INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING

Int. J. Adapt. Control Signal Process. 2014; 28:444–463

Published online 9 November 2012 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/acs.2367

Offset-free multistep nonlinear model predictive control under

plant–model mismatch

Xuemin Tian 1 , Ping Wang 1 , Dexian Huang 2 and Sheng Chen 3,4, * ,†
1 College of Information and Control Engineering, China University of Petroleum (East China), Dongying, Shandong
257061, China
2 Department of Automation, Tsinghua University, Beijing 10084, China
3 Electronics and Computer Science, Faculty of Physical and Applied Sciences, University of Southampton,

Southampton SO17 1BJ, UK

4 Faculty of Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia

SUMMARY
A multistep nonlinear model predictive control (MPC) framework is developed to achieve steady-state offset-
free control in the presence of plant–model mismatch. Our formulation explicitly accounts for the effect of
plant–model mismatch by involving the output feedback error, which is expressed as the difference between
the measured process output and the predicted model output at the previous sampling instance, in the multi-
step model recursive prediction. The proposed scheme is capable of improving the performance of nonlinear
MPC, because the plant–model mismatch is effectively compensated through the recursive prediction prop-
agation. We prove that this formulation is able to remove the steady-state error to achieve offset-free control.
The proposed nonlinear MPC framework is applied to a highly nonlinear two-input two-output continuous
stirred tank reactor, in comparison with other MPC implementations. The results obtained demonstrate that
the proposed technique outperforms some existing popular MPC schemes and can realise offset-free control
even under significant plant–model mismatch and unmeasured disturbances. Copyright © 2012 John Wiley
& Sons, Ltd.

Received 10 October 2011; Revised 10 October 2012; Accepted 19 October 2012

KEY WORDS: nonlinear model predictive control; offset free; multistep model prediction; plant–model
mismatch

1. INTRODUCTION

Model predictive control (MPC), which is referred to as a class of control algorithms utilising an
explicit process model to predict the future response of a plant, has become popular in process
industries [1, 2]. In an MPC scheme, a dynamic model is first developed to predict the future pro-
cess outputs in the prediction horizon, based on which future control actions are then computed to
minimise a prespecified cost function. Therefore, the effectiveness of MPC relies heavily on the
availability of a reasonably accurate process model [3]. MPC algorithms based on linear models
are unable to control effectively the nonlinear systems operating over a wide range of operational
conditions and with distinctly different input–output behaviours. Analysis and synthesis of MPC
schemes based on nonlinear models directly has been an important area of research over the past
decade [4]. The introduction of nonlinear models in MPC formulations may solve some problems
concerning the nonlinear nature of processes, but the mismatch between the model and the actual

*Correspondence to: Sheng Chen, Electronics and Computer Science, Faculty of Physical and Applied Sciences,
University of Southampton, Southampton SO17 1BJ, UK.
† E-mail: [email protected]

Copyright © 2012 John Wiley & Sons, Ltd.

 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 445

process inevitably exists [5, 6]. Furthermore, unmeasured disturbances will widen the gap between
the model and the process which in turn may lead to degradation in MPC performance. Although
the inherent feedback mechanism can tolerate the plant–model mismatch to a certain extent, conven-
tional MPC formulations cannot guarantee steady-state offset-free control [6–8]. The steady-state
offset error refers to a persistent discrepancy between the desired set-point value and the actual
output of the process.
To combat the offset problem caused by plant–model mismatch or unmeasured disturbances,
various methods have been proposed [9–18]. The most widely used scheme in industrial MPC
implementations [1, 3] generates the output targets by shifting the set points to compensate for
plant–model mismatch, assuming that the plant–model mismatch is due to step disturbances in the
output and the disturbance remains constant throughout the prediction horizon. As this correcting
scheme is originally employed in the dynamic matrix control (DMC) algorithm, it is known as the
DMC-like correcting scheme for offset-free control [1, 3]. This method has been shown to eliminate
steady-state offset in some cases, but it cannot effectively deal with the disturbances in the inputs.
On the basis of the steady-state target optimisation, the works [12–14] augment the plant model with
a disturbance model and use the combined model to estimate the size of the disturbance. Although
a number of studies [4–8] have shown that different disturbance models will lead to the different
levels of closed-loop performance in the presence of different unmeasured disturbances or plant–
model mismatch, there exist insufficient practical guidelines for designing nonlinear MPC (NMPC)
disturbance model [6, 10]. Using state observers in MPC formulation is another common method for
removing steady-state offset. For nonlinear process, Huang et al. [17] have proposed an offset-free
NMPC formulation that integrates both the state and output disturbances from the extended Kalman
filter. Recently, the authors of [18] have used the moving horizon estimation instead of extended
Kalman filter to deal with state inequality constraints. However, the accuracy of a state estimator
also heavily depends on the accuracy of the prediction model and, moreover, the dynamics of the
state estimator is often required to be much faster than the dynamics of the state feedback control
law, which means that the error signal between the observer and the actual system should con-
verge to zero at a much faster rate than the convergence rate of the system state. Sometimes, a state
observer may still be used even when the state itself is measurable, which will inevitably impose
some additional computational complexity in MPC implementation.
The plant–model mismatch problem is particularly critical to black-box model-based MPC
algorithms, because a black-box model may be unduly extrapolated into the regions of the state
space where identification data were scarce or nonexistent [19]. The emphasis of black-box model-
based NMPC algorithms has mainly focused on developing the models with good long-range
prediction capability and/or finding the efficient methodologies to reduce the on-line computa-
tional burden. However, there exist relatively few publications in the area of black-box model-based
NMPC that can effectively deal with the issue of plant–model mismatch. Psichogis and Ungar [20]
have utilised a feed-forward neural network (FNN) model in the conventional MPC scheme, where
steady-state offsets are observed to exist during set-point tracking. The corrections to the output
are then made using a constant disturbance estimate to obtain offset-free tracking. The authors of
[21] have studied the effect of two common types of neural network, FNN and external recurrent
network, on the performance of multistep NMPC, and their results show that the steady-state offset
of the NMPC algorithm using FNN models will always occur when the prediction horizon is longer
than the control horizon. Lu and Tsai [22] have presented a design methodology for the generalised
predictive control via recurrent fuzzy neural network. The generalised predictive control law with
integral action is derived on the basis of the minimisation of a modified predictive performance cri-
terion, and the simulations and experiments have shown that this method is capable of controlling
the processes with satisfactory performance under set point and load changes. Zhang et al. [23] have
developed an alternative offset-free output feedback NMPC approach based on fuzzy models and
an integrating disturbance model, where an augmented piecewise observer is used to estimate the
system states and lumped disturbances. However, all the aforementioned studies rely heavily on the
special structures of predictive model and are difficult to extend to other forms of model.
Our motivation is to provide an efficient design methodology to achieve the steady-state offset-
free control in multistep NMPC. We note the fact that, for multistep prediction, the predictive model

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
446 X. TIAN ET AL.

is recursively used and the accuracy of the previous-step predictions directly affects the quality of the
next-step prediction. Therefore, we propose a scheme in which the previous predictions are firstly
corrected by the output feedback error and then used as the input of the predictive model to provide
further prediction. The incorporation of the feedback error into multistep prediction will improve
the MPC performance, because the information about plant–model mismatch and/or unmeasured
disturbances are directly and effectively included in the MPC multistep prediction formulation.
Specifically, the plant–model mismatch and/or unmeasured disturbances, manifested as the pre-
vious output feedback error, are successively corrected or compensated in the multistep recursive
prediction propagation. The advantages of this new formulation are that the steady-state offset can
always be eliminated once the process reaches a steady state without imposing additional com-
putational complexity and the only measurement required is the process output. In addition, the
proposed offset-free MPC strategy is very practical, as it only needs slightly change to the tra-
ditional MPC formulation, and the resulting offset-free NMPC algorithm can adopt a variety of
prediction model forms. We use the highly nonlinear continuously stirred tank reactor (CSTR)
process [24–30], as a benchmark to compare the performance of the proposed multistep NMPC
strategy with the most widely adopted multistep NMPC scheme based on DMC-like correcting. Our
results obtained demonstrate that the proposed method, which offers better performance than this
existing scheme, is capable of achieving the offset-free control in the presence of significant model
mismatch and unmeasured disturbances.
Before presenting the details of our new scheme, we highlight our novel contributions and empha-
sise the differences of our method with typical related works. Various techniques for achieving
steady-state offset-free control may be divided into two classes.

(1) A classical approach to avoid steady-state offset is by introducing an integrator in the control
loop, similar to the case of PI control. For nonlinear systems, this can be realised by introduc-
ing the control variable increments and augmenting or expanding the system state variables,
for example, [9,31,32]. However, this will inevitably increase the dimension of the controlled
system and therefore significantly increases the computational complexity of NMPC on-line
optimisation. Furthermore, to ensure the closed-loop stability of the augmented system, it is
generally required to compute and set the steady-state values of the input and state variables
corresponding to the desired equilibrium (see Assumption in Lemma 2 of [9]). This is very
difficult if not impossible to achieve in the presence of plant–model mismatch. This seriously
restricts the application of this approach to practical industrial processes. Moreover, the inte-
gral action, which is introduced in front of the plant, does not generally guarantee that the
controller can achieve the steady-state zero-error regulation if a state-feedback control law
is used [32], and therefore, an observer needs to be introduced even if the system states are
available.
By contrast, our proposed scheme does not require the computation of the steady-state
values of the input and state variables at the desired equilibrium, and the steady-state offset-
free control is achieved entirely by output feedback without the need of relying on state
observer. Moreover, our scheme is computationally more attractive than the schemes of
[9, 31, 32].
(2) The other approach to avoid steady-state offset in the presence of plant–model mismatch
and external disturbance is to model the plant–model mismatch or external disturbance and
then use the output of this model to compensate the effects of plant–model mismatch or
external disturbance to the controller. Examples of this second class include the DMC-like
correcting scheme widely adopted in practice, the work [33], the approach of [18] and our
proposed scheme. The scheme [33] is based on an adaptive disturbance estimation method
with time-varying forgetting, where the disturbance is modelled as an integrated moving aver-
age process with one lag and is estimated using an adaptive technique that discounts old
measurements. However, a linear model, namely, an integrated moving average process with
one lag, does not provide an adequate representation for the plant–model mismatch or distur-
bance in practice, especially when the plant is nonlinear. By contrast, our proposed method is
applicable to the generic case of nonlinear plant–model mismatch or disturbance. We use the

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 447

plant–model mismatch or the error between the plant output and model prediction to com-
pensate the controller. However, unlike the conventional multistep prediction method that first
makes predictions and then corrects them, our method first corrects the previous predictions
by the output feedback error and then uses the compensated previous predictions as the inputs
of the predictive model to provide further prediction. This approach is very simple to imple-
ment, and our analysis and simulation results confirm that it is very effective in achieving
steady-state offset control.
The moving horizon estimation-based formulation [18] requires both nonlinear state-
space equation model and nonlinear output measurement equation model of the underlying
nonlinear process. If the state-space equation is known and is substituted into the out-
put/measurement equation, then the formulation in [18] appears to ‘reduced’ to our formula-
tion. In reality, however, our input–output model-based formulation has significant practical
and implementation advantages. Firstly, our input–output-based nonlinear auto-regressive
exogenous (NARX) model can be easily obtained with black-box data-drive identification
procedures. By contrast, it is very difficult to obtain a state-space nonlinear model with
data-drive identification algorithms. It is also very difficult to recast an identified nonlin-
ear input–output model into a state-space one. Secondly, our scheme of achieving zero offset
is computationally much simpler than the scheme of [18]. Let us assume that the state-space
model can be obtained. Every step of the on-line optimisation for the NMPC scheme of [18]
requires to solve the on-line state observer estimation optimisation, and it must wait for this
state observer to converge. This inevitably increases on-line optimisation complexity dramat-
ically, which may even make the scheme of [18] impractical, particularly if the state observer
convergence horizon is long.
Our proposed approach does have a limitation. If the plant has some physical variables
that are unmeasurable but are required to meet strict constraints, then it is difficult for our
method to guarantee that these unmeasured ‘state variables’ meet their constraints because
these variables are not presented in the input–output model. In this case, an observer-based
state estimation approach may have to be used.
The structure of this contribution is organised as follows. Multistep NMPC formulation is intro-
duced in Section 2, whereas Section 3 discusses the most widely used DMC-like correcting scheme
to deal with the offset problem and presents our new multistep NMPC scheme for offset-free con-
trol. The simulation results are shown in Section 4 to demonstrate the effectiveness of the proposed
methodology, and our conclusions are given in Section 5.

2. MULTISTEP NONLINEAR MODEL PREDICTIVE CONTROL FORMULATION

Consider the controlled nonlinear process described by the following discrete-time nonlinear
equation
 
y.k/ D f y.k  1/,    , y.k  ny /, u.k  1/,    , u.k  nu / , (1)
where k denotes the sampling instant, f./ represents the unknown my -dimensional nonlinear vector
mapping of the controlled process, y.k/ 2 Rmy and u.k/ 2 Rmu are the process’s my -dimensional
output and mu -dimensional input variables, respectively, whereas ny and nu refer to the maximum
lags in the process output and input, respectively.

2.1. Multistep model prediction

Nonlinear MPC involves using a nonlinear model to predict future process behaviour so that the
controlled variables can be manipulated in order for the controlled process to meet the desired tar-
get. The performance of MPC relies crucially on the quality of the process model, which is used
to generate these predictions [6]. A number of researchers have developed nonlinear models using
a variety of technologies [34], including first-principle and black-box approaches. A first-principle
model can be valid globally and can predict the system dynamics over the entire operating range.
However, development of a reliable first-principle model is a difficult and time-consuming task.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
448 X. TIAN ET AL.

On the other hand, the nonlinear black-box models have certain advantages over the first-principle
models in terms of development time and modelling efforts. We adopt a general NARX model
structure [35, 36] to model the process described in (1)
 
ym .k/ D fm y.k  1/,    , y.k  nO y /, u.k  1/,    , u.k  nO u / , (2)

where fm ./ denotes the constructed my -dimensional model vector mapping, ym .k/ 2 Rmy is the
model output, and nO y and nO u denote the maximum lags of the output and input in the model, respec-
tively. Note that nO y and nO u are not necessarily equal to ny and nu . The model structure (2) is
widely adopted in data-driven nonlinear modelling and control applications [36, 37]. Many exist-
ing NMPC algorithms, which employ FNN [38] or support vector regression (SVR) [39] based
nonlinear prediction models, also adopt this model structure to model the process.
Accurate multistep prediction is essential if MPC is to be successfully executed. However, the
model (2) is identified in the form of the one-step-ahead prediction for the process output. To realise
multistep-ahead prediction, the past process output samples in the model input vector are gradually
replaced by their predicted values. That is, by feeding back the model outputs, the one-step-ahead
predictive model can be recursively cascaded to itself to generate the future predictions for the pro-
cess output [40]. This procedure is repeated until the predicted output trajectory over the whole
prediction horizon is obtained. More specifically, by defining ym .k C i jk/ as the prediction for the
process output y.k C i / based on the available measurements at the sampling instant k, then this
multistep model prediction procedure is described as follows:
ym .k C i  j jk/ D y.k C i  j /, 8j > i ,

ym .k C i jk/ D fm ym .k C i  1jk/,    , ym .k C i  nO y jk/, (3)
u.k C i  1/,    , u.k C i  nO u // , 1 6 i 6 P ,
where P represents the prediction horizon.

2.2. Nonlinear model predictive control formulation

On the basis of the multistep prediction (3), the MPC design problem at the sampling instant k is
defined as the constrained optimisation problem whereby the future manipulated input moves are
determined by minimising the objective function:
P
X  T  
min ysp  yp .k C i jk/ Q ysp  yp .k C i jk/
u.kjk/,,u.kCM 1jk/
iD1
M
X 1
C .u.k C j jk//T Ru.k C j jk/, (4)
j D0

subject to the following constraints:

yp .k C i jk/ D ym .k C i jk/, 1 6 i 6 P,
(5)
ymin 6 yp .k C i jk/ 6 ymax , 1 6 i 6 P,

umin 6 u.k C j jk/ 6 umax , 0 6 j 6 M  1,

(6)
umin 6 u.k C j jk/ 6 umax , 0 6 j 6 M  1,

where ysp 2 Rmy is the set point of the controlled variable, Q is the my  my positive-definite
output error weighting matrix, M denotes the control horizon, R is the mu  mu positive-definite
input move weighting matrix and u.k C j jk/ D u.k C j jk/  u.k C j  1jk/ is the future control
increment, whereas u.k Cj jk/ denotes the control action at the future sampling instant k Cj based
on the available measurements at k. It is assumed that u.k C j jk/ D 0 for j > M . The vector
inequality in (5) and (6) is defined as the element-wise inequality operation.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 449

Given the multistep prediction model (3), the two weighting matrices Q and R and the constraints
on the output, control input and control input move, ymin , ymax , umin , umax , umin and umax, the
resulting constrained nonlinear optimisation problem can be solved using standard nonlinear pro-
gramming techniques. The controller is implemented in a moving horizon framework. Thus, after
solving the optimisation problem, only the first control action, denoted as uopt .kjk/, is implemented
on the plant, and the optimisation problem is reformulated at the next sampling instant k C 1, based
on the updated information from the process. The weighting parameters, Q and R, and the predic-
tion and control horizons, P and M , can be tuned to achieve the desired closed-loop performance,
and different tuning criteria can be found in [1].
Under the assumption that the actual system is identical to the model used for prediction, the
asymptotic stability of the NMPC controller may be guaranteed with several techniques (see, e.g.
[41–43]). In [44], the robustness properties of the NMPC have been investigated with respect to the
process gain and additive perturbations. Assuming Lipschitz continuity of the MPC law, the authors
of [45] have investigated the perturbed asymptotic stability of the NMPC in the face of decaying
perturbations. In [46], a robust feasible NMPC is obtained for a given bound on the admissible
uncertainties. Some results have been also proposed for the solution of the tracking problem with
guaranteed stability. In [47], for example, the tracking of constant signals is solved for the system
described by an input/output NARX model. Generally, some additional stability constraints should
be introduced to ensure the stability of the closed-loop system [9, 31, 32]. However, it is noted
that the introduction of additional stability constraints would always result in some conservatism
in the design and performance of MPC controller, particularly in most real-world applications.
Comprehensive review on stability and robustness of NMPC can be found in [48, 49]. In indus-
trial applications of NMPC, however, the stability of the closed-loop system and the performance
of the control system are traditionally achieved by adjusting the prediction and control horizons
and the weighting matrices Q and R via trial and error, which is also the approach we adopted in
this work.

3. OFFSET-FREE NONLINEAR MODEL PREDICTIVE CONTROL

In the standard NMPC formulation discussed in Section 2, control feedback is manifested as

follows: the manipulated input computed for the present time step is implemented on the plant,
then the prediction horizon is moved forward one step and the problem is resolved using the new
process measurements. Although this formulation exhibits certain robustness to disturbances [7, 8],
the robustness of this NMPC does not mean that it can achieve steady-state offset-free control
in practical situations. Therefore, some form of correcting mechanism is required to improve the
closed-loop performance in the presence of plant–model mismatch and/or unmeasured disturbances.

3.1. Dynamic matrix control-like correcting scheme for offset-free control

As discussed in the Introduction section, most current NMPC implementations use a constant bias
calculated on the basis of the current measurement to correct the model–process mismatch [1, 3].
The bias is generated by comparing the measured process output y.k/ and the predicted process
output ym .k/ at the current sampling instant k as
e.k/ D y.k/  ym .k/, (7)
where ym .k/ D ym .kjk/ is the model prediction generated by (3) at k. This bias is assumed to
remain constant in the future and is utilised to correct the prediction model (3) as
yp .k C i jk/ D ym .k C i jk/ C e.k/, 1 6 i 6 P . (8)
In other words, the NMPC algorithm based on the DMC-like correcting scheme is obtained by
simply replacing the constraints (5) in the standard NMPC algorithm with
yp .k C i jk/ D ym .k C i jk/ C e.k/, 1 6 i 6 P ,
(9)
ymin 6 yp .k C i jk/ 6 ymax , 1 6 i 6 P .

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
450 X. TIAN ET AL.

This approach is widely used in industrial MPC implementations, as it can eliminate steady-state
offset in some cases [1, 3]. However, we point out that it cannot eliminate the steady-state offset in
general. This issue will be further discussed later. In the simulation study, it will also be confirmed
that this DMC-like correcting scheme cannot guarantee offset-free control in the presence of strong
plant–model mismatch and/or unmeasured disturbances.

3.2. Proposed correcting scheme for offset-free control

Because the aforementioned DMC-like correcting scheme cannot guarantee steady-state offset-free
behaviour in general, alternative correcting scheme is required to eliminate steady-state offset under
the generic plant–model mismatch and unmeasured disturbances. Unlike the DMC-like correcting
scheme that only uses the constant bias or output feedback error (7) to correct the multistep predic-
tions generated by the multistep predictive model (3), we propose a correcting scheme in which the
previous model predictions are first corrected by the output feedback error term (7) and then used
as the input of the multistep predictive model to provide the further prediction. Specifically, given
the output feedback error (7), the multistep prediction procedure of (3) is modified as

yQ m .k C i  j jk/ D y.k C i  j /, 8j > i ,


yQ m .k C i jk/ D fm yQ m .k C i  1jk/,    , yQ m .k C i  nO y jk/, (10)
u.k C i  1/,    , u.k C i  nO u // C e.k/, 1 6 i 6 P .

The resulting new multistep NMPC design is obtained by minimising the objective function (4)
subject to the constraints (6) and

yp .k C i jk/ D yQ m .k C i jk/, 1 6 i 6 P,
(11)
ymin 6 yp .k C i jk/ 6 ymax , 1 6 i 6 P.

The key difference to the DMC-like correcting scheme is that the output feedback error e.k/ is
incorporated into the multistep model prediction procedure. Therefore, the plant–model mismatch
and/or unmeasured disturbances are continuously compensated in the multistep recursive model
prediction propagation (10), and this improves the accuracy of the multistep predictions which in
turn results in an enhanced performance of the NMPC algorithm. By contrast, the DMC-like feed-
back scheme only works in the situation where the plant–model mismatch manifests as an additive
output disturbance. We now show that the NMPC algorithm with the proposed correcting scheme
can eliminate the steady-state error to achieve offset-free behaviour.

Proposition 1
Assume that
(1) The closed-loop system is stable.
(2) The set point ysp is reachable.
Then, for asymptotically constant set points and disturbances, there will be no offset.

Proof
Assumption 1 implies that the closed-loop system can reach an asymptotically stable equilibrium
point defined by

y1 D f .y1 ,    , y1 , u1 ,    , u1 / ,

where the superscript 1 denotes the steady state ,whereas y1 and u1 are the process’s steady-state
output and input, respectively. From the identified model (2), the model output

y1 1 1 1 1
m D fm .y ,    , y , u ,    , u /

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 451

is a constant at the process’s steady state fy1 , u1 g. Then, the bias e1 D y1  y1 m ¤ 0 is

always a constant. According to the multistep prediction procedure (10), we have the one-step-ahead
prediction
 1 
yQ 1
m .k C 1jk/ D fm yQ m .kjk/,    , yQ 1
m .k C 1  n O y jk/, u1 ,    , u1 C e1
D fm .y1 ,    , y1 , u1 ,    , u1 / C e1 D y1 m Ce
1
D y1 ,

the two-step-ahead prediction

 1 
yQ 1
m .k C 2jk/ D fm yQ m .k C 1jk/, y1 ,    , y1 , u1 ,    , u1 C e1
D fm .y1 ,    , y1 , u1 ,    , u1 / C e1 D y1 m Ce
1
D y1 ,

and recursively the three or more step-ahead predictions

yQ 1 1
m .k C i jk/ D y , i D 3,    , P .

When the process reaches a steady-state,

u.k C j jk/ D u.k C j jk/  u.k C j  1jk/ D 0, 8j > 0. (12)

Thus, the inequality constraints on u.k C j jk/ in (6) can be removed, and the NMPC formulation
described by (4), (11) and (6) is reduced to
P 
P T  
min ysp  yp .k C i jk/ Q ysp  yp .k C i jk/ ,
u.kjk/,,u.kCM 1jk/ iD1
s.t. yp .k C i jk/ D y1 , 1 6 i 6 P , (13)
ymin 6 y1 6 ymax , 1 6 i 6 P ,
umin 6 u1 6 umax , 1 6 j 6 M  1.

Assumption 2 indicates that the set point ysp and the corresponding control input usp asymptotically
fulfil
ymin 6 ysp 6 ymax ,
(14)
umin 6 usp 6 umax.

From the solution of (13), we immediately obtain y1 D ysp . 

Remark 1
The main advantage of our approach is that the computation of the steady-state value of the con-
trol input associated with the set point can be avoided. If the set points are not reachable, then
 T  
the proposed approach minimises the steady-state output difference ysp  y1 Q ysp  y1 .
Assumption 1 indicates that the NMPC controller can asymptotically stabilise the process (1)
with the given multistep-ahead predictive model (10). This type of assumption is often seen in the
literature. For example, this assumption was also adopted in [18, 50, 51], in order to show that the
controller considered in [18, 50, 51] yields a zero steady-state offset.

Remark 2
Under the same assumptions as given in Proposition 1, the DMC-like method cannot guarantee the
same offset-free control performance. This may be explained as follows. When the process reaches
the steady state fy1 , u1 g, from (3) and (9), we have the error-corrected one-step-ahead prediction

yp .k C 1jk/ D y1
m .k C 1jk/ C e
1
D fm .y1 ,    , y1 , u1 ,    , u1 / C e1 D y1
m Ce
1
D y1 ,

the error-corrected two-step-ahead prediction

 
yp .k C 2jk/ D fm y1 1 1 1
m,y ,,y ,u ,,u
1
C e1 ¤ y 1

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
452 X. TIAN ET AL.

and the error-corrected three or more step-ahead predictions

yp .k C i jk/ ¤ y1 , i D 3,    , P .
Thus, the NMPC formulation described by (4), (9) and (6) is reduced to
P 
P T  
min ysp  yp .k C i jk/ Q ysp  yp .k C i jk/ ,
u.kjk/,,u.kCM 1jk/ iD1
s.t. yp .k C i jk/ D y1 1
m .k C i jk/ C e , 1 6 i 6 P ,
(15)
ymin 6 yp .k C i jk/ 6 ymax , 1 6 i 6 P ,
umin 6 u1 6 umax , 1 6 j 6 M  1.
It then becomes obvious that the solution of the optimisation problem (15) can only be obtained in
the sense of least-squares error, and in general, the offset will occur, namely, y1 ¤ ysp .

Remark 3
In practical applications of MPC, if the desired set point changes significantly or large persistent
disturbances occur, the feasibility of the controller may be lost. This is because in these cases, there
may be a conflict between the constraints (6) and the constraints (11) that may turn the control prob-
lem infeasible. Two approaches can be employed to recover the feasibility of the NMPC whenever
the controller feasibility is lost due to large set-point changes or persistent disturbances.
The first approach introduces an artificial set point to the MPC problem as a new decision variable
[52]. The distance between the artificial set point yO sp 2 Rmy and the real one ysp is then penalised
in the cost function, and the resulting optimisation problem is expressed as
P
X  T  
min yO sp  yp .k C i jk/ Q yO sp  yp .k C i jk/
u.kjk/,,u.kCM 1jk/,Oysp
iD1
M
X 1
 T  
C .u.k C j jk//T Ru.k C j jk/ C yO sp  ysp Qs yO sp  ysp , (16)
j D0

subject to the constraints (11) and (6), where Qs is the weighting matrix for penalising the difference
between the artificial set point and the actual set point. The introduction of the artificial set point
as the new decision variable can explicitly address the feasibility of the controller and ensure the
convergence to the target if admissible. Furthermore, the domain of attraction of this controller is
potentially larger than those of reference governors and standard predictive controllers [53].
In the second approach, the soft constraints are introduced to extend the constraint handling capa-
bilities of MPC [54, 55]. The output predictions are softened by means of slack variables and the
MPC optimisation problem is reformulated as following
P
X  T  
min ysp  yp .k C i jk/  " Q ysp  yp .k C i jk/  "
u.kjk/, ,u.kCM 1jk/,","min ,"max
iD1
M
X 1
C .u.k C j jk//T Ru.k C j jk/ C "T Q" " C min k"mink2 C max k"maxk2 , (17)
j D0

subject to the constraints (6) and

yp .k C i jk/ D yQ m .k C i jk/, 1 6 i 6 P ,
ymin  "min 6 yp .k C i jk/ 6 ymax C "max , 1 6 i 6 P , (18)
"min > 0, "max > 0,

where " 2 Rmy is a slack variable, "min and "max 2 Rmy are the lower-bound and upper-bound slack
variables, respectively, Q" is the weighting matrix for the slack variable " and min and max > 0
are the two weightings for the lower-bound and upper-bound slack variables, respectively. With the

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 453

help of the slack variables, the original hard constraint (11) posed on the magnitude of the output
variable can be temporarily violated. This enlarges the admissible set and prevents the infeasibility
of the solution from happening. The degree of output constraint violation is minimised by the last
three terms of the objective function (17). In addition, it provides a flexible mechanism for shaping
the transient response by tuning Q and Q" . The controller obtained from the solution to the prob-
lems (17), (6) and (18) shares the same properties related to the feasibility of the set-point tracking
problem as the controller defined in (16), (6) and (11).

4. SIMULATION STUDY

The CSTR [27, 29] was employed to compare the closed-loop performance of several MPC
algorithms and to demonstrate that the proposed NMPC scheme can achieve offset-free control in
the presence of significant plant–model mismatch and unmeasured disturbances. A Van der Vusse
chemical reaction takes place in the CSTR with a cooling jacket. The main reaction is given by
the transformation of cyclopentadiene, A, to the product cyclopentenol, B, which in turn produces
cyclopentanediol, C, in an unwanted consecutive reaction. The initial reactant, A, also reacts in an
unwanted parallel reaction to give the by-product dicyclopentadiene, D. Symbolically, this chemical
reaction is described as
1 2
A ! B ! C
3
2A ! D
The dynamics of the reactor can be described by the following nonlinear differential equations that
are derived from the component balances for substances A and B and from the energy balances for
the reactor and cooling jacket
CPa D q .Ca0  Ca /  1 .T /Ca  3 .T /Ca2 ,
CP b D qCb C 1 .T /Ca  2 .T /Cb ,
1  
TP D q.T0  T /  1 .T /Ca H1 C 2 .T /Cb H2 C 3 .T /Ca2 H3
CP (19)
w AR
C .TK  T / ,
CP VR
1
TPK D .QK C w AR .T  TK // .
mK CPK
The reaction velocities i are assumed to depend on the temperature via the Arrhenius law
 
Ei
i .T / D i0 exp  , i D 1, 2, 3. (20)
R.T C 273.15/
The states of the system consist of the concentrations of components A and B, Ca > 0 and Cb > 0,
as well as the temperatures, T and TK , occurring in the reactor and cooling jacket, respectively.
This is a two-input two-output process. The two controlled variables are y1 D Cb and y2 D T , the
concentration of substance B and the temperature in the reactor, whereas the two control inputs are
u1 D q and u2 D QK , the flow rate (scaled to the volume of the reactor) and the cooling power,
which are subject to the constraints
q 2 Œ3, 35h1 , QK 2 Œ9000, 0 kJ/h, jqj 6 5h1 , jQK j 6 2000 kJ/h.
The values for the physical and chemical parameters in (19) and (20) are listed in Table I.
The reactor exhibits input multiplicities, that is, more than one set of the manipulated variable
values can produce an identical steady-state output. Input multiplicities exist when the steady-state
gain matrix becomes singular in the operating region. During the actual operation of the plant, it
is always required to operate at the optimal operating point that can produce the maximum con-
centration of B. Although control at such an operating point appears to be economically the most

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
454 X. TIAN ET AL.

Table I. Physical and chemical parameters of the CSTR process.

Name of parameter Symbol Value of parameter
Pre-exponential factor 10 1.287  1012 h1
Pre-exponential factor 20 1.287  1012 h1
Pre-exponential factor 30 9.043  109 h1
Activation energy for reaction 1 E1 81.1344 kJ/mol
Activation energy for reaction 2 E2 81.1344 kJ/mol
Activation energy for reaction 3 E3 71.1712 kJ/mol
Enthalpies of reaction 1 H1 4.2 kJ/(moll)
Enthalpies of reaction 2 H2 11.0 kJ/(moll)
Enthalpies of reaction 3 H3 41.85 kJ/(moll)
Density  0.9342 kg/l
Heat capacity CP 3.01 kJ/(kgK)
Heat transfer coefficient for cooling jacket w 4032 kJ/(hm2 K)
Surface of cooling jacket AR 0.215 m2
Reactor volume VR 0.01 m3
Coolant mass mK 5.0 kg
Heat capacity of coolant CPK 2.0 kJ/(kgK)
Universal gas constant R 8.314 J/(molK)
Concentration of initial reactant Ca0 5.10 mol/l
Feed temperature T0 104.9 ı C

desirable, the resulting control problem is difficult to handle because there exists a singularity at
the optimum operating point. The existence of such singular points renders the task of controlling a
system exhibiting input multiplicities extremely difficult [56].

4.1. Modelling and model evaluation

A dynamic SVR model with the NARX structure (2) was used to model the CSTR on the basis of
an input–output data set obtained from the CSTR process. The model took the form
ym,i .k/ D fm,i .y1 .k  1/, y2 .k  1/, u1 .k  1/, u2 .k  1//, i D 1, 2,
with the input and output lags set to nO y D nO u D 1. The tuning parameters of the SVR model with
the Gaussian kernel were chosen as follows: the maximum tolerable error  D 0.05, the regulari-
sation parameter C D 50 and the kernel width  D 0.005 were used for identifying fm,1 , whereas
 D 0.05, C D 20 and  D 0.01 were used for identifying fm,2 . In the simulation, the sampling
period was set to 18 s, and 800 input–output data were gathered by injecting the random step signals
with the uniform distributions in Œ3, 35h1 and Œ9000, 0 kJ/h as the control inputs q and QK ,
respectively, with the switching probability of 0.2. The first 200 training data were used to train
the two SVR models with LIBSVM toolbox [57], which led to 43 support vectors for fm,1 and
31 support vectors for fm,2 . The last 600 data points were employed to validate the two obtained
models.
Figure 1 compares the two process outputs with those of the two identified SVR models over
both the training and validation data sets, where it is observed that the two one-step-ahead predic-
tion curves almost coincide with the actual process outputs. The root mean squares error (RMSE)
between the actual process output yi .k/ and the j -step-ahead model prediction ym,i .k C j jk/
v
uN
uX
RMSEi .j / D t .yi .k/  ym,i .k C j jk//2 =N , i D 1, 2
kD1

was evaluated over the validation data set, and the values of RMSE obtained for the two process
outputs Cb and T , respectively, are listed in Table II. The one-step-ahead RMSE values given in
Table II together with the results of Figure 1 indicate that the two identified SVR models achieved
very small one-step-ahead prediction errors. However, the multistep predictive error accumulated

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 455

Figure 1. Comparison of the actual CSTR process outputs (solid) and the one-step-ahead predictions from
the identified SVR models (dashed) over the training (first 200 points) and validation (last 600 points) data.
The sampling period is 18 s.

Table II. RMSE between the actual process output and the j -step-ahead model prediction for
the two process outputs Cb and T .
j 1 2 3 4 5 6 7 8
RMSE for Cb 0.0214 0.0277 0.0320 0.0353 0.0390 0.0402 0.0413 0.0422
RMSE for T 0.2220 0.4245 0.6171 0.7947 0.9573 1.1033 1.2355 1.3542

and increased as the prediction horizon increased. This can be clearly observed from Table II.
Figure 2 compares the steady-state behaviours of the process outputs Cb and T as the function of
the control input q, when fixing the other control input to QK D 1118 kJ/h, with the steady-state
behaviours generated by the identified SVR model. It is observed from Figure 2 that the steady-state
operating locus of the actual plant displays a gain change at the flow rate q D 14.19h1 , which
corresponds to the operation point Cb D 1.09 mol/l and T D 114.2 ı C. By contrast, the gain change
for the SVR model occurs at q D 6.88h1 , and the model steady-state outputs at this point are
Cb D 1.171 mol/l and T D 113.6 ı C. The existence of such significant plant–model mismatch
served well our purpose of testing the performance of various NMPC schemes in the presence of
plant–model mismatch.

4.2. Set-point tracking

The three NMPC schemes implemented in the simulation study are first summarised:
 NMPC1: the NMPC algorithm based on the ‘perfect’ process model without any plant–model
mismatch. This served as the ‘ideal’ NMPC algorithm.
 NMPC2: the NMPC algorithm based on the proposed correcting scheme. This was the proposed
‘new’ NMPC algorithm.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
456 X. TIAN ET AL.

Figure 2. Steady-state behaviours of the actual CSTR process (solid) and the identified SVR models
(dashed) as the function of the control input q, when setting the other control input to QK D 1118 kJ/h.

 NMPC3: the NMPC algorithm based on the DMC-like correcting scheme. This was the ‘most
widely’ implemented NMPC algorithm.
The identified SVR model as given in Section 4.1 was used for designing the NMPC2 and NMPC3.
The prediction horizon and the control horizon were set to P D 13 and M D 1, respectively. The
output and input weighting matrices in the cost function were chosen to be Q D diagf200, 10g and
R D diagf1, 0.02g, respectively. An artificial set point yO sp 2 R2 was introduced to the NMPC optimi-
sation problem to ensure the feasibility of the NMPC during the change of set point. The weighting
matrix for penalising the difference between the artificial set point and the actual set point was set
to Qs D diagf400, 50g. Ability of the proposed NMPC scheme to achieve steady-state offset-free
control was investigated when the CSTR went through a series of set-point changes:

t D 0  0.6 h: process set point .1.09 mol/l, 114.2 ı C/

t D 0.6  1.2 h: process set point .1.06 mol/l, 113.0 ı C/
t D 1.2  1.8 h: process set point .1.09 mol/l, 114.2 ı C/

Figures 3–5 illustrate the closed-loop performance of the three NMPC controllers for set-point
tracking, where the performance of the NMPC1 serves as a lower bound. As can be seen from
Figure 5, the output feedback error e.k/ D 0 for the NMPC1 because the exact process model was
used, whereas e.k/ ¤ 0 for the other two MPC systems as the plant–model mismatch existed in
these two cases. Whereas the output feedback error is the difference between the actual process out-
put and the model output, which reflects the plant–model mismatch, the offset error is the difference
between the actual process output and its set point, and it is highly desired that a control system
can achieve the steady-state zero offset error. From Figure 3, it can be seen that the set-point track-
ing performance of the NMPC2 with the new multistep prediction procedure (10) was very close
to the ideal performance of the NMPC1. This demonstrates that the proposed NMPC2 can effec-
tively correct the plant–model mismatch to achieve offset-free control. By contrast, the NMPC3
with the DMC-type feedback correction could not achieve the offset-free control, as can be clearly
seen from Figure 3. Note that the steady-state gain of the CSTR reduces to zero at the set point
(1.09 mol/l, 114.2 ı C), which can lead to controller output fluctuations at the neighbourhood of this
point. The ability of the proposed correcting scheme to ‘predict’ the system dynamics close to the
singular points accurately, however, ensured that the control input fluctuations were not large even
when the process’s steady-state gain was close to zero. By contrast, the controller response of the
NMPC3 exhibited large overshoot behaviour, as is clearly shown in Figure 4.
The sum of squared errors (SSE) between the desired set point and the actual process output over
the sampling interval ŒN1 , N2 , defined as
N2
X  2
SSEi D ysp,i  yi .k/ , i D 1, 2,
kDN1

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 457

Figure 3. Set-point tracking performance of various NMPC controllers: process output y.k/.

Figure 4. Set-point tracking performance of various NMPC controllers: controller output u.k/.

was used to quantify the offset error performance of a controller. Table III compares the SSE values
over the six sampling intervals of each length 0.3 h for the three NMPC algorithms investigated. As
can be seen from Table III, the NMPC3 cannot achieve offset-free control. By contrast, the NMPC2

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
458 X. TIAN ET AL.

Figure 5. Set-point tracking performance of various NMPC controllers: output feedback error e.k/.

Table III. Comparison of the SSEs between the set point and the actual
process output for the three NMPC controllers in set-point tracking.
Sampling interval (h) SSE for Cb SSE for T
NMPC1 0 – 0.3 0.0083 15.6026
0.3 – 0.6 1.3689e  4 1.0046
0.6 – 0.9 7.7499e  4 3.5356
0.9 – 1.2 3.7521e  7 1.8926e  4
1.2 – 1.5 5.9986e  4 4.0840
1.5 – 1.8 8.7761e  6 0.0047
NMPC2 0 – 0.3 0.0101 29.4195
0.3 – 0.6 5.8316e  4 1.2510
0.6 – 0.9 0.0014 7.3578
0.9 – 1.2 4.3571e  7 0.0025
1.2 – 1.5 0.0011 7.1895
1.5 – 1.8 6.4498e  6 1.8934e  4
NMPC3 0 – 0.3 0.0129 109.6343
0.3 – 0.6 6.1421e  4 299.5152
0.6 – 0.9 0.0213 164.5560
0.9 – 1.2 0.0181 11.4718
1.2 – 1.5 5.3736e  4 97.7208
1.5 – 1.8 1.7602e  4 288.0552

can realise the offset-free control after 0.3 h of each set-point change. This again demonstrates that
the new NMPC controller can effectively utilise the available output feedback error information to
remove the influence of plant–model mismatch in the proposed multistep-ahead predictive proce-
dure. The results of Table III also confirm that the set-point tracking performance of the NMPC2 is
close to that of the ideal NMPC1.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 459

4.3. Unmeasured disturbance rejection

During the operation of a chemical process, unmeasured disturbances may arise owing to var-
ious factors, such as changes in the quality of raw material or interactions between different
plant units. We evaluated the ability of the proposed NMPC formulation, the NMPC2, for dis-
turbance rejection. In the simulation, the feed of the reactor was assumed to come from an
upstream unit, and the feed temperature T0 could vary in the range Œ100 ı C, 115 ı C. The
objective of this part of simulation was to shift the process from an initial operation point
.Ca , Cb , T , TK / D .1 mol/l, 0.5 mol/l, 100 ı C, 100 ı C/ at k D 0 to the nominal optimal oper-
ating point .2.14 mol/l, 1.09 mol/l, 114.2 ı C, 112.9 ı C/ with the set point Cb D 1.09 mol/l and
T D 114.2 ı C when the process was subject to the unmeasured disturbance described by
8
< 104.9 ı C, k < 20,
T0 .k/ D 104.9 C k20 ı
 5.1 C, 20 6 k < 50,
: 30
110.0 ı C, k > 50.
In addition, the measurements of the process outputs, Cb and T , were corrupted by the zero-mean
Gaussian white noise signals with the standard deviations of 0.08 and 0.3, respectively. This control
problem was especially challenging, as the controller attempted to drive the reactor to the opera-
tion point that is characterised by a change of sign in the steady-state gain. Moreover, the process
suffered from unmeasured disturbance, and the measurements were corrected by noise.
Closed-loop performance of the three NMPC controllers are depicted in Figures 6–8. The tracking
results shown in Figure 6 confirm that the performance of the NMPC2 was still close to that of the
ideal NMPC1, under the unmeasured disturbance and measurement noise. It can be observed from
Figure 7 that, for the NMPC1 and NMPC2, the manipulated variables QK decreased appropriately
in responding to the effects of the ramp disturbance T0 .k/ for 20 6 k < 50. For this challeng-
ing control problem, the process outputs of the closed-loop systems under the controllers NMPC1
and NMPC2 achieved relative smooth transitions to the optimal operation point, and they remained

Figure 6. Disturbance rejection performance of various NMPC controllers: process output y.k/.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
460 X. TIAN ET AL.

Figure 7. Disturbance rejection performance of various NMPC controllers: controller output u.k/.

Figure 8. Disturbance rejection performance of various NMPC controllers: output feedback error e.k/.

at the set point without obvious steady-state error. The NMPC3, by contrast, exhibited overshoot
behaviour and experienced clear steady-state error for the set point T . As the process output
measurements were corrupted by the random noise, the output feedback error was no longer

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 461

Table IV. Comparison of the SSEs between the set point and the actual
process output for the three NMPC controllers, averaged over 50 runs,
in disturbance rejection.
Sampling interval (h) SSE for Cb SSE for T
NMPC1 0 – 0.3 5.2780 2.2183e C 3
0.3 – 0.6 6.0816e  4 0.4160
NMPC2 0 – 0.3 6.1125 2.3207e C 3
0.3 – 0.6 6.4123e  4 0.5139
NMPC3 0 – 0.3 6.2801 2.5753e C 3
0.3 – 0.6 0.0046 192.7004

identical to zero for the ideal NMPC1, which is confirmed in Figure 8. The simulation was then
repeated for 50 runs, and the means of the SSEs over the two sampling intervals, Œ0, 0.3 and
Œ0.3, 0.6 h, are listed in Table IV for the three NMPC systems.

5. CONCLUSIONS

The performance of the standard NMPC is critically affected by the accuracy of the prediction
model employed. However, an accurate process model can rarely be found with affordable efforts.
Moreover, in practical applications, unmeasured disturbances and changes in the process param-
eters inevitably result in significant mismatch in the plant and model behaviours. Under these
circumstances, most current NMPC implementations often experience steady-state offset errors,
namely, a persistent discrepancy between the desired set-point value and the actual output of
the process, leading to a significant degradation in the closed-loop performance. In this contri-
bution, a methodology for achieving steady-state offset-free control has been proposed for the
NMPC implementation under significant plant–model mismatch and/or unmeasured disturbances.
The methodology is based on incorporating the available output feedback error in the multistep
recursive prediction propagation, which explicitly accounts for the uncertainties arising from the
plant–model mismatch and unmeasured disturbances. In the proposed multistep prediction pro-
cedure, the prediction can effectively correct itself over the prediction horizon based only on the
previous output feedback error information. We have proved that the new NMPC formulation based
on the proposed correcting scheme is capable of realising steady-state offset-free control. The effec-
tiveness of the proposed NMPC scheme has been demonstrated in the simulation study involving
control of the CSTR process. The results obtained have confirmed that the new NMPC formula-
tion with the proposed multistep prediction procedure is able to achieve the satisfactory closed-loop
performance and offset-free control even in the presence of significant plant–model mismatch and
unmeasured disturbances.

ACKNOWLEDGEMENTS
This work was supported by the National Natural Science Foundation of China under grant 60974008 and
the National High Technology Research and Development Program of China under grant 2007AA04Z193.
The authors would like to thank anonymous reviewers for their constructive comments that strengthened the
manuscript significantly.

REFERENCES
1. Qin SJ, Badgwell TA. A survey of industrial model predictive control technology. Control Engineering Practice
2003; 11(7):733–764.
2. Engell S. Feedback control for optimal process operation. Journal of Process Control 2007; 17(3):203–219.
3. Bequette BW. Non-linear model predictive control: a personal retrospective. Canadian Journal of Chemical
Engineering 2007; 85(4):408–415.
4. Magni L, Raimondo MD, Allgöwer F (eds). Nonlinear Model Predictive Control: Towards New Challenging
Applications, Lecture Notes in Control and Information Sciences, Vol. 384. Springer-Verlag: Berlin, 2009.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
462 X. TIAN ET AL.

5. Badwe AS, Patwardhan RS, Shah SC, Patwardhan SC, Gudi RD. Quantifying the impact of model-plant mismatch
on controller performance. Journal of Process Control 2010; 20(4):408–425.
6. Tenny MJ, Rawlings JB, Wright SJ. Closed-loop behavior of nonlinear model predictive control. AIChE Journal
2004; 50(9):2142–2154.
7. Faanes A, Skogestad S. Offset-free tracking of model predictive control with model mismatch: experimental results.
Industrail & Engineeering Chemistry Research 2005; 44(11):3966–3972.
8. Gerkšič S, Strmčnik S, van den Boom T. Feedback action in predictive control: an experimental case study. Control
Engineering Practice 2008; 16(3):321–332.
9. Magni L, De Nicolao G, Scattolini R. Output feedback and tracking of nonlinear systems with model predictive
control. Automatica 2001; 37(10):1601–1607.
10. González AH, Adam EJ, Marchetti JL. Conditions for offset elimination in state space receding horizon controllers:
a tutorial analysis. Chemical Engineering and Processing: Process Intensification 2008; 47(12):2184–2194.
11. Maeder U, Morari M. Offset-free reference tracking with model predictive control. Automatica 2010;
46(9):1469–1476.
12. Rao CV, Rawlings JB. Steady states and constraints in model predictive control. AICHE Journal 1999;
45(6):1266–1278.
13. Muske KR, Badgwell TA. Disturbance modeling for offset-free linear model predictive control. Journal of Process
Control 2002; 12(5):617–632.
14. Pannocchia G, Rawlings JB. Disturbance models for offset-free model predictive control. AIChE Journal 2003;
49(2):426–437.
15. Pannocchia G. Robust disturbance modeling for model predictive control with application to multivariable
ill-conditioned processes. Journal of Process Control 2003; 13(8):693–701.
16. Pannocchia G, Bemporad A. Combined design of disturbance model and observer for offset-free model predictive
control. IEEE Transactions on Automatic Control 2007; 52(6):1048–1053.
17. Huang R, Patwardhan SC, Biegler LT. Robust extended Kalman filter based nonlinear model predictive con-
trol formulation. Proceeding of Joint 48th IEEE CDC and 28th CCC, Shanghai, China, December 15–18, 2009;
8046–8051.
18. Huang R, Biegler LT, Patwardhan SC. Fast offset-free nonlinear model predictive control based on moving horizon
estimation. Industrial & Engineering Chemistry Research 2010; 49(17):7882–7890.
19. Lee LM, Lee JH. Approximate dynamic programming-based approaches for input-output data-driven control of
nonlinear processes. Automatica 2005; 41(7):1281–1288.
20. Psichogios DM, Ungar LH. Direct and indirect model based control using artificial neural networks. Industrail &
Engineeering Chemistry Research 1991; 30(12):2564–2573.
21. Chu JZ, Tsai PF, Tsai WY, Jang SS, Shieh SS, Lin PH, Jiang SJ. Multi-step model predictive control based on
artificial neural network. Industrail & Engineeering Chemistry Research 2003; 42(21):5215–5228.
22. Lu CH, Tsai CC. Generalized predictive control using recurrent fuzzy neural networks for industrial processes.
Journal of Process Control 2007; 17(1):83–92.
23. Zhang T, Feng G, Zeng XJ. Output tracking of constrained nonlinear processes with offset-free input-to-state stable
fuzzy predictive control. Automatica 2009; 45(4):900–909.
24. Uppal A, Ray WH, Poore AB. On the dynamic behavior of continuous stirred tank reactors. Chemical Engineering
Science 1974; 29(4):967–985.
25. Calvet JP, Arkun Y. Feedforward and feedback linearization of nonlinear systems and its implementation using
internal model control (IMC). Industrail & Engineeering Chemistry Research 1988; 27(10):1822–1831.
26. Hernández E, Arkun Y. Study of the control-relevant properties of back propagation neural network models of
nonlinear dynamical systems. Computers & Chemical Engineering 1992; 16(4):227–240.
27. Chen H, Kremling A, Allgöwer F. Non-linear predictive control of a benchmark, CSTR. Proceeding of 3rd European
Control Conference, Rome, Italy, September 5–8, 1995; 3247–3252.
28. Zhan J, Ishida M. The multi-step predictive control of nonlinear SISO processes with a neural model predictive
control (NMPC) method. Computers & Chemical Engineering 1997; 21(2):201–210.
29. Chen H, Allgöwer F. A computationally attractive nonlinear predictive control scheme with guaranteed stability for
stable systems. Journal of Process Control 1998; 8(5–6):475–485.
30. Venkateswarlu C, Rao KV. Dynamic recurrent radial basis function network model predictive control of unstable
nonlinear processes. Chemical Engineering Science 2005; 60(23):6718–6732.
31. Magni L, Scattolini R. On the solution of the tracking problem for non-linear systems with MPC. International
Journal of Systems Science 2005; 36(8):477–484.
32. Magni L. On robust tracking with non-linear model predictive control. International Journal of Control 2002;
75(6):399–407.
33. Huusom JK, Poulsen NK, Jørgensen SB, Jørgensen JB. ARX-model based model predictive control with offset-free
tracking. Proceeding of 20th European Symposium on Computer Aided Process Engineering, Ischia, Naples, Italy,
June 6–9, 2010; 601–606.
34. Hong X, Mitchell RJ, Chen S, Harris CJ, Li K, Irwin GW. Model selection approaches for non-linear system
identification: a review. International Journal of Systems Science 2008; 39(10):925–946.
35. Chen S, Billings SA. Representation of non-linear systems: the NARMAX model. International Journal of Control
1989; 49(3):1013–1032.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs
 10991115, 2014, 3-5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/acs.2367 by IME - Instituto Militar de Engenharia, Wiley Online Library on [06/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OFFSET-FREE MODEL PREDICTIVE CONTROL 463

36. Billings SA, Chen S, Korenberg MJ. Identification of MIMO non-linear systems using a forward-regression
orthogonal estimator. International Journal of Control 1989; 49(6):2157–2189.
37. Chen S, Billings SA, Luo W. Orthogonal least squares methods and their application to nonlinear system
identification. International Journal of Control 1989; 50(5):1873–1896.
38. Chen S, Billings SA, Grant PM. Non-linear systems identification using neural networks. International Journal of
Control 1990; 51(6):1191–1214.
39. Vapnik VN. The Nature of Statistical Learning Theory. Springer-Verlag: New York, 1995.
40. Narendra KS, Parthasarathy K. Identification and control of dynamical systems using neural networks. IEEE
Transactions on Neural Networks 1990; 1(1):4–27.
41. Scokaert POM, Mayne DQ, Rawlings JB. Suboptimal model predictive control (feasibility implies stability). IEEE
Transactions on Automatic Control 1999; 44(3):648–654.
42. Chen H, Allgöwer F. A quasi-infinite horizon nonlinear model predictive control scheme with guaranteed stability.
Automatica 1998; 34(10):1205–1217.
43. Mayne DQ, Rawlings JB, Rao CV, Scokaert POM. Constrained model predictive control: stability and optimality.
Automatica 2000; 36(6):789–814.
44. De Nicolao G, Magni L, Scattolini R. On the robustness of receding-horizon control with terminal constraints. IEEE
Transactions on Automatic Control 1996; 41(3):451–453.
45. Scokaert POM, Rawlings JB, Meadows ES. Discrete-time stability with perturbations: application to model
predictive control. Automatica 1997; 33(3):463–470.
46. Limón Marreudo D, Álamo T, Camacho EF. Input-to-state stable MPC for constrained discrete-time nonlinear sys-
tems with bounded additive uncertainties. Proceeding of 41st IEEE CDC, Las Vegas, USA, December 10–13, 2002;
4619–4624.
47. De Nicolao G, Magni L, Scattolini R. Stabilizing predictive control of nonlinear ARX models. Automatica 1997;
33(9):1691–1697.
48. De Nicolao G, Magni L, Scattolini R. Stability and robustness of nonlinear receding horizon control. In Nonlinear
Model Predictive Control, Vol. 26, Allgöwer F, Zheng A (eds). Birkhauser: Verlag, 2000; 3–22.
49. Imsland L, Findeisen R, Bullinger E, Allgöwer F, Foss BA. A note on stability, robustness and performance of output
feedback nonlinear model predictive control. Journal of Process Control 2003; 13(7):633–644.
50. Meadows E, Rawlings J. Model predictive control. In Nonlinear Process Control, Henson MA, Seborg DE (eds).
Prentice Hall: Upper Saddle River, NJ, 1997; 233–310.
51. Economou CG, Morari M, Palsson BO. Internal model control: extension to nonlinear systems. Industrial &
Engineering Chemistry Process Design and Development 1986; 25(2):403–411.
52. Limon D, Alvarado I, Alamo T, Camacho EF. MPC for tracking of piecewise constant references for constrained
linear systems. Automatica 2008; 44(9):2382–2387.
53. Ferramosca A, Limon D, Alvarado I, Alamo T, Camacho EF. MPC for tracking with optimal closed-loop
performance. Automatica 2009; 45(8):1975–1978.
54. Rodrigues MA, Odloak D. MPC for stable linear systems with model uncertainty. Automatica 2003; 39(4):569–583.
55. Gonzalez AH, Odloak D. Robust model predictive controller with output feedback and target tracking. IET Control
Theory and Applications 2010; 4(8):1377–1390.
56. Sistu PB, Bequette BW. Model predictive control of processes with input multiplicities. Chemical Engineering
Science 1995; 50(6):921–936.
57. Chang CC, Lin CJ. LIBSVM: a library for support vector machines, September 2001. (Available from: http:
//www.csie.ntu.edu.tw/~cjlin/libsvm) [Accessed on July 1, 2011].

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2014; 28:444–463
DOI: 10.1002/acs