1997 - Adaptive Predictive Control - Limits of Stability
1997 - Adaptive Predictive Control - Limits of Stability
1997 - Adaptive Predictive Control - Limits of Stability
SUMMARY
This paper summarizes the stability results already derived for predictive and adaptive predictive control,
discusses them from an intuitive and practical implementation perspective and, from the same perspective,
illustrates them by means of two simulated examples. In this way it recalls the limits of stability when
applying predictive control and how they are related to the modelling errors, which may change as the
process dynamics changes. Also it recalls how, by adding adaptation to the predictive scheme, this source of
instability may be compensated for. Already within the adaptive predictive formulation it considers the
limits of stability for different scenarios, particularly when a reduced-order adaptive predictive model cannot
account for unmodelled process dynamics. ( 1997 by John Wiley & Sons, Ltd.
Int. J. Adapt. Control Signal Process., 11, 263283 (1997)
No. of Figures: 4
No. of Tables: 0
No. of References: 5
1. INTRODUCTION
During the last decade the increasing number of industrial implementations of adaptive predictive control systems (APCSs) has proved that this methodology is an advanced solution able to
answer satisfactorily the requirements demanded by the nature of the industrial processes and the
optimization of their operation.
A recent book1 has presented and summarized the concepts involved in adaptive predictive
control, the mathematical methodology derived from them, the stability theory, the first applications to unit plants and industrial processes, the formulation of a new concept of optimization
and the development of systems able to apply this technique systematically in an industrial
environment.
Predictive control, which is basically a subset of APCSs in which adaptation of the predictive
model does not take place, has also received considerable attention by the research community,
and many applications, under the generic name of model-based predictive control, have been
reported in academic, experimental and industrial contexts in a variety of areas. When predictive
control is applied, the fixed parameter predictive model must be obtained prior to the control
application and its performance may be satisfactory as long as certain linear time-invariant
*Correspondence to: J. M. Martin Sanchez, School of Mining Engineering, Technical University of Madrid and SCAP
Europa, Rosario Pino 1416, E-28020 Madrid, Spain
CCC 08906327/97/04026321$17.50
( 1997 by John Wiley & Sons, Ltd
264
process dynamic behaviour is maintained within the bounds allowed by the experimentally
proven intrinsic robustness of predictive control. However, in the so frequent presence of model
mismatches, the theoretical questions are: What are the limits of stability? When and how may
adaptation be a solution to overcome these stability limits? And, in a more difficult setting, can
adaptation ensure stability when we are in the presence of a difference in structure between the
process dynamics and the predictive model? Finally, in this last case, what are the limits for
stability? All these questions should be answered in a process time-invariant context as well as in
a time-variant one.
Taking the above considerations into account, this paper deals with these issues and summarizes the corresponding stability theory. Although this paper formally describes the different
scenarios and states the corresponding stability results, it does not include the corresponding
proofs already published,1 but focuses on their practical interpretation. Thus Section 2 of this
paper presents a general description of adaptive predictive control systems in which, under
a general time-varying formulation, the two most significant kind of processes considered by
APCS stability theory are defined. Section 2 also describes an adaptation mechanism, which has
been used to derive the most advanced stability results, and the concepts of projected desired
trajectory (PDT) and driving desired trajectory (DDT), which are of paramount importance in
the unification of the stability theory here considered. Section 3 presents and discusses the
stability results for predictive and adaptive predictive control, while Section 4 considers two
simulation examples that illustrate these results.
265
(c) he adaptation mechanism adjusts the predictive model parameters from the prediction
errors in order to make these errors tend towards zero efficiently. Likewise, it informs the
driver block of the deviation of its outputs with respect to the desired trajectory. In this way
the driver block may redefine the desired trajectory in a manner that is consistent with the
actual process output.
2.2. Description of the process
2.2.1. General description. Let us consider a general process description where the discretetime relation between the inputs and outputs may be given by the time-variant model
y (k)"h(k)T/ (k!d)#m(k)
(1)
!
!
Here h(k) represents the process parameter vector, m(k) represents the effect of the unmeasured
disturbances on the process output at instant k and
/ (k!d)T"[ y (k!d), y (k!d!1), . . . , u (k!d), u (k!d!1), . . . ,
!
!
!
!
!
w (k!d ), w (k!d !1), . . . ]
!
1 !
1
where y , u and w are the actual, present or previous values of the process output, input and
! !
!
measurable disturbances respectively, d represents the time delay related to the process input and
d is the equivalent for the measurable disturbances.
1
The measured values of the process variables differ from their actual values owing to measurement errors, noise, etc., as expressed by the equations
y (k)#n (k),
u(k)"u (k)#n (k),
w(k)"w (k)#n (k)
(2)
!
y
!
u
!
w
where, for each variable, subscript a denotes its actual value and n denotes the additional
corrupting signal. As a result, the corresponding meaasured vector (also called the input/output or
regression vector) / becomes
/(k)"/ (k)#n (k)
!
Substituting (2) and (3) into the process equation (1), we obtain
y(k)"h(k)T/(k!d)#*(k)
(3)
(4)
where
*(k)"n (k)!h(k)Tn (k!d)#m(k)
y
which can be referred to as the perturbation signal and represents the effect of the unmeasured
disturbances and measurement noise acting on the process.
This general process description has been extensively used in the literature and is appropriate
for the theoretical analysis that will follow in this paper. As previously shown in the literature, the
extension of the results obtained for the single-input/single-output case to the multivariable case
is straightforward. We will use the above single-input/single-output description for simplicity
reasons.
2.2.2. Input/output properties associated with the stability nature of the process. In this subsection we will consider two different kind of processes which cover practically without exception
all the potential real applications and whose input/output properties are essential for the stability
analysis.
( 1997 by John Wiley & Sons, Ltd.
266
The first class of processes to be considered will be referred to as those of a linear and stable
nature and is defined in the following.
Definition 1
A process is of a linear and stable nature when it can be described by an equation such as (1) and
when, for any bounded input sequence, it verifies the property
max Du( j)D'q max D y( j#d)D!q k*0
(5)
1
2
0xjxk
0xjxk
where 0(q (R and 0)q (R.
K
1
2
In Reference 1 it is proven that discrete-time stable linear processes with constant parameters
(i.e. with all poles inside the unit circle) verify property (5). Also, the constants q and q are
1
2
calculated and it is shown that while the value of q depends solely on the process dynamics, the
1
value of q depends on the initial conditions, i.e. on the state of the process at k"0. However,
2
property (5) is verified by a class of processes which is much wider than that of discrete-time stable
linear processes. This wider class has been termed of a linear and stable nature.1 For instance,
within this class there will be linear processes with time-varying parameters, whose dynamics will
be able to describe the behaviour of a wide variety of industrial processes.
Given a process, its inverse process results from considering the original inputs as if they were
outputs and the original outputs as if they were inputs. The concept of inverse process leads us to
define the following second class of processes.
Definition 2
A process is of a linear and stable inverse nature when it can be described by an equation such as
(1) and when, for any bounded input sequence, it verifies the property
max D y( j#d)D'o max Du( j)D!o k*0
(6)
1
2
0xjxk
0xjxk
where 0(o (R and 0)o (R.
K
1
2
As can be seen, this definition is derived directly from Definition 1 simply by considering the
input/output exchange that the concept of inverse process involves.
By inverting the transfer function of discrete-time stable linear processes with constant
parameters and by a proof such as that given in Reference 1, it can be shown that linear processes
with constant parameters and with a stable inveerse (i.e. with all zeros inside the unit circle) verify
property (6). However, the class of processes with property (6) is much wider and closer to
industrial reality.
Although in this paper we are dealing with processes belonging either to the class characterized
by Definition 1 or to the one characterized by Definition 2, another class of processes described
by equation (1) can be referred to as unstable processes with an unstable inverse, which clearly do
not verify either Definition 1 or Definition 2.
2.3. Adaptive predictive model and predictive control principle
A general description of the adaptive predictive model may be given by
yL (kDk)"hK (k)T/ (k!d)
3
3
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
(7)
( 1997 by John Wiley & Sons, Ltd.
267
where / (k!d) contains a subset of /(k!d), hK (k)T is an estimation of the process parameter
3
3
vector and yL (k D k) represents the estimation at instant k of the process output y(k) using the model
parameters also estimated at instant k. The dimensions of / and h are usually less than or equal
3
3
to the dimensions of / and h. This reduced dimension accounts for the unmodelled dynamics.
The principle of predictive control essentially consists of calculating the control action that will
make the predicted output equal to a conveniently selected desired output. By using the model in
(7), we may express the output predicted at instant k#d in the form
yL (k#dDk)"hK (k)T/ (k)
(8)
3
3
When d'd , some terms within / (k) related to the disturbance w will not have been
1
3
measured at instant k yet and therefore they should be conveniently estimated. For the sake of
simplicity we will assume in the following that d)d .
1
Denoting now the desired output as y (k#d), the principle of predictive control can be
$
translated into the equation
y (k#d)"hK (k)T/ (k)
$
3
3
(9)
max D/ (k!d)D, c
(10)
i
1xixn1
where / is the ith component of /, c is a positive constant that can be chosen arbitrarily and n is
i
1
the dimension of the process input/output vector /(k!d) and parameter vector h, which is
bigger than the dimension (denoted n ) of the reduced-order input/output vector / (k!d) and
3
3
parameter vector hK (k).
3
The input/output variables in the normalized system are defined as
n(k)"max
y(k)
y/(k)"
,
n(k)
1
x(k!d)"
/(k!d),
n(k)
1
x (k!d)"
/ (k!d)
3
n(k) 3
(11)
Using these variables in (1), the normalized process equation can be written in the form
*(k)
y/(k)"h(k)Tx(k!d)#
n(k)
(12)
*(k)
y/(k)"h (k)Tx (k!d)#h (k)Tx (k!d)#
3
3
6
6
n(k)
(13)
where
h(k)T"[h (k)T, h (k)T],
x(k!d)T"[x (k!d)T, x (k!d)T]
3
6
3
6
The dimension of h is n , as for x ; x (k!d) contains the terms of x(k!d) not included in
3
3
3 6
x (k!d) and thus its dimension is n !n . The last two terms in (13) can be jointly interpreted as
3
1
3
( 1997 by John Wiley & Sons, Ltd.
268
a normalized perturbation signal acting on the normalized process output y/(k). Thus we may
write
y/(k)"h (k)Tx (k!d)#*/(k)
3
3
(14)
*(k)
*/(k)"h (k)Tx (k!d)#
6
6
n(k)
(15)
with
It can be proved1 that the sequence of normalized perturbation signals M*/(k)N defined by
equation (15) is bounded if the sequence of perturbation signals M*(k)N is bounded.
By comparing the normalized process output with the estimations given by the normalized AP
model, we can define normalized a priori and a posteriori estimation errors as
e/(k D k!1)"y/(k)!hK (k!1)Tx (k!d)
3
3
e/(k Dk )"y/(k)!hK (k)Tx (k!d)
3
3
The adaptation mechanism is now defined in the form
(16)
(17)
(18)
269
Lemma 1
Along with the operation of the adaptive system described by equations (10)(23), the following
property is satisfied at each sampling instant h at which the criterion (19)(21) allows the
adaptation of the AP model parameters (t(h)"1):
EhI (hDh)E2!EhI (h D h!1)E2(0
3
3
(24)
K
(25)
where
y (k#1!i)"y(k#1!i) (i"1, . . . , p)
$
At instant k the driver block selects the corresponding value of the PDT
k
p
q
y (k#1)" + a y(k#1!i)# + b y (k#1!i)
(26)
$
i
i 41
i/1
i/1
as the desired output for k#1.
The fact that the PDT is redefined at each sampling instant k introduces a level of feedback of
k
the process output in the generation of the desired output, which has been shown to be a key
factor in practical applications.1 In fact, the desired output trajectory could be calculated using
just the model of (25) recursively from an initial instant k"0 without any kind of redefinition at
the following instants k"1, 2, . . .. In an ideal situation the output measured at instant k#1
would be equal to the predicted output and to the desired output as planned. However, this ideal
situation will not occur in general owing, among other things, to modelling errors and control
limits. In such a case it is advisable to redefine the projected desired trajectory from the actual
process output.
This redefinition introduces the concept of the driving desired trajectory (DDT). This is
produced from the first values of each of the projected desired trajectories that are defined at the
consecutive control instants k. Then the DD will be generated point-by-point in real time and
consequently this trajectory is the one which has to guide the process output to the setpoint in the
desired way: rapidly, without oscillations and, moreover, compatible with a bounded control
action.
( 1997 by John Wiley & Sons, Ltd.
270
2.5.2. Basic strategy. When the desired output y (k#d) is explicitly given, the control action
$
can be computed from the predictive control principle stated in (9) in the form
y (k#d)!hK (k)T/ (k)
30
30
u(k)" $
hK (k)
1
(27)
where hK (k) is the parameter included in the vector hK (k) in equation (7) that corresponds
1
3
to the control signal u(k) in the inner product; hK (k) and / (k) result from excluding the
30
30
parameter hK (k) and the control signal u(k) respectively from hK (k) and / (k). The adaptation
1
3
3
mechanism may be accommodated to formally guarantee that the parameter h (k) is not zero for
1
any instant k.2
This way of generating the desired output and the control action implies the use of the
predictive model in a single-step time horizon and defines what is referred to as the basic strategy
of predictive control. This strategy has been shown to be efficient in processes of a linear and
stable inverse nature in which the time delay d is constant and known. However, this basic
strategy is not recommendable for the class of processes with an unstable inverse and/or with
unknown or variable time delays, since it may lead to unbounded control actions.1
2.5.3. Extended strategy. A general form of application of predictive control, referred to as the
extended strategy, consists of generating the projected desired trajectory not with a single-step
criterion but to satisfy, together with the control sequence that produces it, a performance
criterion over a multistep prediction horizon [k, k#j]. The control applied at instant k is the
first value of this control sequence and the problem is redefined at each sampling instant k. This
means that this strategy includes the basic strategy as a particular case for j"1. When j'1, the
principle of predictive control is also maintained, which means that (9) is verified, y (k#d) being
$
the value of the DDT at instant k#d, but the computation of the control signal u(k) is not made
explicitly by means of (27).
Different performance criteria can be proposed to generate the PDTs and the associated
controls. In any case the objective is to produce a driving desired trajectory (DDT) that is able to
guide the process output to the setpoint in a satisfactory manner. This objective can be stated by
introducing the concept of physical realizability, which plays an important role in the stability
analysis that follows in Section 3.
Definition 3
It is said that the DDT is physically realizable if for any instant k its value y (k#d) can be
$
realized through the application at instant k of a bounded control action.
K
By way of example we consider here the performance criterion defined by the two
conditions
yL (k#j D k)"y (k#j)
3
(28a)
(28b)
In (28a), y (k#j) is a desired value for the output at instant k#j that is explicitly computed at
3
instant k belonging to a reference trajectory defined by a model like the one in (25). Condition
(28b) means that the control sequence is constant in the prediction interval. To make clear how
this performance criterion is used, consider a model of the form (7) but now extended over the
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
271
(29a)
(29b)
i"1, . . . , nL
j"2, . . . , j
i"1, . . . , mL , j"2, . . . , j
(31a)
with
and
eL (1)"aL ,
i
i
gL (1)"b ,
i
i
eL (j~1)
"0,
nL `1
gL (j~1)
"0,
mL `1
i"1, . . . , nL
i"1, . . . , mL
j"2, . . . , j
(31b)
j"2, . . . , j
(31c)
h (j)(k)"gL (j) (k)#gL (j~1) (k)#. . .#gL (1) (k)
1
1
1
The operations to implement the control law in (30) are indeed very simple, even considering
the on-line adaptation of the parameters. This makes this particular implementation of the
extended strategy of predictive control very attractive for real applications. This control law
appears as a simple generalization for j'1 of the control law in (27) for j"1, y (k#j) being the
3
value of the projected desired traectory (PDT) for instant k#j. It will be specifically used in the
second illustrative example presented later on in this paper.
3. STABILITY RESULTS
The stability results for adaptive predictive control are based on (i) the predictive control
principle in (9), which is common to the basic and extended strategies of predictive control, (ii)
boundedness or physical realizability of the DDT, the latter as stated in Definition 3, and (iii) the
input/output properties of the process, which are inherent to its stability nature.
Historically, the stability results have been formally derived for different process and AP model
descriptions, all of them included as particular cases in the process description in (4) and the AP
model description in (9) respectively. Thus the following scenarios have been considered:
(a) the ideal case, in which there is no perturbation signal *(k), the process parameter vector
h is time-invariant and has the same dimension as the model parameter vector, which in this
case is denoted as hK (k)
( 1997 by John Wiley & Sons, Ltd.
272
(b) the real case with no difference in structure, which differs from the ideal case in that the
process description includes the perturbation signal
(c) the real case with difference in structure, which differs from the previous real case in that the
process and model equations may have different orders
(d) the real case with time-varying parameters, which includes the previous real cases and
considers that the process parameters may vary with time.
In this section we will present the stability results derived for the two kinds of processes
considered in Section 2.
3.1. Processes of a linear and stable inverse nature
3.1.1. Predictive control. Where the process is of a linear and stable inverse nature, if the model
describing the dynamic behaviour of the process is assumed to be known, the application of
predictive control as has been described in Sections 2.5.2 and 2.5.3 leads to the desired control
objective. However, in practice we will not know that model exactly a priori and, moreover, the
process may be variable over time. For this reason our analysis in this section will focus on the
effects that dynamic or parametric differences between the predictive model and the process, i.e.
modelling errors, have on the stability and on the control objective.
The stability analysis is based on the input/output property for the kind of processes here
considered stated in the following lemma.
Lemma 2
In a process of a linear and stable inverse nature, where Du(k)D(R k*0, the sequence
ME/(k)EN will only be unbounded if there exists a subsequence Mk N of MkN such that
4
(a) Dy(k )D'a E/(k !d)E!a k *0, with 0(a (R and 0)a (R
4
1
4
2 4
1
2
(b) lim E/(k !d)E"R
4
k4?=
where a depends directly on the dynamic nature of the process, specifically on the parameter
1
o considered in Definition 2, and a depends on the initial conditions, particularly on o and
1
2
1
o .
K
2
Where predictive control is applied, there is no adaptation and therefore the control law in (9)
becomes
y (k#d)"hK T / (k)
(32)
$
3 3
By considering hK "[hK , 0, . . . , 0], if the dimension of the model parameter vector hK is lower than
3
3
the dimension of the process parameter vector hK , equation (9) may also be expressed in the form
y (k#d)"hK T/(k)
$
The stability results considered in this section use the following assumption.
(33)
Assumption 1
The driving desired output (DDT) always remains bounded, i.e. Dy (k#d)D(u2(R k*0,
$
and the delay d is known.
K
The following theorem gives a stability condition for the above predictive control law.
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
273
Theorem 1
Let a process be of a linear and stable inverse nature described by equation (4) and controlled
by the predictive control law (33). Under Assumption 1 the sequence ME/(k)EN is bounded if there
exists a time instant k '0 for which the following condition is met:
&
a'Eh(k)!hK E k*k
(34)
&
where a is the greatest of all possible a verifying condition (a) of Lemma 2.
K
1
From this basic result the following corollaries are derived.
Corollary 1
If condition (34) of Theorem 1 is satisfied, the sequence MDe(k)DN is also bounded in the form
De(k)D(aE/(k!d)E#*
"
k*d
where
e(k)"y(k)!y (k)"[h(k)!hK (k)]T/(k!d)#*(k)
$
* 'D*(k)D k*0
"
Corollary 2
If the process parameter vector is equal to the parameter vector of the predictive model, then
(a) E/(k)E()2 k*0
(b) e(k)"*(k) k*d.
Corollary 3
The DDT is physically realizable for the conditions considered in Theorem 1.
According to the stability results of Theorem 1, which are valid for the ideal and real cases
previously considered, if the norm of the modelling error, Eh(k)!hK E, is less than a certain value a,
which depends on the process dynamics, the predictive control system will be stable and will keep
both the I/O vector and the control error bounded. However, if the modelling error is beyond a,
the stability cannot be guaranteed. From the above results it is also derived that the bound on the
control error will decrease with the modelling error and, if the process parameters were known
and used in the predictive model, the control error would be equal to the perturbation signal in
the real cases and zero in the ideal case.
Finally, Corollary 3 proves that in the stability conditions of Theorem 1 it is sufficient to
choose a bounded DDT in order to render it physically realizable.
3.1.2. Adaptive predictive control. The results derived in the preceding subsection are intuitive
and realistic. In fact, intuitively, the stability result depends on how well the predictive model
estimates the process dynamics and this stability cannot be guaranteed if the modelling error is
beyond a certain bound. The basic motivation for an adaptation mechanism arises from the need
to avoid the possibility that the stability guaranteed by Theorem 1 may fail owing to modelling
errors that may occur either initially or throughout the process operation.
( 1997 by John Wiley & Sons, Ltd.
274
In the presence of an adaptation mechanism for the AP model parameters, stability condition
(34) is easily replaced by
a'Eh(k)!hK (k)E k*k
(35)
&
where the AP model parameter vector hK (k) is generated by the adaptation mechanism.
However, condition (35) has the important practical drawback of being formulated in terms of
the modelling error, i.e. the difference between the process parameters, which are a priori
unknown, and the AP model parameters. Fortunately, a general condition for the stability result
may be formulated in terms of the increments of the AP model parameters and the a posteriori
estimation error as stated in the following theorem.
Theorem 2
Let a process be of a linear and stable inverse nature described by equation (4) and controlled
by the adaptive predictive control law of (9). Under Assumption 1 the sequence ME/(k)EN is
bounded if there exists a time instant k '0 for which the following condition is met:
&
De(kDk)D
a'EhK (k)!hK (k!d)E#
k*k
(36)
3
3
&
maxME/(k!d)E, cN
where a is the greatest of all possible a considered in Lemma 2 and c is any positive constant.
1
K
As can be observed, this general result relaxes the stability condition, as the convergence of the
AP model parameters towards the process parameters is not required. This means that the
desired control objective could be reached in spite of more or less significant modelling errors.
Since the stability result is formulated in terms of the increments of the AP model parameters and
the a posteriori estimation error, it may be combined with the convergence properties of adaptive
systems satisfying the result (24) of Lemma 1 in order to complete the analysis of global stability
for adaptive predictive control. In the following we present and discuss the results of this analysis
for the ideal and the real cases.
Ideal case
In this case the process is described by
y(k)"hT/(k!d)
(37)
and the AP model parameter vector hK (k)"hK (k) has the same dimension as h(k). Additionally,
3
from (24) it is derived that the adaptive system also verifies
lim e(k D k)"0
(38)
k?=
lim [hK (k)!hK (k!1)]"0
(39)
k?=
From these two properties the result of global asymptotic stability for the ideal case is stated in
the following theorem.
Theorem 3
If the process described by equation (37) is of a linear and stable inverse nature, the application
of the adaptive predictive control law of (9), where hK (k)"hK (k), and of the adaptive system
3
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
275
(40)
The dimension of the AP model is identical with that of the process and from condition (24) it is
derived that the adaptive system satisfies the convergence properties
&k such that De(kDk)D(2* k*k '0
&
"
&
lim [hK (k)!hK (k!1)]"0
k?=
From these properties the global stability result is stated in the following theorem.
(41)
(42)
Theorem 4
If the process described by equation (40) is of a linear and stable inverse nature, the application
of the adaptive predictive control law (9), where hK (k)"hK (k), and of the adaptive system satisfying
3
condition (24) guarantees the following properties under Assumption 1:
(a) E/(k)E()(R
k*0
276
From these properties the global stability result is stated in the following theorem.
Theorem 5
If the process described by equation (40) is of a linear and stable inverse nature, the application
of the adaptive predictive control law (9) and the adaptive system satisfying condition (24)
guarantees the following properties under Assumption 1:
(a) E/(k)E()(R
k*0
277
the process dynamics. Theorem 2 states stability condition (36), which is based on the boundedness of the a posteriori estimation error and of the increments of the AP model parameters. The
verification of any of these two conditions, which admit a permanent variation in the AP model
parameters, will depend basically on the type of variation that the process parameters undergo.
The parametric variations observed in the operation of industrial processes are usually due to the
action of perturbations on or changes in the operating points, which could produce rapid changes
in the parameters, or to continuous mild variations of the operating conditions, which could
produce slow and sustained changes in the parameters. In any case it seems reasonable to expect
that the operation of the adaptive system is able to retain the above stability conditions, as has
been repeatedly confirmed by industrial practice.
3.2. Processes of a linear and stable nature
A general condition for stability
In this subsection we deal with the class of processes of a linear and stable nature, which is
characterized by Definition 1, irrespective of their inverse stability or instability. The case where
the inverse is unstable is the one that renders this class of processes not a subset of the class
considered in the preceding subsection. When the inverse is unstable, the property stated in
Lemma 1 is not true and thus the stability results derived in the preceding section are not valid.
The stability results that we will present below for the class of processes considered here are
based on the input/output property stated in the following lemma.
Lemma 3
In a process of a linear and stable nature, where Du(k)D(R k*0, the sequence ME/(k)EN will
only be unbounded if there exists a subsequence Mk N of MkN such that
4
(a) Du(k )D'c E/(k )E!c
k *0, with 0(c (R and 0)c (R
4
1
4
2
4
1
2
(b) lim E/(k )E"R
4
k4?=
where c depends on the dynamic nature of the process, specifically on the coefficient q in
1
1
Definition 1, and c depends on the initial conditions, particularly on q and q
K
2
1
2
The stability results of the preceding subsection are still valid for those processes considered
here that have a stable inverse. Predictive or adaptive predictive control can be applied to them
through the basic or the extended strategy. However, for stable processes with an unstable
inverse, as has already been mentioned, the application of the basic strategy is not possible as it
can generate unbounded control actions.
There is a wide variety of forms of application of the predictive control extended strategy, and
the rigorous stability analysis for each case, in both the predictive and the adaptive predictive
contexts, is a broad and complex open area for theoretical research. However, in Reference
1 a unified stability analysis was presented based on the implicit execution of the principle of
predictive control as defined in (9) and on the general condition defined in the following
assumption.
Assumption 2
The driving desired output y (k#d) is physically realizable k*0 and the time delay d is
$
known.
K
( 1997 by John Wiley & Sons, Ltd.
278
As discussed in the previous section, the objective of the driver block design under the extended
strategy is to produce a DDT able to guide the process output to the setpoint in a satisfactory
manner, which implies that the DDT must be physically realizable and the design must take this
fact into account as a first priority.
For instance, in the driver block design example considered in Section2.5.3, the fact of
imposing the condition that the predicted process output must reach a desired value at the end of
the prediction horizon, with a step-shaped control sequence, allows the predicted trajectory to
reach the desired objective according to the actual process dynamics and by only calculating the
magnitude of the control step sequence. We may intuitively think in this case that the PDT may
be the response of the process dynamics to a control step sequence, which implies that the first
value of this trajectory and therefore the DDT will be physically realizable, as has been repeatedly
demonstrated in practice. If we assume that the driver block design is correct and therefore the
DDT satisfies Assumption 2, then the results of the unified stability analysis presented in
Reference 1 are those summarized in the following.
Predictive control
The stability result in the context of predictive control is analogous to that obtained in the
preceding subsection and is stated in the following theorem.
Theorem 6
Let a process be a linear and stable nature described by an equation such as (4) and controlled
by the predictive control law (33). Under Assumption 2 the sequence ME/(k)EN is bounded if there
exists a time instant k '0 for which the following condition is met:
&
c'Eh(k)!hK E k*k '0
(46)
&
where c"h c , c being the greatest of all possible c verifying condition (a) of Lemma 3 and
1. 0 0
1
0(h )Dh (k)D k*0.
K
1.
1
From this theorem, valid for the ideal and the different real cases, stability results similar to
those stated in Corollaries 1 and 2 may be derived for the class of processes of a linear and stable
nature. Thus the boundedness on the control error will decrease with the modelling error and, if
the process parameters were known and used in the predictive model, the control error would be
equal to the perturbation signal in the real cases and zero in the ideal case. Additionally it is
proven that the physical realizability of DDT implies its boundedness.
Adaptive predictive control
The motivation and objectives for adaptive predictive control are in this case similar to those
considered for the application of adaptative predictive control to processes of a linear and stable
inverse nature. Likewise, the stability result for predictive control stated in Theorem 6 can be
directly transferred to the adaptive predictive context by replacing the stability condition (46) by
the condition
c'Eh(k)!hK (k)E k*k '0
(47)
&
where the AP model parameter vector hK (k) is generated by the adaptation mechanism.
The main result of stability in the context of adaptive predictive control is also analogous to
that obtained in the preceding subsection and is stated in the following theorem.
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
279
Theorem 7
Let a process be of a linear and stable nature described by an equation such as (4) and
controlled by the adaptive predictive control law of (9). Under Assumption 2 the sequence
ME/(k)EN will be bounded if there exists a time instant k '0 for which the following condition
&
holds:
De(k D k)D
c'EhK (k)!hK (k!d)E#
k*k
3
3
&
maxME/(k!d)E, cN
(48)
K
A stability analysis for the ideal case and the different real cases can be derived from the above
theorem in terms that are completely analogous to the one carried out in the preceding subsection
from Theorem 2 for processes of a linear and stable inverse nature, but in this case for processes of
a linear and stable nature and based on Assumption 2.
4. ILLUSTRATIVE EXAMPLES
Example 1
Let us consider the following third-order plant that has second-order, high-frequency
dynamics:
458
y(s)"
u(s)
(s#1)(s2#30s#229)
(49)
For a sampling period of 025 s the plant has the discrete-time representation
y(k)"0820y(k!1)!00327y(k!2)#0000431y(k!3)
#0234u(k!1)#0185u(k!2)#000479u(k!3)
(50)
(51)
where
hK (k)T"[aL (k), b (k)],
/ (k!1)T"[ y(k!1), u(k!1)]
(52)
3
1
1
3
This means that, among the scenarios described in Section 3, we are in the real case with
difference in structure, i.e. in the presence of unmodelled dynamics.
The adaptation mechanism is the one described in Section 2.4, where the parameter required in
the normalization (10) is c"1. The control action is computed using the basic strategy described
in Section 2.5.2. This is appropriate since the plant has a stable inverse.
Figure 2 shows the results of the application of adaptive predictive control to follow an
explicitly given setpoint sequence starting with parameters at arbitrary initial values. These
results clearly illustrate the three properties stated in Theorem 5: (a) the system is stable, i.e. the
input/output vector / is bounded; (b) the parameters converge in a finite time; (c) the norm of the
tracking error e is bounded. The plots of the parameters of the adaptive predictive model include
( 1997 by John Wiley & Sons, Ltd.
280
the squared norm of the parameter identification error vector (22). According to property (24),
this squared norm in the time-invariant case is a non-increasing sequence along the operation of
the adaptation mechanism. This property is clearly illustrated by the corresponding plot in
Figure 2.
This example has been taken from the previous work described in Reference 4, where a deeper
insight into the application of different adaptive predictive control algorithms was considered.
Particularly, when adaptive predictive control is applied through an unnormalized adaptation
mechanism5 that guarantees stability in the ideal case but does not guarantee the results of
Theorem 5, it is shown that the control system becomes unstable and the AP model parameters
diverge.
Example 2
Let us consider the following stable process with an unstable inverse:
1!4s
y(s)"
u(s)
(s#1)(4s#1)
(53)
(54)
For the application of adaptive predictive control we use a second-order model, so that we are
within a scenario with no difference in structure. Since the process inverse is unstable, we use
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
281
the extended strategy described in Section 2.5.3 to generate a control action able to ensure a
physically realizable desired output. Figure 3 shows the input/output behaviour when applying
this strategy with a prediction horizon j"6 with a predictive model having the same parameters
as the process in (53), i.e. hK "[07419, !00821, !0391, 07321]. This means that we are in the
ideal case and no adaptation is performed.
Once we have guaranteed a satisfactory performance in the ideal case, the purpose of the
experiment shown in Figure 4 is to illustrate the effects of the modelling errors in deteriorating
this performance and how adaptation can help to counteract this deterioration. Between instants
0 and 59 the same control law is used without parameter adaptation, but considering a predictive
model with erroneous parameters hK "[07419, !00821, !0391, 07321]#25[!01, 03,
09, !1]. In this case, as we may see in Figure 4, the system becomes unstable with an
unbounded input/output vector /. The instability is clearly understood in terms of Theorem 6,
which states that modelling errors beyond a certain value do not ensure boundedness of the
vector /.
After instant 60 the adaptation mechanism is turned on. This mechanism is the one described in
Section 2.4 with the normalization factor n(k)"1. We may observe how adaptation leads to
stability in terms of ensuring boundedness of the vector /, convergence of parameters and
boundedness of the tracking error e, which accords with the stability results derived for processes
of a linear and stable nature.
( 1997 by John Wiley & Sons, Ltd.
282
Figure 4. Predictive and adaptive predictive control of system (53) in presesnce of modelling errors
5. CONCLUSIONS
This paper has summarized the main results of a theoretical body of stability, for processes of
a linear and stable inverse nature and processes of a linear and stable nature, both in the context
of predictive control and adaptive predictive control. Real processes generally fall within one of
these classes. These results are based on the boundedness and/or physical realizability of the
driving desired trajectory and may be summarized as follows.
1. Stability for predictive control and adaptive predictive control in the general time-variant
case is guaranteed as long as the modelling errors are within a certain limit, which depends
on the dynamic nature of the process.
2. Stability for adaptive predictive control may also be guaranteed by a limiting condition on
the increments of the AP model parameters and the a posteriori estimation error.
3. Adaptation, when there is no difference in structure, guarantees stability unconditionally in
the time-invariant case.
4. In the time-invariant case, when there is a difference in structure, adaptation guarantees
stability if the unmodelled dynamics is within a certain limit, which, as in point 1, depends
on the dynamic nature of the process.
5. In the time-variant case, adaptation guarantees the reduction of the modelling error, which,
as practice has shown, generally derives to stability.
Int. J. Adapt. Control Signal Process., Vol. 11, 263283 (1997)
283
6. The stability condition of physical realizability of the driving desired trajectory implies that
the extended strategy of predictive control has to be applied to processes with an unstable
inverse; otherwise, if the process inverse is stable, both the basic and the extended strategy
would guarantee stability.
REFERENCES
1. Mart n Sanchez, J. M. and J. Rodellar, Adaptive Predictive Control: from the Concepts to Plant Optimization,
Prentice-Hall, Englewood Cliffs, NJ, 1996.
2. Mart n Sanchez, J. M., S. L. Shah and D. G. Fisher, A stable adaptive predictive control system, Int. J. Control, 39,
215234 (1984).
3. Mart n Sanchez, J. M., A globally stable APCS in the pressence of bounded unmeasured noises and perturbations,
IEEE rans. Automatic Control, AC-29, 461464 (1984).
4. Cluett, W. R., J. M. Mart n Sanchez, S. L. Shah and D. G. Fisher, Stable discrete time adaptive control in the presence
of unmodeled dynamics, IEEE rans. Automatic Control, AC-33, 410414 (1988).
5. Mart n Sanchez, J. M., A new solution to adaptive control, Proc. IEEE, 64, 12091218 (1976).