INTERNATIONAL

 JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    

«authorheader»  et  al.,  Vol.4,  No.3,  2014  

PEM Fuel Cell Modelling Using Artificial Neural

Networks (ANN)

K. Belmokhtar*‡, M.L Doumbia*, K. Agboussou*

* Department electrical and computer engineering

‡  
Corresponding Author; «corrauthor», University of Quebec at Trois-Rivieres, Quebec, Canada 3351, Boul Des Forges, Trois-
Rivières, QC G9A 5H7

Received: 16.08.2014 Accepted: 07.09.2014

Abstract- Fuel cells (FC) convert directly into a dc electrical energy the chemical energy of a reaction of hydrogen and
oxygen. Proton Exchange Membrane (PEMFC) is a suitable alternative for both electrical transportation and stationary
applications. This article deals with an Artificial Neural Network (ANN) modelling method of a PEMFC. This modelling
approach permits to describe both transient and steady state behaviours of the PEMFC voltage. Furthermore, the prediction of
the operating temperature of a PEMFC based only on its measured voltage and current is proposed and tested successfully.
Indeed, experimental data from a 1.2 kW Nexa Ballard PEMFC is used to validate the proposed method.
Keywords- Neural network; Thermal model; Polymer electrolyte fuel cells; green energy; distributed sources.

1. Introduction without any risk for the stack, and a low system operating
cost. Modelling of PEMFC met a growing interest in the
Proton Exchange Membrane Fuel Cell (PEMFC) is literature, where it is usually done with complex models
considered as a potential future green power source for both based on knowledge of the physicochemical phenomena [1-
electrical transportation and stationary application, due to its 3]. These models need a good knowledge of the parameters
high efficiency, zero emission if it runs with pure hydrogen describing the behavior of the process [4-7]. Generally, these
and its low operating temperature. The main challenge of parameters are no easy to establish for the PEMFC systems.
PEMFC systems is the design of a power converter suitable A model describing the transient behavior of a PEMFC stack
for conditioning the output power with high efficiency and with equations is given in [8]. However, the internal
reliability. Indeed, about 80% of the damage occurred in the parameters should be defined such as the ohmic resistance,
PEMFC system are involved by the power converters. which determines the humidity of the membrane, as well as
During the design phase, the power converter must be tested overflowing and drying of electrodes. These internal
and adjusted with a real PEMFC, and thereafter, it must be parameters are significant when the cell voltage is
validated. However, the design and development of PEMFC considered, but they were not considered in this
including auxiliaries such as testing an air compressor mathematical model, as the required parameters are difficult
control, power and energy management and performance to calculate. Hybrid models could overcome these issues. In
optimization can damage a PEMFC easily. In addition, the [9], a PEMFC model has been developed, which is able to
cost of testing (hydrogen consumption and the secure characterize the cell either steady state or transient.
facilities requirements) is still relatively high, for Combined electrical circuit based model and the empirical
experiments with a real PEMFC. These disadvantages model, the proposed model presents a good agreement with
demonstrate the great importance of the design of a PEMFC the experimental results. However, this model is correct only
emulator in real time based on a model material for in a small range.
applications such as HIL (Hardware In the Loop). During the Nevertheless, it is possible to obtain a behavioral
design process of the power system of PEMFC, power modeling and without the identification of all these
converters and the auxiliaries development can be initially parameters through a models so called "black box". These
verified and increased with a PEMFC emulator in real time models are based on readily measurable variables such as
INTERNATIONAL  JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    
«authorheader»  et  al.,  Vol.4,  No.3,  2014  

temperature, pressure, or the current of the cell and are able function, which is the error between the network output and
to estimate the output voltage of the PEMFC. the desired output as expressed as:
Today, dynamic models of PEMFC systems based on
Artificial Neural Networks (ANN) are rare in the literature. 1 1
However, a lot of substantially stationary models have been e= ∑
2 j 2 j
[
e 2j (k ) = ∑ y *j − y j (k ) ] 2

established with good results [10-13].

In [13], a static and dynamic model of the PEMFC based (1)
on ANN has been proposed and experimental results were Where yj(k) is the output of jth neuron and yj*(k) is the
presented. This proposed model uses the measured desired output.
temperature, the fuel cells current, the stoichiometry of the To achieve a neuronal dynamic modelling the PEM fuel
two gases and humidity. This modeling approach gives good cell, the neuronal structure illustrated in Fig. 1 was chosen,
results, even if it suffers from a disadvantage concerning the which represents a typical processing element which forms a
evolution of the temperature which is not considered. In [14], weighted sum of its inputs and puts the result via a nonlinear
neuronal modeling of a high power PEMFC is presented transfer function to the output. These transfer functions can
where the evolution of temperature was considered. also be linear, and then the weighted sum is transmitted
However, due to the not recurring structure of the ANN, the directly to the output path. The multi-layer perceptron (MLP)
model presented is static [13]. In addition the temperature used in this work is composed of three main layers. The first
was measured at the anode circuit and in the water tank part is the input layer where experimental data are presented,
which does not provides the accurate operating temperature then are processed and propagated via a hidden layer, to the
of the fuel cell. Indeed, in [15], an infrared camera was used output layer. Training a network consist to modify
in order to improve acquisition of the temperature in the continuously the weights of the connecting links between
cells. processing elements as patterns of inputs and corresponding
In this paper, a novel modelling approach of the PEMFC desired outputs are presented to the network [22].
to provide stack voltage characteristic and the operating Equ. (2) and Equ. (3) represent respectively the weighted
temperature is presented based on ANN and its performances summation of the inputs and the non linear transformation
are analyzed. This method called "black box" consists of the (transfer function) to the output of the neuron.
dynamic neural modeling with recurrent ANN structure to
predict with a good accuracy the PEMFC voltage and σ = ∑ xi wi
operating temperature by the use an experimental data. i (2)

2. Development of the ANN Modelling Approach

Y = f (σ ) (3)

ANN is commonly considered as an attractive and

powerful tool to provide the relation of complex and
nonlinear dynamic model based only on input-output data
mapping [16-18]. In the literature, they are several models
have been proposed for different applications [19].
Generally, the neural network has two types of structure, a
feed-forward architecture networks where the direction of the
signals is only from input to output. Unlike, for networks Figure 1. Schematic diagram of a basic formal neuron.
feed-back structure, the direction of the signals is from input
A recurrent neural structure allows, with adding time
to output and vice versa [13]. ANN learning can be classified
delays to take into account the time variation of input
in two categories [20]: the supervised (learning with a
parameters to analyze the dynamics of output [13]. In order
teacher) or unsupervised (learning without a teacher) [13].
to speed up the convergence of the learning process of ANN,
For the supervised training which assumes the availability of
we choose to use the Levenberg-Marquardt (LM) method
a supervisor, each element in input vector represents an
which improves the gradient decent method of back-
explicit element in the output target vector. While in the
propagation [23].
unsupervised training, the ANN model is not trained to any
The neural electrochemical model the PEMFC consists of
certain output target, and the reinforcement learning is
three layers, the input layer, hidden layer and the output layer
accomplished via a trial and error learns [20, 21].
(see Fig. 2). The activation function used for the first two
A typical feed–forward neural network perceptron with
layers is the tangent sigmoid function (tansig), while the
Back-propagation (BP) algorithm has been used in this study.
The BP algorithm which has a large classification capacity is
widely used in the area of identification and control [10]. The
algorithm uses the technique of gradient descent search in
order to reduce a cost function mean squared error (MSE).
The minimization process is carried out by modifying the
weight vector of the neural networks. Some training
algorithms have been presented in order to adapt the weight
values in the dynamic recurrent network. The minimized cost

726  
 
INTERNATIONAL  JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    
«authorheader»  et  al.,  Vol.4,  No.3,  2014  

linear activation function is used for the output layer.

Figure 3. A block diagram of an ANN PEMFC operating

Figure 2. Schematic diagram of the neural electrochemical temperature.
model the PEM fuel cell.
The structure of neural networks used to provide the 3. Experimental Setup Details
operating temperature of the PEMFC is illustrated in Fig. 3.
For training the ANN, having regard to the electrochemical The all experimental data which are used in this paper are
performance of the PEMFC model presented previously, we derived from [24]. A 1200 W Nexa Ballard with 47 cells
used the same configuration in terms of type of activation connected in series. In order to provide an accuracy cells
and the number of neurons in each layer function. We took operating temperature, an infrared camera is used. The
into account the dynamics of the current and voltage by current profile used in this analyze is depicted in Fig. 4.
introducing time delays to the input of ANN.

Figure 4. A very high dynamic applied current profile [24].

4. Simulation Results and Discussion

To validate the effectiveness of the neural modelling

approach of PEMFC a training process using experimental
Table 1. PEM fuel cell output voltage model parameters
data is achieved using Matlab/Simulink® and Neural
Network Toolbox™. A number of epochs between Epochs between displays 10
displaying, maximum number of epochs, performance goal,
and other training parameters are listed in Table 1. Fig. 5 Maximum number of epochs to train 25000
depicts a training error of PEMFC stack voltage prediction Performance goal 3e-8
during the learning process. Training performance goal was
3e-8 and the maximum training epochs was less than 25000.
After 25000 epochs, the error was 3.117e-8, then the assigned Learning rate 0.001
performance goal was not reached, but the trained neural Maximum validation failures 5
network has a good performances. To validate the proposed
neural modelling method, a very high variation of the current Ratio to increase learning rate 1.5
is used (see Fig. 4). This current profile includes the total
operating range from 0 to 45 A [24]. Ratio to decrease learning rate 0.7

727  
 
INTERNATIONAL  JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    
«authorheader»  et  al.,  Vol.4,  No.3,  2014  

Figure 5. Training error of AE voltage prediction during the learning process.

Figure 6. Neural PEM fuel cell stack voltage prediction performances.

As we can see in Fig. 6, the neural predict with a good Table 2. PEM fuel cell operating temperature prediction
precision the voltage of the 47 cells stack of the PEMFC than model parameters
the model presented in [24]. Indeed, with the neural method,
Epochs between displays 10
the tracking error is less than 0.5%, while it is around 20%
with the method given in [24]. Maximum number of epochs to train 25000
Table 2 summarizes the number of epochs between Performance goal 4.3e-9
displaying, maximum number of epochs, performance goal,
and other training parameters for the neural operating
temperature prediction. Fig. 7 shows a training error of PEM Learning rate 0.001
fuel cell operating temperature prediction during the learning
Maximum validation failures 5
process. Training performance goal was 4.3e-9 and the
maximum training epochs was less than 25000. After 12048 Ratio to increase learning rate 1.5
epochs, the error was 4.299e-9, then the assigned
performance goal was reached, and the neural network Ratio to decrease learning rate 0.7
trained.
728  
 
INTERNATIONAL  JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    
«authorheader»  et  al.,  Vol.4,  No.3,  2014  

Figure 7. Training error for operating tempeature prediction during the learning process.
As we can see in Fig. 8, the neural predict with a good accuracy the voltage of the 47 cells stack of the PEM fuel cell than
the model presented in [24]. Indeed, with the neural method, the tracking error is less than 0.2%, while it is around -10% and
+2% with the method given in [24].

Figure 8. Neural PEM fuel cell stack voltage prediction performances.

The tracking error is less than 0.2% while it is around -

5. Conclusion
10% and +2% with the model derived from the literature.
With this approach, a simulation model which describes
A dynamical model of a PEMFC stack voltage behavior both the 47 cells stack voltage and operating temperature
and operating temperature prediction based on artificial with a good accuracy is presented and validate with an
neural networks was proposed in this paper. Firstly, the feed- experimental data. This model can be used  during the phase
forward neural network (FFNN) was trained offline by using
of development of PEMFC auxiliaries such as the power
experimental data, to predict a PEM fuel cell voltage without
converter without any risk to damage PEMFC, and with a
using any analytical relations. The neural model
less cost.
approximates with a good accuracy a stack voltage of a
PEMFC with comparison with a model existing in the
literature. Indeed, with the proposed approach, a tracking Acknowledgements
error less than ±0.5%, while it is around ±20% with a
compared method. The authors would like to acknowledge the funding
Thereafter, a neural network operating temperature of received from LTE of Hydro-Québec, H2CAN Network and
cells based on the feed-forward neural network (FFNN) was Natural Resources Canada and the Natural Sciences and
presented and tested successfully using experimental data. Engineering Research Council of Canada.

729  
 
INTERNATIONAL  JOURNAL  of  RENEWABLE  ENERGY  RESEARCH    
«authorheader»  et  al.,  Vol.4,  No.3,  2014  

References networks," International Journal of Hydrogen Energy,

vol. 29, pp. 961-966, 2004.
[1] R. F. Mann, J. C. Amphlett, M. A. Hooper, H. M.
Jensen, B. A. Peppley, and P. R. Roberge, [13] S. Jemeï, D. Hissel, M.-C. Pera, and J. M. Kauffmann,
"Development and application of a generalised steady- "A new modeling approach of embedded fuel-cell power
state electrochemical model for a PEM fuel cell," generators based on artificial neural network," Industrial
Journal of Power Sources, vol. 86, pp. 173-180, 2000. Electronics, IEEE Transactions on, vol. 55, pp. 437-447,
2008.
[2] M. Ceraolo, C. Miulli, and A. Pozio, "Modelling static
and dynamic behaviour of proton exchange membrane [14] A. U. Chávez-Ramírez, R. Muñoz-Guerrero, S. Duron-
fuel cells on the basis of electro-chemical description," Torres, M. Ferraro, G. Brunaccini, F. Sergi, et al., "High
Journal of power sources, vol. 113, pp. 131-144, 2003. power fuel cell simulator based on artificial neural
network," International Journal of Hydrogen Energy,
[3] J. M. Corrêa, F. A. Farret, L. N. Canha, and M. G.
vol. 35, pp. 12125-12133, 2010.
Simoes, "An electrochemical-based fuel-cell model
suitable for electrical engineering automation approach," [15] F. Gao, B. Blunier, M. G. Simoes, and A. Miraoui,
Industrial Electronics, IEEE Transactions on, vol. 51, "PEM fuel cell stack modeling for real-time emulation in
pp. 1103-1112, 2004. hardware-in-the-loop applications," Energy Conversion,
IEEE Transactions on, vol. 26, pp. 184-194, 2011.
[4] [4] S. Pasricha and S. R. Shaw, "A dynamic PEM fuel
cell model," IEEE Transactions on Energy Conversion, [16] V. Elanayar and Y. C. Shin, "Radial basis function
vol. 21, pp. 484-490, 2006. neural network for approximation and estimation of
nonlinear stochastic dynamic systems," Neural
[5] [5] J. Amphlett, R. Mann, B. Peppley, P. Roberge, and
Networks, IEEE Transactions on, vol. 5, pp. 594-603,
A. Rodrigues, "A model predicting transient responses
1994.
of proton exchange membrane fuel cells," Journal of
Power Sources, vol. 61, pp. 183-188, 1996. [17] Y. Li, S. Qiang, X. Zhuang, and O. Kaynak, "Robust and
adaptive backstepping control for nonlinear systems
[6] [6] M. Wöhr, K. Bolwin, W. Schnurnberger, M.
using RBF neural networks," Neural Networks, IEEE
Fischer, W. Neubrand, and G. Eigenberger, "Dynamic
Transactions on, vol. 15, pp. 693-701, 2004.
modelling and simulation of a polymer membrane fuel
cell including mass transport limitation," International [18] K. L. Priddy and P. E. Keller, Artificial neural networks:
Journal of Hydrogen Energy, vol. 23, pp. 213-218, 1998. an introduction vol. 68: SPIE Press, 2005.
[7] [7] J. Haubrock, G. Heideck, and Z. Styczynski, [19] S. Ou and L. E. Achenie, "A hybrid neural network
"Dynamic investigation on proton exchange membrane model for PEM fuel cells," Journal of Power Sources,
fuel cell systems," in Power Engineering Society vol. 140, pp. 319-330, 2005.
General Meeting, 2007. IEEE, 2007, pp. 1-6.
[20] R. Sathya and A. Abraham, "Comparison of supervised
[8] [8] W. Friede, S. Raël, and B. Davat, "Mathematical and unsupervised learning algorithms for pattern
model and characterization of the transient behavior of a classification," Int J Adv Res Artificial Intell, vol. 2, pp.
PEM fuel cell," Power Electronics, IEEE Transactions 34-38, 2013.
on, vol. 19, pp. 1234-1241, 2004.
[21] S. Srinivasulu and A. Jain, "A comparative analysis of
[9] X. Kong, A. M. Khambadkone, and S. K. Thum, "A training methods for artificial neural network rainfall–
hybrid model with combined steady-state and dynamic runoff models," Applied Soft Computing, vol. 6, pp.
characteristics of PEMFC fuel cell stack," in Industry 295-306, 2006.
Applications Conference, 2005. Fourtieth IAS Annual
[22] S. Pillutla and A. Keyhani, "Neural network based
Meeting. Conference Record of the 2005, 2005, pp.
saturation model for round rotor synchronous generator,"
1618-1625.
Energy Conversion, IEEE Transactions on, vol. 14, pp.
[10] S. Tao, C. Guang-yi, and Z. Xin-jian, "Nonlinear 1019-1025, 1999.
modeling of PEMFC based on neural networks
[23] A. Chavez, S. Duron, L. Arriaga, and R. Munoz,
identification," Journal of Zhejiang University
"Dynamic model of a high power PEM fuel cell system
SCIENCE A, vol. 6, pp. 365-370, 2005.
on the basis of artificial neural networks," in Electrical
[11] R. Caponetto, L. Fortuna, and A. Rizzo, "Neural Engineering, Computing Science and Automatic
network modelling of fuel cell systems for vehicles," in Control, CCE, 2009 6th International Conference on,
Emerging Technologies and Factory Automation, 2005. 2009, pp. 1-7.
ETFA 2005. 10th IEEE Conference on, 2005, pp. 6 pp.-
[24] G. Fei, B. Blunier, A. Miraoui, and A. El-Moudni, "A
192.
Multiphysic Dynamic 1-D Model of a Proton-Exchange-
[12] W.-Y. Lee, G.-G. Park, T.-H. Yang, Y.-G. Yoon, and Membrane Fuel-Cell Stack for Real-Time Simulation,"
C.-S. Kim, "Empirical modeling of polymer electrolyte Industrial Electronics, IEEE Transactions on, vol. 57,
membrane fuel cell performance using artificial neural pp. 1853-1864, 2010.
 

730