Anda di halaman 1dari 10

Journal of Process Control 20 (2010) 73–82

Contents lists available at ScienceDirect

Journal of Process Control


journal homepage: www.elsevier.com/locate/jprocont

Adaptive peak seeking control of a proton exchange membrane fuel cell


Ravi N. Methekar a, Sachin C. Patwardhan a,*, Ravindra D. Gudi a, Vinay Prasad b
a
Department of Chemical Engineering, Indian Institute of Technology Bombay, Mumbai 400 076, India
b
Department of Chemical and Materials Engineering, University of Alberta, Edmontan, Canada T6G2V4

a r t i c l e i n f o a b s t r a c t

Article history: The primary aim of operating any fuel cell (PEMFC) system is to produce the power/electricity at maxi-
Received 1 December 2008 mum efficiency. The cell voltage/current manipulation appear to be the most suitable choice for control-
Received in revised form 1 October 2009 ling the power density. However, the power density exhibits a highly nonlinear and complex dynamic
Accepted 4 October 2009
relationship with respect to the cell voltage. Since the process output variable (i.e. power density) itself
is the objective function for the optimization, there exists a singularity at the optimum operating condi-
tion. In addition, the location of the optimum operating point changes with time due to the occurrence of
Keywords:
variety of disturbances and/or changes in the operating conditions. Thus, the need to operate the PEMFC
Adaptive optimizing control
Nonlinear internal model control
at its peak power density and track the shifting optimum turns out to be a challenging control problem.
Wiener model The task of on-line optimizing control of PEMFC poses difficulties in real time control due to its fast
Orthonormal basis filters dynamics and it is impractical to employ a mechanistic model for locating the changing optimum on-line.
Fuel cell In this context the adaptive optimizing control scheme developed by Bamberger and Isermann (1978) [1]
appears interesting. Their scheme is based on on-line adaptation of a nonlinear black box time series
models and facilitates analytical computation of changing optimum. Recently, Bedi et al. (2007) [2] have
developed a closed form multi-step predictive control law under nonlinear internal model control frame-
work using a black-box nonlinear model and employed it for peak power control in PEMFC. From the
viewpoint of PEMFC operation, this nonlinear IMC controller meets the demand on the fast computations
as a closed form solution is obtained for the nonlinear control problem at each time step. In this work, we
propose to develop an adaptive optimizing control scheme, which combines the attractive features of the
on-line optimization approach proposed by Bamberger and Isermann (1978) [1] and closed form control
law developed by Bedi et al. (2007) [2]. We demonstrate the effectiveness of the proposed adaptive opti-
mizing scheme by conducting simulation studies on the distributed an along-the-channel model of PEM-
FC. Analysis of the simulation results indicate that the proposed adaptive optimizing control scheme
satisfactorily tracks the shifting optimum operating point in the face of changing unmeasured
disturbances
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction and complex dynamic relationship with respect to the cell voltage.
Since the process output variable (i.e. power density) itself is the
The primary aim of operating any fuel cell (PEMFC) system is to objective function for the optimization, there exists a singularity
produce the power/electricity at maximum efficiency. There are at the optimum operating condition. The steady state gain of con-
various manipulated variables available for controlling power den- trolled output with respect to the manipulated input becomes zero
sity obtained from a PEMFC, such as, inlet molar flow rate of hydro- at the optimal point. In addition, the system exhibit input multi-
gen and oxygen, cell voltage/current and inlet gas temperatures. plicities, i.e. any steady state power density other than the opti-
Inlet molar flow rate of hydrogen/oxygen have limited gain with mum point power density can be obtained using multiple input
respect to the power density. Manipulation of inlet gas tempera- conditions. Input multiplicities, in general, occur due to the pres-
tures requires extra accessories for maintaining the temperature, ence of competing dynamic effects in a system. Such systems are
and consequently, these variables would not be the best choice difficult to control as they exhibit changes in the sign of steady
as manipulated variable(s). The cell voltage/current manipulation state gain(s) in the desired operating region. Morari [3] observed
appear to be the most suitable choice for controlling the power that input multiplicities of nonlinear systems usually cause robust-
density. However, the power density exhibits a highly nonlinear ness problems which cannot be eliminated using linear controllers.
The phenomenon of change in the sign of the steady state gain
poses a difficulty even for nonlinear control schemes. From a sys-
* Corresponding author. Fax: +91 22 25726895.
E-mail address: sachinp@iitb.ac.in (S.C. Patwardhan). tem theoretic viewpoint, the relative degree of the system becomes

0959-1524/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2009.10.001
74 R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82

undefined and the invertibility is lost at the optimum (singular) However, they, have not mentioned the type of controller used
point where the steady state gain is reduced to zero. As a conse- for peak seeking. As the optimum point is a singular point, the
quence, global input–output linearization based approaches can- strategy used for controlling at the optimum is equally important.
not be applied (in a straightforward manner) in the regions of The task of on-line optimizing control of PEMFC poses difficul-
state space where the relative degree is not well defined [4]. Thus, ties in real time control due to its fast dynamics. As mentioned
in the case of almost all the nonlinear control strategies based on above, it is impractical to employ a mechanistic model for locating
exact linearization, an assumption, often not stated explicitly, is the changing optimum on-line. In this context, from the vast body
that the steady state gain of the model cannot change sign in the of literature available in the control domain on adaptive optimum
operating region [5]. Biegler and Rawlings [6] observe that use of seeking control, the adaptive optimizing control scheme developed
unconstrained nonlinear controller may not be sufficient in such by Bamberger and Isermann [1] appears interesting. Their scheme
cases, even when the model is perfect. When zero process gain sit- is based on on-line adaptation of a nonlinear black box time series
uation is encountered, the perfect model cannot be inverted and model. They proposed to employ a Hammerstein type block ori-
the model inversion based controller can become ill-conditioned. ented model to locate the changing optimum. The model parame-
Thus, the need to operate the PEMFC at its peak power density ters were estimated on-line using the recursive least square (RLS)
turns out to be a challenging control problem. parameter estimation scheme. The optimum operating point was
A nonlinear controller operating the system at its optimum located by using the steady state part of the identified nonlinear
point is faced with yet another difficulty. The location of the opti- dynamic model. In fact, block oriented model structure together
mum operating point changes with time due to the occurrence of with polynomial type nonlinearity facilitated analytical computa-
variety of disturbances and/or changes in the operating conditions. tion of the changing optimum. Efficacy of their scheme were dem-
In practice, the PEMFC is a highly nonlinear system, which is sub- onstrated on a thermal pilot process.
ject to various disturbances such as drying/flooding of membrane/ Recently, Bedi et al. [2] have developed a closed form multi-step
channels, change in inlet gas temperatures and coolant tempera- predictive control law under nonlinear internal model control
ture, nonuniform distribution of reactants, deactivation of catalyst framework using a black-box nonlinear model (Wiener type model
and change in ambient conditions. These disturbances are suffi- parameterized using generalized orthonormal basis filters) and em-
cient to shift the optimum operating point of the process signifi- ployed it for peak power control in PEMFC. Similar to the solution ap-
cantly. If these changes go unattended, the system will either proach suggested by Patwardhan and Madhavan [7], an analytical
operate at sub-optimal point or may face unachievable set point solution was constructed to multi-step predictive control problem
and can even lead to unacceptable behavior of the manipulated using the theory of multi-dimensional quadratic equations. This ap-
variables [7]. Therefore, there is a need to develop a control strat- proach appears promising as the nonlinear time series model re-
egy that can track the changing location of the optimum and con- quired for controller synthesis has been identified directly from
trol the PEMFC at the singular optimum operating point. the input–output data and the model structure is amenable to on-
In recent years, there have been significant activities towards an line recursive parameter estimation. Deshpande and Patwardhan
improving the operation of fuel cells by employing a variety of lin- [11] have demonstrated through experiments on a laboratory scale
ear and nonlinear control approaches. However, the problems of MIMO system (three controlled outputs and three manipulated in-
(a) maintaining the PEMFC at the optimum (singular) operating puts) that a control law of this type can be implemented with aver-
point and (b) tracking the changing location of the peak on-line age computation time of 0.184 ms. Thus, from the viewpoint of
have received very little attention. The importance of maintaining PEMFC operation, this nonlinear IMC controller meets the demand
PEMFC operation at the optimum condition has been recognized by on the fast computations as a closed form solution is obtained for
some researchers. Extensive discussion on the polarization curve the nonlinear control problem at each time step.
characteristics found in the fuel cell literature indicates that the In this work, we propose to develop an adaptive optimizing con-
polarization curve is strongly and nonlinearly depends on the oper- trol scheme, which combines the attractive features of the on-line
ating parameters of the PEMFC. This polarization curve gives rise to optimization approach proposed by Bamberger and Isermann [1]
an unconstrained optimum operating point in the operating region. and closed form control law developed by Bedi et al. [2]. To begin
Marwardi et al. [8] have proposed to improve the performance of with, we develop an adaptive version of the closed form NIMC pro-
PEMFC by operating the system at an optimum conditions. The posed by Bedi et al. [2]. To locate the changing location of the peak,
optimum operating conditions were found by using one dimen- the parameters of nonlinear state-output map of the Wiener model
sional steady state mechanistic model. By employing a steady state are updated at each time step using RLS approach. The steady state
mechanistic model, Misra et al. [9] have found that the peak power map of the dynamic model is then used to locate the current opti-
in PEMFC was influenced by various parameters such as fuel cell mum operating point. We demonstrate the effectiveness of the pro-
temperature, anode and cathode pressure, relative humidity at an- posed adaptive optimizing scheme by conducting simulation
ode and cathode, the anode and cathode gas mole fractions. They studies on the distributed an along-the-channel model of PEMFC.
concluded that high operating temperature, high relative anode This paper is organized in five sections. The details of model iden-
humidity and low anode pressure were required to operate the tification and parameter adaption scheme are explained in Section 2.
PEMFC most efficiently. These are, however, open loop approaches The adaptive NIMC and adaptive optimizing control scheme are
and there is no mechanism to incorporate model plant mismatch. developed in Section 3. Simulation results are presented in Section 4
Also, from the view of online control, a mechanistic model based and the main conclusions are summarized in Section 5.
optimization would take longer time to find the optimum solution
and hence may not be useful for formulating an on-line optimizing 2. Model identification using OBF-Wiener model
control scheme. Zhi-dan et al. [10] operated fuel cell at the maxi-
mum power point (MPP) by using extremum seeking algorithm. Consider a process governed by a set of nonlinear differential
They used adaptive extremum seeking algorithm for finding the algebraic equations
maximum operating point online. The fuel cell was perturbed by dzð1Þ
injecting sinusoidal signals and output response was used to esti- ¼ F1 ½zð1Þ ; zð2Þ ; UT ðtÞ; DðtÞ; p ð1Þ
dt
mate the gradient of power with respective manipulated variables.
0 ¼ F2 ½zð1Þ ; zð2Þ ; UT ðtÞ; DðtÞ; p ð2Þ
After filtering the transient dynamics using a high order filter, the
maximum power point was located using the gradient direction. YðtÞ ¼ G½zð1Þ ; zð2Þ  þ ty ðtÞ ð3Þ
R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82 75
h iT
2 Rn represents the process state vector,
T T
where zð1Þ zð2Þ slowly drifting unmeasured disturbances as additive bias correction
UT 2 BU  Rm represents the true value of manipulated inputs, to the output prediction. The main advantage of choosing quadratic
D 2 BD  Rd represents unmeasured disturbances, y 2 Rr represents polynomial function is that the resulting controller synthesis prob-
the vector of measured outputs corrupted with measurement noise lem can be solved analytically. The above model can be looked upon
ty ðtÞ and p 2 Rm represents parameter vector. We further assume as a truncated second order Volterra series model.
that U 2 Rm represents known (or computed) value of manipulated
inputs which are related to the true values as follows: 2.2. Model parameter estimation
UT ðtÞ ¼ UðtÞ þ tu ðtÞ ð4Þ
The model parameter estimation is carried out in two steps. To
where tu 2 Rm denotes an unknown input disturbances, which are begin with, r MISO models are identified off-line from the data RN
assumed to be a zero mean stationary signals. Also, generated by perturbing the plant. This step is used to select the
It is further assumed that the system under consideration is vectors {nðiÞ:i ¼ 1; 2; . . . ; rg of GOBF poles for each model. This step
also generates an initial estimate of parameters
 a fading memory system [12] and h iT fCðiÞ ð0Þ; DðiÞ ð0Þ:i ¼ 1; 2; . . . ; rg of the state to output map of each
T T
 all the steady state operating points, zð1Þ zð2Þ ; obtained by MISO model. In the off-line parameter estimation, the parameter
arbitrarily choosing ðUT ðtÞ ¼ UT 2 BU Þ and ðDðtÞ ¼ D 2 BD Þ are bi is set to zero, i.e. bi ð0Þ ¼ 0. The parameters of the state-output
locally asymptotically stable. map, fbi ðkÞ; CðiÞ ðkÞ; DðiÞ ðkÞ:i ¼ 1; 2; . . . ; rg, are later updated online
using recursive least square parameter estimation step and further
used in formulating the proposed adaptive peak seeking control
2.1. OBF-Wiener model scheme.

In practice, the dynamic model given by Eqs. (1)–(3) is difficult 2.2.1. Off-line parameter estimation
to develop and can prove to be complex for developing a controller. The off-line estimation of OBF poles and the parameters of
The informations available from the plant are the sampled se- state-output map can be carried out using a nested optimization
quence of input and output vectors RN ¼ fðYðkÞ; UðkÞÞ:k ¼ approach as proposed by Srinivasrao et al. [15]. Thus, with state-
1; 2; . . . ; Ng. Given input and output data set RN collected from a output map given by Eq. (10), the output can be expressed as
plant, the problem of identifying a nonlinear time series model yi ðkÞ ¼ ½uðiÞ ðkÞT ðHðiÞ Þ þ v i ðkÞ ð11Þ
can be stated as finding a nonlinear operator N½:
where
yðkÞ ¼ N½uðkÞ; h þ eðkÞ ð5Þ  T
uðkÞ ¼ u½uðk  1Þ; . . . ; uð1Þ; yðk  1Þ; . . . ; yð1Þ ð6Þ uðiÞ ðkÞ ¼ 1 ½xðiÞ ðkÞT ½vðiÞ ðkÞT ð12Þ
h iT
2
such that a suitable norm of model residuals feðkÞ:k ¼ 1; . . . ; Ng is vðiÞ ðkÞ ¼ ðxðiÞ
1 ðkÞÞ
ðiÞ ðiÞ
2x1 ðkÞx2 ðkÞ . . . ð13Þ
minimized with respect to parameter vector h. Here, ðyðkÞ; uðkÞÞ
and
represents perturbation variables defined as
h iT
ðiÞ ðiÞ
yðkÞ ¼ YðkÞ  Y and uðkÞ ¼ UðkÞ  U HðiÞ ¼ bi CðiÞ D11 D12 . . . DðiÞ
ni ;ni
ð14Þ

in the neighborhood of a steady state, say ðY; UÞ. In this work we ðiÞ
Here, HðiÞ is a Ni  1 vector with N i ¼ 1 þ ni  ðni þ 3Þ=2; xl ðkÞ rep-
ðiÞ
propose to develop r MISO Wiener type state space models of the resents lth element of vector xðiÞ ðkÞ and Djl represents ðj; lÞth ele-
form [15] ment of matrix DðiÞ . Thus, the main advantage of choosing
polynomial function is that the resulting state-output map is linear
xðiÞ ðk þ 1Þ ¼ UðiÞ xðiÞ ðkÞ þ CðiÞ uðkÞ ð7Þ
in parameters.
yi ðkÞ ¼ N ½x ðkÞ þ v i ðkÞ
ðiÞ ðiÞ
ð8Þ Given an input–output data set RN , the least squares estimate of
ðiÞ ni the parameters can be obtained by solving the following minimiza-
where i ¼ 1; 2; . . . ; r. Here, x ðkÞ 2 R represents the state vector of
tion problem [15]:
the ith MISO model and NðiÞ ½: represents some nonlinear static map
N h i2
relating states with the outputs for the ith MISO model. These MISO
b ðiÞ ; b 1 X
models are then stacked to generate an overall MIMO model. ðH n ðiÞ Þ ¼ arg min vb i ðk; HðiÞ ; nðiÞ Þ ð15Þ
HðiÞ ;nðiÞ
N k¼1
We choose to parametrize matrices ðUðiÞ ; CðiÞ Þ using Generalized
Orthonormal Basis Filters [13], which represent an orthonormal subject to
basis for the set of strictly proper stable transfer functions (de- ðiÞ
noted as H2 Þ. Ninness and Gustafson [14] have shown that a com- jnj j < 0 for j ¼ 1; 2 . . . ; ni ð16Þ
plete orthogonal set in H2 can be constructed as follows: The parameter estimation problem is formulated in terms of two
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  2  nested optimization problems as follows:
 ðiÞ 
1  nk  N h i2
Y
k1 ðiÞ
ð1  nk zÞ b ðiÞ ; b 1 X b ðiÞ ; nðiÞ Þ
F k ðz; nðiÞ Þ ¼ ð9Þ ðH n ðiÞ Þ ¼ arg min vb i ðk; H ð17Þ
ðz 
ðiÞ
nk Þ ðz 
ðiÞ
nj Þ nðiÞ
N k¼1
j¼1

ðiÞ
where fnk :k ¼ 1; 2; . . .g is an arbitrary sequence of poles inside the subject to constraint (16). Here, given a guess of pole vector e
n ðiÞ , the
parameter vector H b ðiÞ is estimated by solving another optimization
unit circle appearing in complex conjugate pairs. The nonlinear
state-output map NðiÞ ½::Rni ! R is chosen to be simple multi-dimen- problem
sional quadratic polynomial functions of the form [15] N h i2
b ðiÞ ½e 1 X
T
H n ðiÞ  ¼ arg min vb i ðk; HðiÞ ; en ðiÞ Þ ð18Þ
ðiÞ ðiÞ ðiÞ
N ½: ¼ bi þ C x ðkÞ þ ðx ðkÞÞ D ðx ðkÞÞ ðiÞ ðiÞ ðiÞ
ð10Þ HðiÞ
N k¼1

Here, CðiÞ represents a ð1  ni Þ vector and DðiÞ represents a ni  ni Since the state-output map is linear in parameters, we can exploit
symmetric matrix. The term bi is included to capture the effect of b ðiÞ can be estimated analytically
the fact that the parameter vector H
76 R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82

by a simple linear regression scheme. Thus, the parameter vector xðiÞ ðk þ pjkÞ ¼ ðUðiÞ Þp xðiÞ ðkÞ þ XðiÞ uðkÞ ð26Þ
b ðiÞ can be estimated as follows:
H ðiÞ ðiÞ p1 ðiÞ p2 ðiÞ
X ¼ ½ðU Þ þ ðU Þ þ    þ IC
h i1 h
b ðiÞ ½e
H n ðiÞ  ¼ EðuðiÞ ðkÞðuðiÞ ðkÞÞ Þ T
EðuðiÞ ðkÞyðiÞ ðkÞÞ ð19Þ y i ðk þ pjkÞ ¼ yi ðk þ pjkÞ þ uT ðkÞ ðXðiÞ ÞT ½DðiÞ ðkÞXðiÞ uðkÞ
e 
h iT 
þ ½CðiÞ ðkÞXðiÞ þ 2½ðUðiÞ Þp xðiÞ ðkÞ ½DðiÞ ðkÞXðiÞ uðkÞ ð27Þ
and Eð:Þ represents the expected value operator.

 i ðk þ pjkÞ is defined as
for i ¼ 1; 2; . . . ; r where y
2.2.2. Online recursive parameter update
The model parameters, which are estimated from input–output i ðk þ pjkÞ ¼ bi ðkÞ þ ½CðiÞ ðkÞðUðiÞ Þp xðiÞ ðkÞ
y
data set using the above off-line identification procedure, captures h i
the process dynamics under nominal operating conditions. When þ ½ðUðiÞ Þp xðiÞ ðkÞT ½DðiÞ ðkÞ½ðUðiÞ Þp xðiÞ ðkÞ
the plant under consideration is subjected to known/unknown dis-
turbance, its dynamic and steady state characteristics undergo Defining matrix KðkÞ as
changes. To predict the plant behavior accurately under changing 2 3
circumstances, it becomes necessary to update the model parame- ½Cð1Þ ðkÞXð1Þ þ 2½ðUð1Þ Þp xð1Þ ðkÞT ½Dð1Þ ðkÞXð1Þ
6 7
ters on-line. In this work, we propose to update the model param- KðkÞ ¼ 4 ......... 5
eters of the nonlinear state-output map using recursive least ½CðrÞ ðkÞXðrÞ þ 2½ðUðrÞ Þp xðrÞ ðkÞT ½DðrÞ ðkÞXðrÞ
square (RLS) estimation approach as follows [16]:
and fWðkÞg as a r  n  n bilinear matrix representation of the fol-
b ðiÞ ðk  1Þ
vb i ðkÞ ¼ yi ðkÞ  ½uðiÞ ðkÞT H ð20Þ lowing three dimensional array (see Appendix A):
Hb ðiÞ ðkÞ ¼ Hb ðiÞ ðk  1Þ þ LðiÞ ðkÞ vb i ðkÞ ð21Þ 2hh i i3
ðiÞ ðiÞ ðiÞ ðiÞ T ðiÞ ðiÞ T 1 ðXð1Þ ÞT ½Dð1Þ ðkÞXð1Þ ½0 . . . ½0
L ðkÞ ¼ P ðk  1Þu ðkÞ½ki þ u ðkÞ P ðk  1Þu ðkÞ  ð22Þ 6 7
fWðkÞgrnn 4 ......... 5
ðiÞ ðiÞ ðiÞ
P ðkÞ ¼ ½I  L ðkÞu ðkÞ P ðk  1Þ=ki T ðiÞ
ð23Þ  
½0 . . . ½0 ½ðXðrÞ ÞT ½DðrÞ ðkÞXðrÞ 
where PðiÞ ðkÞ represents the covariance matrix of the estimation er- the above r quadratic output prediction equations given by Eq. (27)
ror, LðiÞ ðkÞ represents the Kalman gain and ki represents the forget- can be combined into a single multi-dimensional quadratic equa-
ting factor ð0 < ki 6 1Þ. Using ki < 1 makes RLS algorithm more tion of the form
sensitive towards recent estimation error.
It may be noted that the proposed parameter updation scheme e ðk þ pjkÞ ¼ y
y ðk þ pjkÞ þ ½KðkÞuðkÞ þ fWðkÞg½uðkÞ; uðkÞ ð28Þ
does not alter the location of poles of matrix UðiÞ . These poles are
fixed at the values obtained in the off-line parameter estimation To account for plant-model mismatch and unmeasured disturbance
exercise. The proposed off-line parameter estimation makes sure mismatch, the output predictions are corrected as follows:
that the origin of each identified MISO is a stable node. The recur- ^ðk þ pjkÞ ¼ e
y ^ þ pjkÞ
y ðk þ pjkÞ þ dðk ð29Þ
sive parameter update only adapts the nonlinear state-output map
associated with each MISO model. where the estimation of future disturbances is generated as
^ þ j þ 1jkÞ ¼ dðk
dðk ^ þ jjkÞ for j ¼ 0; . . . ; p  1 ð30Þ
3. Design of on-line adaptive optimizing controller ^
dðkjkÞ ¼v
b ðkÞ ð31Þ

The development of the proposed on-line adaptive optimizing where ith element of residual signal v b ðkÞ is given by Eq. (20).
controller (AOC) is presented in two steps (a) development of mul- Let rðkÞ represent the desired final setpoint for the process.
ti-step adaptive predictive control law under nonlinear internal Then, imposing the constraint that the setpoint should be reached
model control (NIMC) framework for given setpoint and (b) locat- ^ ðk þ pjkÞ ¼ rðkÞ, gives the following
after p -steps in the future, i.e. y
ing the changing optimum operating point (i.e. adaptive peak controller design equation Q ðuðkÞÞ ¼ 0, where
seeking).  ðk þ pjkÞ  ðkÞ
QðuðkÞÞ ¼ fWðkÞg½uðkÞ; uðkÞ þ ½KðkÞuðkÞ þ y ð32Þ
where ðkÞ ¼ rðkÞ  vb ðkÞ. The above equation is a multi-dimen-
3.1. Adaptive NIMC control law
sional quadratic operator polynomial, which can be solved analyti-
cally using an approach developed by Rall [17] for solving quadratic
In this section, we develop an adaptive version of the NIMC law
operator polynomials in Banach spaces. Patwardhan and Madhavan
developed by Bedi et al. [2]. It is assumed that the system under
[18] have used Rall’s analytical solution strategy to develop a one
consideration is a square system, i.e. number of manipulated in-
step ahead control law based on an approximation of a mechanistic
puts equals the number of controlled outputs. Consider the p-steps
model. Proceeding in a similar manner, a multi-step adaptive qua-
ahead prediction obtained using MISO model with time varying
dratic control law can be derived as follows.
state-output map e denote some input vector such that the matrix
Let u
xðiÞ ðk þ 1Þ ¼ UðiÞ xðiÞ ðkÞ þ CðiÞ uðkÞ ð24Þ e Þ ¼ 2½fWðkÞgð u
rU ½Q ð u e Þ þ KðkÞ ð33Þ
ðiÞ ðiÞ ðiÞ T ðiÞ ðiÞ
~ i ðkÞ ¼ bi ðkÞ þ C ðkÞx ðkÞ þ ðx ðkÞÞ ½D ðkÞx ðkÞ
y ð25Þ is nonsingular. Then, Eq. (32) can be transformed as
e
f WðkÞgðuðkÞ e ; uðkÞ  u
u e Þ þ ðuðkÞ  u
e Þ þ E0 ðkÞ ¼ 0 ð34Þ
under the constraint that manipulated inputs held constant over
the prediction horizon, i.e. where
uðk þ ijkÞ ¼ uðkjkÞ for i ¼ 1; 2; . . . ; p  1 e
f WðkÞg e ÞÞ1  fWðkÞg
¼ ðrU ½Qð u ð35Þ
1
e ÞÞ ½Q ð u
E0 ðkÞ ¼ ðrU ½Q ð u e Þ ð36Þ
The p-steps ahead output prediction generated can be expressed as
R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82 77

Here, symbol () denotes the ‘left dot product’ between matrix f ðkÞ ¼ FIMC ðq1 ÞðkÞ ð40Þ
e ÞÞ1 and the bilinear matrix fWðkÞg (see Appendix A).
ðrU ½Q ð u 1
The solution of the above transformed multi-dimensional quadratic The IMC filter FIMC ðq Þ is typically selected as a diagonal matrix
equation can be written as with first or higher order and unit gain transfer functions appearing
on the main diagonal.
n o1
1 1
e
uðkÞ ¼ u I þ ðDðkÞÞ2 E0 ðkÞ ð37Þ 3.2. On-line estimation of optimum operating point
2
e
where DðkÞ ¼ ½I  4f WðkÞgðE 0 ðkÞÞ ð38Þ To locate the time varying optimum operating point, consider
the state space representation with time varying state-output
Note that, in general, a matrix has multiple square roots and conse- map given by the following set of equations:
quently different values of uðkÞ will be obtained for every choice of
the square root of matrix DðkÞ. Also, even though the original matrix xðiÞ ðk þ 1Þ ¼ UðiÞ xðiÞ ðkÞ þ CðiÞ uðkÞ ð41Þ
has all real elements, the square root can have complex elements ðiÞ
yi ðkÞ ¼ bi ðkÞ þ C ðkÞx ðkÞ þ ðx ðkÞÞ ½D ðkÞðx ðkÞÞ þ v
ðiÞ ðiÞ T ðiÞ
b i ðkÞ ð42Þ ðiÞ
and consequently the resulting uðkÞ can be complex. Patwardhan
and Madhavan [18] have suggested the following remedies to alle- The parameter bi ðkÞ is expected to capture the effect of time varying
ðiÞ
viate these difficulties: unmeasured disturbances on the output as an additive drift. Let xs
represent the steady state and ys;i ðkÞ represent the corresponding
1
 The matrix square root ðDðkÞÞ2 should be selected such that all steady state output obtained by holding uðkÞ ¼ us , which can be
its eigenvalues have non-negative real parts. Specifically, when estimated as follows:
matrix DðkÞ is a positive definite matrix, the positive definite ðiÞ 1 ðiÞ ðiÞ
square root of the matrix should be used for control law xðiÞ
s ¼ ðI  U Þ C us ð43Þ
ðiÞ T ðiÞ
computations. ys;i ðkÞ ¼ bi ðkÞ þ ½C ðkÞxðiÞ
s þ ðxðiÞ ðiÞ
s Þ ½D ðkÞxs ð44Þ
 When the solution vector becomes complex, the real part of the
complex solution vector can be used for manipulation. under the assumption that E½ v
b ðkÞ ¼ 0. The unconstrained optimum
(maximum or minimum) of the ith output can be estimated using
They have also shown that the complex solutions arise when the necessary condition for optimality, i.e. by setting
the specified setpoint is unattainable due to nonlinearity. A de- @ys;i ðkÞ
tailed discussion regarding the rationale behind these recommen- ðiÞ
¼0
@us
dations and related theoretical results can be found in [18]. Thus,
incorporating the above suggestions, the quadratic control law where 0 represents zero vector. The necessary condition for opti-
becomes mality yields
( )
1n 1
o1 ½uðiÞ ðiÞ 1 ðiÞ T ðiÞ ðiÞ 1 ðiÞ 1 ðiÞ
s ðkÞopt ¼ 1=2½ððI  U Þ C Þ D ðkÞððI  U Þ C Þ ½C ðkÞ
e  REAL
uðkÞ ¼ u I þ ðDðkÞÞ2 E0 ðkÞ ð39Þ
2  ðI  UðiÞ Þ1 CðiÞ T

It may be noted that, if the specified setpoint is attainable at steady The optimum setpoint ri ðkÞ can then be computed by substituting
ðiÞ
state and the prediction horizon is selected sufficiently large, then ½us ðkÞopt in place of us in Eqs. (43) and (44). The overall adaptive
the complex solutions are not expected to arise during control optimizing control scheme is shown in Fig. 1.
law implementation. Though the above scheme is derived for the case where the set-
The robustness of the control law in the presence of plant-mod- point is updated at each sampling instant, this may not be desir-
el mismatch can be increased by introducing a diagonal filter able particularly when the model parameters are changing fast.
FIMC ðzÞ in the feedback path. This can be achieved by replacing In practice, the determination of the optimum operating point
ðkÞ in Eq. (32) by f ðkÞ, where f ðkÞ represents the filtered feed- can be carried out periodically at intervals considerably larger than
back signal computed as the sampling interval.

Fig. 1. On- line optimizing control: schematic representation.


78 R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82

4. Simulation studies where the first term on the right hand side indicate the rate of evap-
oration/condensation of the liquid water and the second term is the
We demonstrate the efficacy of the proposed adaptive optimiz- net amount of water transported across the membrane. The water
ing control scheme by simulating peak seeking control problem vapor balance at the cathode is given by
associated with a proton exchange membrane fuel cell (PEMFC). v l
dMw ;c ðxÞ dM w ;c ðxÞ haðxÞ h
To begin with, we describe the mechanistic model used for carry- ¼ þ IðxÞ þ IðxÞ ð49Þ
ing out simulation of PEMFC dynamics. The model used in this arti- dx dx F 2F
cle is an along-the-channel model developed by Yi and Nguyen where the third term on the right hand side indicate the amount of
[19] and extended by Golbert and Lewin [20], which we have de- water vapor generated by the reaction at the cathode. The change in
tailed in [21]. For clarity, we reproduce the details of the model temperature of anode gas, cathode gas and coolant are given by fol-
here. We then proceed to present the results of model identifica- lowing set of ODEs:
tion and the adaptive peak seeking control studies.
dT k ðxÞ U g Ag fT s ðxÞ  T k ðxÞg
¼ P ð50Þ
dx i C pi M i ðxÞ
4.1. PEMFC simulation
k ¼ a; c
The operating principle of the PEMFC involves the transporta- dT cool ðxÞ U w Acool fT s ðxÞ  T cool ðxÞg
¼ ð51Þ
tion of reactants through the electrodes to the reactive sites on dx C p;w Mcool
the catalyst, where the reaction takes place. To develop this model
following assumptions are made, which are given as (i) plug flow where T is the temperature, cool, g (subscripts) indicate coolant and
condition is assumed, (ii) pressure drop along the channels are ne- gas, respectively. U and A are heat transfer coefficient and heat ex-
glected, (iii) for gaseous species ideal gas law is considered, (iv) change area per unit length, respectively. Quantities appearing in
water transport in and out of the electrodes is assumed to be in the model, such as the ratio of water molecules per proton, mem-
the form of vapor only, (v) electrode layers are considered as ultra- brane conductivity and the enthalpy of water evaporation, and their
thin and hence resistance due to thickness is neglected, and (vi) an- temperature dependency, are specified using empirical equations
ode and cathode gas flows are assumed to be parallel. To avoid taken from the work of Golbert and Lewin [20].
dehydration at the anode, the anode and cathode streams are sat- The solid temperature of the PEMFC is affected by various
urated with water vapor. This saturation is maintained with inlet mechanisms such as (a) heat transfer by conduction, (b) heat trans-
streams. The model accounts for (i) the reactions taking place at fer to the fuel and oxidant flows and coolant channels, (c) heat gen-
the anode and cathode, (ii) heat transfer taking place between eration by the reactions, (d) heat of evaporation/condensation. The
the solid and the two gas channels, and (iii) heat transfer between energy balance is given by
the solid and the coolant. The dynamics of the system are captured @T s @ 2 T s U g Ag U w Acool
through the energy balance equation. The electrochemical reac- qs C ps ¼ ks 2 þ ðT a þ T c  2T s Þ þ ðT cool
@t @x f f
tions are known to be faster than the temperature dynamics, hence  
e DH
all other equations except the energy balance are assumed to be at  TsÞ  þ V cell IðxÞ
quasi-steady state for a given solid temperature profile. f 2F
!
l l
The consumption equations for the fuel and oxidant are given as 1 dM w ;a ðxÞ dMw ;c ðxÞ
follows: þ DHv ap ðT s Þ þ ð52Þ
f dx dx
dM H2 ðxÞ h
¼  IðxÞ ð45Þ with boundary conditions
dx 2F

dM O2 ðxÞ h @T s 
¼  IðxÞ ð46Þ ks ¼ U c ðT s  T inf Þ
dx 4F @t x¼0

where h is channel length in cm, F is Faraday’s constant and IðxÞ is @T s 
ks ¼ U c ðT s  T inf Þ
the local current density. MH2 and MO2 are the molar flow rate of @t  x¼L
hydrogen and oxygen, respectively. Evaporation and condensation
influence the balance for liquid water, which is given by the follow- The voltage obtained from the PEMFC is given by
ing set of equations: !  
( ) RT s PH2 P0:5
O2 RT s IðxÞ IðxÞt m
l v V cell ¼ V 0oc þ ln  ln  ð53Þ
dM w ;k ðxÞ K c hd M w ;k ðxÞ ðsatÞ nF P H2 O F i0 PO2 ðxÞ rm ðxÞ
¼ PNk Pk  Pw ðT k Þ
dx RT k ðxÞ i M i ;k
ð47Þ
where V 0oc is standard open circuit potential, i0 is the exchange cur-
k ¼ a; c rent density and rm is the membrane conductivity.
where M indicate molar flow rate, l; v (superscripts) indicate liquid The data for the base case considered in this paper is given in
and vapor, respectively, w; k (subscripts) indicate water and elec- Table 1. The modeling equations are solved along the channel.
trode, respectively. K c is conduction rate constant and P is the Eq. (53) is solved for current density after setting the cell voltage.
pressure. After obtaining the current density, the quasi-steady state Eqs.
The water vapor flow rates in the channels are affected by var- (45)–(51) are integrated along the channel (in the positive x-direc-
ious phenomena taking place in the PEMFC, such as (a) water vapor tion) by using a standard ODE solver. Here, it is important to note
being dragged through the membrane by migrating protons, (b) that Eq. (47) is used only when liquid water is present at the cath-
water being generated due to reaction at the cathode, (c) water va- ode or anode and/or the partial pressure is greater than the satura-
por diffusing through the membrane, and (d) liquid water con- tion pressure of water vapor. This type of situation has been
densing and evaporating depending upon the difference in partial handled in MATLAB using the ‘event’ function. In the present work,
pressure and saturation pressure. Based on all these phenomena, the PEMFC is operated on pure hydrogen and air at the anode and
the water vapor balance at the anode is given by cathode, respectively. The energy balance equation, which captures
v l
the dynamics of the system, is discretized using the Crank–Nichol-
dM w ;a ðxÞ dMw ;a ðxÞ haðxÞ son method as given in [20]. The set of algebraic equations ob-
¼  IðxÞ ð48Þ
dx dx F tained by discretizing the energy balance equation along the
R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82 79

Table 1 gas temperature at anode increases from 353 K to 355 K, the opti-
PEMFC simulation: nominal operating condition. mum operating point shift in positive direction from 0.6832 to
Input variables Value 0:7328 W cm2 . An increase in the inlet gas temperature helps in
1 increasing the amount of water molecule carried by inlet hydrogen
M H2 ðmol s Þ 1:34  105
1 gas from humidifier. Increase in amount of water molecules in-
M O2 ðmol s Þ 7:6  104
creases humidity of membrane, which, in turn, increases the
M vw;a ðmol s Saturated
1
Þ
amount of current produced. As a consequence, the optimum oper-
M vw;c ðmol s Saturated
1
Þ
ating power density is increased. Operating the PEMFC at the nom-
M lw;a ðmol s
1
Þ 0.0
inal optimum set point under such conditions leads to sub-optimal
M lw;c ðmol s
1
Þ 0.0
operation. On the other hand, for negative change of 2 K in inlet gas
T a ðKÞ 353
temperature at anode, the maximum attainable power output de-
T c ðKÞ 353
T cool ðKÞ 320 creases to approx. 0:6411 W cm2 and the nominal optimum set-
T inf ðKÞ 343 point becomes unattainable. Thus, if shifts in inlet gas
P a ðatmÞ 1 temperature at anode go unnoticed, then a conventional feedback
P c ðatmÞ 1 controller will operate PEMFC either at a sub-optimal operating
point or it may have to deal with an unachievable set point. Thus,
to maximize the efficiency of PEMFC operation, it becomes neces-
channel are solved to obtain the solid temperature profile for the sary to estimate the shifting optimum operating point on-line and
next time instant. move the operation to the current optimum.
The controlled outputs are average power density ðy1 Þ and aver-
age solid temperature ðy2 Þ. Cell voltage ðu1 Þ and inlet molar flow 4.2. Off-line OBF-Wiener model development
rate of coolant ðu2 Þ are selected as manipulated inputs. The inlet
gas temperature at anode is treated as unmeasured disturbance. To estimate the control oriented OBF-Wiener model, the cell
While simulating the process dynamics, the uncertainties in voltage and inlet molar flow rate of coolant are simultaneously
manipulated inputs are assumed to be colored stationary noise se- perturbed using multilevel random signals with standard devia-
1
quence governed by the following equation: tions of ru1 ¼ 0:1421 V and ru2 ¼ 1:1914  104 mol s , respec-
2 3 tively. Fig. 4 presents the input perturbations in the validation
0:01q1
0  data set, which are qualitatively similar to the perturbations in
5 w1 ðkÞ
10:99q 1
tu ðkÞ ¼ 4 1 the identification data set. These signals are obtained by modifying
0 0:01q w2 ðkÞ
10:99q 1
pseudo random binary sequence (PRBS) signals generated using
the ‘idinput’ function in the system identification toolbox of MAT-
where w1 ðkÞ and w2 ðkÞ are zero mean normally distributed random LAB with switching time of 1 s. Initial 2500 data points in the data
variables with standard deviations of 0.006 V and 2:27  106 set are used for model identification while the remaining 400 data
1
mol s , respectively. In addition, measurements of power density samples are used for model validation. The standard deviations of
and average temperature are assumed to be corrupted by zero the outputs (ry1 ¼ 0:0887 W cm2 and ry2 ¼ 1:9808 K) are used
mean normally distributed sequences with standard deviations of for scaling the output data. Table 2 reports the optimum GOBF
0:008 W cm2 and 0.28 K, respectively. poles ni in terms of continuous time poles ai where
It is well known in the literature that PEMFC are vulnerable to
various disturbances as explained the earlier section. As a conse- ni ¼ expðai TÞ
quence of level changes in unmeasured disturbances, the transient
and steady state characteristic (polarization curve) of the PEMFC Table 2
Optimum poles of identified GOBF model for the PEMFC system.
changes. For example, Fig. 2 presents the steady state behavior of
average power density with respect to manipulated cell voltage u1 u2
at different levels of inlet gas temperature at anode. When inlet y1 ½ 0:0067 0:0067  ½ 0:8863 0:8863 0:8863 
y2 ½ 0:9350 0:9350  ½ 0:8867 0:8867 0:8867 

0.75
Anode temp 355 K
Anode temp 353 K
0.7 Anode temp 351 K 0.8
density (W cm-2)

0.7
Avg. power
Avg. power densi ty (W cm )
-2

0.65 0.6
0.5 Measured output
0.6 Estimated output
0.4

0.55 0 5 10 15 20
360
0.5
temperature (K)
Avg. solid

355
0.45
350
Measured output
0.4
0.4 0.5 0.6 0.7 0.8 0.9 Estimated output
345
Cell Voltage (V) 0 5 10 15 20
Time (s)
Fig. 2. Steady state input–output behavior: shift in peak in response to level
changes in disturbance. Fig. 3. Model validation: comparison of process and simulated outputs.
80 R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82

The prediction horizon p is chosen as 60 for non-adaptive as


Cell voltage (V)

0.8 well as adaptive NIMC. In the adaptive NIMC formulation, RLS algo-
rithms with fixed forgetting factor of 0.975 (for average power
0.6 density) and 0.978 (for average solid temperature) have been em-
ployed. To kickstart the parameter estimation when the distur-
0.4 bance occurs, the covariance matrices are reset to aI where a is a
large value. In practice, this can be automated based on controller
0 5 10 15 20 and estimator errors. In the adaptive case, to avoid oscillations
-4 caused due to frequent changes in set point, the estimation of set
x 10
rate of coolant (mol s-1)

8 point (i.e. optimum operating point) is carried after every 20 sam-


ples. Also, the estimated value of set points are filtered through
Inlet molar flow

7
first order unity gain filter
6

5 rf ðkÞ ¼ ½aIrf ðk  1Þ þ ½ð1  aÞIrðkÞ


4 with filter parameter a ¼ 0:9.
3 The task of the non-adaptive NIMC controller is to maintain the
0 2 4 6 8 10 12 14 16 18 20
operation at the nominal peak point (r1 ¼ 0:6902 W cm2 and
Time (s)
r2 ¼ 352:68 K). The results for closed loop simulation when non-
Fig. 4. Model validation: input perturbations. adaptive NIMC is employed for control are reported in Figs. 7 and
8. When the inlet gas temperature at anode increases, the optimum
operating point shifts to a higher location. Unaware of this shift,
Fig. 3 compares the results of model simulation with measured out- the non-adaptive NIMC maintains the operation at the nominal
puts for the validation data set. From Fig. 3 it may be inferred that operating point without causing any offset. When subsequently
the identified model adequately captures the output variability the inlet gas temperature at anode decreases, the setpoint based
caused by manipulated input variations. Fig. 5 compares the steady on the nominal set point on power density becomes unattainable
state behavior of the identified model with that of the plant. The and this results in offset as can be seen from Fig. 7. In the presence
identified model capture the steady state behavior with reasonable of such offset, any other controller with integral action can be ex-
accuracy in the desired operating range of the PEMFC. pected to lead to input saturation or large oscillations in input [7].
It may be noted that, even though there is an offset that cannot be
eliminate, the NIMC controller does not lead to input saturation.
4.3. Adaptive optimizing control This can be attributed to a peculiar feature of the NIMC controller.
When the solutions of the quadratic controller design equation be-
From Fig. 5 it is clear that the peak voltage point is the most come complex then the setpoint become unattainable. In fact, after
desirable operating point. However, as shown in Fig. 2, the location the initial transient, the solutions turn out to be purely a complex
of the peak shifts when levels of unmeasured disturbances change. number with zero real part. As a consequence, the NIMC controller
Thus, to make sure that the controller operates PEMFC at the opti- given by Eq. (39) does not lead to input saturation or oscillations
mum conditions in the face of disturbances, it becomes necessary [7].
to track the shifting peak. In this subsection, we compare the per- Fig. 9 shows the results of the proposed adaptive optimizing
formances of the proposed adaptive peak seeking NIMC controller control scheme. Using time varying parameters, the optimum
with non-adaptive NIMC controller when an unmeasured distur- power density is recomputed on-line as described above and given
bance (inlet gas temperature at anode) changes as shown in to the adaptive NIMC controller as a set point. When gas inlet tem-
Fig. 6. The corresponding variation of optimum operating point is perature at anode changes in positive direction, the proposed on-
shown in Fig. 2. line optimization scheme is quickly able to locate the changing

0.8
0.7
density (W cm-2)
density (W cm-2)

Estimated
0.75
Avg. power
Avg. power

Measured
0.6 0.7
Estimated
Measured 0.65
0.5
0.6
0.3 0.4 0.5 0.6 0.7 3 4 5 6 7 -4
x 10
360 356
Estimated
temperature (K)

temperature (K)

355 354
Avg. solid

Avg. solid

Measured

350 Estimated 352


Measured
345 350
0.3 0.4 0.5 0.6 0.7 3 4 5 6 7
-1
Voltage (V) Flow rate of coolant (mol s ) -4
x 10

Fig. 5. Model validation: comparison of steady state behavior w.r.t. manipulated inputs.
R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82 81

356 0.75

density (W cm-2)
Avg. power
Set point (Adaptive NIMC)
0.7
Inlet gas temperature at anode (K)

355
0.65 Measured output

354 0 5 10 15 20 25 30 35
Disturbance
353

temperature (K)
Avg. solid
353
352
Set point

351 Measured output


352
350
0 5 10 15 20 25 30 35
351 Time (s)

Fig. 9. Adaptive optimizing NIMC: closed loop behavior in response to unmeasured


350 disturbance variation.
0 5 10 15 20 25 30 35
Time (s)

Fig. 6. Variation of unmeasured disturbance.

Cell voltage (V)


0.65
0.6
0.55
0.7 0.5
density (W cm-2)
Avg. power

0.68 0.45
Set point (Non Adaptive NIMC) 0 5 10 15 20 25 30 35
-4
0.66 x 10
rate of coolant (mol s-1)

Measured output
0.64 7
Inlet molar flow

0.62
0 5 10 15 20 25 30 35 6
353
5
temperature (K)
Avg. solid

352.5
4
0 5 10 15 20 25 30 35
352
Time (s)
Set point (Non Adaptive NIMC)
Measured output
Fig. 10. Adaptive optimizing NIMC: variation of manipulated inputs.
351.5
0 5 10 15 20 25 30 35
Time (s)
optimum operating point. Also, the adaptive NIMC controller
Fig. 7. Non-adaptive NIMC: closed loop behavior in response to disturbance tracks the time varying setpoint trajectory quite well. However,
variation. when gas inlet temperature at anode changes in negative direction,
it takes longer time for the on-line optimizer to locate the new
optimum. The process moves towards maximum attainable power
density under changed circumstances before the RLS can find the
0.7 correct setpoint. As the proposed NIMC is able to deal with unat-
tainable setpoint, this does not lead to any problem while locating
Cell voltage (V)

0.65 the new optimum point. The manipulated variable profiles gener-
ated by adaptive NIMC are shown in Fig. 10. The real concern in
0.6
adaptive optimizing control is to track the optimum when it be-
0.55 comes higher than the current setpoint, which satisfactorily car-
ried out by proposed approach.
0.5
0 5 10 15 20 25 30 35
-4
5. Conclusions
x 10
8
rate of coolant (mol s-1)

In this work, an adaptive optimizing control scheme has been


Inlet molar flow

developed, which combines the attractive features of the on-line


optimization approach proposed by Bamberger and Isermann [1]
6
and closed form control law developed by Bedi et al. [2]. The effec-
tiveness of the proposed adaptive optimizing scheme have been
demonstrated by conducting simulation studies on the distributed
4 an along-the-channel model of PEMFC. It is observed that sus-
0 5 10 15 20 25 30 35
tained changes in the unmeasured disturbance levels can cause
Time (s)
significant change in optimum power density attainable using a
Fig. 8. Non-adaptive NIMC: manipulated input variation. PEMFC. While non-adaptive control is able to maintain the
82 R.N. Methekar et al. / Journal of Process Control 20 (2010) 73–82

operation at the nominal setpoint when the setpoint is feasible, it References


results in sub-optimal operation of the PEMFC. The proposed on-
line optimizing control scheme estimates the changing location [1] W. Bamberger, R. Isermann, Adaptive on-line steady state optimization of slow
dynamic processes, Automatica 14 (1978) 223–230.
of the optimum operating point in response to variations in distur- [2] P. Bedi, R. Methekar, S.C. Patwardhan, V. Prasad, R.D. Gudi, Nonlinear Internal
bances quite well particularly when the operation at the nominal Model Control of PEM Fuel Cell, DYCOPS, 2007.
setpoint becomes sub-optimal. The proposed adaptive NIMC con- [3] M. Morari, Robust stability of systems with integral control, in: Proc. of
Decision and Control Conf., San Antonia, TX, 1983, pp. 865–868.
troller satisfactorily tracks the changing optimum operating point. [4] J.E. Slotine, W. Li, Feedback Linearization, Applied Nonlinear Control, Prentice
As the computation of the shifting optimum operating point and Hall, NJ, 1991.
the control law computations are carried out analytically, the pro- [5] M.A. Henson, D.E. Seborg, Nonlinear Process Control, Prentice Hall PTER, NJ,
1997.
posed scheme meets the fast computing time requirement and is a
[6] L.T. Biegler, J.B. Rawlings, Optimization approaches to nonlinear model
suitable candidate for implementing on-line optimizing control of predictive control, in: Y. Arkun, W.H. Ray (Eds.), Fourth Int. Conf. on Chem.
PEMFC in real time. Proc. Control, Padre Island, Texas, CACHE-AICHE, New York, 1991, pp. 543–
571.
[7] S.C. Patwardhan, K.P. Madhavan, Nonlinear internal model control using
Appendix A. Bilinear matrix operations quadratic prediction models, Comp. Chem. Eng. 22 (4/5) (1998) 587–601.
[8] A. Marwardi, F. Yang, R. Pitchumani, Optimization of the operating parameters
of the proton exchange membrane fuel cell for maximum power density, J. Fuel
Definition A.1 (Bilinear matrix). A bilinear matrix B of dimension Cell Sci. Technol. 2 (2005) 121–135.
ðr  n  mÞ is ordered collection of numbers babc ; a ¼ 1; 2; . . . ; r; [9] V. Mishra, F. Yang, R. Pitchumani, Analysis and design of PEM fuel cells, J.
Power Sources 141 (2005) 47–64.
b ¼ 1; 2; . . . ; n; c ¼ 1; 2; . . . ; m. It is highlighted by inclusion in the [10] Z. Zhi-dan, H. Hai-bo, Z. Xin-jian, C. Guang-yi, R. Yuan, Adaptive maximum
curly brackets as {B} or fbabc g. power point tracking control of fuel cell power plants, J. Power Sources 176
(2008) 259–269.
[11] S. Deshpande, S.C. Patwardhan, Unconstrained NMPC based on a class of
Definition A.2. A ðr  n  mÞ bilinear matrix {B} operating on a Wiener models: a closed form solution, in: International workshop on
ðn  1Þ vector v is represented as A ¼ fBgðv Þ where A is a Assessment and Future Directions of Nonlinear Model Predictive Control
ðr  n  mÞ matrix with elements (NMPC’08), Pavia, Italy, 2008.
[12] S. Boyd, L. Chua, Fading memory and the problem of approximating nonlinear
X
n
operators with Volterra series, IEEE Trans. Circ. Syst. CAS-31 (11) (1985) 1150–
aac ¼ babc v b 1161.
b¼1 [13] S.C. Patwardhan, S.L. Shah, From data to diagnosis and control using
generalized orthonormal basis filters. Part I: development of state observers,
J. Process Control 15 (7) (2005) 819–835.
Definition A.3. A ðr  n  mÞ bilinear matrix {B} operating on a
[14] B.M. Ninness, F. Gustafson, A unifying construction of orthonormal basis for
ðn  1Þ vector v and a ðm  1Þ vector w is represented as system identification, IEEE Trans Autom Control 42 (4) (1997) 515–521.
z ¼ fBgðv ; wÞ where z is a ðr  1Þ vector with elements [15] M. Srinivasarao, S.C. Patwardhan, R.D. Gudi, From data to nonlinear predictive
control: I Identification of multivariable nonlinear state observers, Ind. Eng.
X
m X
n
Chem. Res. 45 (2006) 1989–2001.
za ¼ babc v b wc [16] L. Ljung, T. Soderstrom, Theory and Practice of Recursive Identification, MIT
c¼1 b¼1 Press, Cambridge, MA, 1983.
[17] L.B. Rall, Computational Solutions of Nonlinear Operator Equations, John
A ðr  n  nÞ bilinear matrix is called symmetric if Wiley, New York, 1969.
fBgðv ; wÞ ¼ fBgðw; v Þ for every v ; w 2 Rn . [18] S.C. Patwardhan, K.P. Madhavan, Nonlinear predictive control using
approximate second order perturbation model, Ind. Eng. Chem. Res. 32
(1993) 334–344.
Definition A.4 (Left dot product). The left dot product of a [19] J.S. Yi, T.V. Nguyen, An along-the-channel model for proton exchange
ðr  n  mÞ bilinear matrix {B} with a ðk  rÞ matrix A is repre- membrane fuel cells, J. Electrochem. Soc. 145 (1998) 1149–1159.
sented as [20] J. Golbert, D. Lewin, Model-based control of fuel cells: (1) Regulatory control, J.
Power Sources 135 (2004) 135–151.
fDg ¼ A  fBg [21] R.N. Methekar, V. Prasad, R.D. Gudi, Dynamic analysis and linear control
strategies for proton exchange membrane fuel cell using a distributed
where {D} is a ðk  n  mÞ bilinear matrix with elements parameter model, J. Power Sources 165 (2007) 152–170.

X
r
dabc ¼ aag bgbc
g¼1

Anda mungkin juga menyukai