Robust Nonlinear Model Predictive Control With Reduction of Uncertainty Via Robust Optimal Experiment Design
Robust Nonlinear Model Predictive Control With Reduction of Uncertainty Via Robust Optimal Experiment Design
Robust Nonlinear Model Predictive Control With Reduction of Uncertainty Via Robust Optimal Experiment Design
Abstract: This paper studies the reduction of the conservativeness of robust nonlinear
model predictive control (NMPC) via the reduction of the uncertainty range using guaranteed
parameter estimation. Optimal dynamic experiment design is formulated in the framework of
robust NMPC in order to obtain probing inputs that maximize the information content of
the feedback and simultaneously to guarantee the satisfaction of the process constraints. We
propose a criterion for optimal experiment design which provides a minimization of parameter
uncertainty in the direction of improved performance of the process under robust economic
NMPC. A case study from the chemical engineering domain is studied to show the benefits of
the proposed approach.
Keywords: Nonlinear model predictive control, Robust control, Optimal experiment design,
Guaranteed parameter estimation
1. INTRODUCTION
Many studies have been devoted to the problem of realtime optimal process control. The most challenging problems involve handling of nonlinearities and uncertainties
of the processes. This paper is focused on the reduction
of the parametric uncertainties of the processes for performance improvement of nonlinear model-based control
approaches.
We consider a model of the process described by a set
of nonlinear ordinary differential equations (ODEs) and
output equations in sampled form as
xk+1 = f (xk , uk , dk ),
(1a)
y k = g(xk ),
(1b)
where k stands for the sampling instant, x denotes the nx dimensional vector of process states, u represents the nu dimensional vector of process control variables (inputs),
d is the nd -dimensional vector of uncertain parameters
of the process model with a priori known bounds D k :=
U
[dL
k , dk ], and y denotes the ny -dimensional vector of model
outputs (predictions of measurable state combinations).
The superscripts L and U representing the lower and upper
bounds of an interval box are understood component-wise
throughout the paper. We want to control the process,
i.e. determine the control inputs for (1), such that
Np 1
J (y k+1 , uk ) :=
L(y k+1 , uk ),
(2)
k=0
1 The authors thank Professor Sebastian Engell for fruitful discussions and useful suggestions. This research was funded by the
European Commission under grant agreement number 248940 (EMBOCON), 291458 (MOBOCON) and from the Deutsche Forschungsgemeinschaft under grant agreement number EN 152/39-1.
1904
xjk+1
(4b)
at stage k + 1 and position
p(j)
subject to:
p(j)
xjk+1 = f (xk
r(j)
, ujk , dk
),
(j, k + 1) I, (5b)
y jk
= g(xjk ),
(j, k) I, (5c)
r(j)
j
j
j
0 h(y k+1 , xk+1 , uk , dk ),
(j, k) I, (5d)
p(j)
p(l)
(j, k), (l, k) I, (5e)
ujk = ulk if xk = xk ,
PN
j
j
In order to represent the real-time decision problem correctly, the control inputs cannot anticipate the realization
of the uncertainty at the instant k. This is modeled by the
non-anticipativity constraints (5e) that force all the control
p(j)
inputs ujk that branch at the same parent node xk to be
equal. The definition of problem (5) allows the uncertainty
to vary over the time. In this study we will assume that
the uncertain parameters are constant for simplicity, hence
U
the notation d0 D 0 := [dL
0 , d0 ].
3. ROBUST OPTIMAL DYNAMIC EXPERIMENT
DESIGN
1905
u11 d11
u12 d12
u13 d13
1
1
x
u
x0
1
0
d 01
2
0
u d
2
0
u03 3
d0
u14 d11
u15 d12
u16 d13
x12
u17 d11
u18 d12
u19 d13
3
1
x12
u12 d 21
x31
u31 d 31
x14
x22
u32 d32
x42
u22 d 22
x32
3
2
u23 d 23
3
3
x24
u d
3
3
x43
u24 d 21
x34
u34 d 31
x44
x25
u25 d 22
x35
u35 d 32
x45
x26
u26 d 23
x36
u36 d 33
x46
x27
u27 d 21
x37
u37 d 31
x47
x28
u28 d 22
x38
u38 d 32
x48
x29
u29 d 23
x39
u39 d 33
x49
3
3
Robust horizon = 2
Prediction horizon = 4
where Ne stands for the horizon where the optimal experiment is realized, Q is the inverse of the covariance
matrix of the measurement noise, and sy,k+1 represents
the matrix of parametric output sensitivities, here defined
in the fully relative form suggested by Munack (1991):
sx,k+1 = f s (sx,k , xk , uk , d),
(8a)
sy,k = g s (sx,k , xk ),
(8b)
where the sensitivity of the ith state w.r.t. the jth parameter obeys the dynamics
fi
fi
dj
s
,
(9)
{sx,k }j +
fi,j =
xj,k xT
dj
where {}j represents the jth column of a matrix. The
function g s is derived analogously respecting (1b).
Among the several different possible experiment design
criteria (Munack, 1991), we choose a modified E-design
criterion:
mE (F) = max i (F) / min i (F) ,
(10)
i
i=1
In contrast, guaranteed (bounded-error) parameter estimation accounts explicitly for the fact that the actual process
outputs, y p , are only known to be corrupted by some
bounded measurement errors e E := [eL , eU ], so that
L U
y pk y m
(13)
k + [e , e ] =: Y k .
Here, the main objective is to estimate the set D e of all
possible parameter values d such that y k Y k for every
k = 1, . . . , Ne ; that is,
x, y :
xk+1 = f (xk , uk , d),
. (14)
De := d D0
y k = g(xk ),
y Y , k {1, . . . , N }
k
k
e
Depicted in red in Fig. 2 is the set of parameters D e
projected in the (d1 , d2 ) space that generate trajectories
satisfying y k Y k , k = 1, . . . , Ne .
1906
d2
D Dint
D Dout
De
D Dbnd
D0
d1
DDint
DDbnd
(15)
Further details on possible implementation variants of
the described procedure can be found in Paulen et al.
(2013a,b).
5. PROPOSED ALGORITHM
1907
0.05
Economic
Modified E
Proposed
6.2
6.6
Tcf [ C]
0
0
0.2
0.4
0.6
0.8
Economic
Modified E
Proposed
Real value
6.8
80
78
7
7.2
7.4
76
0
0.2
0.4
0.6
0.8
7.6
0.1
u [l/h]
x 10
6.4
H [J/mol]
cC [mol]
0.1
7.8
8
0.056
0.05
0
0
0.2
0.4
0.6
Time [h]
0.8
0.057
0.058
k [l/mol h]
0.059
0.06
1908
cC [mol]
1
0.5
0
0
Tcf [ C]
Economic
Modified E
Proposed
10
15
20
25
10
15
20
25
10
15
Time [h]
20
25
80
78
76
0
u [l/h]
0.1
0.05
0
0
Astr
om, K. and Wittenmark, B. (1971). Problems of
identification and control. Journal of Mathematical
Analysis and Applications, 34, 90113.
Bertsekas, D.P. (2000). Dynamic Programming and Optimal Control. Athena Scientific, 2nd edition.
Feldbaum, A.A. (1960). Dual control theory I. Automation
Remote Control, 21, 874880.
Franceschini, G. and Macchietto, S. (2008). Model-based
design of experiments for parameter precision: State of
the art. Chem. Eng. Sci., 63(19), 48464872.
Gevers, M., Bombois, X., Hildebrand, R., and Solari,
G. (2011). Optimal experiment design for open and
closed-loop system identification. Communications in
Information and Systems, 11, 197224.
Houska, B., Ferreau, H.J., and Diehl, M. (2011). An autogenerated real-time iteration algorithm for nonlinear
MPC in the microsecond range. Automatica, 47(10),
22792285.
1909