A Robust Multi-Model Predictive Controller for

Distributed Parameter Systems

Mı́riam R. Garcı́aa,b , Carlos Vilasb , Lino O. Santosc , Antonio A. Alonsob,∗
a
Hamilton Institute – National University of Ireland, Maynooth, Ireland
b
Process Engineering Group, IIM-CSIC, Eduardo Cabello 6, 36208 Vigo, Spain
c
CIEPQPF, Department of Chemical Engineering,
Faculty of Sciences and Technology, University of Coimbra, R. Sı́lvio Lima, 3030-790
Coimbra, Portugal

Abstract

In this work a robust nonlinear model predictive controller for nonlinear

convection-diffusion-reaction systems is presented. The controller makes use
of a collection of reduced order approximations of the plant (models) re-
constructed on-line by projection methods on POD (Proper Orthogonal De-
composition) basis functions. The model selection and model update step
is based on a sufficient condition that determines the maximum allowable
process-model mismatch to guarantee stable control performance despite pro-
cess uncertainty and disturbances. Proofs on the existence of a sequence of
feasible approximations and control stability are given.
Since plant approximations are built on-line based on actual measure-
ments the proposed controller can be interpreted as a multi-model nonlin-
ear predictive control (MMPC). The performance of the MMPC strategy is

∗
Corresponding author
Email addresses: [email protected] (Mı́riam R. Garcı́a),
[email protected] (Carlos Vilas), [email protected] (Lino O. Santos),
[email protected] (Antonio A. Alonso)

Preprint submitted to Elsevier February 3, 2012

illustrated by simulation experiments on a problem that involves reactant
concentration control of a tubular reactor with recycle.
Keywords: Plant Model Mismatch; Proper Orthogonal Decomposition;
Controller Stability; Projection methods.

1. Introduction

The success of Model Predictive Control (MPC) industrial applications in the

1970s [1, 2] led to a wide spread of this control technology in industry [3]. The
MPC algorithm relies on two main concepts: a predictive horizon to perform
process dynamics model based predictions, and the solution of an open-loop
control problem at every sampling-time. MPC proved over the years to be
an efficient control tool for a wide class of multivariable nonlinear dynamic
systems. Its appeal must be found essentially in its conceptual simplicity
although, behind this formulation, a number of “not that simple” questions
and problems hide. The vast and rich literature, with several books, e.g.,
[4, 5], available on this subject testifies the research efforts undertaken so
far. It also reflects the degree of success attained in overcoming theoretical
and practical issues related with stability, robustness and implementation.
An early summary on predictive control research and applications can be
found in [6]. Surveys on MPC theory and practice are given for instance
in [7, 8], while [9, 10] provide an overview on robust MPC. The developments
and challenges in nonlinear model predictive control (NMPC) are discussed
in [11, 12, 13, 14, 15]. Optimality and stability issues related to MPC can
be found in [16]. Robustness assessment of the controller has been recently
addressed in the context of “Input to State Stability” concepts by [17, 18].

2
 Distributed parameter systems (DPS) are examples of infinite dimen-
sional nonlinear dynamic systems that challenge the efficiency of MPC im-
plementations due to a number of reasons directly or indirectly related with
the following issues: the high level of spatial discretization required to ap-
proximate the original set of partial differential equations; the dependence of
the discretization scheme on the dynamic properties of the plant; the poten-
tially stiff nature of the resulting nonlinear differential algebraic or ordinary
equations or the usually large times of convergence associated with the sub-
sequent optimization problems.
This may be tackled by formulating simpler ODE (Ordinary Differen-
tial Equation) models to describe the behavior of the DPS. For instance,
in [19, 20] NMPC of a fixed-bed water-gas shift reactor is formulated using
a simplified model with fewer equations and states. Also a lumped model-
ing approach is considered in [21] where a lexicographic optimization based
MPC is applied to control a continuous pulp digester. In a strategy to apply
on-line MPC to an experimental drying process [22], the partial differential
equations are solved off-line, and then a linearized PDE (Partial Differen-
tial Equation) model around the previous off-line behavior is used to find
the optimal variations for the on-line predictive control. This MPC strategy
combines a two phase approximation of the PDE model in an internal model
control (IMC) structure [23, 24]. Also, in [25, 26] to avoid the spatial dis-
cretization of the system the authors develop a neural network based model
for the predictive controller. From another perspective, the development of
more efficient NMPC algorithm strategies can greatly reduce the computa-
tional burden associated to the large-scale problems that arise in distributed

3
parameter systems when using a full discretization of the PDE model [27].
Besides the various approaches taken in the aforementioned works, these
challenges have been usually overcome in control and optimization applica-
tions by the so-called projection techniques [28, 29, 30, 31]. These techniques
stand on the dissipative nature of the parabolic (diffusion based) PDE set
and the time scale separation of dynamic modes, transforming the original
PDE model into a low dimensional dynamic system capturing the most repre-
sentative (slow) dynamics [32]. This approach has been successfully applied
in the context of dynamic matrix control [33, 34] and MPC [35, 36, 37].
The so-called POD (Proper Orthogonal Decomposition) technique lies
into this category. It was firstly proposed by Sirovich [38] and exploited by
the group of Holmes and Lumley [39, 40, 41] in the context of fluid dynamics
as a way to explore the routes to turbulence phenomena, and was rapidly
extended to other fields such as chemical reaction and biological systems. In
this technique, measurements of the spatio-temporal evolution of the field are
employed to derive the ROM (Reduced Order Model). We can envisage the
application of these techniques in processes with process tomography devices
to measure spatial state fields, that is, to capture the so called snapshots of
the spatial distribution in order to track the states spatial variations inside
the process. In fact, a recent review on control of DPS recognizes that the
POD has emerged as a popular tool for model reduction towards model-
based control applications [42] and it has been exploited as an off-line model
reduction technique for MPC [43].
The main disadvantage of PODs however, lies in that their accuracy is
limited to a certain region of the state space defined by the field measure-

4
ments employed to derive the ROM [44]. In the context of Model Predictive
Control, this limitation translates into a divergence between the plant and
its reduced order representation known as plant-process mismatch that if not
properly assessed, for instance at the control design stage, it will deteriorate
control performance or even drive the plant unstable. A detailed discus-
sion about different approaches to waive this limitation can be found in [45]
and references therein. Among those, particularly interesting is the work by
[46] which presented a methodology to update the snapshots based on an
optimization procedure which is able to find those regions where the ROM
significantly differs from the full model. Although the aforementioned tech-
niques increase the validity region of the POD technique, problems may arise
when plant perturbations lead the system far away from such region. From
another point of view, Atwell and King [47, 48] approached the problem for
closed loop systems by using an alternative type of input collection which
takes into account the derivation of the infinite dimensional feedback law.
Using this input collection does not require guesswork and, as pointed out
by the authors, controllers designed in this way perform better as compared
with those designed using the traditional snapshot approach.
Independently of the order reduction technique, however, the dissipative
nature of convection-diffusion-reaction systems ensures the existence of a
reduced order representation capable of approximating at arbitrary accuracy
the real plant dynamics for a given operation region [32]. Therefore, it is
desirable to update the ROM currently in use by an MPC controller whenever
closed loop performance and/or stability can not be guaranteed.
In this work a sufficient condition for robust stability of nominal MPC

5
with modeling error similar to the one proposed in [49, 50] will be extended
to the application of MPC of dissipative distributed process systems. It will
be employed to select a reduced order model of the plant among a sequence
of possible approximations so that the stabilizing properties of MPC are
preserved despite uncertainties and disturbances. The authors in [49, 50]
address the question of how much mismatch a given MPC algorithm can
tolerate and how it can be assessed for stability of the MPC controller. Then
an off-line procedure is developed to evaluate constants which determine
sufficient conditions for robust stability. In this work, the proposed sufficient
condition is applied to trigger the needed on-line updates of the ROM model
used by the MPC controller in order to ensure robust stability. Therefore,
when the sufficient condition is not satisfied, new plant measurements are
taken to update on-line the POD basis, thus overcoming the reduced order
model accuracy limitations discussed above.
Since the model is built on-line, based on actual measurements from the
plant, the proposed approach can be interpreted as an MPC controller mak-
ing use of multiple models, that is, a multi-model predictive control (MMPC).
Early versions of the proposed strategy have been reported in [51, 52] and
were applied to the control of simulated moving bed (SMB) separation pro-
cesses [53]. The present contribution concentrates on the theoretical of the
approach related with the model selection criterion and the convergence of
the resulting control configuration.
The paper is organized as follows: The class of systems under consider-
ation in this work is described in Section 2. This includes the role of dissi-
pation in model reduction, the POD method and its main properties to be

6
exploited later in the MMPC strategy. Section 3 summarizes the robust MPC
approach to tackle plant/reduced order model mismatch and discusses con-
ditions under which the closed-loop system remains stable. The final MMPC
configuration comprising the reduced model update scheme and MPC, and
its convergence properties, are described along Section 5. The performance
of the MMPC strategy is illustrated by simulation experiments on a problem
that involves the reactant concentration control of a tubular reactor with
recycle (Section 6). Finally, main conclusions are drawn in Section 7.

2. Dissipative systems and its dynamic approximations

2.1. System description and properties

The class of systems we consider is that described by quasi-linear partial

differential equations representing convection-diffusion-reaction mechanisms
for mass or energy transport with (possibly non-linear) chemical reactions.
For the sake of clarity we will concentrate in the present work on one di-
mensional partial differential equations. Note however that methods are not
restricted to 1D but can be extended in a straightforward manner to 2D or
3D cases using arguments presented in [32, 54]. Let the field z(ξ, t) represent
the evolution in space and time of a physical property, such as species con-
centration or temperature, in deviation form with respect to a given reference
state. Formally, the resulting quasi-linear PDE takes the form:

∂z(ξ, t) ∂j(z)
+ = f (z) + p(ξ, t) , (1)
∂t ∂ξ

where ξ ∈ V = [0, L] ⊂ R denotes spatial position with L being the length of

the spatial domain and t ∈ [0, ∞) denotes time. Function j(z) includes both

7
convective and diffusive fluxes so that:
∂z(ξ, t)
j(z) = vz(ξ, t) − k . (2)
∂ξ
As it is standard in tubular reactors, we will assume that the velocity field
∂v
v will be independent of space, thus = 0. Note however that equivalent
∂ξ
conclusions can be drawn for the more general case of v ≡ v(ξ) as it is
described in [55]. The terms f (z) and p(ξ, t) represent the nonlinear reaction
term and the, possibly distributed, control input, respectively. The system
is completed with boundary conditions of the form:
   
∂z(ξ, t)
n· k + hz(ξ, t) = 0, (3)
∂ξ ξ=B

where n is a unit normal vector pointing outwards of the boundary B and h >
0 is a constant representing a property (mass or energy) transfer coefficient
at the boundary. In what follows some of the function arguments will be
omitted for the sake of clarity. We assume that the system is dissipative in
the sense of references [56, 55]. This implies, among other things, that k in
(2) must be positive and that the nonlinear term f (z) satisfies the following
assumption:

Assumption 1 (Lipschitz condition). There exist some nonnegative param-

eters µ and  and a positive definite function `(z, µ) > 0, with `(z, 0) > 0 for
k z k> , such that the following relation holds for f(z) in (1):

zf (z) + `(z, µ) = µz 2 . (4)

Integrating (4) over the spatial domain V leads to

Z
hz, f iV + Lµ = µ hz, ziV , Lµ = `(z, µ)dξ , (5)
V

8
where hg, hiV should be understood as integration over V, i.e., hg, hiV =
R T
V
g h dξ. Given a field z(ξ, t), let us define a L2 norm as follows: kzk22 =
hz, ziV .
As shown in [32, 55] a field that obeys equations (1)-(4) is dissipative
and therefore bounded in L2 , provided that the control input field p(ξ, t) is
bounded. On the other hand, since f (z) is Lipschitz, it is also bounded and
thus belongs to L2 . Because of this, both functions accept Fourier series
expansions of the form [57]:
∞
X ∞
X
z(t, ξ) = mi (t)φi (ξ) , f (z) = σi (t)φi (ξ) , (6)
i=1 i=1

where mi (t) and σi (t) are the system dynamic modes, and φi (ξ) are eigen-
functions that constitute the solutions of the so-called Euler-Lagrange equa-
tion [58] (see also [32] for a discussion in the context of distributed passive
systems):
∂ 2 φi (ξ)
= −λi φi (ξ) . (7)
∂ξ 2
Equation (7) must be solved with the appropriate boundary conditions (in
our case equation (3)) to produce a complete set of orthonormal eigen-
functions {φi (ξ)}∞ ∞
i=1 with their corresponding positive eigenspectrum {λi }i=1

where each λi is an eigenvalue. As proved elsewhere (see for instance [58]),

the eigenvalues are ordered such that λi 6 λj for all integers i, j whenever
i < j and λn → ∞ as n → ∞. To simplify the notation we denote φi ≡ φi (ξ).
Note that projecting relation (7) over the eigenfunctions leads to:
 
∂ 2 φj
φi , = −λj hφi , φj iV = −λj δij , (8)
∂ξ 2 V

where δij is the Kronecker delta.

9
 The background material we have just outlined in this section will allow us
to construct a dynamic approximation of the plant capable of reproducing
the behavior of the system at arbitrary precision (reduced order models).
These will constitute the models employed by the predictive controllers to
command the operation. A brief description of reduced order models (ROMs)
is presented in Appendix A. The dynamics of the error associated to the
approximation and its bounds are discussed next.

2.2. Dynamic approximations

Let us make use of relation (2) to rewrite system (1) as:
 
∂z(ξ, t) ∂z(ξ, t) ∂ ∂z(ξ, t)
+v = k + f (z) + p(ξ, t) . (9)
∂t ∂ξ ∂ξ ∂ξ
By what has been mentioned previously, fields in (9) accept expansion in
Fourier series of the form of (6). Such series can be split into two contribu-
tions: one containing a finite number of elements N and the other containing
the remaining so that:
∞
X N
X ∞
X
z(ξ, t) = mi (t)φi (ξ) = mi (t)φi (ξ) + mi (t)φi (ξ) , (10)
i=1 i=1 i=N +1
∞
X N
X ∞
X
f (z) = σi (t)φi (ξ) = σi (t)φi (ξ) + σi (t)φi (ξ) , (11)
i=1 i=1 i=N +1
∞
X N
X ∞
X
p(ξ, t) = πi (t)φi (ξ) = πi (t)φi (ξ) + πi (t)φi (ξ) . (12)
i=1 i=1 i=N +1
In the same way first spatial derivatives of the field can be expressed by
convergent series expansions since, as shown in Appendix B, they belong to
L2 so that:
∞ N ∞
∂z(ξ, t) X X X
= τi (t)φi (ξ) = τi (t)φi (ξ) + τi (t)φi (ξ) . (13)
∂ξ i=1 i=1 i=N +1

10
In the remaining of the paper, we will employ the subindex A to denote the
part of the field related to the finite dimensional contribution in Eqns (10)-
(13) and a subindex B to refer to the part of the field corresponding to the
infinite dimensional contribution. In this way,
N
X ∞
X
rA (ξ, t) = αi (t)φi (ξ); rB (ξ, t) = αi (t)φi (ξ)
i=1 i=N +1

∂z(ξ, t)
with r(ξ, t) representing any of the fields z(ξ, t), f (ξ, t), p(ξ, t) or ,
∂ξ
and α(t) denoting the time dependent coefficients m(t), σ(t), π(t) or τ (t).
In all the expressions above, N + 1 will turn out to be the mode number
after which one can assure that the remaining modes will relax exponentially
fast. This will be shown next together with the fact that under Assumption
1 such N can always be found.
To simplify the notation we denote z ≡ z(ξ, t), zA ≡ zA (ξ, t), zB ≡
zB (ξ, t). In addition, by orthogonality of the eigenfunctions and relations
(10) we have that hzA , zB iV = 0, so that:

kzk22 = kzA k22 + kzB k22 . (14)

By the same token, and using expansions (10) and (13), it follows that:
       
∂z ∂zB ∂z ∂zB
zB , = zB , and zB , = zB , .
∂t V ∂t V ∂ξ V ∂ξ V

With these preliminary statements, let us define a Lyapunov function as

W = 12 kzB k22 and compute its time derivative along (9):
      
dW ∂zB ∂zB ∂ ∂zB
= zB , = −v zB , + zB , k
dt ∂t V ∂ξ V ∂ξ ∂ξ V (15)
+ hzB , fB iV + hzB , pB iV .

11
Let us now examine the different terms in the RHS of equation (15). Using
the first Green’s identity, the convection term in (15) is given by:
  Z
∂zB 1 ∂
1

v zB , = v zB2 dξ = v zB2 (L) − zB2 (0) , (16)
∂ξ V 2 V ∂ξ 2

or in terms of the Lyapunov function

 
∂zB
v zB , = v(W (L) − W (0)) . (17)
∂ξ V

By assuming no disturbance at ξ = 0, zB2 (0) = 2W (0) = 0 which makes the

left hand side of (17) positive definite.
Expanding the field of the diffusion term in (15) in series of the form (10)
we have:
   * ∞ ∞
+
∂ ∂zB X X ∂ 2 φi
zB , k = mi φi , k mi 2 .
∂ξ ∂ξ V i=N +1 i=N +1
∂ξ
V

Using (7) and orthonormality of eigenfunctions we get:

   * ∞ ∞
+ ∞
∂ ∂zB X X X
zB , k = mi φi , k mi (−λi )φi =k (−λi )m2i .
∂ξ ∂ξ V i=N +1 i=N +1 i=N +1
V

Taking into account that the smallest eigenvalue in the previous series is
P
λN +1 and kzB k22 = ∞ 2
i=N +1 mi , the diffusive terms can be bounded as
  
∂ ∂zB
zB , k 6 −kλN +1 kzB k22 = −2kλN +1 W . (18)
∂ξ ∂ξ V

Finally, invoking the Lipschitz condition (4) in its integral form (5), we derive
a bound on the nonlinear term in (15), being of the form:

hzB , fB iV 6 −Lµ + 2µW 6 2µW . (19)

12
Combining (17), (18), and (19), with (15), we obtain the following inequali-
ties:
dW
6 −vW (L) − 2kλN +1 W + 2µW + hzB , pB iV
dt
6 −2εN W + hzB , pB iV , (20)

where εN = kλN +1 − µ. Note that because of the ordering property of

eigenvalues λi mentioned in the previous section, for a given µ and k it is
always possible to find a number of modes large enough such that εN > 0.
In fact, let N be such number. Following these preliminary developments we
establish a bound on the error associated to the N -mode approximation of
the field, that is, a bound on kzB k22 quite connected to the notion of input to
state stability [59]. This is formally stated in the following proposition:

Proposition 1. Let N be such that εN = kλN +1 − µ > 0 and the input

√
control field be bounded as kpB k2 6 β 2 over a time interval T with β ∈ R+ .
Then the error associated to the approximation is bounded as follows:
2β 2
kzB (T )k22 6 kzB (0)k22 exp (−εN T ) + (1 − exp (−εN T )) . (21)
ε2N
Proof: First note that since the control field pB is assumed to be bounded,
√
using Hölder’s inequality, we have that hzB , pB iV 6 2kzB k2 β. Substituting
√
in (20) and using the fact that kzB k2 = 2W 1/2 we have that:
dW
6 −2εN W + 2βW 1/2 . (22)
dt
Because the second term on the RHS (square root of W ) is a concave function,
it is upper bounded by its supporting hyperplane [56], what allows us to write:

1/2 1
W 1/2 6 W1 + 1/2
(W − W1 ) . (23)
2W1

13
  2
1 β
Let W1 be of the form W1 = ω2 2εN
, with 0 < ω < 1 an auxiliary
parameter. Substituting W1 in (23), inequality (22) can be rewritten as:
!
dW β 1/2
6 −2 εN − 1/2
W + βW1
dt 2W1
β2
6 −2εN (1 − ω)W + .
2ωεN

Multiplying both sides of the above inequality by exp(rt), with r = 2εN (1 −

ω), and integrating over time in the interval (0, T ) we obtain:

β2
W (T ) 6 exp (−rT )W (0) + (1 − exp (−rT )) . (24)
4ε2N ω(1 − ω)

The second term on the right hand side attains a minimum for ω = 1/2 for
which r = εN . Consequently, a minimum bound for the propagation of the
approximation error kzB k22 takes the form given by Eqn (21). 4
Note that the larger N (and hence εN ), the smaller the bound on the
field zB . Thus the error associated to neglecting the field zB can be made
arbitrarily small by increasing the number of modes in the field zA .

3. Model Predictive Control. Problem Statement

As it is well known in process control, the essential idea behind any

model predictive controller is that of finding at regular time intervals those
control actions which minimize a given cost function indicative of the process
performance. To that purpose a reliable dynamic description of the process
to be controlled, usually in the form of algebraic and difference or differential
equations (DAES) is needed together with appropriate sensors to determine
the present state of the process.

14
 For the case considered in this work, and in order to construct the cost
function to be minimized during operation we define the state and control
variables in deviation form with respect to a reference which is chosen as the
controller setpoint, i.e.,

z = ẑ − zsp , p = p̂ − pref , (25)

where z ∈ Rq and p ∈ Rr are the vector fields, and q and r are the number
of states and control variables, respectively. From (25) it follows that at the
setpoint we have that z = 0, and p = 0.
• Cost function at time tk
In the MPC problem formulation we consider a predictive horizon M and a
control horizon Mc , with Mc 6 M . Let zk , and pk represent the deviation
variables at every sampling time tk , k > 0. Let us define a positive definite
function h(z, p) : Rq × Rr → R+ , and denote by h(zk , pk ) the value of such
function at time tk . A possible candidate cost function, frequently employed
in the MPC formulations, is the quadratic function:

1 T

h(zk , pk ) = zk Qz,k zk + pTk Qp,k pk , (26)
2

where Qz,k and Qp,k are appropriate positive definite symmetric matrices.
We define the functional (cost function) to be minimized over the spatial
domain V with boundary B subject to dynamic system (9) as follows:
Z
J(zk , Pk ) = H(zk , Pk )dξ , (27)
V

where zk is the field initial condition and Pk collects the control actions
(control fields) over a control horizon of Mc sampling times, Pk = {pi }k+M
i=k
c −1
.

15
Accordingly, Pk produces over the predictive horizon the state sequence Zk =
{zi }k+M
i=k+1 . Function H(zk , Pk ) in (27) takes the form:

k+M
X−1
H(zk , Pk ) = h(zi , pi ) + hF (zk+M ) , (28)
i=k

T
with hF (zk+M ) = zk+M Qz,k+M zk+M . It is important to remark that since
functions h(z, p) and hF (z) are lower bounded there exists a positive constant
αh such that:

h(z, p) > αh z T z , and hF (z) > αh z T z . (29)

• Optimal Control problem at time tk

The optimal control problem to be solved at every k is that of finding

the control profile P∗k over a given control horizon Mc which minimizes the
deviation between the state trajectory and the state reference or setpoint
over the predictive horizon M > Mc . In this formulation, the plant – Eq. (9)
– will be approximated by a number N of slow modes. Each approximation
constitutes a model of the plant described by a set of nonlinear differential
equations (see Appendix A for details), we formally express as:

dmA
= AA mA + FA , (30)
dt
P
with each field is approximated as z ≈ zA = N
i=1 φi mi . With some abuse of

notation, m should be understood as a vector associated to each field within

the vector z. Accordingly, a plant approximation would require as many
equations of the form (30) as fields considered. The corresponding NMPC

16
problem to be solved at every k is then stated as:

min. J(zk , Pk ) (31a)

Pk
dmA
s.t. = AA mA + FA (31b)
dt
zk+M = 0 . (31c)

Optional constraints may be added to problem (31) for Mc 6 M . This for-

mulation comprises two types of constraints as described in [50] to enforce
the nominal stability of the NMPC controller, the terminal cost term in (28)
and the terminal state constraint (31c). At this point we are not considering
any additional bounds and/or constraints on the problem solution. Given
the NMPC problem with a field initial condition zk , we assume that Pk
is a feasible solution for (31) and that there exists a sufficiently long hori-
zon that ensures an admissible trajectory to satisfy the terminal constraints.
Therefore, from solving (31) we obtain an optimal solution P∗k which when
implemented in the system will produce an optimal state sequence Z∗k , and
the corresponding optimal objective function value J(zk , P∗k ).

3.1. Assumptions

Regarding the developments of the robust stability analysis in the next

sections it is essential to list here the main assumptions on which that analysis
is based [49, 50].

Assumption 2. All the states z can be measured (or estimated) at every

time index k.

Assumption 3. Both the plant model (9) and the reduced order model (30)

17
satisfy zk+1 = 0 for the steady state reference values (zk , pk ) = (0, 0). Note
that because of cost function definition (26), it follows that J(0, P∗k ) = 0.

Assumption 4. There exists a sufficiently long horizon to ensure an ad-

missible trajectory (i.e., with zk+M = 0). This is a controllability property
termed Property C in [60].

4. Robustness conditions for Multi-Model Predictive Control

The stability analysis follows from familiar arguments developed by [61].

We first consider the perfect model case in order to discuss some of the
arguments that will be employed to address the mismatch case, resulting
from using the reduced order approximation (30).

4.1. Perfect model case

Consider the optimal solution of problem (31) at time index k, P∗k , with
cost J(zk , P∗k ), and hF (zk+M ) = 0, and apply the first element of the optimal
control sequence (p∗k ). As a result, the system will evolve to a state zk+1
∗

at time index k + 1. The optimal solution computed in the previous step

(P∗k ) is also a feasible candidate for the optimization problem at time index
∗
k + 1 starting from the initial condition zk+1 = zk+1 . Let P0k+1 = P∗k \p∗k
be the sequence computed at time index k without the first element, and
J(zk+1 , P0k+1 ) be its corresponding cost (which of course might not be the

18
optimal one). Then, we can write:
Z k+M
X−1 Z
∗ 0 ∗
J(zk , Pk ) − J(zk+1 , Pk+1 ) = h(zi , pi )dξ + hF (zk+M )dξ
V i=k V
Z k+M
X Z
− h(zi , p∗i )dξ − hF (zk+M +1 )dξ
V i=k+1 V
Z Z
= h(zk , p∗k )dξ − hF (zk+M +1 )dξ (32)
V V

According to Assumption 4 and Eq (31c) the last term of Eq (32) becomes

zero. Thus it simplifies to
Z
J(zk , P∗k ) − J(zk+1 , P0k+1 ) = h(zk , p∗k )dξ . (33)
V

Consider now the NMPC problem at time index k + 1. As mentioned above,

P0k+1 is a feasible candidate for the optimization problem at time index
k + 1, however it might not be the optimal one. Note that, in the worst
case, P∗k+1 = P0k+1 which implies J(zk+1 , P∗k+1 ) = J(zk+1 , P0k+1 ). In other
words, it is always possible to find a sequence P∗k+1 such that J(zk+1 , P∗k+1 ) 6
J(zk+1 , P0k+1 ). Therefore from (33) it follows that
Z
∗ ∗
J(zk , Pk ) − J(zk+1 , Pk+1 ) > h(zk , p∗k )dξ . (34)
V

Thus the sequence J(zk , P∗k ) over n time indices decreases, and because h(·)
is bounded from below by zero it converges, and h(zk , p∗k ) → 0 with k → n.

4.2. Mismatch case

In order to ensure stability of MPC under mismatch we essentially will
follow the arguments given by [49, 50]. The outcome of the approach will be
a criterion on the objective function value of problem (31) which when satis-
fied it will ensure its decrease along the computed trajectories and therefore

19
stability of the closed-loop system. In essence, the criterion will be employed
to show that approximations of the slow dynamics for system (9) can always
be found such that the system is closed-loop stable in the presence of the
plant/reduced order model mismatch.

Assumption 5. In what follows we assume that the optimal control problem

at every time index k is feasible in the sense that for any bounded initial
condition zk there exists a bounded set of control actions P∗k ≡ P∗ (zk ) which
minimizes (31a) subject to (31b) and (31c), being J(zk , P∗k ) the minimum of
1
the cost function for the initial condition zk .

In this way, for any pair of initial conditions z1,k and z2,k there exists a
positive constant αP such that:

kP∗ (z1,k ) − P∗ (z2,k )k22 6 αP kz1,k − z2,k k22 . (35)

Since function (28) is Lipschitz in its arguments we also have that:

|J (z1,k , P∗ (z1,k )) − J (z2,k , P∗ (z2,k ))| 6 αJ kz1,k − z2,k k22 . (36)

With these preliminary observations we are in the position to state next the
main robustness condition.

Proposition 2. A predictive controller that makes use of the solution to

problem (31) will be stable provided that, for each k, the following inequality
holds: Z
J(zk+1 , P∗k+1 ) − J(zA,k+1 , P∗A,k+1 ) 6ρ h(zk , p∗k )dξ , (37)
V

1
Note that the set of optimal control actions must be itself function of the initial
condition

20
where ρ ∈ [0, 1).
Moreover, at each k it is always possible to find a number N of slow modes
so to ensure that the above inequality (mismatch condition) holds.

Proof: To prove the first part we add and subtract J(zA,k+1 , P∗A,k+1 ) to the
LHS of (34) so that:

J(zk , P∗k ) − J(zk+1 , P∗k+1 ) =J(zk , P∗k ) − J(zA,k+1 , P∗A,k+1 )

− J(zk+1 , P∗k+1 ) − J(zA,k+1 , P∗A,k+1 ) . (38)

Note that since at time index k state measurements are available, zA,k ≡ zk .
Thus, applying (34) to the first term on the RHS of equation (38) leads to:
Z
∗ ∗
J(zk , Pk ) − J(zk+1 , Pk+1 ) > h(zk , p∗k )dξ
V


− J(zk+1 , P∗k+1 ) − J(zA,k+1 , P∗A,k+1 ) . (39)

Combining (37) with (39) we obtain:

Z
J(zk , P∗k ) − J(zk+1 , P∗k+1 ) > (1 − ρ) h(zk , p∗k )dξ . (40)
V

Since h(zk , p∗k ) is bounded from below by a quadratic function of the field
(Eqn (29)) we also have:

J(zk , P∗k ) − J(zk+1 , P∗k+1 ) > (1 − ρ)αh kzk k22 ,

what leads to limk→∞ kzk k22 = 0, since 1 − ρ > 0.

To prove the second part of the proposition we make use of (36) and the
orthogonality property of the eigenfunctions to show that the left hand side
of (37) is bounded as follows:
 
J(zk+1 , P∗k+1 ) − J(zA,k+1 , P∗A,k+1 ) 6 αJ kzB,k+1 k22 . (41)

21
According to proposition 1, kzB,k+1 k22 is bounded by its corresponding norm
at k as:

2β 2
kzB,k+1 (T )k22 6 kzB,k k22 exp (−εN T ) + (1 − exp (−εN T )) . (42)
ε2N

and the result follows since for any kzB,k k22 , εN → ∞ as N → ∞ so the right
hand side can be made arbitrarily small by increasing N . 4

5. The Multi-Model Predictive Control framework

We demonstrated in the previous section that approximations of the slow

dynamics can always be found such that the resulting system is closed-loop
stable in the presence of the plant/reduced order model mismatch. Those ap-
proximations, however, are not known a priori or even can change depending
on the range of operation. In this section, we describe the MMPC method-
ology to assure close-loop stability and illustrate the main steps with a flow
chart (Figure 1).
The basic idea consists of using a given system approximation of the
form of equation (31b) to compute the optimal solution of problem (31a).
Then, when the sufficient condition for stability described in Proposition 2
(Eq (37)) is not satisfied a new model of the form (30) is recomputed by
updating the POD basis. Let us rewrite (37) such that:
Z
∗ ∗
J(zk , Pk ) − J(zA,k , PA,k ) 6 ρ h(zk−1 , p∗k−1 )dξ . (43)
V

In order to check this condition, it is necessary to know at any time tk the

term in the RHS and the optimal cost function of two optimizations that we
denote here as “standard optimization” – since standard MPC also needs its

22
calculation – and “non-standard optimization”. The only difference between
both is the initial condition. In the former such initial condition is the current
measure of the field zk while in the “non-standard optimization” it is given
by zA,k which is to be estimated by using the previous field measurement
zk−1 and the approximation of the system.
In the case that (43) is not satisfied, we need to update the model, i.e., to
recalculate the set of PODs representative of the current range of operation.
To this purpose we make use of the last Nm measurements obtained from the
plant, i.e. {zk−Nm , ..., zk }. The technique to re-calculate the PODs is outlined
in Appendix A, and the criterion to choose the dimension of the POD set
is determined by the percentage of energy captured by the approximation.
It must be highlighted that considering different sets of snapshots will, in
general, lead to different sets of PODs. Therefore, the number of PODs nec-
essary to capture a given percentage of energy will depend on the considered
set of snapshots. A nominal criterion for the energy captured could be 99%
as proposed by [38]. A good selection of number of measurements, Nm , and
the energy captured by the PODs, E(%), could accelerate considerably the
computation of the optimal profiles. When Nm and E(%) are small, numer-
ous updates must be done. On the contrary, when these indices are large, few
updates of the model are required but its dimension is higher. Consequently
it leads to an increase of the computational effort to solve the optimization
problem.
Simulation studies with the system described in Section 6, indicate that
these parameters must be selected with a compromise between the quality of
the model approximation and the computational cost to solve the optimiza-

23
tion problem.
When implementing this technique, there are also other important as-
pects which will speed up substantially the calculations (see Figure 1). For
example, the RHS term required at time tk in (43) is implicitly calculated
by the “standard optimization” at time tk−1 . The schematic representation
of the algorithm in Figure 1 gives also the basic steps to follow taking into
account scenarios where numerical errors may result in unnecessary updates
of the model. Note that when the plant is close to the reference point, the
terms in (43) are close to zero and therefore very small values may trigger
unnecessary model updates. As it is illustrated in Section 6, this can be
avoided by permitting model updates only when the cost function value is
higher than a given tolerance , i.e.,

J(zk , P∗A,k ) >  . (44)

Finally, we recall that in this work we focus on nonlinear diffusion-convection-

reaction systems, and that their dissipative nature ensures the existence of a
reduced order representation capable of approximating at arbitrary accuracy
the real plant dynamics for a given operation region. Nevertheless, we empha-
size that it is not possible to define precisely the typical range of N values for
this class of systems. Previous works on this class of systems ([62, 32, 54, 53])
show that for most systems the range of the number of ODEs per field is not
larger than 102 (even for the most complex hydrodynamic systems in 2D or
3D) in order to obtain a reasonably accurate description of the dynamics. A
priori, this range of values renders NMPC practical. Among other aspects
to take into consideration regarding the practicability of NMPC are the for-
mulation, algorithm, and numerical implementation strategies. For example,

24
[27] report the application of their advanced-step NMPC to a distributed
system described by a DAE model with 350 states and differential-algebraic
equations. Thus, the adoption of the reduced order methodology described
in this work can contribute to the application of existing advanced NMPC
technology to distributed systems of higher dimension.

6. Case study

The performance of the multimodel MPC approach is demonstrated by sim-

ulation experiments on a non-isothermal tubular reactor with recycle. The
reactant concentration, C ≡ C(ξ, t), and the temperature inside the reactor,
T ≡ T (ξ, t), are described by the following set of PDEs [63, 64]:
∂C 1 ∂ 2 C ∂C
= − − f (C, T ) ,
∂t PeC ∂ξ 2 ∂ξ
(45)
∂T 1 ∂ 2T ∂T
= 2
− + BT f (C, T ) + βT (Tc − T ) ,
∂t PeT ∂ξ ∂ξ
where Tc is the control variable and represents the cooling jacket temperature
which is assumed to be uniform along the reactor. The nonlinear kinetic term
f (C, T ) is given by
!
γT
f (C, T ) = BC (1 + C) exp .
1+T
The reactant of the outlet stream is recycled to the reactor feed at a ratio r
which results into the following boundary conditions:
 h i

 ∂C
 = PeC (1 − r) C0 + r C(L, t) − C(0, t)
∂ξ
ξ=0: h i ,


 ∂T = PeT (1 − r) T0 + r T (L, t) − T (0, t)
∂ξ
∂C ∂T
ξ≡L=1: = = 0.
∂ξ ∂ξ

25
The values of the parameters are chosen from [63, 65]:

PeC = 7 , BC = 0.1 , γ = 10 ,

PeT = 7 , BT = 2.5 , βT = 2 .

Under the operating conditions C0 = T0 = 0, Tc = 0, and with a recycle ratio

r = 0.5, the reactor exhibits limit cycle behavior (Figure 2). The objective
of the MMPC strategy (Figure 1) is to stabilize the limit cycle around a
sp
given outlet concentration setpoint, Cout , using Tc as the manipulated input.
The objective function consists of a weighted quadratic cost function with
two terms. One term is defined as the square of the deviation between the
output concentration, Cout ≡ C(L, t), and the concentration setpoint, and
the other one penalizes the square of deviation of Tc with respect to a given
reference value Tcref . The optimal control problem becomes:
X−1 h
k+M
sp

2
2 i
ref
min. Jk = α Cout,i − Cout,i + β Tc,i − Tc,i (46a)
Tc
i=k
Z  
dmCA C 1 ∂ 2 C ∂C
s.t. = φA − − f (C, T ) dξ , (46b)
dt V PeC ∂ξ 2 ∂ξ
Z  
dmTA T 1 ∂ 2T ∂T
= φA − + BT f (C, T ) + βT (Tc − T ) dξ ,
dt V PeT ∂ξ 2 ∂ξ
(46c)
where α and β are the weighting parameters, and mxA = [mx1 , mx2 , ..., mxNx ]T ,
sp
φxA = [φx1 , φx2 , ..., φxNx ]T , x = [C, T ]. We consider in this example that Cout,k =
ref
−0.9 and Tc,k = −0.01 for every tk although other values (even a time depen-
sp
dent function) could be also considered for Cout,k . The weighting parameters
in (46a) are α = 100 and β = 100.
In this simulation experiment the real plant will consist of a finite ele-
ment implementation of model (45) using a uniform spatial grid. The degree

26
of discretization (31 nodes in this example) is selected such as further dis-
cretizations do not significantly alter the solution. The model employed to
approximate the plant behavior is constructed using the POD technique (see
(46b)-(46c)) where the number of slow modes in the approximation (NC , NT )
is computed using the energy criteria (see Appendix A for more details).
Some relevant aspects of the MMPC implementation are outlined next (see
also Figure 1):

1. At the beginning of the process operation there are no snapshots avail-

able to derive the POD basis. Hence, the reactor operates in open-loop
mode until t = 1, and several snapshots are taken during this period
of time. It should be noted that this is a limitation of the methodol-
ogy that can be mitigated by building an off-line reduced model based
on measurements of a similar plant or on data obtained from a model
available in the literature.
2. Using these snapshots the POD basis that captures 99.9 % of the energy
is computed and the ROM is This value of energy was selected to
compensate for the poor information available from the snapshots at
this point.
3. This ROM is employed to solve the nonlinear MPC problem (46a)
with a predictive horizon M = 6 and control horizon Mc = 2. It is
noted that long predictive horizons are required to satisfy Assumption
4. However, as it will be shown, short horizons can be used as well.
In our case study, off-line tests of the optimal control problem were
carried out to estimate such values.
4. New concentration and temperature measurements are taken at every

27
 sampling time, ∆t = 0.5 time units.
5. Before each new optimization problem, criteria (43) and (44) are checked.
In case they do not fulfill, new POD basis that captures 99.5 % of the
energy is computed using the snapshots obtained from the measure-
ments in step 4, and a new ROM is derived. Note that the value of
energy in this step is lower than that one considered in step 2 because
a richer information (snapshots) is available at this point. Also it was
selected larger than the one recommended in [38] because it ensures
a good representation of the plant behavior while keeping the model
dimensionality adequate for MPC purposes.
6. Steps 3-5 are repeated.

It must be noted that from a practical implementation point of view, the

solution of the optimal control problem should remain small as compared
to the process time constants for the controller to operate properly. This
implies a ratio `/v large enough as compared with the time to complete each
optimization2 . In our case, where optimizations where performed in the order
of 10 seconds, ratios of about 100 seconds should suffice.
First of all, in order to emphasize the importance of the updating pro-
cedure, a conventional MPC scheme (which employs a ROM of the form
(30) without model updating) was applied. It must be noted that since the
model employed in this scheme may not be accurate enough to represent the
oscillatory behavior of the plant (due to, for instance, not enough process
information is available), the conventional MPC scheme is neither able to

2
` and v correspond with physical reactor length and fluid velocity, respectively

28
target the concentration setpoint nor to stabilize the system. This is the
case illustrated in Figures 3 and 5. Alternatively a more representative sam-
pling could capture the precise behavior but at the expenses of exhaustive
off-line experiments to be carried out beforehand.
On the contrary, the proposed MMPC methodology is able to detect
through (43) and (44) when the current ROM is not accurate enough to
track the outlet concentration setpoint producing new representative POD
sets in the sense of ensuring controller stability. The circle marks in Figures
5 and 6 indicate when the model was updated using the new available state
measurements. Figures 4 and 5 show that the MMPC approach is able to
stabilize the system around the desired outlet concentration setpoint. Four
model updates were required, respectively at t = 3, 10, 12, and 12.5 units
of time. It is important to note that the ROM obtained with 99.5 % of the
energy always consisted of less than 6 ODEs per field.
Finally, the evolution of objective function (46a) as well as inequality
(43) are represented in Figure 6 illustrating that, for the MMPC approach,
asymptotic convergence is ensured. Besides, the evolution of the mismatch
term, expressed in terms of the L2 norm of the differences between the mea-
sured (real) states, y = [Cout , Tout ], and the predicted states, ŷ = [Ĉout , T̂out ])
is also represented in Figure 6. When applying the conventional MPC scheme
this term keeps oscillating whereas the MMPC approach is able to reduce it
to zero.

29
7. Conclusions

Robustness in MPC is at a high extent conditioned by the available model

of the plant on which the accuracy of predictions rely. It is well known that
although process-model mismatch can lead to serious control performance
deterioration, a certain amount can be tolerated in the sense that it will not
destroy the stability of the control loop. However the question is on how to
determine such a bound on the allowable amount of mismatch. In this work
we have shown that a positive answer to this question is possible at least
for dissipative systems as those described by convection-diffusion-reaction
mechanisms.
In that purpose, arguments similar as those proposed by [49, 50] have
been employed to demonstrate that it is always possible to choose within a
collection of reduced approximations (models) one that preserves the stabi-
lizing properties of MPC despite uncertainties and disturbances.
The practical implication of this evidence is a robust nonlinear model
predictive controller that makes use of a collection of reduced order approx-
imations of the plant (models) reconstructed on-line by projection methods
on POD (Proper Orthogonal Decomposition) basis functions. Since plant ap-
proximations are built on-line based on actual measurements the proposed
controller can be interpreted as a multi-model nonlinear MPC or MMPC.
The performance of the MMPC strategy has been illustrated by simula-
tion experiments on a problem that involves temperature and concentration
control of a tubular reactor with recycle. Extensions of this methodology to
more general classes of systems can be foreseen as long as their representa-
tions can be cast into complete sets.

30
Acknowledgments

This work has been also partially founded by the Spanish Ministry of Sci-
ence and Innovation (SMART-QC, AGL2008-05267-C03-01), the FP7 CAFE
project (KBBE-2007-1-212754), the Project PTDC/EQU-ESI/73458/2006
from the Portuguese Foundation for Science and Technology and PI grant
07/IN.1/I1838 by Science Foundation Ireland. Also, the authors acknowledge
financial support received by a collaborative grant GRICES-CSIC. Finally,
we want to thank the reviewers for their useful comments.

Appendix A. Reduced order models

Let us start the derivation of reduce order models (ROM) by expanding

the field into a Fourier series of the form (6):
∞
X
z(t, ξ) = mi (t)φi (ξ) (A.1)
i=1

where {φi (ξ)}∞

i=1 is the set of basis functions containing the spatial infor-

mation of the solution. The time information is encoded in the so-called

mode set {mi (t)}∞
i=1 . Each element φi (ξ) of the set of basis functions is com-

puted off-line as the solution of the following integral eigenvalue problem

[38, 63, 66]: Z
λi R(ξ, ξ 0 )φi (ξ 0 )dξ 0 = φi (ξ) (A.2)
V

where λi corresponds with the eigenvalue associated with each global eigen-
function φi . In this work, the proper orthogonal decomposition (POD) will
be selected as the technique to compute the basis functions due to its effi-
ciency [54]. In the POD, the kernel R(ξ, ξ 0 ) in equation (A.2) corresponds

31
with the two point spatial correlation function, defined as follows [54]:
1 X̀
R(ξ, ξ 0 ) = z(ξ, tj )z(ξ 0 , tj ). (A.3)
` j=1

where z(ξ, tj ) corresponds with the value of the field at each instant tj and the
summation extends over a sufficiently rich collection of uncorrelated snap-
shots at j = 1, · · · , ` [38]. The basis functions obtained by means of the
POD technique are also known as empirical basis functions or POD basis.
The dissipative nature of the kind of systems considered in this work
allows us to separate the system dynamics in two subsets [67, 68, 31]: one
stable with fast dynamics and infinite dimensional (zB ) and the other (zA )
composed by a finite number of elements with slow dynamics which may be
unstable and will capture the relevant dynamics of the system. The number
of elements (N ) in the stable subset is usually chosen using a criteria based
on the energy captured by the POD basis. Such energy is connected to the
eigenspectrum {λi }`i=1 or, to be more precise, to the inverse of the eigenvalues
µi = 1/λi as follows (for a deeper insight see [63, 38]):
N
X
µi
i=1
E(%) = 100 × (A.4)
X̀
µi
i=1

Note that µi → 0 as i → ∞ since, as pointed out in section 2.1, λi → ∞ as

i → ∞.
In this work, the numerical solution of Eqn (A.2) has been computed by
taking advantage of the finite element structure. A detailed explanation of
such procedure is provided in [69]. It is worth mentioning that the compu-
tational time for solving the eigenvalue problem (A.2) rapidly increase with

32
the number of discretization points. For most of 1D problems this should
not be an issue. However, when considering 2D or 3D systems, solving (A.2)
may render impractical for MPC purposes. In those cases, the method of
snapshots or strobes proposed in [38] can be employed.
Finally, in order to recover the field zA (ξ, t), the original PDE system (9)
is projected onto the slow basis set3 . As a result, the following set of ODEs
is obtained:
dmA
= AA mA + FA (A.5)
dt
   
T ∂ ∂φA ∂φA
where mA = [m1 , m2 , · · · , mN ] , AA = φA , k −v , FA =
∂x ∂x ∂x V
hφA , f (z) + p(x, t)iV and φA = [φ1 , φ2 , · · · , φN ]T .
It is important to highlight that, as shown in section 2, the number of
elements N in the slow subset zA can be increased to approximate the original
state z with an arbitrary degree of accuracy.

Appendix B. Boundedness of the convective term

In order to show the conditions under which the first derivative can be
expanded in convergent series of the form (13) let us start with a definition:

Definition 1 (Passive systems). A system is said to be passive if there exists

a function W(t) bounded from below so that
Z T +t
W(T + t) − W(t) > hy, piV ∀t, T > 0 (B.1)
t

with p and y being, respectively, the input and the output of the system.

3
R
The basis set has been previously normalized so that V
φi φi dξ = 1, ∀i = 1, .., N .

33
 1
Consider the quadratic function W(t) = 2
hz, ziV = 12 kzk22 . Multiplying
Eq (9) by z and integrating over the spatial coordinates, we obtain:
    
dW ∂ ∂z ∂z
= z, k − z, v + hz, f iV + hz, piV (B.2)
dt ∂ξ ∂ξ V ∂ξ V

The first two terms on the RHS of Eq (B.2) were proved to be non positive
in [55], therefore using Eq (5) with µ = 0 leads to:

dW
6 −L0 + hz, piV (B.3)
dt

Taking into account that L0 is a positive quantity, time integration of Eq

(B.3) yields: Z T
W(T ) 6 W(0) + hz, piV ds (B.4)
0

Note that this inequality implies that W(T ) is upper bounded for all T > 0
provided that the controls are also bounded so that hz, piV < ∞.
Let us now consider the diffusive term in Eq (B.2):
Z   Z   
∂ ∂z ∂z ∂z ∂ ∂z
zk dξ = k dξ + z, k
V ∂x ∂ξ V ∂ξ ∂ξ ∂ξ ∂ξ V

or, by means of the divergence theorem

Z   Z   
∂z ∂z ∂z ∂ ∂z
n zk dξ = k dξ + z, k
B ∂ξ V ∂ξ ∂ξ ∂ξ ∂ξ V

It should be noted that, making use of boundary condition (3), the LHS is
negative and we obtain
Z   
∂z ∂z ∂ ∂z
− k dξ > z, k ⇒
V ∂ξ ∂ξ ∂ξ ∂ξ V
2   
∂z ∂ ∂z
−k
∂ξ > z, ∂ξ k ∂ξ (B.5)
2 V

34
Substituting expression (B.5) into (B.2) and using the Lipschitz condition
(5) with µ = 0:
2  
dW ∂z ∂z

6 −k − z, v − L0 + hz, piV
dt ∂ξ 2 ∂ξ V
 
∂z
Taking into account that z, v > 0 – see Eqn (17) –
∂ξ V
2
dW ∂z
6 −k
∂ξ − L0 + hz, piV
dt 2

Finally, time integration of the former expression leads to:

Z T 2 Z T
∂z
k
∂ξ ds 6 W(0) − W(T ) + (−L0 + hz, piV )ds
0 2 0

Since the all the terms in the RHS are upper bounded – see Eq. (B.4) – we
conclude that the first derivative is bounded.

References

[1] J. Richalet, A. Rault, J. L. Testud, J. Papon, Model predictive heuristic

control: Applications to industrial processes, Automatica 14 (5) (1978)
413–428.

[2] C. R. Cutler, B. L. Ramaker, Dynamic matrix control - a computer algo-

rithm, in: 86th National Meeting of the American Institute of Chemical
Engineers, Houston, TX, 1979, 1979, pp. 13–15.

[3] S. Qin, T. Badgwell, A survey of industrial model predictive control

technology, Control Engineering Practice 11 (7) (2003) 733–764.

35
 [4] E. F. Camacho, C. Bordons, Model Predictive Control, 2nd Edition,
Advanced Textbooks in Control and Signal Processing, Springer-Verlag,
London, 2004.

[5] J. B. Rawlings, D. Q. Mayne, Model Predictive Control: Theory and

Design, Nob Hill Publishing, LLC, 2009.

[6] W. H. Ray, Multivariable process control - a survey, Computers & Chem-

ical Engineering 7 (4) (1983) 367 – 394.

[7] C. Garcı́a, D. Prett, M. Morari, Model predictive control: Theory and

practice a survey, Automatica 25 (3) (1989) 335–348.

[8] M. Morari, J. H. Lee, Model predictive control: Past, present and future,
Computers & Chemical Engineering 23 (4-5) (1999) 667–682.

[9] A. Bemporad, M. Morari, Robust model predictive control: A survey,

in: A. Tesi, A. Vicino (Eds.), Robustness in Identification and Control,
no. 245 in Lecture Notes in Control and Information Sciences, Springer-
Verlag, 1999, pp. 207–226.

[10] J. Rawlings, Tutorial overview of model predictive control, Control Sys-

tems Magazine, IEEE 20 (3) (2000) 38–52.

[11] B. Bequette, Nonlinear control of chemical processes: A review, Indus-

trial & Engineering Chemistry Research 30 (7) (1991) 1391–1413.

[12] M. Henson, Nonlinear model predictive control: current status and fu-
ture directions, Computers and Chemical Engineering 23 (2) (1998) 187–
202.

36
[13] M. Cannon, Efficient nonlinear model predictive control algorithms, An-
nual Reviews in Control 28 (2) (2004) 229–237.

[14] L. Magni, D. M. Raimondo, F. Allgöwer (Eds.), Nonlinear Model Pre-

dictive Control – Towards New Challenging Applications, Vol. 384 of
Lecture Notes in Control and Information Sciences, Springer, 2009.

[15] S. Qin, T. Badgwell, An overview of nonlinear model predictive control

applications, Nonlinear model predictive control (2000) 369–392.

[16] D. Q. Mayne, J. B. Rawlings, C. V. Rao, P. O. M. Scokaert, Constrained

model predictive control: Optimality and stability, Automatica 36 (6)
(2000) 789–814.

[17] D. Limon, T. Alamo, F. Salas, E. Camacho, Input to state stability of

min-max mpc controllers for nonlinear systems with bounded uncertain-
ties, Automatica 42 (5) (2006) 797–803.

[18] D. Limon, I. Alvarado, T. Alamo, E. Camacho, Robust tube-based mpc

for tracking of constrained linear systems with additive disturbances,
Journal of Process Control 20 (3) (2010) 248–260.

[19] A. A. Patwardhan, G. T. Wright, T. F. Edgar, Nonlinear model-

predictive control of distributed-parameter systems, Chemical Engineer-
ing Science 47 (4) (1992) 721 – 735.

[20] G. T. Wright, T. F. Edgar, Nonlinear model predictive control of a

fixed-bed water-gas shift reactor: An experimental study, Computers &
Chemical Engineering 18 (2) (1994) 83–102.

37
[21] N. Padhiyar, S. Bhartiya, Profile control in distributed parameter sys-
tems using lexicographic optimization based MPC, Journal of Process
Control 19 (1) (2009) 100 – 109.

[22] P. Dufour, Y. Touré, D. Blanc, P. Laurent, On nonlinear distributed

parameter model predictive control strategy: on-line calculation time
reduction and application to an experimental drying process, Computers
& Chemical Engineering 27 (11) (2003) 1533 – 1542.

[23] P. Dufour, Y. Touré, Multivariable model predictive control of a cat-

alytic reverse flow reactor, Computers & Chemical Engineering 28 (11)
(2004) 2259 – 2270.

[24] J. De Temmerman, P. Dufour, B. Nicolaı̈, H. Ramon, MPC as control

strategy for pasta drying processes, Computers & Chemical Engineering
33 (1) (2009) 50 – 57.

[25] E. Aggelogiannaki, H. Sarimveis, D. Koubogiannis, Model predictive

control in long ducts by means of a neural network approximation tool,
Applied Thermal Engineering 27 (14-15) (2007) 2363 – 2369.

[26] E. Aggelogiannaki, H. Sarimveis, Nonlinear model predictive control for

distributed parameter systems using data driven artificial neural net-
work models, Computers & Chemical Engineering 32 (6) (2008) 1225–
1237.

[27] V. M. Zavala, L. T. Biegler, Optimization-based strategies for the op-

eration of low-density polyethylene tubular reactors: nonlinear model

38
 predictive control, Computers & Chemical Engineering 33 (10) (2009)
1735 – 1746.

[28] C. Antoniades, P. D. Christofides, Integrated optimal actuator/sensor

placement and robust control of uncertain transport-reaction processes,
Computers & Chemical Engineering 26 (2) (2002) 187–203.

[29] J. A. Burns, B. B. King, A reduced basis approach to the design of low-

order feedback controllers for nonlinear continuous systems, Journal of
Vibration and Control 4 (3) (1998) 297 – 323.

[30] C. W. Gear, I. G. Kevrekidis, C. Theodoropoulos, “coarse” integra-

tion/bifurcation analysis via microscopic simulators: micro-galerkin
methods, Computers & Chemical Engineering 26 (7-8) (2002) 941–963.

[31] A. A. Alonso, I. G. Kevrekidis, J. R. Banga, C. E. Frouzakis, Optimal

sensor location and reduced order observer design for distributed process
systems, Computers and Chemical Engineering 28 (1-2) (2004) 27–35.

[32] A. A. Alonso, C. V. Fernández, J. R. Banga, Dissipative systems: from

physics to robust nonlinear control, Int. J. Robust Nonlinear Control
14 (2) (2004) 157–179.

[33] D. Zheng, K. A. Hoo, Low-order model identification for implementable

control solutions of distributed parameter systems, Computers & Chem-
ical Engineering 26 (7-8) (2002) 1049–1076.

[34] D. Zheng, K. A. Hoo, System identification and model-based control

for distributed parameter systems, Computers & Chemical Engineering
28 (8) (2004) 1361–1375.

39
[35] H. Shang, J. F. Forbes, M. Guay, Model predictive control for quasilinear
hyperbolic distributed parameter systems, Industrial and Engineering
Chemistry Research 43 (9) (2004) 2140–2149.

[36] S. Dubljevic, P. D. Christofides, Predictive control of parabolic PDEs

with boundary control actuation, Chemical Engineering Science 61 (18)
(2006) 6239 – 6248.

[37] S. Dubljevic, Model predictive control of kuramoto-sivashinsky equation

with state and input constraints, Chemical Engineering Science 65 (15)
(2010) 4388 – 4396.

[38] L. Sirovich, Turbulence and the dynamics of coherent structures. Part

I: Coherent structures, Quaterly of Appl. Math. 45 (3) (1987) 561–571.

[39] G. Berkooz, P. Holmes, L. Lumley, The Proper Orthogonal Decompo-

sition in the analysis of turbulent flows, Ann. Rev. Fluid Mech. 25 (1)
(1993) 539–575.

[40] P. Holmes, J. L. Lumley, G. Berkooz, Turbulence, Coherent Structures,

Dynamical Systems and Symmetry., Cambridge University Press, 1996.

[41] P. J. Holmes, J. L. Lumley, G. Berkooz, J. C. Mattingly, R. W. Wit-

tenberg, Low-dimensional models of coherent structures in turbulence,
Physics Reports 287 (4) (1997) 337–384.

[42] R. Padhi, S. F. Ali, An account of chronological developments in control

of distributed parameter systems, Annual Reviews in Control 33 (1)
(2009) 59–68.

40
[43] W. Xie, C. Theodoropoulos, An off-line model reduction-based tech-
nique for on-line linear mpc applications for nonlinear large-scale dis-
tributed systems, Computer Aided Chemical Engineering 28 (2010) 409–
414.

[44] S. Shvartsman, C. Theodoropoulos, R. Rico-Martinez, I. Kevrekidis,

E. Titi, T. Mountziaris, Order reduction for nonlinear dynamic models of
distributed reacting systems, Journal of Process Control 10 (2-3) (2000)
177–184.

[45] C. Theodoropoulos, Optimisation and linear control of large scale non-

linear systems: a review and a suite of model reduction-based tech-
niques, Coping with Complexity: Model Reduction and Data Analysis
(2011) 37–61.

[46] K. Veroy, A. Patera, Certified real-time solution of the parametrized

steady incompressible navier–stokes equations: rigorous reduced-basis a
posteriori error bounds, International Journal for Numerical Methods in
Fluids 47 (8-9) (2005) 773–788.

[47] J. Atwell, B. King, Proper orthogonal decomposition for reduced basis

feedback controllers for parabolic equations, Mathematical and com-
puter modelling 33 (1-3) (2001) 1–19.

[48] J. Atwell, B. King, Reduced order controllers for spatially distributed

systems via proper orthogonal decomposition, SIAM Journal on Scien-
tific Computing 26 (2004) 128.

41
[49] L. O. Santos, L. T. Biegler, A tool to analyze robust stability for model
predictive controllers, Journal of Process Control 9 (11) (1999) 233–246.

[50] L. Santos, L. Biegler, J. Castro, A tool to analyze robust stability for

constrained nonlinear mpc, Journal of Process Control 18 (3-4) (2008)
383–390.

[51] M. R. Garcı́a, Identification and real time optimization in the

food processing and biotechnology industries, Ph.D. thesis, Univer-
sity of Vigo ”http://digital.csic.es/bitstream/10261/4662/1/
MiriamRGarcia_thesis.pdf” (2008).

[52] M. R. Garcı́a, C. Vilas, L. O. Santos, A. A. Alonso, A robust and sta-

bilizing multi-model predictive control approach to command the oper-
ation of distributed process systems, in: AIChE 2006 Anual Meeting,
San Francisco, CA, 2006, p. 455b.

[53] C. Vilas, A. V. Wouwer, Combination of multi-model predictive control

and the wave theory for the control of simulated moving bed plants,
Chemical Engineering Science 66 (4) (2011) 632–641.

[54] M. R. Garcı́a, C. Vilas, J. R. Banga, A. A. Alonso, Optimal field re-

construction of distributed process systems from partial measurements,
Industrial & Engineering Chemistry Research 46 (2) (2007) 530–539.

[55] C. Vilas, M. R. Garcı́a, J. R. Banga, A. A. Alonso, Robust feed-back

control of distributed chemical reaction systems, Chemical Engineering
Science 62 (11) (2007) 2941–2957.

42
[56] A. A. Alonso, B. E. Ydstie, Stabilization of distributed systems using
irreversible thermodynamics, Automatica 37 (11) (2001) 1739–1755.

[57] B. Reddy, Introductory functional analysis: with applications to bound-

ary value problems and finite elements, Springer Verlag, 1998.

[58] J. Smoller, Shock Waves and Reaction-Diffusion Equations, 2nd Edition,

Springer-Verlag, New york, 1994.

[59] E. Sontag, The ISS philosophy as a unifying framework for stability-like

behavior, Vol. 259 of Lecture Notes in Control and Information Sciences,
CNRS, 2001, pp. 443–467.

[60] S. S. Keerthi, E. G. Gilbert, Optimal infinite-horizon feedback laws for

general class of constrained discrete-time systems: Stability and moving-
horizon approximations, Journal of Optimization Theory and Applica-
tions 57 (2) (1988) 265–293.

[61] J. B. Rawlings, K. R. Muske, The stability of constrained receding hori-

zon control, Automatic Control, IEEE Transactions on 38 (10) (1993)
1512–1516.

[62] E. Balsa-Canto, A. A. Alonso, J. R. Banga, Reduced-order models for

nonlinear distributed process systems and their application in dynamic
optimization, Industrial & Engineering Chemistry Research 43 (13)
(2004) 3353–3363.

[63] A. A. Alonso, C. E. Frouzakis, I. G. Kevrekidis, Optimal sensor place-

ment for state reconstruction of distributed process systems, AIChE
Journal 50 (7) (2004) 1438–1452.

43
[64] C. Antoniades, P. D. Christofides, Integrating nonlinear output feedback
control and optimal actuator/sensor placement for transport-reaction
processes, Chemical Engineering Science 56 (15) (2001) 4517–4535.

[65] C. Antoniades, P. D. Christofides, Studies on nonlinear dynamics and

control of a tubular reactor with recycle, Nonlinear Analysis 47 (9)
(2001) 5933–5944.

[66] S. S. Ravindran, A reduced-order approach for optimal control of fluids

using proper orthogonal decomposition, International journal for numer-
ical methods in fluids 34 (5) (2000) 425–448.

[67] C. Antoniades, P. D. Christofides, Computation of optimal actuator

locations for nonlinear controllers in transport reaction processes, Comp.
Chem. Eng. 24 (2-7) (2000) 577–583.

[68] P. D. Christofides, Nonlinear and Robust Control of PDE Systems:

Methods and Applications to Transport-Reaction Processes, Birkhäuser,
Boston, 2001.

[69] M. Garcı́a, C. Vilas, J. Banga, A. Alonso, Exponential observers for

distributed tubular (bio) reactors, AIChE Journal 54 (11) (2008) 2943–
2956.

44
 Plant (Distributed process)

zk
kth “standard opti-
mization” J(zk , P∗k )
P∗k

Store P∗k ,
k > 1 and No R kth “standard opti-
h(zk , p∗k )dξ
J(zk , P∗A,k ) >  V
mization” J(zk , P∗k )
and zk−Nm , ..., zk

Load Yes Load

R
Yes V
h(zk−1 , p∗k−1 )dξ zk−Nm , ..., zk

kth “non-standard
check sufficient No Update PODs
optimization”
condition (43) (see Appendix A)
J(zA,k , P∗A,k )

Figure 1: Illustrative flow chart of the MMPC algorithm.

45
C(ξ, t)

0.0

-0.2

-0.4
0.0
-0.6
0.2
-0.8
0.4
-1.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0
T (ξ, t)

1.6

1.2

0.8 0.0

0.2
0.4
0.4
0.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0

Figure 2: Open-loop concentration and temperature responses.

46
 C(ξ, t)

0.0

-0.2

-0.4
0.0
-0.6
0.2
-0.8
0.4
-1.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0
T (ξ, t)

1.6

1.2

0.8 0.0

0.2
0.4
0.4
0.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0

Figure 3: Closed-loop concentration and temperature responses with a conventional MPC

strategy using a ROM that captures 99.9 % of the energy.

47
 C(ξ, t)

0.0

-0.2

-0.4
0.0
-0.6
0.2
-0.8
0.4
-1.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0
T (ξ, t)

1.6

1.2

0.8 0.0

0.2
0.4
0.4
0.0
0.6 ξ
0
5
10 0.8
15
t 20
25 1.0

Figure 4: Closed-loop concentration and temperature responses with the MMPC strategy
using updated ROMs that capture 99.5 % of the energy.

48
 0.0
a)
-0.2 b)
-0.4 Csp
Cout

MMPC model update

-0.6
-0.8
-1.0
0 5 10 15 20 25

1.6 a)
b)
1.2 MMPC model update
Tout

0.8

0.4

0.0
0 5 10 15 20 25

0.10
a) b)
MMPC model update
0.05
Tc

0.00

-0.05

-0.10
0 5 10 15 20 25
t

Figure 5: Outlet reactor concentration and temperature closed-loop responses, and con-
troller actuation: a) conventional MPC; b) MMPC strategy. The dots on the MMPC
profiles indicate when a model update occurred.

49
 a) b)
Jk∗

MMPC model update

1.6
Cost function ,

1.2

0.8

0.4

0.0
0 5 10 15 20 25

1.5
k−1 )dξ

1.0 a) b) MMPC model update

(h∗

0.5
V
R

0.0
Jk∗ − JA,k −

-0.5
∗

-1.0
-1.5
0 5 10 15 20 25

0.15
a) b) MMPC model update
2

yk − ŷk

0.10

0.05

0.00
0 5 10 15 20 25
t

Figure 6: Evolution of the optimal cost function in (46a), inequality (43), and of the
states plant/model mismatch: a) conventional MPC; b) MMPC strategy. The dots on the
MMPC profiles indicate when a model update occurred.

50