
2eF 60Rð298:15Þ101:33 ð15Þwhere N s is the number of cells in the stack, R is the universal gas
constant (8.314 J mol1 K1), and eF denotes the Faraday constant
(96,485 C mol1).
5.2. Air blower
The air blower consumes a considerable amount of power in a
fuel cell systemand is identical to air compressors in terms of ther-
modynamics. Han and Han [31] proposed a hybrid model for an
air-compression system that combined a thermodynamic com-
pression equation with empirical equations. In this study, the air
blower was modeled by employing the same approach as proposed
by Han and Han [31]. The power consumption of the air blower can
be estimated from the following equation which divides the mini-
mum power consumption ðW minA Þ of the air blower by its overall
efficiency (gA).
W A ¼ 100W minA =gA ð16Þ
The minimum power required for the air blower in an adiabatic
and reversible compression process can be calculated by
W minA ¼ F aRðT a þ 273:15Þca50; 680ðca 1ÞMWa
P a þ 101:33
P e þ 101:33
ca 1ca
ð Þ 1
" # ð17Þ
The overall efficiency of the air blower is a function of both the
air flow rate (F a) and the discharge pressure of the air (P a). A mul-
tivariable linear regression analysis was performed using the per-
formance measurement data gathered from the initial testing of
the air blower to finally obtain the following correlation:
gA ¼ 0:0688 F a 1:1740 Pa þ 1:2199 ð18Þ
The discharge pressure of the air is an unknown state variable
and must be predicted to compute the overall efficiency of the
air blower. A simple linear regression using the modeling data
set was satisfactory to obtain the following relationship between
the discharge pressure and the air flow rate:
P a ¼ 0:03057 F a 5:2792 ð19Þ
The R2 for the above regression is as high as 0.9947, and Eqs.
(18) and (19) are valid only when F a ranges from 358 to 1430 SLPM.
5.3. Power converter
The power converter elevates the voltage of the DC power from
the stack to the desired voltage level [32,33]. The following equa-
tion obtained by a linear regression analysis of the modeling data
set (R2 = 0.9997) was used to predict the power output from the
power converter as a function of the power input:
W outP ¼ 0:9428 W
inP 0:1202 ð20Þ
5.4. Pump and other balance of plants
The cooling water pump is one of the major power consuming
components in the fuel cell system. Power consumption by the
other BOPs, including control boards, cell voltage monitoring
board, sensors, and valves, should be taken into account in model-
ing the fuel cell system. The cooling water pump and other BOPs
were considered as a single unit in the model. The following non-
linear equation was used to predict the power consumption of the
pump and other BOPs:
W B ¼ 0:00152 F 3c 0:00093 F
2c þ 0:57 F c þ 386:4
10
3 ð21Þ
The equation above was obtained from a polynomial curve fit-
ting using the modeling data set in which the measured values of
the cooling water flow rate varied between 12 and 53 LPM. The
cooling water flow rate is manipulated to control the temperaturerise of the cooling water through the stack, and therefore, must be
predicted to calculate the power consumption of the pump and
other BOPs. The following equation obtained from a simple heat
balance on the stack was used to calculate the cooling water flow
rate:
F c ¼ 60W S1:482
V cell 1
qwC pwDT c
ð22Þ
6. Results and discussion
6.1. Model validation
The stack–humidifier model predicts the following fourunknownvariables:the averagecell voltageof thestack, the temper-
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 57
-
8/19/2019 2016 Review (in-Sun)
7/14
ature of thestack exhaust gas, thetemperature of thehumidified air
from the humidifier, and the fuel utilization efficiency. The neural
network model described in Section 5.1 was used to predict these
unknown variables, except for the fuel utilization efficiency. Before
training the neural network, the modeling and testing data were
scaled and mean-centered. The number of hidden nodes in the hid-
den layer is the only tuningparameter in the neural networkmodel.
To determine the tuning parameter, the neural network was trained
using the modeling data set while increasing the number of hidden
nodes over the range from 1 to 20 at intervals of one node. The root
mean squared error (RMSE) between the predicted and measured
outputs, obtainedfrom the neural network model andfrom the test-
ing data set, respectively, was used as the criterionto determine the
optimum number of hidden nodes. Finally, the model structure of
the neural network was established with 7 nodes in a single hidden
layer,which resulted in theminimum RMSE amongthe results from
the numbers of hidden nodes examined.
Fig. 5(a)–(c) compare the predicted outputs with the measured
ones (testing data set) that were not used for training the neural
network. In particular, the predicted values of both the average cell
voltage and the temperature of the stack exhaust gas show excel-
lent agreement with the measured values (Fig. 5(a) and (c)). The
excellent predictive performance is confirmed by the low RMSEs,
0.0023 V and 0.2571 C, for the average cell voltage and the tem-
perature of the stack exhaust gas, respectively. For the temperature
of the humidified air, the predicted values exhibit somewhat larger
deviations from the measured values than those of the other two
output variables, as shown in Fig. 5(b), resulting in a low RMSE
of 0.3892 C and a maximum deviation of 1.1 C.
The fuel utilization efficiency was predicted by the semi-
empirical models, Eqs. (13)–(15). The calculated RMSE between
the predicted and measured values is as small as 0.0160%. Fig. 5
(d) compares the measured and predicted values of the fuel utiliza-
tion efficiency. The predicted values of this variable are generally in
accordance with the measured ones, even though small fluctua-
tions of the fuel utilization efficiency are not predicted well. The
prediction errors arising from these small fluctuations are trivial
if one considers that the fuel utilization efficiency varies within a
maximum range of about 0.06%, which is negligibly small and con-
tributes a maximum prediction error of only about 0.03% in the
system efficiency.
The models for the air blower, the power converter, and the
pump and other BOPs were verified by comparing the predicted
and measured results. Fig. 6(a)–(d) shows the prediction results
for the power consumption of the air blower, the power output
of the converter, the power consumption of the pump and other
BOPs, and the power output of the fuel cell system, respectively.
The power output of the system was calculated from the power
balance equations (Eqs. (3)–(5)). As can be seen in Fig. 6(a)–(d),
all the predicted values match the measured values quite well,
even though there are some small deviations in certain regions.
All the RMSEs for the predicted variables are less than 0.05 kW
(which corresponds to only 0.23% of the maximum power output
of the system): 0.0319, 0.0482, 0.0292, and 0.0489 kW for the
Fig. 5. Prediction results fromthe model for the stack–humidifier:(a) average cell voltage of the stack, (b) temperature of humidified air fromthe humidifier, (c) temperatureof stack exhaust gas, and (d) fuel utilization efficiency of the stack.
58 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
8/14
power consumption of the air blower, the power output of the con-
verter, the power consumption of the pump and other BOPs, and
the power output of the system, respectively. The results from
the model validations indicate that the empirical and semi-
empirical models developed for the fuel cell system demonstrate
satisfactory predictive performance and can be used for the
optimization.
6.2. Sensitivity analysis
Prior to conducting the optimization of the operating conditions
of the fuel cell system, it was worth analyzing the extent of the
effect exerted by a decision variable on the objective function
(i.e., the system efficiency). Sensitivity analyses were performed
to elucidate the effects of the decision variables on the objective
function under the constraints given by Eqs. (2)–(11). Among the
four decision variables, the stack current cannot be freely changed
for the sensitivity analyses because it must meet the equality con-
straints (Eqs. (2)–(5)). Thus, the remaining three decision variables
(F a, T c, and DT c) were varied within their operating bounds shown
in Table 2. Each of the three decision variables was gradually
increased from the lower to the upper bound while the other
two decision variables were fixed at constant values, as given in
Table 3, to identify variations in the objective function.
Fig. 7 illustrates the effects of the air flow rate on both the sys-
tem efficiency and the stack power when the power output of the
system is fixed at the following power loads applied to the system
in sequence: 10%, 50%, 75%, and 100%. In general, the power output
Fig. 6. Prediction results from the models for the air blower, power converter, and the pump and other BOPs: (a) power consumption of the air blower, (b) power output of
the power converter, (c) power consumption of the pump and other BOPs, and (d) power output of the system.
Table 2
Constraints imposed on the decision variables and the other two operating variables.
Variable Lower bound Upper bound Remark
I S 0 A 250 A
F a 358 SLPM 1426 SLPM – The lower and upper bounds
vary depending on the stack
current
– The lower and upper bounds
are at stack currents of 0 and
250 A, respectively.
T c 54 C 65 C
DT c 0 C 10 C
F c 12 LPM 53 LPM
DT g–wa N/A 8 C
Table 3
Base operating conditions of the decision variables given for each power load for the
sensitivity analyses.
Power load (%) Decision variablesa
F a (SLPM) T c (C) DT c (C)
10 420 60 1.8
50 805 60 7.0
75 1100 60 8.0
100 1240 60 10.0
a
The stack current is automatically determined by solving the optimizationproblem.
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 59
-
8/19/2019 2016 Review (in-Sun)
9/14
of a PEM fuel cell stack is known to increase as the air flow rate
entering the stack rises, within a certain limit (i.e., increasing the
stoichiometry of the air) [34,35]. As expected, the stack power
increases almost linearly as the air flow rate increases for all the
power loads considered. Changes in the stack power owing to
the increases in the air flow rate from the lower to the upper
bounds are quite large, ranging from 0.15 to 0.47 kW depending
on the power load applied to the system. For power loads of 10%,
50%, and 75%, the system efficiency shows a tendency to decrease
with an increasing air flow rate. However, for the maximum power
load (100%), the tendency is reversed and decreases with an
increasing air flow rate, as can be observed in Fig. 7(d). When the
air flow rate is changed from the lower to the upper bounds, the
system efficiency reveals substantial changes: 1.8%, 0.9%,
0.4%, and 0.8% for power loads of 10%, 50%, 75%, and 100%,
respectively. From these results, therefore, it can be concluded that
a change in the air flow rate has a considerable effect on the systemefficiency, which can be either positive or negative depending on
the power load applied.
Fig. 8 depicts the effects of the cooling water temperature on
both the system efficiency and the stack power when the power
output of the system is fixed at the same power loads as those
applied in Fig. 7. As can be seen in Fig. 8, the stack power remains
almost constant (about 0 kWfor a 10% load) or is slightly decreased
(0.003, 0.006, and 0.16 kW for 50%, 75%, and 100% loads,
respectively) with increases in the cooling water temperature from
the lower to the upper bounds. On the other hand, the system effi-
ciency shows substantial changes with increases in the cooling
water temperature from the lower to the upper bounds. Fig. 8
shows that the increases in the cooling water temperature result
in some gains in system efficiency. These gains become larger as
the power load is increased, attaining about 0.3%, 0.4%, 0.5%, and
0.7% at loads of 10%, 50%, 75%, and 100%, respectively.
Fig. 9 shows the effects of the rise in the cooling water temper-
ature on both the stack power and the system efficiency when the
power output of the system is fixed at the following loads: 10%,
50%, 75%, and 100%. For all the power loads investigated, the stack
power decreases as the temperature rise of the cooling water
increases within its operating bounds. However, the changes in
the stack power are small when the temperature rise of the cooling
water is altered from the lower to the upper bound, ranging from
0.117 to 0.160 kWdepending on the power load applied to the sys-
tem. The system efficiency increases with the increasing cooling
water temperature rise for all the power loads investigated. The
gains in the system efficiency are not trivially small (2.0%, 0.7%,
0.5%, and 0.3% for power loads of 10%, 50%, 75%, and 100%, respec-
tively) when the temperature rise of the cooling water is elevated
from the lower to the upper bound.The sensitivity analyses suggest that the system efficiency is
somewhat more sensitive to both the air flow rate and the cooling
water temperature rise than to the cooling water temperature. The
system efficiency is proportional to both the cooling water temper-
ature and its temperature rise for all the power loads investigated.
On the other hand, the system efficiency can be increased or
decreased as the air flow rate increases: it is proportional to the air
flow rate when the power load is less than or equal to 75%, and is
inversely proportional to the air flow rate when the power load is
100%. That is, an increase in the air flow rate exerts either a positive
or a negative effect on the system efficiency depending on the
applied power load. It was found that a change in one of the three
decision variables within its operating bounds could produce a sub-
stantial change in the system efficiency (e.g., the 2% increase in
Fig. 7. Effects of the air flow rate on the system efficiency and the stack power: (a) 10% load, (b) 50% load, (c) 75% load, and (d) 100% load.
60 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
10/14
Fig. 8. Effects of the cooling water temperature on the system efficiency and the stack power: (a) 10% load, (b) 50% load, (c) 75% load, and (d) 100% load.
Fig. 9. Effects of the temperature rise of the cooling water through the stack on the system efficiency and the stack power: (a) 10% load, (b) 50% load, (c) 75% load, and (d)100% load.
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 61
-
8/19/2019 2016 Review (in-Sun)
11/14
system efficiency with an increasing cooling water temperature
rise). Therefore, if all three decision variables are simultaneously
changed in a positive manner, it is expected that the system effi-
ciency can be increased by as much as several percent.
6.3. Optimization and verification
The operating conditions for the fuel cell system described inFigs. 1–3 were optimized using the described optimization method.
The optimal operating conditions were sought for a total of 11
power loads which ranged from the minimum (10%) to the maxi-
mum (100%) at intervals of 5–15%. The minimum and maximum
loadscorrespond to systempower outputs of 2.2 and22 kW, respec-
tively. Fig. 10 displays a plot of the maximum system efficiency
against thepower output of thesystem. As shown in Fig. 10, the sys-
tem efficiency rises sharply as the power output increases, until
reaching its maximum, and then gradually declines. Typically, the
system efficiency of a PEM fuel cell system is influenced by both
the stack efficiency and the BOP power consumption. Without the
BOP power consumption, the system efficiency would gradually
diminish as thepower output increases, from its maximum at a zero
load to its minimum at a maximum load. For an actual fuel cell sys-tem, however, the system efficiency displays a unimodal curve
against the power output owing tothe effect of the BOPson the sys-
tem efficiency. The lower system efficiencies in the lower power
output range (2.2 kW < the power output < 6 kW) are due to the
high base-power consumption of the BOPs. The maximum system
efficiency is found to reach the least value of 40.9% at the minimum
power output (2.2 kW) and the greatest value of 45.6% at a power
output of 6.6 kW. As shown in Fig. 10, the maximum system effi-
ciencyis quite variable dependingon the power output, andthe lar-
gest efficiency gap reaches about 4.5% from the differences of the
power load.
The optimal decision variables, illustrated in Fig. 11, were deter-
mined to obtain the maximum system efficiencies shown in
Fig. 10. The stack current is the most effective variable for adjusting
the power output of the system, and thus, its optimal values are
obtained at the points where the stack power generation meets
the power requirements of both the system components and the
electricity users. As can be seen in Fig. 11(b), the optimal values
of the air flow rate are equal to the lower bound until the power
output reaches 13.2 kW, and then rise between the lower and
upper bounds. This indicates that, as expected from Fig. 7, the ten-
dency of the system efficiency to decrease with the increasing air
flow rate begins to reverse at power outputs greater than
13.2 kW. For all the power outputs investigated, the optimal values
of the cooling water temperature are obtained at the upper bound,
as expected from the sensitivity analysis results shown in Fig. 8.
The optimal values of the cooling water temperature rise are
obtained between the lower and upper bounds until the power
output approaches 8.8 kW, after which they stay at the upper
bound. In the lower range of the power output, the heat generated
from the stack is not large, so that the optimal values of the tem-
perature rise of the cooling water are sought below the upper
bound.
Fig. 12 plots the proportions of the power consumption by the
BOPs (the blower, the pump and other BOPs, and the power con-
verter) to the power generation by the stack at the optimal operat-
ing conditions when the power output of the system is varied from
2.2 to 22 kW. The proportion of the total power consumption by
the BOPs to the power generation sharply drops with the power
output, from the largest value (27.6%) at the minimum power out-
put (2.2 kW) to about 16% at a power output of around 6 kW, and
then gradually decreases to 13.5% until the maximum power out-
put is reached. This confirms that the lower system efficiencies
in the lower power output range are mainly due to the relatively
large proportion of the total power consumption by the BOPs to
the power generation in this range. Especially, the pump and other
BOPs show the largest contribution to such large proportions of the
total power consumption in the lower power output range. This
means that, to raise the system efficiency in this lower power out-
put range, the power consumption of the pump and other BOPs
should be reduced. When the power output is greater than about
4.4 kW, the power converter consumes the largest portion of the
total power among the BOPs. This result emphasizes that an
improvement in the efficiency of the power converter will be a
more effective way to increase the system efficiency across the
power output range.It is possible to estimate achievable improvements in the sys-
tem efficiency by calculating the differences between the mini-
mum and maximum system efficiencies obtained from the worst
and best operating conditions, respectively, in the adjustable
ranges of the operating variables. An additional optimization was
carried out to find the minimum system efficiencies owing to the
worst operating conditions by placing a minus sign before the
objective function with the same constraints as those used for
the maximization of the system efficiency. Fig. 13 illustrates the
minimum and maximum system efficiencies along with the effi-
ciency gap between the two against the power output of the sys-
tem. The minimum system efficiency ranges from a low of 35.4%
to a high of 42.3% over the power output range. The efficiency
gap attains the largest value (5.5%) at the minimum power outputand then diminishes with increasing power output until it reaches
the smallest value (1.2%) at 19.8 kW. In the lower power output
range, the efficiency gap is quite large, ranging from 3.6% to 5.5%,
while in the higher power output range (power output > 13.2 kW),
it is smaller, ranging from 1.2% to about 2%. Therefore, the opti-
mizations and operations of the fuel cell system in the lower power
output range should be given a higher priority in handling than
those in the higher power output range to obtain a substantial
increase in the system efficiency.
To verify the operation optimization method proposed in this
study, the operating conditions obtained from the maximizations
of the system efficiency were applied to the fuel cell system. The
decision variables of the system were adjusted to their optimal val-
ues while regulating the power output at each of the followingpower loads in sequence: 10%, 50%, and 100%. After the system
38
40
42
44
46
48
0 5 10 15 20 25
M a x i m u m s
y s t e m e f f
i c i e n c y / %
Power output of the system / kW
Fig. 10. Maximum system efficiency obtained from the optimization across thepower output of the system.
62 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
12/14
reached a steady state under each optimal set of operating condi-
tions, the data for the operating variables were collected for
10 min and then averaged. Table 4 compares the measured
optimization results with the expected optimal values. Overall,
the discrepancies between the measured and expected values are
trivially small. The maximum deviation of the system efficiency
is as small as 0.6% when a power load of 10% was applied to the
system. These results indicate that the operation optimization
method is quite successful for a substantial increase in the
efficiency of the fuel cell system.
Fig. 11. Optimal decision variables obtained from the maximization of system efficiency: (a) stack current, (b) air flow rate; (c) cooling water temperature, and (d)
temperature rise of the cooling water through the stack.
0
5
10
15
20
25
30
0 5 10 15 20 25
P r o p o r t i o n o f t h e p o w e r c o n s u m p t i o n
t o t h e p o w e r g e n e r a t i o n / %
Power output of the system / kW
Blower
Pump & other BOPs
Converter
Total power consumption
Fig. 12. Proportion of the power consumption by the BOPs (the blower, the pump
and other BOPs, and the power converter) to the power generation by the stack
under the optimal operation conditions when the power output of the system
varies from 2.2 to 22 kW.
0
1
2
3
4
5
6
34
36
38
40
42
44
46
0 5 10 15 20 25
E f f i c i e n c y g a p / %
S y s t e m e
f f i c i e n c y / %
Power output of the system / kW
Maximum efficiencyMinimum efficiencyEfficiency gap
Fig. 13. Minimum and maximum system efficiencies against the power output of
the system, obtained from the optimizations under the constraints.
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 63
-
8/19/2019 2016 Review (in-Sun)
13/14
7. Conclusions
An operation optimization method and its application to an
actual PEM fuel cell system were demonstrated to maximize the
system efficiency. The empirical and semi-empirical models used
in this study covered most of the system components including
the stack, humidifier, blower, power converter, and a pump andother balance of plants. The models are based on neural networks
and semi-empirical equations. Model validations results showed
that the developed models had satisfactory predictive performance
for the optimizations. Sensitivity analyses were carried out to elu-
cidate the effects on the system efficiency of the major operating
variables, including the air flow rate, the cooling water tempera-
ture, and the temperature rise of the cooling water through the
stack. Optimizations of the fuel cell system were performed to seek
the best operating conditions and were verified by comparing the
expected optimal values with the measured ones. The optimization
results showed that the efficiency gaps between the best and worst
performance of the system can reach 1.2–5.5%, depending on the
power load.
The proposed operation optimization method can be easilyextended to similar PEM fuel cell systems for stationary power
generators or vehicular applications, by incorporating an appropri-
ate data set for the targeted components, although the optimiza-
tion model is based on empirical and semi-empirical modeling
for a specific fuel cell system. The optimization method may
require a model update for a specific fuel cell system or a new
operating region. However, this model can be more easily updated
with higher accuracy than first-principles models if a sufficient
amount of operational data is obtained.
Acknowledgement
This work was funded by the Ministry of Trade, Industry &
Energy (MOTIE) through the Korea Institute for the Advancementof Technology (KIAT) under grant number R0001668.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.enconman.2016.
01.045.
References
[1] Sharaf OZ, Orhan MF. An overview of fuel cell technology: fundamentals and
applications. Renew Sustain Energy Rev 2014;32:810–53.
[2] Mekhilef S, Saidur R, Safari A. Comparative study of different fuel cell
technologies. Renew Sustain Energy Rev 2012;16:981–9.
[3] Kirubakaran A, Jain S, Nema RK. A review on fuel cell technologies and powerelectronic interface. Renew Sustain Energy Rev 2009;13:2430–40.
[4] Larminie J, Dicks A. Fuel cell systems explained. 2nd ed. Chichester, UK: Wiley;
2003.
[5] Wang Y, Chen KS, Mishler J, Cho SC, Adroher XC. A review of polymer
electrolyte membrane fuel cell: technology, applications, and needs on
fundamental research. Appl Energy 2011;88:981–1007.
[6] Zhu J. Optimization of power system operation. 2nd ed. Hoboken, NJ,
USA: Wiley; 2015.
[7] Mawardi A, Yang F, Pitchumani R. Optimization of the operating parameters of
a proton exchange membrane fuel cell for maximum power density. J Fuel CellSci Technol 2005;2:121–35.
[8] Meidanshahi V, Karimi G. Dynamic modeling, optimization and control of
power density in a PEM fuel cell. Appl Energy 2012;93:98–105.
[9] Zhang C, Zhou W, Ehteshami MM, Wang Y, Chan SH. Determination of the
optimal operating temperature range for high temperature PEM fuel cell
considering its performance, CO tolerance and degradation. Energy Convers
Manage 2015;105:433–41.
[10] Kanani H, Shams M, Hasheminasab M, Bozorgnezhad A. Model development
and optimization of operating conditions to maximize PEMFC performance by
response surface methodology. Energy Convers Manage 2015;93:9–22.
[11] Ni M, Leung MKH, Leung DYC. Parametric study of solid oxide fuel cell
performance. Energy Convers Manage 2007;48:1525–35.
[12] Tafaoli-Masoule M, Bahrami A, Elsayed EM. Optimum design parameters and
operating condition for maximum power of a direct methanol fuel cell using
analytical model and genetic algorithm. Energy 2014;70:643–52.
[13] Subramanyan K, Diwekar UM,Goyal A. Multi-objective optimizationfor hybrid
fuel cells power system under uncertainty. J Power Sources 2004;132:99–112.
[14] CaliandroP, TockL, Ensinas AV, Marechal F. Thermo-economic optimization of
a solid oxide fuel cell–gas turbine system fuelled with gasified lignocellulosicbiomass. Energy Convers Manage 2014;85:764–73.
[15] Ranjbar F, Chitsaz A, Mahmoudi SMS, Khalilarya S, Rosen MA. Energy and
exergy assessments of a novel trigeneration system based on a solid oxide fuel
cell. Energy Convers Manage 2014;87:318–27.
[16] Godat J, Marechal F. Optimization of a fuel cell system using process
integration techniques. J Power Sources 2003;118:411–23.
[17] Bao C, Ouyang M, Yi B. Modelingand optimization of theair systemin polymer
exchange membrane fuel cell systems. J Power Sources 2006;156:232–43.
[18] Wu J, Liu Q, Fang H. Toward the optimization of operating conditions for
hydrogen polymer electrolyte fuel cells. J Power Sources 2006;156:388–99.
[19] Hasikos J, Sarimveis H, Zervas PL, Markatos NC. Operational optimization and
real-time control of fuel-cell systems. J Power Sources 2009;193:258–68.
[20] Wishart J, Dong Z, Secanell M. Optimization of a PEM fuel cell system based on
empirical data and a generalized electrochemical semi-empirical model. J
Power Sources 2006;161:1041–55.
[21] Mert SO, Ozcelik Z, Ozcelik Y, Dincer I. Multi-objective optimization of a
vehicular PEM fuel cell system. Appl Therm Eng 2011;31:2171–6.
[22] Frangopoulos CA, Nakos LG. Development of a model for thermoeconomicdesign and operation optimization of a PEM fuel cell system. Energy
2006;31:1501–19.
[23] Pukrushpan JT, Stefanopoulou AG, Peng H. Control of fuel cell power
systems. New York: Springer; 2004.
[24] Han I-S, Jeong J, Shin HK. PEM fuel-cell stack design for improved fuel
utilization. Int J Hydrog Energy 2013;38:11996–2006.
[25] Rao SS. Engineering optimization: theory and practice. 4th ed. Hoboken, NJ,
USA: John Wiley & Sons; 2009.
[26] Optimization toolbox user’s guide. [accessed 12.29.2015].
[27] Hagan MT, Demuth HB, Beale M, De Jesus O. Neural network design. 2nd
ed. Boston: Cengage Learning; 2014.
[28] Saengrung A, Abtahi A, Zilouchian A. Neural network model for a commercial
PEM fuel cell system. J Power Sources 2007;172:749–59.
[29] Han I-S, Shin HK. Modeling of a PEM fuel cell stack using partial least squares
and artificial neural networks. Korean Chem Eng Res 2015;53:236–42.
[30] Hornik K. Approximation capabilities of multilayer feedforward networks.
Neural Networks 1991;4:251–7.
[31] Han I-S, Han C. Modeling of multistage air-compression systems in chemicalprocesses. Ind Eng Chem Res 2003;42. 2209–8.
Table 4
Implementation results of the optimal operating conditions at various power loads.
Load (%) Values System efficiency (%) System power output (kW) Stack power (kW) Decision variables
I S (A) F a (SLPM) T c (C) DT c (C)
10 Expected 40.9 2.20 3.04 24.8 358 65.0 2.8
Measured 40.3 2.27 3.17 25.9 357 64.9 2.8
Deviation 0.6 0.07 0.13 1.1 1 0.1 0.0
50 Expected 44.6 11.00 12.76 113.5 658 65.0 10.0Measured 44.7 11.11 12.83 114.6 657 64.9 10.3
Deviation 0.1 0.11 0.07 1.1 1 0.1 0.3
100 Expected 41.1 22.00 25.44 246.8 1334 65.0 10.0
Measured 41.0 22.09 25.62 248.0 1326 65.0 10.3
Deviation 0.1 0.09 0.18 1.2 8 0.0 0.3
64 I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65
-
8/19/2019 2016 Review (in-Sun)
14/14
[32] Kolli A, Gaillard A, De Bernardinis A, Bethoux O, Hissel D, Khatir Z. A review on
DC/DC converter architectures for power fuel cell applications. Energy Convers
Manage 2015;105:716–30.
[33] De Bernardinis A. Synthesis on power electronics for large fuel cells: From
power conditioning to potentiodynamic analysis technique. Energy Convers
Manage 2014;84:174–85.
[34] SantarelliMG, Torchio MF, Cali M, Giaretto V. Experimental analysis of cathode
flow stoichiometry on the electrical performance of a PEMFC stack. Int J
Hydrog Energy 2007;32:710–6.
[35] KimB, Cha D, KimY. The effects of air stoichiometryand air excessratioon the
transient response of a PEMFC under load change conditions. Appl Energy
2015;138:143–9.
I.-S. Han et al. / Energy Conversion and Management 113 (2016) 52–65 65