An Introduction to Nonlinear Model Predictive Control

Rolf Findeisen, Frank Allgöwer,

Institute for Systems Theory in Engineering,
University of Stuttgart, 70550 Stuttgart, Germany,
findeise,allgower
@ist.uni-stuttgart.de

Abstract
While linear model predictive control is popular since the 70s of the past century, the 90s have witnessed a steadily
increasing attention from control theoretists as well as control practitioners in the area of nonlinear model predictive
control (NMPC). The practical interest is driven by the fact that today’s processes need to be operated under tighter
performance specifications. At the same time more and more constraints, stemming for example from environmental
and safety considerations, need to be satisfied. Often these demands can only be met when process nonlinearities and
constraints are explicitly considered in the controller. Nonlinear predictive control, the extension of well established
linear predictive control to the nonlinear world, appears to be a well suited approach for this kind of problems. In this
note the basic principle of NMPC is reviewed, the key advantages/disadvantages of NMPC are outlined and some
of the theoretical, computational, and implementational aspects of NMPC are discussed. Furthermore, some of the
currently open questions in the area of NMPC are outlined.

1 Principles, Mathematical Formulation and Properties of

Nonlinear Model Predictive Control
Model predictive control (MPC), also referred to as moving horizon control or receding horizon control, has become
an attractive feedback strategy, especially for linear processes. Linear MPC refers to a family of MPC schemes in
which linear models are used to predict the system dynamics, even though the dynamics of the closed-loop system is
nonlinear due to the presence of constraints. Linear MPC approaches have found successful applications, especially in
the process industries. A good overview of industrial linear MPC techniques can be found in [64, 65], where more than
2200 applications in a very wide range from chemicals to aerospace industries are summarized. By now, linear MPC
theory is quite mature. Important issues such as online computation, the interplay between modeling/identification
and control and system theoretic issues like stability are well addressed [41, 52, 58].
Many systems are, however, in general inherently nonlinear. This, together with higher product quality specifications
and increasing productivity demands, tighter environmental regulations and demanding economical considerations in
the process industry require to operate systems closer to the boundary of the admissible operating region. In these
cases, linear models are often inadequate to describe the process dynamics and nonlinear models have to be used. This
motivates the use of nonlinear model predictive control.
This paper focuses on the application of model predictive control techniques to nonlinear systems. It provides a review
of the main principles underlying NMPC and outlines the key advantages/disadvantages of NMPC and some of the
theoretical, computational, and implementational aspects. Note, however, that it is not intended as a complete review
of existing NMPC techniques. Instead we refer to the following list for some excellent reviews [4, 16, 22, 52, 58, 68].
In Section 1.1 and Section 1.2 the basic underlying concept of NMPC is introduced. In Section 2 some of the system
theoretical aspects of NMPC are presented. After an outline of NMPC schemes that achieve stability one particular
NMPC formulation, namely quasi-infinite horizon NMPC (QIH-NMPC) is outlined to exemplify the basic ideas to
achieve stability. This approach allows a (computationally) efficient formulation of NMPC while guaranteeing stability
and performance of the closed-loop.
Besides the basic question of the stability of the closed-loop, questions such as robust formulations of NMPC and
some remarks on the performance of the closed-loop are given in Section 2.3 and Section 2.2. Section 2.4 gives some
remarks on the output-feedback problem in connection with NMPC. After a short review of existing approaches one

1
specific scheme to achieve output-feedback NMPC using high-gain observers for state recovery is outlined. Section 3
contains some remarks and descriptions concerning the numerical solution of the open-loop optimal control problem.
The applicability of NMPC to real processes is shown in Section 4 considering the control of a high purity distillation
column. This shows, that using well suited optimization strategies together with the QIH-NMPC scheme allow real-
time application of NMPC even with todays computing power. Final conclusions and remarks on future research
directions are given in Section 5.
In the following,  denotes the Euclidean vector norm in  n (where the dimension n follows from context) or the
associated induced matrix norm. Vectors are denoted by boldface symbols. Whenever a semicolon “;” occurs in a
function argument, the following symbols should be viewed as additional parameters, i.e. f  x; γ  means the value of
the function f at x with the parameter γ.

1.1 The Principle of Nonlinear Model Predictive Control

In general, the model predictive control problem is formulated as solving on-line a finite horizon open-loop optimal
control problem subject to system dynamics and constraints involving states and controls. Figure 1 shows the basic
principle of model predictive control. Based on measurements obtained at time t, the controller predicts the future

past future/prediction

set-point

predicted state x̄

closed-loop
state x open loop input ū

closed-loop
input u
t t δ t Tc t Tp
control horizon Tc
prediction horizon Tp

Figure 1: Principle of model predictive control.

dynamic behavior of the system over a prediction horizon Tp and determines (over a control horizon Tc
Tp ) the input
such that a predetermined open-loop performance objective functional is optimized. If there were no disturbances and
no model-plant mismatch, and if the optimization problem could be solved for infinite horizons, then one could apply
the input function found at time t  0 to the system for all times t  0. However, this is not possible in general. Due to
disturbances and model-plant mismatch, the true system behavior is different from the predicted behavior. In order to
incorporate some feedback mechanism, the open-loop manipulated input function obtained will be implemented only
until the next measurement becomes available. The time difference between the recalculation/measurements can vary,
however often it is assumed to be fixed, i.e the measurement will take place every δ sampling time-units. Using the
new measurement at time t δ, the whole procedure – prediction and optimization – is repeated to find a new input
function with the control and prediction horizons moving forward.
Notice, that in Figure 1 the input is depicted as arbitrary function of time. As shown in Section 3, for numerical
solutions of the open-loop optimal control problem it is often necessary to parameterize the input in an appropriate
way. This is normally done by using a finite number of basis functions, e.g. the input could be approximated as
piecewise constant over the sampling time δ.
As will be shown, the calculation of the applied input based on the predicted system behavior allows the inclusion
of constraints on states and inputs as well as the optimization of a given cost function. However, since in general

2
the predicted system behavior will differ from the closed-loop one, precaution must be taken to achieve closed-loop
stability.

1.2 Mathematical Formulation of NMPC

We consider the stabilization problem for a class of systems described by the following nonlinear set of differential
equations1

ẋ  t  f  x  t  u  t  x  0  x0 (1)

subject to input and state constraints of the form:

u  t  t  0 x  t  t  0 (2)

where x  t  ! n and u  t "# m denotes the vector of states and inputs, respectively. The set of feasible input
values is denoted by  and the set of feasible states is denoted by  . We assume that  and  satisfy the following
assumptions:
Assumption 1 %$& p is compact, '( n is connected and  0  0 )+*, .
In its simplest form,  and  are given by box constraints of the form:

 : .- u & m / umin
u
umax 0  (3a)
 : .- x ( n / xmin
x
xmax 01 (3b)

Here umin , umax and xmin , xmax are given constant vectors.
With respect to the system we additionally assume, that:
Assumption 2 The vector field f :  n *& m 2  n is continuous and satisfies f  0  0 3 0. In addition, it is locally
Lipschitz continuous in x.
Assumption 3 The system (1) has an unique continuous solution for any initial condition in the region of interest and
any piecewise continuous and right continuous input function u 45 : 6 0  Tp 7 2  .

In order to distinguish clearly between the real system and the system model used to predict the future “within” the
controller, we denote the internal variables in the controller by a bar (for example x̄  ū).
Usually, the finite horizon open-loop optimal control problem described above is mathematically formulated as fol-
lows:
Problem 1 Find min J  x  t < ū => ; Tc  Tp 
ū 8:9 ;
with
J  x  t < ū => ; Tp  Tc  : @?
A
t Tp
F  x̄  τ < ū  τB dτ (4)
t

subject to:

x̄˙  τ  f  x̄  τ  ū  τ B) x̄  t  x  t  (5a)

ū  τ CDE τ D6 t  t Tc 7 (5b)
ū  τ F ū  τ Tc G τ D6 t Tc  t Tp 7 (5c)
x̄  τ E τ D6 t  t Tp 7 (5d)
where Tp and Tc are the prediction and the control horizon with Tc
Tp .
The bar denotes internal controller variables and x̄ 45 is the solution of (5a) driven by the input ū 45 : 6 t  t Tp 7 2 
with initial condition x  t  . The distinction between the real system and the variables in the controller is necessary,
since the predicted values, even in the nominal undisturbed case, need not, and in generally will not, be the same as the
actual closed-loop values, since the optimal input is recalculated (over a moving finite horizon Tc ) at every sampling
1 In this paper only the continuous time formulation of NMPC is considered. However, notice that most of the presented topics have dual

counterparts in the discrete time setting.

3
instance.
The function F, in the following called stage cost, specifies the desired control performance that can arise, for example,
from economical and ecological considerations. The standard quadratic form is the simplest and most often used one:
F  x  u H x I xs  T Q  x I xs  J u I us  T R  u I us K (6)

where xs and us denote given setpoints; Q and R denote positive definite, symmetric weighting matrices. In order for
the desired reference  xs  us  to be a feasible solution of Problem 1, us should be contained in the interior of  . As
already stated in Assumption 2 we consider, without loss of generality that  xs  us LM 0  0  is the steady state that
should be stabilized. Note the initial condition in (5a): The system model used to predict the future in the controller is
initialized by the actual system state; thus they are assumed to be measured or must be estimated. Equation (5c) is not
a constraint but implies that beyond the control horizon the predicted control takes a constant value equal to that at the
last step of the control horizon.
In the following an optimal solution to the optimization problem (existence assumed) is denoted by ū NO4 ; x  t < Tp  Tc  :
6 t  t Tp 7 2  . The open-loop optimal control problem will be solved repeatedly at the sampling instances t 
jδ  j  0  1 BB , once new measurements are available . The closed-loop control is defined by the optimal solution of
Problem 1 at the sampling instants:
u N  τ  :  ū N  τ; x  t < Tp  Tc  τ C6 t  δ7 1 (7)

The optimal value of the NMPC open-loop optimal control problem as a function of the state will be denoted in the
following as value function:
V  x; Tp  Tc  J  x  ūPQ4 ; x  t  ; Tp  Tc  1 (8)
The value function plays an important role in the proof of the stability of various NMPC schemes, as it serves as a
Lyapunov function candidate.

1.3 Properties, Advantages, and Disadvantages of NMPC

In general one would like to use an infinite prediction and control horizon, i.e. Tp and Tc in Problem 1 are set to ∞.
5case) to minimize the performance objective determined by the cost However as mentioned, the open-loop optimal
control Problem 1, that must be solved on-line, is often formulated in a finite horizon manner and the input function
is parameterized finitely, in order to allow a (real-time) numerical solution of the nonlinear open-loop optimal control
problem. It is clear, that the shorter the horizon, the less costly the solution of the on-line optimization problem. Thus
it is desirable from a computational point of view to implement MPC schemes using short horizons. However, when
a finite prediction horizon is used, the actual closed-loop input and state trajectories will differ from the predicted
open-loop trajectories, even if no model plant mismatch and no disturbances are present [4]. This fact is depicted in
Figure 2 where the system can only move inside the shaded area as state constraints of the form x  τ R are assumed.
This makes the key difference between standard control strategies, where the feedback law is obtained a priori and

x2
S UV x̄ S 0U
x0
S T δU V x S δU
x̄ 0

S T
x̄ 0 Tp U
W S T
x̄ δ Tp U
x1

Figure 2: The difference between open-loop prediction and closed-loop behavior.

NMPC where the feedback law is obtained on-line and has two immediate consequences. Firstly, the actual goal to
compute a feedback such that the performance objective over the infinite horizon of the closed loop is minimized is not
achieved. In general it is by no means true that a repeated minimization over a finite horizon objective in a receding
horizon manner leads to an optimal solution for the infinite horizon problem (with the same stage cost F) [10]. In fact,
the two solutions differ significantly if a short horizon is chosen. Secondly, if the predicted and the actual trajectories

4
differ, there is no guarantee that the closed-loop system will be stable. It is indeed easy to construct examples for
which the closed-loop becomes unstable if a (small) finite horizon is chosen. Hence, when using finite horizons in
standard NMPC, the stage cost cannot be chosen simply based on the desired physical objectives.
The overall basic structure of a NMPC control loop is depicted in Figure 3. As can be seen, it is necessary to estimate

NMPC controller

dynamic u y
Plant
optimizer

cost function x̂
+ system model state estimator
constraints

Figure 3: Basic NMPC control loop.

the system states from the output measurements.

Summarizing the basic NMPC scheme works as follows:
1. obtain measurements/estimates of the states of the system
2. compute an optimal input signal by minimizing a given cost function over a certain prediction horizon in the
future using a model of the system
3. implement the first part of the optimal input signal until new measurements/estimates of the state are avail-
able
4. continue with 1.
From the remarks given so far and from the basic NMPC setup, one can extract the following key characteristics of
NMPC:
X NMPC allows the use of a nonlinear model for prediction.
X NMPC allows the explicit consideration of state and input constraints.
X In NMPC a specified performance criteria is minimized on-line.
X In NMPC the predicted behavior is in general different from the closed loop behavior.
X The on-line solution of an open-loop optimal control problem is necessary for the application of NMPC.
X To perform the prediction the system states must be measured or estimated.
In the remaining sections various aspects of NMPC regarding these properties will be discussed. The next section
focuses on system theoretical aspects of NMPC. Especially the questions on closed-loop stability, robustness and the
output feedback problem are considered.

2 System Theoretical Aspects of NMPC

In this section different system theoretical aspects of NMPC are considered. Besides the question of nominal stability
of the closed-loop, which can be considered as somehow mature today, remarks on robust NMPC strategies as well as
the output-feedback problem are given.

5
2.1 Stability
One of the key questions in NMPC is certainly, whether a finite horizon NMPC strategy does lead to stability of
the closed-loop. As pointed out, the key problem with a finite prediction and control horizon stems from the fact
that the predicted open and the resulting closed-loop behavior is in general different. Ideally one would seek for a
NMPC strategy that achieves closed-loop stability independent of the choice of the performance parameters in the cost
functional and, if possible, approximates the infinite horizon NMPC scheme as good as possible. A NMPC strategy
that achieves closed-loop stability independent of the choice of the performance parameters is usually referred to a
NMPC approach with guaranteed stability. Different possibilities to achieve closed-loop stability for NMPC using
finite horizon length have been proposed. After giving a short review about these approaches we exemplary present
on specific approach that achieves guaranteed stabilit, the so called quasi-infinite horizon approach to NMPC (QIH-
NMPC) . This approach achieves guaranteed closed loop stability while being computationally feasible.
Here only the key ideas are reviewed and no detailed proofs are given. Furthermore notice, that we will not cover all
existing NMPC approaches, instead we refer the reader to the overview papers [4, 22, 52].
For all the following sections it is assumed that the prediction horizon is set equal to the control horizon, Tp  Tc .

2.1.1 Infinite Horizon NMPC

The most intuitive way to achieve stability is the use of an infinite horizon cost [10, 39, 54], i.e. Tp in Problem 1 is set to
∞. In the nominal case feasibility at one sampling instance also implies feasibility and optimality at the next sampling
instance. This follows from Bellman‘s Principle of Optimality [7], i.e. the input and state trajectories computed as
the solution of the NMPC optimization Problem 1 at a specific instance in time, are in fact equal to the closed-loop
trajectories of the nonlinear system, i.e. the remaining parts of the trajectories after one sampling instance are the
optimal solution at the next sampling instance. This fact also implies closed-loop stability.
Key ideas of the stability proof: Since nearly all stability proofs for NMPC follow along the same basic steps as for
the infinite horizon proof, the key ideas are shortly outlined. In principle the proof is based on the use of the value
function as a Lyapunov function. First it is shown, that feasibility at one sampling instance does imply feasibility at
the next sampling instance for the nominal case. In a second step it is established that the value function is strictly
decreasing and by this the state and input converge to the origin. Utilizing the continuity of the value function at the
origin and the monotonicity property, asymptotic stability is established in the third step. As feasibility thus implies
asymptotic stability, the set of all states, for which the open-loop optimal control problem has a solution does belong
to the region of attraction of the origin.

2.1.2 Finite Horizon NMPC Schemes with Guaranteed Stability

Different possibilities to achieve closed-loop stability for NMPC using a finite horizon length have been proposed,
see for example [3, 17, 20, 23, 34, 35, 38, 39, 44, 51, 53, 55, 56, 60, 62, 63, 70]. Most of these approaches modify the
NMPC setup such that stability of the closed-loop can be guaranteed independently of the plant and performance
specifications. This is usually achieved by adding suitable equality or inequality constraints and suitable additional
penalty terms to the cost functional. These additional constraints are usually not motivated by physical restrictions or
desired performance requirements but have the sole purpose to enforce stability of the closed-loop. Therefore, they
are usually termed stability constraints [49, 50, 52].
The simplest possibility to enforce stability with a finite prediction horizon is to add a so called zero terminal equality
constraint at the end of the prediction horizon [39, 51, 53], i.e. to add the equality constraint

x̄  t Tp; x  t < t  ū  0 (9)

to Problem 1. This leads to stability of the closed-loop, if the optimal control problem possesses a solution at t  0,
since the feasibility at one time instance does also lead to feasibility at the following time instances and a decrease
in the value function. One disadvantage of a zero terminal constraint is that the system must be brought to the origin
in finite time. This leads in general to feasibility problems for short prediction/control horizon lengths, i.e. a small
region of attraction. Additionally, from a computational point of view, an exact satisfaction of a zero terminal equality
constraint does require an infinite number of iterations in the nonlinear programming problem [17]. On the other hand,
the main advantages are the straightforward application and the conceptual simplicity.

6
Many schemes have been proposed (i.e. [17, 20, 34, 38, 51, 56, 60, 63]), that try to overcome the use of a zero terminal
constraint of the form (9). Most of them either use a so called terminal region constraint
x̄  t Tp Y Ω F (10)
and/or a terminal penalty term E  x̄  t Tp B which is added to the cost functional:
A
t Tp
J  x  t < ū => ; Tp Z [ F  x̄  τ < ū  τ  dτ E  x̄  t Tp  1 (11)
t

Note that the terminal penalty term is not a performance specification that can be chosen freely. Rather E and the
terminal region Ω in (10) are determined off-line such that stability is “enforced”. We do not review all these methods
here. Instead we exemplify the basic idea considering one specific approach, the so called quasi-infinite horizon
NMPC approach [17].

2.1.3 Quasi-Infinite Horizon NMPC

In the quasi-infinite horizon NMPC method [15, 17] a terminal region constraint of the form (10) and a terminal
penalty term E  x̄  t Tp B as in (11) are added to the standard setup. As mentioned the terminal penalty term is not
a performance specification that can be chosen freely. Rather E and the terminal region Ω are determined off-line
such that the cost functional with terminal penalty term (11) gives an upper approximation of the infinite horizon cost
functional with stage cost F. Thus closed-loop performance over the infinite horizon is addressed. Furthermore, as
is shown later, stability is achieved, while only an optimization problem over a finite horizon must be solved. The
resulting open-loop optimization problem is formulated as follows:
Problem 2 [Quasi-infinite Horizon NMPC]:
Find
min J  x  t < ū 45 ; Tp  (12)
ū 8:9 ;
with:
A
t Tp
J  x  t  ū 45 ; Tp  :  [ F  x̄  τ < ū  τ  dτ E  x̄  t Tp  1 (13)
t

subject to:
x̄˙  τ \ f  x̄  τ < ū  τ  x̄  t  x  t  (14a)
ū  τ )D% τ D6 t  t Tp 7 (14b)
x̄  τ Y] τ D6 t  t Tp 7 (14c)
x̄  t Tp ) Ω 1 (14d)
If the terminal penalty term E and the terminal region Ω are chosen suitably, stability of the closed-loop can be
guaranteed. To present the stability results we need that the following holds for the stage cost-function.
Assumption 4 The stage cost F :  n *^ 2  is continuous in all arguments with F  0  0 _ 0 and F  x  u Q` 0 Y x  u a
 n *C%bK- 0  0 0 .
Given this assumption, the following result, which is a slight modification of Theorem 4.1 in [14], can be established:
Theorem 1 Suppose
(a) that Assumptions 1-4 are satisfied,
(b) E is C1 with E  0  0 L 0, Ω c is closed and connected with the origin contained in Ω and there exists a
continuous local control law k :  n 2  m with k  0  0, such that:
∂E
f  x  k  x  F  x  k  x B
0 E x  Ω (15)
∂x
with k  x Yd x  Ω

7
 (c) the NMPC open-loop optimal control problem has a feasible solution for t  0.
Then for any sampling time 0 e δ e Tp the nominal closed-loop system is asymptotically stable with the region of
attraction f being the set of states for which the open-loop optimal control problem has a feasible solution.
A formal proof of Theorem 1 can be found in [14, 16] and for a linear local controller as described below in [17].
Loosely speaking E is a local Lyapunov function of the system under the local control k  x  in Ω. As will be shown,
Equation (15) allows to upper bound the optimal infinite horizon cost inside Ω by the cost resulting from a local
feedback k  x  .
Notice, that the result in Theorem 1 is nonlocal in nature, i.e. their exists a region of attraction f of at least the size
of Ω. The region of attraction is given by all states for which the open-loop optimal control problem has a feasible
solution.
Obtaining a terminal penalty term E and a terminal region Ω that satisfy the conditions of Theorem 1 is not easy.
If the linearized system is stabilizable and the cost function is quadratic with weight matrices Q and R, a locally
linear feedback law u  Kx can be used and the terminal penalty term can be approximated as quadratic of the form
E  x  xT Px. For this case, a procedure to systematically compute the terminal region and a terminal penalty matrix
off-line is available [17]. Assuming that the Jacobian linearization  A  B  of (1) is stabilizable, where A :  ∂x
∂f
 0  0  and
B :  ∂u  0  0  , this procedure can be summarized as follows:
∂f

Step 1 : Solve the linear control problem based on the Jacobian linearization  A  B  of (1) to obtain a locally stabi-
lizing linear state feedback u  Kx.
Step 2 : Choose a constant κ 6 0  ∞  satisfying κ eCI λmax  AK  and solve the Lyapunov equation

 AK κI  T P P  AK κI QIhg Q K T RK i (16)

to get a positive definite and symmetric P, where AK :  A BK.

Step 3 : Find the largest possible α1 defining a region

Ω1 : .- x ( n / xT Px
α1 0 (17)

such that Kx c , for all x  Ω1 F .

Step 4 : Find the largest possible α  0  α1 7 specifying a terminal region

Ω : .- x ( n / xT Px
α0 (18)

such that the optimal value of the following optimization problem is non-positive:

max - xT Pϕ  x jI κ  xT Px / xT Px
α 0 (19)
x

where ϕ  x  :  f  x  Kx _I AK x.
This procedure allows to calculate E and Ω if the linearization of the system at the origin is stabilizable. If the terminal
penalty term and the terminal region are determined according to Theorem 1, the open-loop optimal trajectories found
at each time instant approximate the optimal solution for the infinite horizon problem.
The following reasoning make this plausible: Consider an infinite horizon cost functional defined by
∞
J ∞  x  t  ū 45B :  [ F  x̄  τ < ū  τ  dτ (20)
t

with ū => on 6 t  ∞  . This cost functional can be split up into two parts
A
t Tp ∞
min J ∞  x  t < ū => min [ F  x̄  τ < ū  τ  dτ l[ F  x̄  τ  ū  τ B dτm 1 (21)
ū 8:9 ; ū 8:9 ;Yk t A
t Tp

The goal is to upper approximate the second term by a terminal penalty term E  x̄  t Tp  . Without further restrictions,
this is not possible for general nonlinear systems. However, if we ensure that the trajectories of the closed-loop system
remain within some neighborhood of the origin (terminal region) for the time interval 6 t Tp  ∞  , then an upper bound

8
on the second term can be found. One possibility is to determine the terminal region Ω such that a local state feedback
law u  k  x  asymptotically stabilizes the nonlinear system and renders Ω positively invariant for the closed-loop. If
an additional terminal inequality constraint x  t Tp _ Ω (see (14d)) is added to Problem 1, then the second term of
equation (21) can be upper bounded by the cost resulting from the application of this local controller u  k  x  . Note
that the predicted state will not leave Ω after t Tp since u  k  x  renders Ω positively invariant. Furthermore the
feasibility at the next sampling instance is guaranteed dismissing the first part of ū and replacing it by the nominal
open-loop input resulting from the local controller. Requiring that x  t Tp a Ω and using the local controller for
τ 6 t Tp  ∞  we obtain:
A
t Tp ∞
minJ ∞ x  t  ū 45B
min [ F  x̄  τ < ū  τ  dτ l[ F  x̄  τ < k  x̄  τ o dτm 1 (22)
ū 8:9 ; ū 8n9 ;k t A
t Tp

If, furthermore, the terminal region Ω and the terminal penalty term are chosen according to condition b) in Theorem
1 (as for example achieved by the procedure given above), integrating (15) leads to
∞
[ F  x̄  τ < k  x̄  τ o dτ
E  x̄  t Tp  1 (23)
t Tp A
Substituting (23) into (22) we obtain

min J ∞  x  t  ū 45B

min J  x  t < ū => ;t Tp  1 (24)
ū 8:9 ; ū 8:9 ;
This implies that the optimal value of the finite horizon problem bounds that of the corresponding infinite horizon
problem. Thus, the prediction horizon can be thought of as extending quasi to infinity which gives this approach its
name. Equation (24) can be exploited to prove Theorem 1.
Like in the dual-mode approach [56], the use of the terminal inequality constraint gives the quasi-infinite horizon
nonlinear MPC scheme computational advantages. Note also, that as for dual-mode NMPC, it is not necessary to
find optimal solutions of Problem 1 in order to guarantee stability. Feasibility also implies stability here [17, 70]. In
difference to the dual-mode controller, however, the local control law u  k  x  is never applied. It is only used to
compute the terminal penalty term E and the terminal region Ω.
Many generalizations and expansions of QIH-NMPC exist. For example discrete time variants can be found in [21, 33].
If the nonlinear system is affine in u and feedback linearizable, then a terminal penalty term can be determined such
that (23) is exactly satisfied with equality [14], i.e. the infinite horizon is recovered exactly. In [18, 19, 44] robust
NMPC schemes using a min-max formulation are proposed, while in [27] an extension to index one DAE systems is
considered. A variation of QIH-NMPC for the control of varying setpoints is given in [28, 30].

2.2 Performance of Finite Horizon NMPC Formulations

Ideally one would like to use an infinite horizon NMPC formulation, since in the nominal case, the closed-loop
trajectories do coincide with the open-loop predicted ones (principle of optimality). The main problem is, that infinite
horizon schemes can often not be applied in practice, since the open-loop optimal control problem cannot be solved
sufficiently fast. Using finite horizons, however, it is by no means true that a repeated minimization over a finite
horizon objective in a receding horizon manner leads to an optimal solution for the infinite horizon problem (with the
same stage cost F). In fact, the two solutions will in general differ significantly if a short horizon is chosen.
From this discussion it is clear that short horizons are desirable from a computational point of view, but long horizons
are required for closed-loop stability and in order to achieve the desired performance.
The QIH-NMPC strategy outlined in the previous section allows in principle to recover the performance of the infinite
horizon scheme without jeopardizing the closed-loop stability. The value function resulting from Problem 2 can be
seen as an upper bound of the infinite horizon cost. To be more precise, if the terminal penalty function E is chosen
such that a corresponding local control law is a good approximation of the control resulting from the infinite horizon
control law in a neighborhood of the origin, the performance corresponding to Problem 2 can recover the performance
of the infinite horizon cost even for short horizons (assuming the terminal region constraint can be satisfied).

9
2.3 Robustness
So far only the nominal control problem was considered. The NMPC schemes discussed before do require that the
actual system is identical to the model used for prediction, i.e. that no model/plant mismatch or unknown disturbances
are present. Clearly this is a very unrealistic assumption for practical applications and the development of a NMPC
framework to address robustness issues is of paramount importance. In this note the nonlinear uncertain system is
assumed to be given by:

ẋ  t  f  x  t < u  t < d  t  (25)

where the uncertainty d => satisfies d  τ prqs x  u  and q is assumed to be compact. Like in the nominal stability
and performance case, the resulting difference between the predicted open-loop and actual closed-loop trajectory is
the main obstacle. As additional problem the uncertainty d hitting the system now leads not only to one single future
trajectory in the prediction, instead a whole tree of possible solutions must be analyzed.
Even though the analysis of robustness properties in nonlinear NMPC must still be considered as an unsolved problem
in general, some preliminary results are available. In principle one must distinguish between two approaches to
consider the robustness question. Firstly one can examine the robustness properties of the NMPC schemes designed
for nominal stability and by this take the uncertainty/disturbances only indirectly into account [40, 47]. Secondly one
can consider to design NMPC schemes that directly take into account the uncertainty/disturbances.

2.3.1 Inherent Robustness of NMPC

As mentioned above, inherent robustness corresponds to the fact, that nominal NMPC can cope with input model
uncertainties without taking them directly into account. This fact stems from the close relation of NMPC to optimal
control. Assuming that the system under consideration is of the following (input affine) form

ẋ  t  f  x  t ) g  x  t  u  t < x  0  x0 (26)

and the cost function takes the form:

A
t Tp 1
J  x  t < ū => ; Tp  :  [  u 2
q  x  dτ E  x̄  t Tp  (27)
t 2
where q is positive definite, that there are no constraints on the state and the input and the resulting control law and
the value function satisfies further assumptions (uP being continuously differentiable and the value function being
twice continuously differentiable). Then one can show [47] that the NMPC control law is inverse optimal, i.e. it is
also optimal for a modified optimal control problem spanning over an infinite horizon. Due to this inverse optimality,
the NMPC control law inherits the same robustness properties as infinite horizon optimal control assuming that the
sampling time δ goes to zero. In particular, the closed-loop is robust with respect to sector bounded input uncertainties;
the nominal NMPC controller also stabilizes systems of the form:

ẋ  t  f  x  t ) g  x  t  φ  u  t  (28)

where φ 45 is a nonlinearity in the sector  1 t 2  ∞  .

2.3.2 Robust NMPC Schemes

At least three different robust NMPC formulations exist:
X Robust NMPC solving an open-loop min-max problem [18, 45]:
In this formulation the standard NMPC setup is kept, however now the cost function optimized is given by the
worst case disturbance “sequence” occurring, i.e.
A
t Tp
J  x  t < ū => ; Tp  :  max [ F  x̄  τ  ū  τ B dτ E  x̄  t Tp  1 (29)
d̄ 8:9 ; t

subject to

x̄˙  t  f g x̄  t  ū  t  d̄  t  i  x̄  t  x  t  1 (30)

10
 The resulting open-loop optimization is a min-max problem. The key problem is, that adding stability constraints
like in the nominal case, might lead to the fact that no feasible solution can be found at all. This mainly stems
from the fact, that one input signal must “reject” all possible disturbances and guarantee the satisfaction of the
stability constraints.
X H∞ -NMPC [11, 18, 45, 46]: Another possibility is to consider the standard H∞ problem in a receding horizon
framework. The key obstacle is, that an infinite horizon min-max problem must be solved (solution of the
nonlinear Hamilton-Jacobi-Isaacs equation). Modifying the NMPC cost functions similar to the H∞ problem
and optimizing over a sequence of control laws robustly stabilizing finite horizon H∞ –NMPC formulations can
be achieved. The main obstacle is the prohibitive computational time necessary. This approach is in close
connection to the first approach.
X Robust NMPC optimizing a feedback controller used during the sampling times [45]:
The open-loop formulation of the robust stabilization problem can be seen as very conservative, since only
open-loop control is used during the sampling times, i.e. the disturbances are not directly rejected in between the
sampling instances. To overcome this problem it has been proposed not to optimize over the input signal. Instead
of optimizing the open-loop input signal directly, a feedback controller is optimized, i.e. the decision variable
ū is not considered as optimization variable instead a “sequence” of control laws ui  ki  x  applied during the
sampling times is optimized. Now the optimization problem has as optimization variables the parameterizations
of the feedback controllers - k1  1B11  kN 0 . While this formulation is very attractive since the conservatism is
reduced, the solution is often prohibitively complex.

2.4 Output Feedback NMPC

So far it was assumed, that the system state necessary for prediction is (perfectly) accessible through measurements. In
general this is not the case and a state observer, as already shown in Figure 3 must be implicitly or explicitly included in
the control loop. Two main questions arise from the use of a state observer. Firstly the question occurs, if the closed-
loop including the state observer possesses the same stability properties as the state feedback contribution alone.
Secondly the question arises, what kind of observer should be used to obtain a good state estimate and good closed
loop performance. The second point is not considered in detail here. It is only noted, that a dual of the NMPC approach
for control does exist for the state estimation problem. It is formulated as an on-line optimization similar to NMPC
and is named moving horizon estimation (MHE). It is dual in the sense, that a moving window of old measurement
data is used to obtain an optimization based estimate of the system state, see for example [1, 57, 66, 67, 69, 75].

2.4.1 Possible Solutions to the Output Feedback NMPC Problem

The most often used approach for output-feedback NMPC is based on the “certainty equivalence principle”. The
estimate state x̂ is measured via a state observer and used in the model predictive controller. Even assuming, that
the observer error is exponentially stable, often only local stability of the closed-loop is achieved [42, 43, 71], i.e.
the observer error must be small to guarantee stability of the closed-loop and in general nothing can be said about the
necessary degree of smallness. This is a consequence of the fact that no general valid separation principle for nonlinear
systems exists. Nevertheless this approach is applied successfully in many applications.
To achieve non-local stability results of the observer based output-feedback NMPC controller, different possibilities
to attack the problem exist:
X Direct consideration of the observer error in the NMPC controller: One could in principle consider the
observer error as disturbance in the controller and design a NMPC controller that can reject this disturbance.
The hurdle of this approach is the fact, that so far an applicable robust NMPC scheme is not available.
X Separation of observer error from the controller [31, 37, 57]: In this approach the observer error is “decou-
pled”/separated from the controller by either a time scale separation, i.e. making the observer much faster than
the other system parts or by projection of the observer error. For example using special separation principles
based on high-gain observers, semi-regional stability results for the closed-loop can be established. The key
component is, that the speed of the observer can be made as fast as necessary.
X Usage of I/O models [65]: One could use suited I/O models that have no internal states for prediction.

11
In the following we shortly review one possible approach for output-feedback NMPC using a time-scale separation of
the observer and controller.

2.4.2 Output Feedback NMPC using High-Gain Observers

We propose to combine high-gain observers with NMPC to achieve semi-regional stability. I.e. if the observer gain is
increased sufficiently, the stability region and performance of the state feedback is recovered. The closed loop system
is semi-regionally stable in the sense, that for any subset u of the region of attraction f of the state-feedback controller
(compare Theorem 1, Section 2.1.3) there exists an observer parameter (gain), such that u is contained in the region
of attraction of the output-feedback controller.
The results are based on “nonlinear separation principles [6, 72]” and it is assumed, that the NMPC feedback is
instantaneous (see below). We will limit the presentation to a special SISO systems class and only give the main
result. The more general MIMO case considering the NMPC inherent open loop control parts (i.e. no instantaneous
feedback) can be found [31, 37].
In the following we consider the stabilization of SISO systems of the following form:

ẋ  Ax bφ  x  u  (31a)
y  x1 1 (31b)

with u  t !v$w and y  t Y& . The output y is given by the first state x1 . The n * n matrix A and the n * 1 vector b
have the following form:
xyy |>}
yy 0 1 0 B 0 }}
zy 0 0 1 B 0 }}
A ..
.
..
.  b € 0 B 0 1 ‚ Tn  1  (32a)
0 BB 0 1~
0 BB{B 0 n n 
Additional to the Assumptions 1-4 we assume, that:
Assumption 5 The function φ :  n *C 2  is locally Lipschitz in its arguments over the domain of interest. Fur-
thermore φ  0  0  0 and φ is bounded in x everywhere.
Note that global boundedness can in most cases be achieved by saturating φ outside a compact region of  n of interest.
The proposed output feedback controller consists of a high-gain observer to estimate the states and an instantaneous
variant of the full state feedback QIH-NMPC controller as outlined in Sect. 2.1.3. By instantaneous we mean that
the system input at all times (i.e. not only at the sampling instances) is given by the instantaneous solution of the
open-loop optimal control problem:

u  x  t B :  uP  τ  0 ; x  t < Tp  1 (33)

This feedback law differs from the standard NMPC formulation in the sense that no open-loop input is implemented
over a sampling time δ. Instead u  t  is considered as a “function” of x  t  .
To allow the subsequent result to hold, we have to require that the feedback resulting from the QIH-NMPC is locally
Lipschitz.
Assumption 6 The instantaneous state feedback (33) is locally Lipschitz.
The observer used for state recovery is a high-gain observer [6, 72, 73] of the following form:

x̂˙  Ax̂ bφ  x̂  u  H  x1 I x̂1  (34)

where H ƒs € α1 t ε  α2 t ε2  11B1  αn t εn ‚ 1 The αi ’s are chosen such that the polynomial

sn α1 sn „ 1 ^B… αn „ 1 s αn  0 

12
is Hurwitz. Here 1ε is the high-gain parameter and can be seen as a time scaling for the observer dynamics (34). A, b
and φ are the same as in (31).
Notice that the use of an observer makes it necessary that the input also be defined (and bounded) for (estimated) states
that are outside the feasible region of the state feedback controller. We simply define the open-loop input for x ‡
† f as
fixed to an arbitrary value u f c :
u  x \ u f   x p
† f 1 (35)
This together with the assumption that  is bounded separates the peaking of the observer from the controller/system [26].
Using the high-gain observer for state recovery, the following result, which establishes semi-regional stability of the
closed-loop can be obtained [36, 37]:
Theorem 2 Assume that the conditions a)-c) of Theorem 1 and Assumption 5-6 hold. Let u be any compact set
contained in the interior of f (region of attraction of the state feedback). Then there exists a (small enough) ε N` 0
such that for all 0 e ε
ε N , the closed-loop system is asymptotically stable with a region of attraction of at least u .
Further, the performance of the state feedback NMPC controller is recovered as ε decreases.
By performance recovery it is meant that the difference between the trajectories of the state feedback and the output
feedback can be made arbitrarily small by decreasing ε. The results show that the performance of the state feedback
scheme can be recovered in the output feedback case, if a state observer with a suitable structure and a fast enough
dynamics is used.
Figure 5 shows the simulation result for an illustrative application of the proposed output feedback scheme to a two
dimensional pendulum car system as depicted in Figure 4 and presented in [36]. The angle between the pendulum

z1 ˆ t ‰

uˆ t‰

Figure 4: Sketch of the inverted pendulum/car system

and the vertical is denoted by z1 , while the angular velocity of the pendulum is given by z2 . The input u is the force
applied to the car. The control objective is to stabilize the upright position of the pendulum. To achieve this objective,
a QIH-NMPC scheme with and without (state-feedback case) a high-gain observer is used. For the results shown in
Figure 5 the pendulum is initialized with an offset from the upright position, while the high-gain observer is started
with zero initial conditions. The figure shows the closed loop trajectories for state feedback QIH-NMPC controller and
the output-feedback controller with different observer gains. The gray ellipsoid around the origin is the terminal re-
gion of the QIH-NMPC controller. The outer “ellipsoid” is an estimate of the region of attraction of the state-feedback
controller. As can be seen for small enough values of the observer parameter ε the closed loop is stable. Furthermore
the performance of the state feedback is recovered as ε tends to zero. More details can be found in [36].
Furthermore the recovery of the region of attraction and the performance of the state-feedback is possible up to any
degree of exactness. In comparison to other existing output-feedback NMPC schemes [42, 71] the proposed scheme
is thus of non-local nature. However, the results are based on the assumption that the NMPC controller is time
continuous/instantaneous. In practice, it is of course not possible to solve the nonlinear optimization problem instanta-
neously. Instead, it will be solved only at some sampling instants. A sampled version of the given result, in agreement
with the “usual” sampled NMPC setup can be found in [31]. Notice also, that the use of a high gain observer is critical,
if the output measurements are very noise, since the noise will be amplified due to the high gain nature of the observer.

3 Computational Aspects of NMPC

NMPC requires the repeated on-line solution of a nonlinear optimal control problem. In the case of linear MPC the
solution of the optimal control problem can be cast as the solution of a (convex) quadratic program and can be solved

13
 1
terminal region
0.5
border region of attraction
state feedback NMPC
0

−0.5

−1

−1.5
z2

−2

−2.5

−3

−3.5 ε=0.01
ε=0.05
−4 ε=0.09
ε=0.1
state feedback
−4.5
−1 −0.5 0 0.5 1 1.5
z1

Figure 5: Phase plot of the pendulum angle (z1 ) and the angular velocity (z2 )

efficiently even on-line. This can be seen as one of the reasons why linear MPC is widely used in industry. For the
NMPC problem the solution involves the solution of a nonlinear program, as is shown in the preceding sections. In
general the solution of a nonlinear (non-convex) optimization problem can be computational expensive. However in
the case of NMPC the nonlinear program shows special structure that can be exploited to still achieve a real-time
feasible solution to the NMPC optimization problem.
For the purpose of this Section the open-loop optimal control Problem 2 of Section 2.1.2 will be considered in a more
optimization focused setting. Especially it is considered, that the state and input constraints x D , u w can be
recasted as a nonlinear inequality constraint of the form l  x  u 
0. Furthermore for simplicity of exposition it is
assumed that the control and prediction horizon coincide and that no final region constraint is present, i.e. we consider
the following deterministic optimal control problem in Bolza form that must be solved at every sampling instance:
Problem 3: Find

minū 8:9 ; J  x  t < ū => ; Tp  (36)

with J  x  t  ū 45 ; Tp  :  ? tt A Tp

F  x̄  τ  ū  τ  dτ E  x̄  t Tp B (37)

subject to: x̄˙  τ  f  x̄  τ < ū  τ ) x̄  t  x  t  (38a)

l  ū  τ < x̄  τ 
0  τ ]6 t  t Tp 7 1 (38b)

3.1 Solution Methods for the Open-Loop Optimal Control Problem

In principle three different approaches to solve the optimal control Problem 3 exist (see for example [9, 48]):
X Hamilton-Jacobi-Bellmann partial differential equations/dynamic programming: This approach is based
on the direct solution of the so called Hamilton-Jacobi-Bellmann partial differential equations. Rather than just
seeking for the optimal u  τ  trajectory the problem is approach as finding a solution for all x  t  . The solution
derived is a state feedback law of the form uP  k  x  and is valid for every initial condition. The key obstacle
of this approach is, that since the “complete” solution is considered at once, it is in general computationally

14
 intractable and suffers from the so called curse of dimensionality, i.e. can be only solved for small systems.
Ideally one would like to obtain such a closed loop state feedback law. In principle the intractability of the
solution can be seen as the key motivation of receding horizon control.
X Euler-Lagrange differential equations/calculus of variations/maximum principle: This method employs
classical calculus of variations to obtain an explicit solution of the input as a function of time u  τ  and not as
feedback law. Thus it is only valid for the specified initial condition x  t  . The approach can be thought of
as the application of the necessary conditions for constrained optimization with the twist, that the optimiza-
tion is infinite dimensional. The solution of the optimal control problem is cast as a boundary value problem.
Since an infinite dimensional problem must be solved, this approach can normally not be applied for on-line
implementation.
X Direct solution using a finite parameterization of the controls and/or constraints: In this approach the input
and/or the constraints are parametrized finitely, thus an approximation of the original open-loop optimal control
problem is seeked. The resulting finite dimensional dynamic optimization problem is solved with “standard”
static optimization techniques.
For an on-line solution of the NMPC problem only the last approach is normally used. Since no feedback is obtained,
the optimization problem must be solved at every sampling instance with the new state information. In the following
only the last solution method is considered in detail.

3.2 Solution of the NMPC Problem Using a Finite Parameterization of the Controls
As mentioned the basic idea behind the direct solution using a finite parameterization of the controls is to approxi-
mate/transcribe the original infinite dimensional problem into a finite dimensional nonlinear programming problem.
In this note the presentation is limited to a parameterization of the input signal as piecewise constant over the sampling
times. The controls are piecewise constant on each of the N  δp predicted sampling intervals: ū  τ Š ūi for τ 
T

6 τi  τiA 1  , τi  t iδ, compare also Figure 6. Thus in the optimal control problem the “input vector” - ū1  11B1 ūN 0 is

past future/prediction

set-point

predicted state x̄

closed-loop
state x open loop input sequence Œ ūk 
ūN
ū2
closed-loop ū1
input u
t δ
t
‹ t t TNδ
p

control/prediction horizon Tp

Figure 6: Piecewise constant input signal for the direct solution of the optimal control problem.

optimized, i.e. the optimization problem takes the form

min Ž ū1     ūN ‘ J  x  t <4- ū1  1B11 ūN 0 ; Tp  (39)

subject to the state and input constraints and the system dynamics. Basically two different solution strategies to this
optimization problem exist [8, 9, 13, 48, 74]:

15
 X Sequential approach: In this method in every iteration step of the optimization strategy the differential equa-
tions (or in the discrete time case the difference equation) are solved exactly by a numerical integration, i.e. the
solution of the system dynamics is implicitly done during the integration of the cost function and only the input
vector - ū1  11B1 ūN 0 appears directly in the optimization problem.
X Simultaneous approach: In this approach the system dynamics (38a) at the sampling points enter as nonlinear
constraints to the optimization problems, i.e. at every sampling point the following equality constraint must be
satisfied:

s̄iA 1  x̄  tiA 1 ; s̄i  ūi  1 (40)

Here s̄i is introduced as additional degree in the optimization problem and describes the “initial” condition for
the sampling interval i, compare also Figure 7. This constraint requires, once the optimization has converged,

x ’ τ“
x̄ ’ t ” Nδ; s̄N — 1 • ūN — 1 “

s̄N

x̄ ’ t ” δ; s̄1 • ū1 “

– x’ t “
s̄2
s̄1

t t” δ t ” Tp time

Figure 7: Simultaneous approach.

that the state trajectory pieces fit together. Thus additionally to the input vector - ū1  111 ūN 0 also the vector of
the s̄i appears as optimization variables.
For both approaches the resulting optimization problem is often solved using sequential quadratic programming tech-
niques (SQP). Both approaches have different advantages and disadvantages. For example the introduction of the “ini-
tial” states s̄i as optimization variables does lead to a special banded-sparse structure of the underlying QP-problem.
This structure can be taken into account to lead to a fast solution strategy [8, 24, 74]. In comparison the matrices for
the sequential approach are often dense and thus the solution is expensive to obtain. A drawback of the simultane-
ous approach is, that only at the end of the iteration a valid state trajectory for the system is available. Thus if the
optimization cannot be finished in time, nothing can be said about the feasibility of the trajectory at all.

3.2.1 Remarks on State and Input Equality Constraints

In the description given above, the state and input constraints were not taken into account. The key problem is, that
they should be satisfied for the whole state and input vector. While for a suitable parametrized input signal (e.g.
parametrized as piecewise constant) it is not a problem to satisfy the constraints since only a finite number of points
must be checked, the satisfaction of the state constraints must in general be enforced over the whole state trajectory.
Different possibilities exist to consider them during the optimization:
X Satisfaction of the constraints at the sampling instances: An approximated satisfaction of the constraints can
be achieved by requiring, that they are at least satisfied at the sampling instances, i.e. at the sampling times it is
required:

l  ū  ti < x̄  ti 
01 (41)

16
 Notice, that this does not guarantee that the constraints are satisfied for the predicted trajectories in between the
sampling instances. However, since this approach is easy to implement it is often used in practice.
X Adding a penalty in the cost function: An approach to enforce the constraint satisfaction exactly for the whole
input/state trajectory is to add an additional penalty term to the cost function. This term is zero as long as the
constraints are satisfied. Once the constraints are not satisfied the value of this term increases significantly, thus
enforcing the satisfaction of the constraints. The resulting cost function may look as following:
A
t Tp
J  x  t  ū 45 ; Tp  :  [  F  x̄  τ< ū  τ p  l  x̄  τ  ū  τ B dτ E  x̄  t Tp  (42)
t

where p in the case that only one nonlinear constraint is present might look like shown in Figure 8. A drawback

p ’ l ’ x• u“

0 l ’ x• u“

Figure 8: Constraint penalty function for one nonlinear constraint.

.

of this formulation is, that the resulting optimization problem is in general difficult to solve for example due to
the resulting non-differentiability of the cost function outside the feasible region of attraction.

3.2.2 Efficient NMPC Formulations

One should notice, that besides an efficient solution strategy of the occurring open-loop optimal control problem the
NMPC problem should be also formulated efficiently. Different possibilities for an efficient NMPC formulation exist:
X Use of short horizon length without loss of performance and stability [17, 34, 63]: As was outlined in
Section 2 short horizons are desirable from a computational point of view, but long horizons are required for
closed-loop stability and in order to achieve the desired performance in closed-loop. The general NMPC scheme
outlined in Section 2.1.2 offers a way out of this dilemma. It uses a terminal region constraint in combination
with a terminal penalty term. The terminal penalty term can be used to given a good approximation of the
infinite horizon cost utilizing a local control law. Additionally the terminal region constraint is in general not
very restrictive, i.e. does not complicate the dynamic optimization problem in an unnecessary manner, as for
example in the zero terminal constraint approach. In some cases, e.g. stable systems, feedback linearizable
systems or systems for which a globally valid control Lyapunov function is known it can even be removed.
Thus such an approach offers the possibility to formulate a computationally efficient NMPC scheme with a
short horizon while not sacrificing stability and performance.
X Use of suboptimal NMPC strategies, feasibility implies stability [17, 34, 56, 70]: In general no global minima
of the open-loop optimization must be found. It is sufficient to achieve a decrease in the value function at
every time to guarantee stability. Thus stability can be seen as being implied by feasibility. If one uses an
optimization strategy that delivers feasible solutions at every sub-iteration and a decrease in the cost function,
the optimization can be stopped if no more time is available and still stability can be guaranteed. The key

17
 obstacle is that optimization strategies that guarantee a feasible and decreasing solution at every iteration are
normally computationally expensive.
X Taking the system structure into account [2, 60, 61]: It is also noticeable, that the system structure should be
taken into account. For example for systems for which a flat output is known the dynamic optimization problem
can be directly reduced to a static optimization problem. This results from the fact that for flat systems the
input and the system state can be given in terms of the output and its derivatives as well as the system initial
conditions. The drawback however is, that the algebraic relation between the output and the derivatives to the
states and inputs must be known, which is not always possible.
Combining the presented approaches for an efficient formulation of the NMPC problem and the efficient solution
strategies of the optimal control problem, the application of NMPC to realistically sized applications is possible even
with nowadays computational power. Besides the problem of stability of the closed-loop and the output-feedback
problem, the efficient solution of the resulting open-loop optimal control problem is important for any application of
NMPC to real processes. Summarizing, a real-time application of NMPC is possible [8, 29, 59] if: a) NMPC schemes
that do not require a high computational load and do not sacrifice stability and performance, like QIH-NMPC, are
used and b) the resulting structure of the open-loop optimization problem is taken into account during the numerical
solution.

4 Application Example–Real-Time Feasibility of NMPC

To show that nonlinear predictive control can be applied to even rather large systems if efficient NMPC schemes and
special tailored numerical solution methods are used, we give some results from a real-time feasibility study of NMPC
for a high-purity distillation column as presented in [5, 24, 25, 59]. Figure 9 shows the in this study considered 40
tray high-purity distillation column for the separation of Methanol and n-Propanol. The binary mixture is fed in the

L xD
40

28

F, xF
21

14

V
1

xB

Figure 9: Scheme of the distillation column

column with flow rate F and molar feed composition xF . Products are removed at the top and bottom of the column
with concentrations xB and xD respectively. The column is considered in L/V configuration, i.e. the liquid flow rate L
and the vapor flow rate V are the control inputs. The control problem is to maintain the specifications on the product
concentrations xB and xD . For control purposes, models of the system of different complexity are available. As usual
in distillation control, xB and xD are not controlled directly. Instead an inferential control scheme which controls the
deviation of the concentrations on tray 14 and 28 from the setpoints is used, i.e. only the concentration deviations from
the setpoint on trays 14 and 28 plus the inputs are penalized in the cost-function. The QIH-NMPC control scheme is
used for control. The terminal region and terminal penalty term have been calculated as suggested in Sect. 2.1.3.
In Table 1 the maximum and average CPU times necessary to solve one open-loop optimization problem for the QIH-
NMPC scheme in case of a disturbance in xF with respect to different model sizes are shown. Considering that the

18
Table 1: Comparison of the average and maximum CPU time in seconds necessary for the solution of one open-loop
optimal control problem. The results are obtained using MUSCOD-II [12] and QIH-NMPC for models of different
size. The prediction horizon of is 10 minutes and a controller sampling time δ  30sec is used
model size max avrg
42 1.86s 0.89s
164 6.21s 2.48s

sampling time of the process control system connected to the distillation column is 30sec, the QIH-NMPC using the
appropriate tool for optimization is even real-time feasible for the 164th order model. Notice, that a straightforward
solution of the optimal control problem for the 42nd order model using the optimization-toolbox in Matlab needs in
average 620sec to find the solution and is hence not real-time implementable. Also a numerical approximation of the
infinite horizon problem by increasing the prediction horizon sufficiently enough is not real-time feasible as shown
in [32]. More details and simulation results for the distillation column example can be found in [5, 24, 59]. First
experimental results on a pilot scale distillation column are given in [25].
The presented case study underpins, that NMPC can be applied in practice already nowadays, if efficient numerical
solution methods and efficient NMPC formulations (like QIH-NMPC) are used.

5 Conclusions
Model predictive control for linear constrained systems has been shown to provide an excellent control solution both
theoretically and practically. The incorporation of nonlinear models poses a much more challenging problem mainly
because of computational and control theoretical difficulties, but also holds much promise for practical applications.
In this note an overview over the theoretical and computational aspects of NMPC is given. As outlined some of the
challenges occurring in NMPC are already solvable. Nevertheless many unsolved questions remain. Here only a few
are noticed as a guide for future research:
X Output feedback NMPC: While some first results in the area of output feedback NMPC exist, none of them
seem to be applicable to real processes. Especially the incorporation of suitable state estimation strategies in the
NMPC formulation must be further considered.
X Robust NMPC Formulations: By now a few robust NMPC formulations exist. While the existing schemes in-
crease the general understanding they are computationally intractable to be applied in practice. Further research
is required to develop implementable robust NMPC strategies.
X Industrial Applications of NMPC: The state of industrial application of NMPC is growing rapidly and seems
to follow academically available results more closely than linear MPC. However, none of the NMPC algorithms
provided by vendors include stability constraints as required by control theory for nominal stability; instead they
rely implicitly upon setting the prediction horizon long enough to effectively approximate an infinite horizon.
Future developments in NMPC control theory will hopefully contribute to making the gap between academic
and industrial developments even smaller.

Acknowledgements
The authors gratefully acknowledge Lars Imsland and Bjarne Foss from the Department of Engineering Cybernetics,
NTNU Trondheim, Norway for the ongoing cooperation in the area of output-feedback NMPC. Parts of Sect. 2.4 are
based on this cooperation. Furthermore the authors would like to acknowledge the fruitful cooperation with Moritz
Diehl, Hans Georg Bock and Johannes Schlöder from the Center for Scientific Computing (IWR) at the University of
Heidelberg, Germany in the area of efficient solution of NMPC as outlined in Sect. 3 and Sect. 4.

19
References
[1] M. Alamir. Optimization based nonlinear observers revisited. Int. J. Contr., 72(13):1204–1217, 1999.
[2] M. Alamir and P. Bonard. Globally stabilising discontinuous feedback law for a class of nonlinear systems.
European J. of Control, 1(2):44–50, 1996.
[3] M. Alamir and G. Bornard. Stability of a truncated infinite constrained receding horizon scheme: The general
discrete nonlinear case. Automatica, 31(9):1353–1356, 1995.
[4] F. Allgöwer, T. A. Badgwell, J. S. Qin, J. B. Rawlings, and S. J. Wright. Nonlinear predictive control and
moving horizon estimation – An introductory overview. In P. M. Frank, editor, Advances in Control, Highlights
of ECC’99, pages 391–449. Springer, 1999.
[5] F. Allgöwer, R. Findeisen, Z. Nagy, M. Diehl, H.G. Bock, and J.P. Schlöder. Efficient nonlinear model predictive
control for large scale constrained processes. In Proceedings of the Sixth International Conference on Methods
and Models in Automation and Robotics, pages 43–54. Miedzyzdroje, Poland, 2000.
[6] A.N. Atassi and H.K. Khalil. A separation principle for the stabilization of a class of nonlinear systems. IEEE
Trans. Automatic Control, 44(9):1672–1687, 1999.
[7] R. Bellman. Dynamic Programming. Princeton University Press, Princeton, New Jersey, 1957.
[8] L. Biegler. Efficient solution of dynamic optimization and NMPC problems. In F. Allgöwer and A. Zheng,
editors, Nonlinear Predictive Control, pages 219–244. Birkhäuser, 2000.
[9] T. Binder, L. Blank, H.G. Bock, R. Burlisch, W. Dahmen, M. Diehl, T. Kronseder, W. Marquardt, J.P. Schlöder,
and O. von Stryk. Introduction to model based optimization of chemical processes on moving horizons. In
M. Groetschel, S.O. Krumke, and J. Rambau, editors, Online Optimization of Large Scale Systems: State of the
Art, pages 295–339. Springer, 2001.
[10] R. R. Bitmead, M. Gevers, and V. Wertz. Adaptive Optimal Control – The Thinking Man’s GPC. Prentice Hall,
New York, 1990.
[11] R. Blauwkamp and T. Basar. A receding-horizon approach to robust output feedback control for nonlinear
systems. In Proc. 38th IEEE Conf. Decision Contr., pages 4879–4884, San Diego, 1999.
[12] H.G. Bock, M. Diehl, D. Leineweber, and J. Schlöder. A direct multiple shooting method for real-time optimiza-
tion of nonlinear DAE processes. In F. Allgöwer and A. Zheng, editors, Nonlinear Predictive Control, pages
245–268. Birkhäuser, 2000.
[13] H.G. Bock, M. Diehl, J.P. Schlöder, F. Allgöwer, R. Findeisen, and Z. Nagy. Real-time optimization and nonlinear
model predictive control of processes governed by differential-algebraic equations. In Proc. Int. Symp. Adv.
Control of Chemical Processes, ADCHEM, pages 695–703, Pisa, Italy, 2000.
[14] H. Chen. Stability and Robustness Considerations in Nonlinear Model Predictive Control. Fortschr.-Ber. VDI
Reihe 8 Nr. 674. VDI Verlag, Düsseldorf, 1997.
[15] H. Chen and F. Allgöwer. A quasi-infinite horizon predictive control scheme for constrained nonlinear systems.
In Proc. 16th Chinese Control Conference, pages 309–316, Qindao, 1996.
[16] H. Chen and F. Allgöwer. Nonlinear model predictive control schemes with guaranteed stability. In R. Berber
and C. Kravaris, editors, Nonlinear Model Based Process Control, pages 465–494. Kluwer Academic Publishers,
Dodrecht, 1998.
[17] H. Chen and F. Allgöwer. A quasi-infinite horizon nonlinear model predictive control scheme with guaranteed
stability. Automatica, 34(10):1205–1218, 1998.
[18] H. Chen, C.W. Scherer, and F. Allgöwer. A game theoretic approach to nonlinear robust receding horizon control
of constrained systems. In Proc. Amer. Contr. Conf., pages 3073–3077, Albuquerque, 1997.

20
[19] H. Chen, C.W. Scherer, and F. Allgöwer. A robust model predictive control scheme for constrained linear
systems. In 5th IFAC Symposium on Dynamics and Control of Process Systems, DYCOPS-5, pages 60–65,
Korfu, 1998.
[20] G. De Nicolao, L. Magni, and R. Scattolini. Stabilizing nonlinear receding horizon control via a nonquadratic ter-
minal state penalty. In Symposium on Control, Optimization and Supervision, CESA’96 IMACS Multiconference,
pages 185–187, Lille, 1996.
[21] G. De Nicolao, L. Magni, and R. Scattolini. Stabilizing receding-horizon control of nonlinear time-varying
systems. In Proc. 4rd European Control Conference ECC’97, Brussels, 1997.
[22] G. De Nicolao, L. Magni, and R. Scattolini. Stability and robustness of nonlinear receding horizon control. In
F. Allgöwer and A. Zheng, editors, Nonlinear Predictive Control, pages 3–23. Birkhäuser, 2000.
[23] S. de Oliveira Kothare and M. Morari. Contractive model predictive control for constrained nonlinear systems.
IEEE Trans. Automat. Contr., 45(6):1053–1071, 2000.
[24] M. Diehl, R. Findeisen, Z. Nagy, H.G. Bock, J.P. Schlöder, and F. Allgöwer. Real-time optimization and nonlinear
model predictive control of processes governed by differential-algebraic equations. J. Proc. Contr., 2002. In
press.
[25] M. Diehl, I. Uslu, R. Findeisen, S. Schwarzkopf, F. Allgöwer, H.G. Bock, T. Bürner, E.D. Gilles, A. Kienle,
J.P. Schlöder, and E. Stein. Real-time optimization of large scale process models: Nonlinear model predictive
control of a high purity distillation column. In M. Groetschel, S.O. Krumke, and J. Rambau, editors, Online
Optimization of Large Scale Systems: State of the Art, pages 363–384. Springer, 2001.
[26] F. Esfandiari and H.K. Khalil. Output feedback stabilization of fully linearizable systems. Int. J. Control,
56(5):1007–1037, 1992.

[27] R. Findeisen and F. Allgöwer. Nonlinear model predictive control for index–one DAE systems. In F. Allgöwer
and A. Zheng, editors, Nonlinear Predictive Control, pages 145–162. Birkhäuser, Basel, 1999.
[28] R. Findeisen and F. Allgöwer. A nonlinear model predictive control scheme for the stabilization of setpoint
families. Journal A, Benelux Quarterly Journal on Automatic Control, 41(1):37–45, 2000.
[29] R. Findeisen, F. Allgöwer, M. Diehl, H.G. Bock, J.P. Schlöder, and Z. Nagy. Efficient nonlinear model predictive
control. In 6th International Conference on Chemical Process Control – CPC VI, pages 454–460, 2000.

[30] R. Findeisen, H. Chen, and F. Allgöwer. Nonlinear predictive control for setpoint families. In Proc. Amer. Contr.
Conf., pages 260–4, Chicago, 2000. ACC.
[31] R. Findeisen, L. Imsland, F. Allgöwer, and B.A. Foss. Output feedback nonlinear Predictive control -a separation
principle approach. To appear in proceedings of 15th IFAC World Congress, 2001.
[32] R. Findeisen, Z. Nagy, M. Diehl, F. Allgöwer, H.G. Bock, and J.P. Schlöder. Computational feasibility and
performance of nonlinear model predicitve control. In Proc. 6st European Control Conference ECC’01, pages
957–961, Porto, Portugal, 2001.
[33] R. Findeisen and J.B. Rawlings. Suboptimal infinite horizon nonlinear model predictive control for discrete time
systems. Technical Report # 97.13, Automatic Control Laboratory, Swiss Federal Institute of Technology (ETH),
Zürich, Switzerland, 1997. Presented at the NATO Advanced Study Institute on Nonlinear Model Based Process
Control.
[34] F.A. Fontes. A general framework to design stabilizing nonlinear model predictive controllers. Syst. Contr. Lett.,
42(2):127–143, 2000.
[35] E. Gurkovics. Receding horizon control via bolza-type optimization. Syst. Contr. Lett., 35(2):195–200, 1998.

21
[36] L. Imsland, R. Findeisen, E. Bullinger, F. Allgöwer, and B.A. Foss. A note on stability, robustness and per-
formance of output feedback nonlinear model predictive control. Accepted for publication in J. Proc. Contr.,
2001.
[37] L. Imsland, R. Findeisen, E. Bullinger, F. Allgöwer, and B.A. Foss. On output feedback nonlinear model predic-
tive control using high gain observers for a class of systems. In 6th IFAC Symposium on Dynamics and Control
of Process Systems, DYCOPS-6, pages 91–96, Jejudo, Korea, 2001.
[38] A. Jadbabaie, J. Yu, and J. Hauser. Unconstrained receding horizon control of nonlinear systems. IEEE Trans.
Automat. Contr., 46(5):776 –783, 2001.
[39] S.S. Keerthi and E.G. Gilbert. Optimal infinite-horizon feedback laws for a general class of constrained discrete-
time systems: Stability and moving-horizon approximations. J. Opt. Theory and Appl., 57(2):265–293, 1988.
[40] D.L. Kleinman. An easy way to stabilize a linear constant system. IEEE Trans. Automat. Contr., 15:692, 1970.
[41] J.H. Lee and B. Cooley. Recent advances in model predictive control and other related areas. In J.C. Kantor, C.E.
Garcia, and B. Carnahan, editors, Fifth International Conference on Chemical Process Control – CPC V, pages
201–216. American Institute of Chemical Engineers, 1996.
[42] L. Magni, D. De Nicolao, and R. Scattolini. Output feedback receding-horizon control of discrete-time nonlinear
systems. In Preprints of the 4th Nonlinear Control Systems Design Symposium 1998 - NOLCOS’98, pages
422–427. IFAC, July 1998.
[43] L. Magni, G. De Nicolao, and R Scattolini. Output feedback and tracking of nonlinear systems with model
predictive control. Automatica, 37(10):1601–1607, 2001.
[44] L. Magni, G. De Nicolao, and R Scattolini. A stabilizing model-based predicitve control algorithm for nonlinear
systems. auto, 37(10):1351–1362, 2001.
[45] L. Magni, G. De Nicolao, R. Scattolini, and F. Allgöwer. Robust receding horizon control for nonlinear discrete-
time systems. To appear in Proceedings 15th IFAC World Congress, 2002, 2001.
[46] L. Magni, H. Nijmeijer, and A.J. van der Schaft. A receding-horizon approach to the nonlinear h∞ control
problem. auto, 37(5):429–435, 2001.
[47] L. Magni and R. Sepulchre. Stability margins of nonlinear receeding–horizon control via inverse optimality.
Syst. Contr. Lett., 32(4):241–245, 1997.
[48] D.Q. Mayne. Optimization in model based control. In Proc. IFAC Symposium Dynamics and Control of Chemical
Reactors, Distillation Columns and Batch Processes, pages 229–242, Helsingor, 1995.
[49] D.Q. Mayne. Nonlinear model predictive control: An assessment. In J.C. Kantor, C.E. Garcia, and B. Carna-
han, editors, Fifth International Conference on Chemical Process Control – CPC V, pages 217–231. American
Institute of Chemical Engineers, 1996.
[50] D.Q. Mayne. Nonlinear model predictive control. In F. Allgöwer and A. Zheng, editors, Nonlinear Predictive
Control, pages 24–44. Birkhäuser, 2000.
[51] D.Q. Mayne and H. Michalska. Receding horizon control of nonlinear systems. IEEE Trans. Automat. Contr.,
35(7):814–824, 1990.
[52] D.Q. Mayne, J.B. Rawlings, C.V. Rao, and P.O.M. Scokaert. Constrained model predictive control: stability and
optimality. Automatica, 26(6):789–814, 2000.
[53] E.S. Meadows, M.A. Henson, J.W. Eaton, and J.B. Rawlings. Receding horizon control and discontinuous state
feedback stabilization. Int. J. Contr., 62(5):1217–1229, 1995.
[54] E.S. Meadows and J.B. Rawlings. Receding horizon control with an infinite horizon. In Proc. Amer. Contr. Conf.,
pages 2926–2930, San Francisco, 1993.

22
[55] H. Michalska. A new formulation of receding horizon stabilizing control without terminal constraint on the state.
European J. of Control, 3(1):1–14, 1997.
[56] H. Michalska and D.Q. Mayne. Robust receding horizon control of constrained nonlinear systems. IEEE Trans.
Automat. Contr., AC-38(11):1623–1633, 1993.
[57] H. Michalska and D.Q. Mayne. Moving horizon observers and observer-based control. IEEE Trans. Automat.
Contr., 40(6):995–1006, 1995.
[58] M. Morari and J.H. Lee. Model predicitve control: Past, present and future. Comp. and Chem. Eng., 23(4/5):667–
682, 1999.
[59] Z. Nagy, R. Findeisen, M. Diehl, F. Allgöwer, H.G. Bock, S. Agachi, J.P. Schlöder, and D. Leineweber. Real-time
feasibility of nonlinear predictive control for large scale processes – a case study. In Proc. Amer. Contr. Conf.,
pages 4249–4254, Chicago, 2000. ACC.
[60] V. Nevistić and M. Morari. Constrained control of feedback-linearizable systems. In Proc. 3rd European Control
Conference ECC’95, pages 1726–1731, Rome, 1995.
[61] N. Petit, M.B. Miliam, and R.M. Murray. Inversion based contrained trajectory optimization. In Nonlinear
Control Systems Design Symposium 2001 - NOLCOS’01. IFAC, July 2001.
[62] E. Polak and T.H. Yang. Moving horizon control of linear systems with input saturation and plant uncertainty.
Part 1. Robustness. Int. J. Contr., 58(3):613–638, 1993.
[63] J. Primbs, V Nevistić, and J. Doyle. Nonlinear optimal control: A control Lyapunov function and receding
horizon perspective. Asian Journal of Control, 1(1):14–24, 1999.
[64] S.J. Qin and T.A. Badgwell. An overview of industrial model predictive control technology. In J.C. Kantor, C.E.
Garcia, and B. Carnahan, editors, Fifth International Conference on Chemical Process Control – CPC V, pages
232–256. American Institute of Chemical Engineers, 1996.
[65] S.J. Qin and T.A. Badgwell. An overview of nonlinear model predictive control applications. In F. Allgöwer and
A. Zheng, editors, Nonlinear Predictive Control, pages 369–393. Birkhäuser, 2000.
[66] C.V. Rao. Moving Horizon Estimation of Constrained and Nonlinear Systems. PhD thesis, University of
Wisconsin–Madison, 1999.
[67] C.V. Rao, J.B. Rawlings, and D.Q. Mayne. Constrained state estimation for nonlinear discrete-time systems:
Stability and moving horizon approximations. Submitted for publication in IEEE Transactions on Automatic
Control, 1999.
[68] J. B. Rawlings. Tutorial overview of model predictive control. IEEE Contr. Syst. Magazine, 20(3):38–52, 2000.
[69] D.G. Robertson and J.H. Lee. A least squares formulation for state estimation. J. Proc. Contr., 5(4):291–299,
1995.
[70] P.O.M. Scokaert, D.Q. Mayne, and J.B. Rawlings. Suboptimal model predictive control (feasibility implies
stability). IEEE Trans. Automat. Contr., 44(3):648–654, 1999.
[71] P.O.M. Scokaert, J.B. Rawlings, and E.S. Meadows. Discrete-time stability with perturbations: Application to
model predictive control. Automatica, 33(3):463–470, 1997.
[72] A. Teel and L. Praly. Tools for semiglobal stabilization by partial state and output feedback. SIAM J. Control
and Optimization, 33(5):1443–1488, 1995.
[73] A. Tornambè. High-gain observers for non-linear systems. Int. J. of Systems Science, 23(9):1475–1489, 1992.
[74] S. J. Wright. Applying new optimization algorithms to model predictive control. In J.C. Kantor, C.E. Garcia,
and B. Carnahan, editors, Fifth International Conference on Chemical Process Control – CPC V, pages 147–155.
American Institute of Chemical Engineers, 1996.
[75] G. Zimmer. State observation by on-line minimization. Int. J. Contr., 60(4):595–606, 1994.

23