Professional Documents
Culture Documents
201400080
Received May 06, 2014; accepted August 14, 2014; published online October 08, 2014
1 Introduction fore, when dealing with fast dynamic systems, the most chal-
lenging part remains to achieve on-line implementation.
Proton exchange membrane fuel cell (PEMFC) arises as a Recently, several works dealing with benefits of predictive
suitable clean energy generator technology for a wide range of control strategies regarding PEMFC control have been
applications [1–3]. However, several bolts, especially related reported. Wu et al. [14] proposed a multi-loop nonlinear pre-
to its control, still remain to be removed to improve the overall dictive control strategy using a reduced order model to regu-
system performance. Indeed, accurate power response, opti- late oxygen excess ratio and stack temperature of a fuel cell.
mal air and hydrogen consumption, and global efficiency However, the controller has only been tested in simulation
improvement remain challenging control goals. environment, and as emphasized by authors some devices
Numerous control strategies have been reported in the lit- and design, such as hydrogen storage or power management,
erature for PEMFC systems, ranging from PID controller [4–6], have not been considered. Shokuhi-Rad et al. [15] designed an
state feedback linearizing or differential flatness approaches approximate predictive control strategy to regulate the output
[7, 8], dynamic neural network controller [9, 10], linear quad- voltage of a PEMFC. This approach, based on an instanta-
ratic Gaussian (LG) controller [11], to model predictive control neous linearization of a neural network model, has been tested
[12]. Due to its ability to take into account dynamic nonlineari- in simulation environment, and appeared to be an interesting
ties of the process, to handle just as well state and input con- alternative to achieve real-time control. Gruber et al. [16]
straints as economical constraints, nonlinear model-based developed a model-based predictive strategy to regulate the
predictive control (NMPC) strategy appears as a promising
candidate regarding PEMFC control. However, despite its suc-
cessful application in numerous industries [13], in its basic –
form, NMPC controller requires high computation cost. There- [*] Corresponding author, michel.benne@univ-reunion.fr
886 ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim FUEL CELLS 14, 2014, No. 6, 886–893
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
FUEL CELLS 14, 2014, No. 6, 886–893 www.fuelcells.wiley-vch.de ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim 887
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
Teflon glass fiber as seal joints and clamped at 11 Nm bolt tor- 3.2 Model Identification and Validation
ORIGINAL RESEARCH PAPER
0.55
the process, and how each variable affects the output. In the
following, load current and hydrogen and air mass flow rates 0.5
are considered as inputs. 0.45
Once inputs of the network are determined, the next step is
0.4
to define its topology, which is the number of layers and neu-
rons. Since no accurate method allows to determine the opti- 0.35
mal number of neurons in the hidden layer, a ‘‘trial-error’’
0.3
method is performed to find the best architecture in terms of
accuracy and computational cost. After comparing several 0.25
architectures, a fully connected three layers feedforward net- 0.2
work (Figure 2) is retained. Three inputs, namely hydrogen 0 50 100 150 200 250 300
Time / sample
and air flow rates, and load current are considered for the
input layer. The output layer of the ANN provides the cell out- Fig. 3 Training data.
put voltage using one neuron, and five neurons compose the
hidden layer. A tangent sigmoid activation function f1, and a 0.7 Experimental data
linear activation function f2 are used for the hidden layer and Simulated data (ANN)
0.65
the output layer, respectively.
0.6
0.55
Output voltage / V
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0 50 100 150 200
Time / samples
Fig. 2 ANN architecture. Fig. 4 Validation data.
888 ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.fuelcells.wiley-vch.de FUEL CELLS 14, 2014, No. 6, 886–893
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
Ny
AE¥, AME, and RMSE denote the absolute maximal error, X ðkþiÞ ðkþiÞ ðkÞ
J ðkÞ ðuÞ ¼ gðysp ym þ yðkÞ ym Þ2
the absolute mean error, and the root mean squared error, i¼1
respectively. n represents the number of data, y and ^y denote
Nu 2
the experimental and the simulated data, respectively. 3 X
X ðkþiÞ
Considering real-time control goal, and seeing that the pro- þ bj ðDuj ðpÞÞ (5)
j¼1 i¼1
posed NMPC strategy handles modeling error, the prediction
accuracy of the ANN model is more than sufficient (Table 1). uðkþiÞ ðpÞ ˛ ½umin ; umax ; umin ˛R3 ; umax ˛R3 for i
The designed ANN model is used to predict the PEMFC out- ¼ 1; 2 . . . ; Nu (6)
put voltage in the parameterized NMPC strategy. h iT
ðkÞ ðkÞ ðkÞ ðkÞ
uðkÞ ðpÞ ¼ u1 ðpÞu2 ðpÞu3 ðpÞ ¼ ½Icell ðpÞm _ ðkÞ _ ðkÞ
h2 ðpÞmair ðpÞ
T
is
the control actions vector at time k, g, and bj are some weight-
4 NMPC Controller Design ðkÞ
ing parameters.ysp denotes the desired reference trajectory, Nu
NMPC is an optimization-based multivariable constrained and Ny denote the control horizon and the predictive horizon,
control method that uses a nonlinear model to predict the respectively. Here, the modeling error at time k is taken into
future behavior of the process. Classical NMPC is defined by consideration using the difference between the measured out-
ðkÞ
solving on-line an optimization problem that leads to an opti- put y(k) and the model output ym . Duj represents the changes
mal sequence of future control actions that minimizes a given in control variables and is defined as follows:
cost function over a certain prediction horizon. NMPC offers a ðkþiÞ ðkþiÞ ðkþi1Þ
Duj ðpÞ ¼ uj ðpÞ uj ðpÞ for j ˛f1; 2; 3g (7)
great flexibility since it can handle input/state constraints,
process nonlinearities, and even economical or operating con- In this work, to obtain a low-dimensional optimization
straints [35]. However, in its basic form, NMPC strategies problem decoupled from the choice of the predictive horizon,
involve heavy computations, which becomes a bolt when the subsequent parameterization is proposed:
dealing with fast dynamic systems. n o
uðkþiÞ ðpÞ ¼ uðkþiÞ þ pEðkþiÞ for ˛ 0; 1; Ð; Ny (8)
In this work, to address this problem and reach real-time
control goal, a parameterized NMPC strategy is designed. The This parameterization involves the steady control actions
main idea lies in reducing the computational time by decreas- vector h obtained iTin steady state at time k
ing the optimization problem dimension. In this aim, the con- ðkÞ ðkÞ ðkÞ
u ðkÞ ¼ u1 u2 u3 , a vector parameter p ˛ [–1, 1]3, and
trol sequence u is defined using a low dimensional parameter the trajectory tracking error E*(k) obtained using u*(k). In the
vector p. Indeed, in parameterized NMPC strategy, the optimi- present case, seeing that the system is open-loop stable, using
zation problem dimension depends on the dimension of the u* allows to reduce the degree of freedom of the control profile:
parameter vector p. In this context, a well-structure parameter- 8 ðkþiÞ ðkþiÞ
ization may lead to a low-dimensional optimization problem >
< u1 ¼ a1 ysp þ b1
whatever the predictive control horizon Ny. A low-dimen- ðkþiÞ ðkþiÞ
u ¼ a2 u1 þ b2 (9)
>
: 2ðkþiÞ
sional optimization problem turns out to be a significant u3
ðkþiÞ
¼ a3 u1 þ b3
advantage to achieve on-line implementation [36]. Let us recall
that when dealing with classical NMPC strategies, the predic- The expressions of ui¼1;2;3 have been determined from
tive horizon sets the optimization problem dimension, which experimental observations of the steady state, and aj = 1,2,3 and
could lead to huge problem dimension. bj = 1,2,3 are constant parameters that must be estimated off-
Here, a parameterized NMPC controller is used to control line.
the PEMFC output voltage, once manipulated variables are Note that there is no universal parameterization. Indeed,
load current (Icell), hydrogen (m_ h2 ) and air (m
_ air ) molar flow each problem requires a particular parameterization. More
rates. In this context, the optimal set of parameters ^ p is details about parameterized NMPC strategies can be
obtained by minimizing the cost function J, which is subject to consulted in Refs. [18, 37, 38].
constraints on manipulated variables:
FUEL CELLS 14, 2014, No. 6, 886–893 www.fuelcells.wiley-vch.de ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim 889
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
ORIGINAL RESEARCH PAPER
Table 2 Control design parameters. Table 3 Experimental setpoint tracking performance of the proposed
NMPC controller.
Parameter Value
Profile 1 Profile 2
a1 –0.3460
Case 1: Ny = Nu = 5
b1 0.3170
3 RMSE (V) 9.4033·10–4 5.8732 · 10–4
a2 1.0423 · 10
AME (V) 0.0045 0.0046
b2 0.0414
3 AE¥ (V) 0.0157 0.0440
a3 8.2798 · 10
Case 2: Ny = Nu = 20
b3 –1.0186
RMSE (V) 0.0015 5.5206 · 10–4
g 1
AME (V) 0.0078 0.0047
bj 1
AE¥ (V) 0.0181 0.0350
0.65
a) Setpoint
PEMFC output voltage (parameterized NMPC)
0.6
0.55
Output voltage / V
0.5
0.45
0.4
0.35
0.3
0.25
5 10 15 20 25 30 35 40 45
Time / s
0.65
b) Setpoint
PEMFC output voltage (parameterized NMPC)
0.6
0.55
Output Voltage / V
0.5
0.45
0.4
0.35
0.3
0.25
0 20 40 60 80 100 120 140 160
Time / s
Fig. 5 Experimental setpoint tracking results obtained with the proposed NMPC controller (a) profile 1 with Ny = Nu = 5; (b) profile 2 with Ny = Nu = 20.
890 ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.fuelcells.wiley-vch.de FUEL CELLS 14, 2014, No. 6, 886–893
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
150
100
0 20 40 60 80 100 120 140 160
Time / s
2000
b) Ny = Nu = 5
Ny = Nu = 20
1800
1600
Air flow / sccm
1400
1200
1000
800
0 20 40 60 80 100 120 140 160
Time / s
c) Ny = Nu = 5
12
N = N = 20
y u
11
10
Current / A
FUEL CELLS 14, 2014, No. 6, 886–893 www.fuelcells.wiley-vch.de ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim 891
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
controller is implemented externally. The NMPC controller, excellent candidate for on-line applications. For fast dynamic
ORIGINAL RESEARCH PAPER
implemented in Matlab environment, runs on-line and com- systems when classical NMPC controller cannot be consid-
municates with the SCDA system of the fuel cell. In this ered, the proposed approach is an up-and-coming strategy.
section, two of the above-mentioned output voltage profiles
are used to illustrate the performance of the NMPC controller
(Figure 6).Hydrogen and air molar flow rates, and load References
current are manipulated variables, whereas the controlled
[1] J. Larminie, A. Dicks, M. S. McDonald, Fuel Cell Systems
variable is the output voltage. The cell temperature is main-
Explained, Wiley, Chichester, United Kingdom, 2003.
tained at 70 C and the sampling period is set to 3 s. The
[2] C. Wang, M. H. Nehrir, IEEE Trans. Energy Convers.
optimization problem is solved on-line using Levenberg–Mar-
2007, 22, 864.
quardt algorithm. All experiments are performed using
[3] C. Wang, M. H. Nehrir, H. Gao, IEEE Trans. Energy Con-
control parameters listed in Table 2.
vers. 2006, 21, 586.
The NMPC controller performance, especially in terms of
[4] S. R. Huang, C. Y. Lin, C. C. Wu, S. J. Yang, Int. J. Hydro-
setpoint tracking accuracy, is assessed using RMSE, AME, and
gen Energy 2008, 33, 5205.
AE¥ as statistical criteria. For both voltage profiles, two differ-
[5] F. C. Wang, C. C. Ko, Int. J. Hydrogen Energy 2010, 35,
ent values of control and predictive horizons are investigated:
10437.
case 1: Ny = Nu = 5, case 2: Ny = Nu = 20.
[6] R. N. Methekar, V. Prasad, R. D. Gudi, J. Power Sources
Table 3 summarizes the setpoint tracking accuracy perfor-
2007, 165, 152.
mance of the NMPC controller for different values of Nu and
[7] Q. Li, W. Chen, Y. Wang, J. Jia, M. Han, J. Power Sources
Ny, and different setpoint voltage profiles.
2009, 194, 338.
The parameterized NMPC controller demonstrates excel-
[8] R. Da Fonseca, E. Bideaux, M. Gerard, B. Jeanneret,
lent performance in terms of setpoint tracking accuracy for
M. Desbois-Renaudin, A. Sari, Appl. Energy 2014, 113,
both voltage profiles, regardless of control and predictive
219.
horizons dimensions (Table 2).
[9] M. Hatti, M. Tioursi, Int. J. Hydrogen Energy 2009, 34,
As illustrated in Figures 5 and 6, the controller demon-
5015.
strates good tracking capability, while offering an entirely suit-
[10] J. Hasikos, H. Sarimveis, P. L. Zervas, N. C. Markatos,
able dynamic for manipulated variables, which underlines,
J. Power Sources 2009, 193, 268.
among other things, its robustness against inescapable model-
[11] R. N. Methekar, S. C. Patwardhan, R. Rengaswamy,
process mismatch. However, it is important to notice that the
R. D. Gudi, V. Prasad, Chem. Eng. Res. Des. 2010, 88, 861.
manipulated variables are smoother for Ny = Nu = 20. There-
[12] J. Golbert, D. R. Lewin, J. Power Sources 2004, 135, 135.
fore, even if the setpoint tracking accuracy is quite similar,
[13] M. Bauer, I. K. Craig, J. Process Control 2008, 18, 18.
Ny = Nu = 20 appears to be the best configuration since
[14] W. Wu, J. P. Xu, J. J. Hwang, Int. J. Hydrogen Energy 2009,
constraints on actuators are decreased. It should be pointed
34, 3953.
out that such length of horizons could not be used with a
[15] A. Shokuhi-Rad, A. Jamali, M. Naghashzadegan,
classical NMPC controller due to its heavy computations.
N. Nariman-zadeh, A. Hajiloo, Int. J. Hydrogen Energy
2012, 37, 11244.
[16] J. Gruber, M. Doll, C. Bordons, Control Eng. Pract. 2009,
6 Conclusion
17, 874.
In this study, a parameterized multivariable NMPC control- [17] C. Ziogou, S. Papadopoulou, M. C. Georgiadis, S. Voute-
ler has been designed and experimentally tested on-line to takis, J. Process Control 2013, 23, 483.
control the PEMFC output voltage. First, to reach real-time [18] C. Damour, M. Benne, A. Kadjo, S. Rosini, B. Grondin-
control goal, a parameterized NMPC strategy has been devel- Perez, Int. J. Hydrogen Energy 2013, 38, 7407.
oped. In this aim, a specific parameterization of the control [19] A. J. J. Kadjo, P. Brault, A. Caillard, C. Coutanceau, J. P.
sequence has been proposed, which led to a low-dimensional Garnier, S. Martemianov, J. Power Sources 2007, 172, 613.
optimization problem. In a second stage, an ANN model, used [20] T. E. Springer, T. A. Zawodzinski, S. Gottesfeld, J. Elec-
as predictor in the NMPC controller, has been developed and trochem. Soc. 1991, 138, 2334.
experimentally validated. Eventually, the performance of the [21] D. M. Bernardi, M. W. Verbrugge, J. Electrochem. Soc.
NMPC controller has been evaluated on a test bench. The 1992, 139, 2477.
original built-in controller has been overridden and the para- [22] T. V. Nguyen, R. E. White, J. Electrochem. Soc. 1993, 140,
meterized NMPC controller has been implemented externally 2178.
and executed on-line. [23] J. S. Yi, T. V. Nguyen, J. Electrochem. Soc. 1998, 145, 1149.
Experimental results validated the efficiency of the NMPC [24] V. Gurau, H. Liu, S. Kakac, AIChE J. 1998, 44, 2410.
controller in terms of robustness against model-process mis- [25] S. Dutta, S. Shimpalee, J. W. Van Zee, Appl. Electrochem.
match and tracking capability. The real-time implementable 2000, 30, 135.
NMPC strategy designed in this study turned out to be an
892 ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.fuelcells.wiley-vch.de FUEL CELLS 14, 2014, No. 6, 886–893
Damour et al.: On-Line PEMFC Control Using Parameterized Nonlinear Model-Based Predictive Control
______________________
FUEL CELLS 14, 2014, No. 6, 886–893 www.fuelcells.wiley-vch.de ª 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim 893