Automatica, VoL 23, No. 2, pp. 137 148, 1987 0005 1098/87 $3.00+0.

0~
Printed in Great Britain. Pergamon Journals Ltd.
~ 1987 International Federation of Automatic Control
Generalized Predi cti ve Control
Al gori thm*
Part I. The Basic
D. W. CLARKEt , C. MOHTADI t and P. S. TUFFS~:
A new member of the family of long-range predictive controllers is shown to be suitable
f or the adaptive control of processes with varying parameters, dead-time and model-
order.
Key Words--Adapt i ve control; predictive control; LQ control; self-tuning control; nonminimum-phase
plant.
Akst raet --Current self-tuning algorithms lack robustness to
prior choices of either dead-time or model order. A novel
method--generalized predictive control or GPC- - i s developed
which is shown by simulation studies to be superior to accepted
techniques such as generalized minimum-variance and pole-
placement. This receding-horizon method depends on predicting
the plant' s output over several steps based on assumptions
about future control actions. One assumpt i on--t hat there is a
"control horizon" beyond which all control increments become
zer o- - i s shown to be beneficial both in terms of robustness and
for providing simplified calculations. Choosing particular values
of the output and control horizons produces as subsets of the
method various useful algorithms such as GMV, EPSAC,
Peterka' s predictive controller (1984, Automatica, 20, 39-50)
and Ydstie's extended-horizon design (1984, IFAC 9th World
Congress, Budapest, Hungary). Hence GPC can be used either
to control a "simple" plant (e.g. open-loop stable) with little
prior knowledge or a more complex plant such as nonminimum-
phase, open-loop unstable and having variable dead-time. In
particular GPC seems to be unaffected (unlike pole-placement
strategies) if the plant model is overparameterized. Furthermore,
as offsets are eliminated by the consequence of assuming a
CARIMA plant model, GPC is a contender for general self-
tuning applications. This is verified by a comparative simulation
study.
1. I NTRODUCTI ON
ALTHOUGH SELF-TUNING and adaptive control has
made much progress over the previous decade,
both in terms of theoretical understanding and
practical applications, no one method proposed so
far is suitable as a "general purpose" algorithm for
the stable control of the majority of real processes.
To be considered for this role a method must be
applicable to:
* Received 2 March 1985; revised 7 July 1985; revised 3 March
1986; revised 22 September 1986. The original version of this
paper was not presented at any IFAC meeting. This paper was
recommended for publication in revised form by Associate
Editor M. Gevers under the direction of Editor P. C. Parks.
t Department of Engineering Science, Parks Road, Oxford
OXI 3PJ, U.K.
Aluminum Company of America, Alcoa Technical Center,
Pittsburgh, PA 15069, U.S.A.
(1) a nonminimum-phase plant: most continuous-
time transfer functions tend to exhibit discrete-time
zeros outside the unit circle when sampled at a fast
enough rate (see Clarke, 1984);
(2) an open-loop unstable plant or plant with
badly-damped poles such as a flexible spacecraft or
robots;
(3) a plant with variable or unknown dead-time:
some methods (e.g. minimum-variance self-tuners,
/~str6m and Wittenmark, 1973) are highly sensitive
to the assumptions made about the dead-time and
approaches (e.g. Kurz and Goedecke, 1981) which
attempt to estimate the dead-time using operating
dat a tend to be complex and lack robustness;
(4) a plant with unknown order: pole-placement
and LQG self-tuners perform badly if the order of
the plant is overestimated because of pole/zero
cancellations in the identified model, unless special
precautions are taken.
The method described in this paper--General -
ized Predictive Control or GPC- - appear s to over-
come these problems in one algorithm. It is capable
of stable control of processes with variable parame-
ters, with variable dead-time, and with a model
order which changes instantaneously provided that
the input/output data are sufficiently rich to allow
reasonable plant identification. It is effective with
a plant which is simultaneously nonminimum-
phase and open-loop unstable and whose model
is overparameterized by the estimation scheme
without special precautions being taken. Hence it
is suited to high-performance applications such as
the control of flexible systems.
Hitherto the principal applied self-tuning
methods have been based on the "Generalized
Minimum-Variance" approach (Clarke and
Gawthrop, 1975, 1979) and the pole-placement
algorithm (Wellstead et al . , 1979; /~str6m and
137
A U T 2 3 / 2 - A
138 1). W. CLARKk et al.
Wi t t enmark, 1980). The implicit GMV self-tuner
(so-called because the controller parameters are
directly estimated wi t hout an intermediate calcul-
ation) is robust against model order assumptions
but can perform badly if the plant dead-time varies.
The explicit pole-placement method, in which a
Di ophant i ne equat i on is numerically solved as a
bridging step between plant identification and the
control calculation, can cope with variable dead-
time but not with a model with overspecified order.
Its good behavi our with variable dead-t i me is due
to overparamet eri zat i on of the numer at or dynami cs
B(q - 1); this means t hat the order of the denomi nat or
dynami cs has to be chosen with great care to avoid
singularities in the resolution of the Di ophant i ne
identity. The GPC approach being based on an
explicit plant formul at i on can deal with variable
dead-time, but as it is a predictive met hod it can
also cope with overparamet eri zat i on.
The type of plant t hat a self-tuner is expected to
cont rol varies widely. On the one hand, many
industrial processes have "simple" models: low
order with real poles and probabl y with dead-time.
For some critical loops, however, the model might
be more complex, such as open-loop unstable,
under damped poles, multiple integrators. It will be
shown t hat GPC has a readi l y-underst andabl e
def aul t operat i on which can be used for a simple
plant wi t hout needing the detailed prior design of
many adaptive methods. Moreover, at a slight
increase of comput at i onal time more complex pro-
cesses can be accommodat ed by GPC within the
basic framework.
All industrial plants are subjected to load-dis-
turbances which tend to be in the form of random-
steps at r andom times in the deterministic case or
of Browni an mot i on in stochastic systems. To
achieve offset-free closed-loop behaviour given
these disturbances the controller must possess
inherent integral action. It is seen t hat GPC adopt s
an i nt egrat or as a nat ural consequence of its
assumpt i on about the basic plant model, unlike the
maj ori t y of designs where integrators are added in
an ad hoc way.
2. THE CARIMA PLANT MODEL AND OUTPUT
PREDICTION
When considering regulation about a particular
operat i ng point, even a non-linear pl ant generally
admits a locally-linearized model:
If the plant has a non-zero dead-time the leading
elements of the pol ynomi al B( q - ~) are zero. In (1),
u(t) is the control input, .~t) is the measured variable
or out put , and x(t ) is a disturbance term.
In the literature x(t ) has been considered to be
of movi ng average form:
x ( t ) = C ( q - ~)~.(t) ( 2 )
where C( q - 1 ) = I + c l q ~ + ' + c,cq -"c.
In this equation, (t) is an uncorrel at ed r andom
sequence, and combi ni ng with (I) we obt ai n the
CARMA (Controlled Auto-Regressive and Mov-
ing-Average) model:
A ( q - l ) y ( t ) = B ( q ~ 1 ) u ( t - I )
+ C( q - l ) ~ ( t ) . (3)
Though much self-tuning t heory is based on this
model it seems to be i nappropri at e for many
industrial applications in which disturbances are
non-st at i onary. In practice, two principal disturb-
ances are encountered: r andom steps occuring at
r andom times (for example, changes in material
quality) and Browni an mot i on (found in plants
relying on energy balance). In bot h these cases an
appropri at e model is:
x(t ) = C( q - 1)~(t)/A (4)
where A is the differencing operat or I - q - 1 .
Coupl ed with (1) this gives the CARI MA model
(integrated moving-average):
A ( q - l )y(t ) = B( q l )u(t - 1)
+ C( q - l)~(t)/A.
This model has been used by Tufts and Clarke
(1985) to derive GMV and pole-placement self-
tuners with inherent integral action. For simplicity
in the development here C( q - 1 ) is chosen to be 1
(alternatively C-1 is t runcat ed and absorbed into
the A and B polynomials; see Part II for the case
of general C) t o give the model:
A( q - 1 ) y ( t ) = B( q 1)u(t - l) + ~(tffA. (5)
A( q l )y(t ) = B( q - 1 ) u ( t - 1) + x( t ) (1)
where A and B are pol ynomi al s in the backward
shift oper at or q-1:
A( q - 1) = 1 + a l q - 1 + . . , + a, ~q -"~
B(q - 1) = b o + b l q - ~ + . . . + bnbq -rib.
To derive a j-step ahead predictor of ~t +j )
based on (5) consider the identity:
1 = E ~ q - 1 ) A A + q - J Fj ~ q - l ) (6)
where Ej and Fj are pol ynomi al s uniquely defined
given A( q - 1 ) and the prediction interval j. If (5) is
Generalized predictive c ont r ol - - Pa r t I 139
multiplied by E j A q j we have:
Subt ract i ng (9) from (10) gives:
E j A A y ( t + j ) = E j B A u ( t + j - 1)
+ E j ( t + j )
and subst i t ut i ng for E j A A from (6) gives:
y ( t + j ) = E j B A u ( t + j - 1)
+ F j y ( t ) + E~(t + j ) . (7)
0 = .4(R - E) + q-J(q-1S - F).
The pol ynomi al R - E is of degree j and may be
split into two parts:
R - E = ~ + r j q - j
so that:
As E j ( q - 1 ) is of degree j - 1 the noise component s
are all in the future so t hat the opt i mal predictor,
given measured out put dat a up to time t and any
given u ( t + i) for i > l, is clearly:
9( t + j I t ) = G j A u ( t + j - 1) + F j y ( t ) (8)
where G j ( q - 1 ) = E j B .
Not e t hat G j ( q - 1) = B ( q - 1)[1 - q - ~ F j ( q - t)]/
A ( q - t)A so t hat one way of comput i ng G~ is simply
to consider the Z-t ransform of the plant' s step-
response and to t ake the first j terms (Clarke and
Zhang, 1985).
In the devel opment of the GMV self-tuning
cont rol l er onl y one prediction ~ ( t + k i t ) is used
where k is the assumed value of the plant' s dead-
time. Here we consider a whole set of predictions
for which j runs from a mi ni mum up to a large
value: these are termed the mi ni mum and maxi mum
"predi ct i on horizons". For j < k the prediction
process .P(t +J l t) depends entirely on available
dat a, but for j >/ k assumpt i ons need to be made
about future cont rol actions. These assumpt i ons
are the cornerst one of the GPC approach.
2. 1. R e c u r s i o n o f t h e D i o p h a n t i n e e q u a t i o n
One way to i mpl ement long-range prediction is
to have a bank of self-tuning predictors for each
hori zon j; this is the approach of De Keyser
and Van Cauwenberghe (1982, 1983) and of the
MUSMAR met hod (Mosca e t al . , 1984). Altern-
atively, (6) can be resolved numerically for E i and
F i for the whole range of j s being considered.
Bot h these met hods are comput at i onal l y expensive.
Inst ead a simpler and more effective scheme is to
use recursion of the Di ophant i ne equat i on so t hat
the pol ynomi al s E j + I and F j + I are obt ai ned given
the values of E j and Fj.
Suppose for clarity of not at i on E = E j , R = E j . 1,
F = F j , S = F j + ~ and consider the two Di ophant i ne
equat i ons with $ defined as AA:
1 = E A + q- i F ( 9 )
1 = R.4 + q- tJ + 1~S. ( 1 0 )
~I R + q - J ( q - t S - F + , 4 r j ) = O.
Clearly t hen /~ = 0 and also S is given
S q ( F - , ~r j).
As ,4 has a unit leading element we have:
by
r j -~ f o (1 la)
S i = f i + 1 - - ~l i + l r i (1 lb)
for i = 0 to the degree of S ( q - 1 ) ;
and: R ( q - i ) = E ( q - 1 ) + q - J r j (12)
G j + I = B ( q - 1 ) R ( q - 1). (13)
Hence given the plant pol ynomi al s A ( q - 1 ) and
B ( q - 1 ) and one sol ut i on E j ( q - a ) and F j ( q - 1 ) t hen
(11) can be used to obt ai n F j + ~ ( q - 1 ) and (12) to
give E j+ l(q- 1) and so on, with little comput at i onal
effort. To initialize the iterations not e t hat f or j = 1:
1 --- E 1 . 4 + q - 1 F t
and as the leading element of ,4 is 1 then:
E1 = 1, F1 = q(1 - ,']).
The calculations involved, therefore, are straightfor-
ward and simpler t han those required when using
a separate predictor for each out put horizon.
3. THE PREDICTIVE CONTROL LAW
Suppose a future set-point or reference sequence
[ w ( t + j); j = 1,2 . . . . ] is available. In most cases
w ( t + j ) will be a const ant w equal to the current set-
point w(t), t hough sometimes (as in batch process
cont rol or robotics) future variations in w ( t + j )
would be known. As in the I DCOM al gori t hm
(Richalet e t al . , 1978) it might be considered t hat a
smoot hed approach from the current out put Xt) to
w is required which is obt ai nabl e from the simple
140 D. W. CLARKE et al.
Y
t - 2 t
u [ - -
j J
/ / -
J
w s e t - p o i n t
P r e d i c t e d o u t p u t
I t + I N u I ! .
t ~ l t + N T i me t
P r o j e c t e d c o n t r o t s
FIG. 1. Set -poi nt , cont r ol and out put s in GPC.
fi rst -order lag model:
w( t ) = y ( t )
w( t + j ) = ~ w( t + j - 1)
+ ( 1 - a ) w j = l , 2 . . . .
where ~ ~ 1 for a slow t ransi t i on from the cur r ent
measured-vari abl e t o the real set -poi nt w. GPC is
capabl e of consi deri ng bot h const ant and varyi ng
fut ure set-points.
The objective t hen of the predictive cont r ol law
is t o drive fut ure pl ant out put s Xt +j ) "close" t o
w( t + j ) in some sense, as shown in Fig. 1, beari ng
in mi nd t he cont r ol activity requi red t o do so. Thi s
is done using a recedi ng-hori zon appr oach for
which at each sampl e-i nst ant t:
(1) the fut ure set -poi nt sequence w(t + j ) is calcu-
lated;
(2) the predi ct i on model of (8) is used to generat e
a set of predi ct ed out put s ~(t + j l t ) with corre-
spondi ng predi ct ed system errors e(t + j )
= w ( t + j ) - p ( t + j l t ) not i ng t hat p ( t + j [ t ) for
j > k depends in par t on fut ure cont r ol signals
u(t + i) which are to be det ermi ned;
(3) some appr opr i at e quadr at i c funct i on of the
fut ure er r or s and cont rol s is minimized, assumi ng
t hat after some "cont r ol hor i zon" furt her incre-
ment s in cont r ol are zero, to pr ovi de a suggested
sequence of fut ure cont rol s u(t + j);
(4) the first el ement u(t) of t he sequence is asserted
and t he appr opr i at e dat a vect ors shifted so t hat t he
cal cul at i ons can be repeat ed at the next sample
instant.
Not e t hat the effective cont r ol law is st at i onary,
unlike a fi xed-hori zon LQ policy. However, in the
self-tuned case new estimates of the pl ant model
par amet er s requires new values for t he par amet er
pol ynomi al s, which means t hat the fast Di ophant i ne
recursi on is useful for adapt i ve cont r ol appl i cat i ons
of the GPC met hod.
Consi der a cost funct i on of the form:
t~j=N 1
}
+ ~ i ( j ) [ A u ( t + j - It ] 2
j = l
(14)
where:
N 1 is the mi ni mum costing hori zon;
N 2 is t he maxi mum costing hori zon, and
).(j) is a cont rol -wei ght i ng sequence.
The expect at i on in (14) is condi t i oned on dat a up
to time t assuming no fut ure measurement s are
available (i.e. the set of cont r ol signals are appl i ed
in open- l oop in the sequel). As ment i oned earlier,
the first cont r ol is appl i ed and the mi ni mi zat i on is
repeat ed at the next sample. The resulting cont r ol
law bel ongs t o the class known as Open- Loop-
Feedback-Opt i mal cont rol (Bertsekas, 1976). Appen-
dix A examines t he rel at i on bet ween GPC and
Cl osed- Loop- Feedback- Opt i mal when the di st urb-
ance process is autoregressive. It is seen t hat costing
on t he cont r ol is over all fut ure inputs which affect
the out put s included in J. In general N2 is chosen
to encompass all the response which is significantly
affected by the cur r ent cont rol ; it is reasonabl e t hat
it shoul d at least be great er t han the degree of
B( q - 1) as t hen all states cont r i but e t o the cost
(Kailath, 1980), but mor e typically N2 is set to
appr oxi mat e t he rise-time of t he plant. N 1 can often
be t aken as 1; if it is known a pri ori t hat the dead-
time of the pl ant is at least k sample-intervals t hen
General i zed predictive c o n t r o l - - Pa r t I
N1 can be chosen as k or mor e t o minimize
comput at i ons. It is found, however, t hat a large
class of pl a nt model s can be stabilized by GP C
with default values of 1 and 10 for N1 and N2. Par t
II provi des a t heoret i cal j ust i fi cat i on for these
choices of hori zon. For simplicity in the deri vat i on,
bel ow 2(j) is set t o t he const ant 2, N1 to 1 and N 2
to N: the "out put hori zon".
Recall t hat (7i model s t he fut ure out put s:
y(t + 1) = GxAu( t ) + F l y ( t ) + E ~ ( t + 1)
y(t + 2) = G2Au( t + 1) + F2y(t) + E2~(t + 2)
y ( t + N) = GNAu( t + N - 1)
+ FNy(t ) + EN~(t + N) .
Consi der Xt +j ) , It consists of t hree terms: one
dependi ng on fut ure cont r ol actiol,s yet to be
det ermi ned, one dependi ng on past known cont rol s
t oget her with filtered measur ed variables and one
dependi ng on fut ure noise signals. The assumpt i on
t hat t he cont r ol s are t o be per f or med in open- l oop
is t ant amount t o i gnori ng t he fut ure noise sequence
{~(t + j)} in cal cul at i ng the predictions. Let f ( t + j )
be t hat component of Xt + j ) composed of signals
which are known at time t, so t hat for example:
f ( t + 1) =
f ( t + 2) =
[ GI ( q - 1) _ g l o ] Au ( t )
+ Fl y ( t ) , and
q [ G2 ( q - 1) _ q - tg21
- - g20] Au( t ) + F2y(t ),
etc.
where Gi( q - 1 ) = gio + g i l q - 1 + . . . .
141
go
g t
G=
0
go
g N- 1 g N- 2 go
Not e t hat if the pl ant dead-t i me is k > 1 the first
k - 1 rows of G will be null, but if instead N1 is
assumed t o be equal t o k the l eadi ng el ement is
non-zero. However, as k will not in general be
known in the self-tuning case one key feat ure of
t he GP C appr oach is t hat a stable sol ut i on is
possible even i f t he l eadi ng r ows o f G are zero.
Fr om the definitions of the vect ors above and
with:
w = [w(t + 1), w(t + 2 ) . . . . . w(t + N) ] T
the expect at i on of the cost -funct i on of (14) can be
written:
J1 ~-- E{J ( I , N)}
= E{(y - w)T(y -- W) + 2iiTii}
(16)
i.e.
Jl = {(Gii + f - w)T(Gii + f - w) + 2iiTii}.
The mi ni mi zat i on of J1 assumi ng no const rai nt s
on fut ure cont r ol s results in the proj ect ed cont r ol -
i ncrement vector:
fi = (GTG + 2I ) - 1Gr(w - f). (17)
Not e t hat the first el ement of fi is Au(t) so t hat
the cur r ent cont r ol u(t) is given by:
Then t he equat i ons above can be wri t t en in the key
vect or form:
9 = Ga + f (15)
where the vect ors are all N x 1:
= [j~(t + 1),j~(t + 2 ) , . . . , ~ t + N)] T
ii _ [Au(t), Au ( t + 1) . . . . . Au ( t + N - 1)] T
f = [ f ( t + 1 ) , f ( t + 2 ) , . . . , f ( t + N ) ] r .
As i ndi cat ed earlier, t he first j t erms in Gj ( q - 1) are
the par amet er s of the st ep-response and t herefore
g~j = g~ for j = 0, I, 2 . . . . < i i ndependent of t he
par t i cul ar G pol ynomi al .
The mat r i x G is t hen l ower - t r i angul ar of di men-
sion N x N:
u(t) = u(t - 1) + ~T(w - f) (I 8)
where ~r is t he first row of ( GTG + 2I ) - tGT. Hence
t he cont r ol includes integral act i on which provi des
zero offset pr ovi ded t hat for a const ant set -poi nt
w( t + i) = w, say, the vect or f i nvol ves a unit steady-
state gain in the feedback path.
Now the Di ophant i ne equat i on (6) for q = 1 gives
1 = Ej(1)A(1)A(I) + Fj(1)
and as A( 1 ) = 0 t hen F j ( 1 ) = I so t hat
f ( t + j ) = Fj)~t) is a signal whose mean value equals
t hat of y(t ). Fur t her mor e, defining F j ( q - i ) t o be
E j ( q - 1). ~(q- 1) gives:
F j ( q - 1)y(t) = (1 - F~( q- l)A)y(t)
= y( t ) - F j Ay ( t )
142 D. W. CLARKE et al.
which shows t hat if y(t) is the const ant y so t hat
A3~t) = 0 t hen the component Fj l q-1)~t } reduces
to f. This, t oget her with the cont rol given by (18),
ensures offset-free behavi our by integral action.
of di mensi on N U and the predi ct i on equat i ons
reduce to:
~ = G~ i i + f
3,1. The control horizon
The di mensi on of the mat ri x involved in (17) is
N x N. Al t hough in the non- adapt i ve case the
inversion need be performed once only, in a self-
t uni ng versi on t he comput at i onal l oad of inverting
at each sample woul d be excessive. Mor eover , if
the wr ong value for dead-t i me is assumed, GVG is
singular and hence a finite non- zer o value of
weighting 2 woul d be requi red for a realizable
cont r ol law, which is i nconveni ent because the
"cor r ect " value for 2 woul d not be known a priori.
The real power of the GPC appr oach lies in the
assumpt i ons made about future cont r ol actions.
Inst ead of allowing t hem to be "free" as for the
above devel opment , GPC bor r ows an idea from
the Dynami c Mat r i x Cont r ol met hod of Cut l er and
Ramaker (1980). Thi s is t hat after an interval
NU < N 2 projected control increments are assumed
to be zero,
i.e. Au(t +j - 1) = 0 j > NU. (19)
The value NU is called the "cont r ol hori zon". In
cost -funct i on t erms this is equi val ent to placing
effectively infinite weights on cont r ol changes after
some fut ure time. For example, i f NU = 1 onl y one
cont r ol change (i.e. Au(t)) is consi dered, after which
the cont rol s u(t + j) are all t aken to be equal t o
u(t). Suppose for this case t hat at time t t here is a
step change in w(t) and t hat N is large. The choice
of u(t) made by GPC is the opt i mal "mean-l evel "
cont rol l er which, if sustained, woul d place the
settled pl ant out put to w with the same dynami cs
as the open- l oop plant. Thi s cont r ol law (at least
for a simple stable plant) gives act uat i ons which
are generally smoot h and sluggish. Larger values
of N U, on the ot her hand, provi de mor e active
controls.
One useful intuitive i nt er pr et at i on of the use
of a cont r ol hor i zon is in the stabilization of
nonmi ni mum- phase plant. If the cont r ol weighting
2 is set to zero the opt i mal cont r ol which minimizes
J 1 is a cancel l at i on law which at t empt s t o remove
the process dynami cs using an inverse pl ant model
in the cont rol l er. As is well known, such (mi ni mum-
variance) laws are in practice unst abl e because
t hey involve growi ng modes in the cont r ol signal
cor r espondi ng to the pl ant ' s nonmi ni mum- phase
zeros. Const r ai ni ng these modes by placing infinite
cost i ng on fut ure cont r ol increments stabilizes the
resulting cl osed-l oop even if the weighting 2 is zero.
The use of NU < N mor eover significantly reduces
the comput at i onal burden, for the vect or fi is t hen
where:
GI =
-go
gl
gN- 1
0
go
gN- 2
0
0
i s N x NU.
go
g N - N u
The cor r espondi ng cont r ol law is given by:
f i = [ G ~ G I + ) . I ] - l G ~ ( w - f ) ( 2 0 )
and the mat ri x i nvol ved in the i nversi on is of the
much reduced di mensi on NU NU. In part i cul ar,
if NU = 1 (as is usefully chosen for a "simple"
plant), this reduces to a scalar comput at i on. An
exampl e of the comput at i ons i nvol ved is given in
Appendi x B.
3.2. Choice of the output and control horizons
Si mul at i on exercizes on a vari et y of pl ant models,
including stable, unst abl e and nonmi ni mum- phase
processes with variable dead-time, have shown how
N1 , N 2 and NU shoul d best be selected, These
studies also suggest t hat the met hod is robust
against these choices, giving the user a wide latitude
in his design.
3.2.1. NI: The minimum output horizon. If the
dead-t i me k is exactly known t here is no poi nt in
setting N1 t o be less t han k since t here would t hen be
superfl uous calculations in t hat the cor r espondi ng
out put s cannot be affected by the first act i on u(t).
If k is not known or is variable, t hen N1 can be set
to 1 with no loss of stability and the degree of
B(q-1) increased to encompass all possible values
of k.
3.2.2. N 2 : The maxi mum output horizon. If the
pl ant has an initially negat i ve-goi ng nonmi ni mum-
phase response, N 2 shoul d be chosen so t hat the
later positive-going out put samples are included in
the cost: in discrete-time this implies t hat N z exceeds
the degree of B( q- 1) as demonst r at ed in Appendi x
B. In practice, however, a r at her larger value of N2
is suggested, cor r espondi ng mor e closely to the rise-
time of the plant.
3.2.3. NU: The control horizon. Thi s is an i mport -
ant design paramet er. For a simple pl ant (e.g. open-
l oop stable t hough with possible dead-t i me and
Generalized predictive c ont r ol - - Pa r t I 143
nonmi ni mum-phasedness) a value of NU of 1 gives
generally acceptable control. Increasing NU makes
the cont rol and the correspondi ng out put response
more active until a stage is reached where any
furt her increase in NU makes little difference. An
increased value of NU is more appropri at e for
complex systems where it is found t hat good control
is achieved when NU is at least equal to the number
of unst abl e or badl y-damped poles.
One i nt erpret at i on of these rules for chosing NU
is as follows. Recall t hat GPC is a receding-
hori zon LQ cont rol law for which the future cont rol
sequence is recalculated at each sample. For a
simple, stable plant a cont rol sequence following a
step in the set-point is generally well-behaved (for
example, it would not change sign). Hence there
would not be significant corrections to the control
at the next sample even if NU = 1. However, it is
known from general state-space consi derat i ons t hat
a plant of order n needs n different control values
for, say, a dead-beat response. Wi t h a complex
system these values might well change sign frequ-
ently so t hat a short control hori zon would not
allow for enough degrees of freedom in the deriv-
at i on of the current action.
General l y it is found t hat a value of NU of 1 is
adequat e for typical industrial plant models,
whereas if, say, a modal model is to be stabilized
NU shoul d be set equal to the number of poles
near the stability boundary. If furt her dampi ng of
the cont rol act i on is then required 2 can be increased
from zero. Not e in part i cul ar that, unlike with
GMV, GPC can be used with a nonmi ni mum-
phase plant even if 2 is zero.
The above discussion implies t hat GPC can be
considered in two ways. For a process cont rol
default setting of N 1 = 1, N 2 equal to the plant rise-
time and NU = 1 can be used to give reasonable
performance. For hi gh-performance applications
such as the cont rol of coupled oscillators a larger
value of NU is desirable.
4. RELATI ONSHI P WI TH OTHER APPROACHES
GPC depends on the i nt egrat i on of five key
ideas: the assumpt i on of a CARI MA rat her t han
a CARMA plant model, the use of long-range
prediction over a finite hori zon greater t han the
dead-t i me of the plant and at least equal to the
model order, recursion of the Di ophant i ne equ-
ation, the consi derat i on of weighting of control
increments in the cost-function, and the choice of
a cont rol hori zon after which projected cont rol
increments are t aken to be zero. Many of these
ideas have arisen in the literature in one form or
anot her but not in the part i cul ar way described
here, and it is their j udi ci ous combi nat i on which
gives GPC its power. Nevertheless, it is useful
to see how previous successful met hods can be
considered as subsets of the GPC approach so t hat
accepted theoretical (e.g. convergence and stability)
and practical results can be extended to this new
met hod.
The concept of using long-range prediction as a
potentially robust control tool is due to Richalet
et al. (1978) in the I DCOM algorithm. This met hod,
t hough report edl y having some industrial success,
is restricted by its assumpt i on of a weighting-
sequence model (all-zeros), with an ad hoc way of
solving the offset probl em and with no weighting
on control action. Hence it is unsuitable for unstable
or nonmi ni mum- phase open-loop plants. The
DMC al gori t hm of Cut l er and Ramaker (1980) is
based on step-response models but does include
the key idea of a control horizon. Hence it is
effective for nonmi ni mum- phase plants but not
for open-l oop unstable processes. Again the offset
problem is dealt with heuristically and moreover
the use of a step-response model means t hat the
savings in paramet eri zat i on using a A( q- 1) poly-
nomi al are not available. Clarke and Zhang (1985)
compare the I DCOM and DMC designs.
The GMV approach of Clarke and Gawt hr op
(1975) for a plant with known dead-time k can be
seen to be a special case of GPC in which bot h the
mi ni mum and maxi mum horizons N1 and N2 are
set to k and onl y one cont rol signal (the current
cont rol u(t) or Au(t)) is weighted. This met hod is
known to be robust against overspecification of
model -order but it can onl y stabilize a certain class
of nonmi ni mum- phase plant for which the control
weighting 2 has to be chosen with reasonable care.
Moreover, GMV is sensitive to varying dead-t i me
unless 2 is large, with correspondi ngl y poor control.
GPC shares the robustness properties of GMV
wi t hout its drawbacks.
By choosing N1 = N2 --- d > k and NU -- 1 with
2 = 0, GPC becomes Ydstie' s (1984) extended-
hori zon approach. This has been shown theoret-
ically to be a stabilizing cont rol l er for a stable
nonmi ni mum- phase plant. Ydstie, however, uses a
CARMA model and the met hod has not been
shown to stabilize open-l oop unstable processes.
Indeed, for a plant with poorl y damped poles
si mul at i on experience shows t hat the extended-
hori zon met hod is unstable, unlike the GPC design.
This is because in this case more t han one future
out put needs to be account ed for in the cost-
function for stability (i.e. N 2 > N1).
Wi t h NI = 1, N z = d > k and NU= 1 with
2 = 0 and a CARI MA model, GPC reduces to
the independently-derived EPSAC al gori t hm (De
Keyser and Van Cauwenberghe, 1985). This was
shown by the aut hors to be a part i cul arl y useful
met hod with several practical applications; this is
verified by the simulations described below where
the "defaul t " settings of GPC t hat were adopt ed
144 I). W. C L AR KE et al.
with N 2 = 10 are similar to those of EPSAC Part
lI of this paper gives a stability result for these
settings of the GPC horizons.
Peterka' s elegant predictive cont rol l er (1984) is
nearest in phi l osophy to GPC: it uses a CARI MA-
like model and a similar cost-function, t hough
concent rat i ng on the case where N2--, ~ . The
algorithmic development is then rat her different
from GPC, relying on matrix factorization and
decomposi t i on for the finite-stage case, instead of
the rather more direct formul at i on described here.
Though Pet erka considers two values of control
weighting (2 x for the first set of controls and 22 for
the final 6B steps), he does not consider the useful
GPC case where 22 = ~,. GPC is essentially a
finite-stage approach with a restricted cont rol hor-
izon. This means (though not considered here) t hat
constraints on bot h future controls and control
rates can be t aken into account by suitable modi-
fications of tile GPC calculations - a generalization
which is impossible to achieve in infinite-stage
designs. Nevertheless, the Peterka procedure is an
i mport ant and applicable approach to adaptive
control.
5, A S I MUL AT I ON S T UDY
The objective of this st udy is to show how an
adaptive GPC i mpl ement at i on can cope with a
plant which changes in dead-time, in order and in
parameters compared with a fixed PID regulator,
with a GMV self-tuner and with a pole-placement
self-tuner. The PI D regulator was chosen to give
good cont rol of the initial plant model which for
this case was assumed to be known. For simplicity
all the adaptive met hods used a st andar d recursive-
least-squares paramet er est i mat or with a fixed for-
getting-factor of 0.9 and with no noise included in
the simulation.
The graphs display results over 400 samples of
si mul at i on for each met hod, showing the cont rol
signal in the range - 100 to 100, and the set-point
w(t) with the out put )~t) in the range - 1 0 to 80.
The control signal to the simulated process was
clipped to lie in [ - 100, 100], but no const rai nt was
placed on the plant out put so t hat the limitations
on ~(t) seen in Fig. 4 are graphical with the actual
out put exceeding the displayed values.
Each si mul at i on was organized as follows. Dur-
ing the first 10 samples the cont rol signal was fixed
at 10 and the est i mat or (initialized with parameters
[1, 0, 0 . . . . ]) was enabled for the adapt i ve control-
lers. A sequence of set-point changes between three
distinct levels was provided with switching every
20 samples. After every 80 samples the simulated
cont i nuous-t i me plant was changed, following the
models given in Table 1. It is seen t hat these changes
in dynami cs are large, and t hough it is difficult to
imagine a real plant varying in such a drastic way,
TABLE l , TRANSIER-FUNCTIONS Ot THE SIMlr
LATED MODELS
Nu mb e r S a mp l e s Mo d e l
1
I 1 79
I + 10s + 40s 2
2 8 0 - 1 5 9 e 2 ....
f + 10s + 40s 2
3 1 6 0 - 2 3 9 e .,7~
I + 10s
l
4 240 -319
1 + 10s
1
5 3 2 0 - 4 0 0
10s(l + 2. 5s)
the simulations were chosen to illustrate the relative
robustness and adapt i vi t y of the methods. The
models are given as Laplace transforms and the
sampling interval was chosen in all cases to be 1 s.
Figure 2 shows the behaviour of a fixed digital
PI D regulator which was chosen to give reasonable
if rat her sluggish cont rol of the initial plant. It
was i mpl ement ed in interacting PID form with
numer at or dynami cs of (I + 10s + 25s z) and with
a gain of 12. For models 1 and 5 this controller
gave acceptable results but its performance was
poor for the other models with evident excessive
gain. Despite integral action (with desaturation)
offset is seen with models 3 and 4 due to persistent
control sat urat i on.
The first adaptive controller to be considered
was the GMV approach of Clarke and Gawt hr op
(1975, 1979) using design transfer-functions:
p(q- 1 ) = ( 1 - 0 . 5 q - 1 ) / 0 . 5 and
Q(q- 1) = (1 - q- ~)/(1-0.5q- ~),
with the "det uned model-reference" interpretation.
This implicit self-tuner used a k-step-ahead predi-
ctor model with 2F(q-1) and 5 G(q-1) parameters
and with an adopt ed fixed value for k of 2. The
det uned version of GMV was chosen as the dead-
time was known to be varying and the use of Q(q- 1)
makes the design less sensitive to the value of k.
Not e in part i cul ar t hat for models 2 and 3 the
real dead-time is greater t han in the two samples
assumed.
The si mul at i on results using GMV are shown in
Fig. 3; reasonable if not particularly "t i ght " cont rol
was achieved for models 1, 3, 4 and 5. The weighting
of control increments (as found in other cases)
contributes to the overshoots in the step responses.
The behavi our with model 2 was, however, less
acceptable with poorl y damped transients. Never-
theless, the adapt at i on mechanism worked well and
the responses are certainly better t han with the
nonadapt i ve PI D controller.
An increasingly popul ar met hod in adaptive
General i zed predictive c o n t r o l - - Pa r t I 145
6 0 o/~
- l a */
Out put ( Y )
Systeml~ change
'~ 400
Syst em chonge
I OO */~
0",
-I00"/@
Cont r ot si gnal
400
FIG. 2. The fixed PID controller with the variable plant.
60 / -
- i o o / ~
Out put ( Y
S y s t e m ~ c h a n g e
, 1 I 1 1
' t ' ' t ' '
Syst em ~ change Syst em change Syst em change ~oo
Cont r o I . signaL
400
FIG. 3. The behaviour of the adaptive GMV controller.
cont r ol is pol e-pl acement (Wellstead et al . , 1979) as
vari at i ons in pl ant dead-t i me can be cat ered for
using an augment ed B( q- 1) pol ynomi al . To cover
the range of possibilities of dead-t i me here 2 A( q - ~)
and 6B(q -1) par amet er s were est i mat ed and the
desired pol e-posi t i on was specified by a pol ynomi al
p ( q - l ) = 1 - 0 . 5 q -1 (i.e. like the GMV case
wi t hout detuning). Fi gure 4 shows how an adapt i ve
pol e-pl acer which solves the cor r espondi ng
Di ophant i ne equat i on at each sampl e coped with
the set of si mul at ed models. For cases 1, 2 and 5
the out put response is good with less over shoot
t han t he GMV appr oach. For t he fi rst -order model s
3 and 4, however, the al gor i t hm is ent i rel y ineffective
due to singularities in t he Di ophant i ne solution.
This verifies the observat i ons t hat pol e-pl acement
is sensitive t o model - or der changes. In ot her cases,
even t hough the response is good, the cont r ol is
r at her "jittery".
The GPC cont r ol l er i nvol ved the same number s
of A and B (2 and 6) par amet er s as in t he pole-
pl acement simulation. Defaul t settings of t he out put
and cont r ol hor i zons were chosen with N 1 = I,
N 2 = 10 and N U = 1 t hr oughout , as it has been
f ound t hat this gives r obust performance. Fi gure 5
shows the excellent behavi our achi eved in all cases
6 0 ,~
-IOO/o
lOG / *
0 */ *
-IOO/o
Output ( Y ]
System change Systerr change
, I
i t ,
System change Syst em change 400
Cont r ol si gnal
146 D. W. CLARKE e t al .
0 400
FIG. 4. The behavi our of the adaptive pole-placer.
6O*/o Out put ( Y )
- i o /o
o
Cont rol signal
o o I I I . . . . .
o
i I
a o o
FIG. 5. The behavi our of the adaptive GPC algorithm.
by the GPC algorithm. For each new model only
at most two steps in the set-point were required for
full adaptation, but more importantly there is no
sign of instability, unlike all the other controllers.
6. CONCLUSI ONS
This paper has described a new robust algorithm
which is suitable for challenging adaptive control
applications. The method is simple to derive and
to implement in a computer; indeed for short
control horizons GPC can be mounted in a micro-
computer. A simulation study shows that GPC is
superior to currently accepted adaptive controllers
when used on a plant which has large dynamic
variations.
Montague and Morris (1985) report a compar-
ative study of GPC, LQG, GMV and pole-place-
ment algorithms for control of a heat-exchanger
with variable dead-time (12-85s, sampled at 6s
intervals) and for controlling the biomass of a
penicillin fermentation process. The controllers
were programmed on an IBM PC in pro-
Fortran/proPascal. Their paper concludes that the
GPC approach behaved consistently in practice
and that "the LQG and GPC algorithms gave the
best all-round performance", the GPC method
Generalized predictive cont r ol - - Par t I 147
being preferred for long time-delay processes, and
it confirms that GPC is simple to implement and
t o u s e .
The GPC method is a further generalization of
the well-known GMV approach, and so can be
equipped with design polynomials and transfer-
functions which have interpretations as in the GMV
case. The companion paper Part II, which follows,
explores these ideas and presents simulations which
show how GPC can be used for more demanding
control tasks.
e = [El ( t + 1),E2~(t + 2) . . . . . EN(t + N)] r.
Cl earl y as e is a st ochast i c pr ocess t he expect ed value of t he
cost subj ect t o appr opr i at e condi t i ons must be mi ni mi zed. The
cost t o be mi ni mi zed t herefore becomes
J l = E{(y - w)T(y - w) + )~firfi}.
Two different assumpt i ons will be made in t he next sect i on
about t he class of admi ssi bl e cont r ol l er s for t he mi ni mi zat i on
of t he cost above. Onl y model s of aut oregressi ve t ype are
consi dered:
A( q- l)y(t) = B( q- ~)u(t - 1) + (t)/A.
REFERENCES
As t r r m, K. J. and B. Wi t t enmar k (1973). On sel f-t uni ng regula-
tors. Aut omat i ca, 9, 185-199.
, ~st r r m, K. J. and B. Wi t t enmar k (1980). Sel f-t uni ng cont r ol l er s
based on pol e-zero pl acement . Proc. l EE, 127D, 120-130.
Bertsekas, D. P. (1976). Dynami c Programmi ng and St ochat i c
Control. Academi c Press, New York.
Cl arke, D. W. (1984). Sel f-t uni ng cont r ol of nonmi ni mum- phas e
syst ems. Aut omat i ca, 20, 501-517.
Cl arke, D. W. and P. J. Ga wt hr op (1975). Self-tuning cont rol l er.
Proc. I EE, 122, 929-934.
Clarke, D. W. and P. J. Ga wt hr op (1979). Sel f-t uni ng cont rol .
Proc. I EE, 126, 633-640.
Clarke, D. W. and L. Zhang (1985). Does l ong- r ange predi ct i ve
cont r ol work? l EE Conf er ence "Cont r ol 85", Cambr i dge.
Cut l er, C. R. and B. L. Ramaker (1980). Dynami c Mat r i x
Co n t r o l - - A Comput er Cont rol Algorithm. JACC, San Fr anci -
sco.
De Keyser, R. M. C. and A. R. Van Cauwenber ghe (1982).
Typical appl i cat i on possi bi l i t i es for sel f-t uni ng predi ct i ve
cont rol . I F A C Syrup. Ident. Syst . Param. Est., Washi ngt on.
De Keyser, R. M. C. and A. R. Van Cauwenber ghe (1983).
Mi cr ocomput er - cont r ol l ed servo syst em based on sel f-adap-
tive l ong- r ange predi ct i on. Advances in Meas ur ement and
Co n t r o l - - - MECO 1983.
De Keyser, R. M. C. and A. R. Van Cauwenber ghe (1985).
Ext ended pr edi ct i on sel f-adapt i ve cont rol . I F A C Syrup. l dent .
Syst . Param. Est., York.
Kai l at h, T. (1980). Li near Syst ems. Prent i ce-Hal l , Engl ewood
Cliffs, NJ.
Kurz, H. and W. Goedecke (1981). Digital par amet er - adapt i ve
cont r ol of processes wi t h unknown dead time. Aut omat i ca,
17, 245-252.
Mont ague, G. A. and A. J. Mor r i s (1985). Appl i cat i on of adapt i ve
cont rol : a heat exchanger syst em and a penicillin f er ment at i on
process. Thi r d wor ks hop on t he t heor y and appl i cat i on of
sel f-t uni ng and adapt i ve cont rol , Oxford, U. K.
Mosca, E., G. Zappa and C. Manf r edi (1984). Mul t i st ep hor i zon
sel f-t uni ng cont rol l ers: t he MUS MAR appr oach. I FAC 9th
Wor l d Congr ess, Budapest , Hungar y.
Pet erka, V. (1984). Pr edi ct or - based sel f-t uni ng cont rol . Aut oma-
tica, 20, 39-50.
Richalet, J., A. Rault, J. L. Test ud and J. Pa pon (1978).
Model predi ct i ve heuri st i c cont rol : appl i cat i ons t o i ndust ri al
processes. Aut omat i ca, 14, 413-428.
Schweppe, F. C. (1973). Uncert ai n Dynami c Syst ems. Pr ent i ce
Hall, Engl ewood Cliffs, NJ.
Tufts, P. S. and D. W. Cl ar ke (1985). Sel f-t uni ng cont r ol of
offset: a unified appr oach. Proc. I EE, 132D, 100-110.
Wellstead, P. E., D. Pr ager and P. Zanker (1979). Pol e assi gnment
sel f-t uni ng regul at or. Proc. l EE, 126, 781-787.
Ydstie, B. E. (1984). Ext ended hor i zon adapt i ve cont rol . I FAC
9th Wor l d Congr ess, Budapest , Hungary.
AP P ENDI X A. MI NI MI ZATI ON PROPERTI ES OF GP C
Cons i der t he per f or mance i ndex
J = (y - w)r(y - w) + 2iiTii wher e y = Gii + f + e
wher e y, w, fi are as defi ned in t he mai n t ext and
In t he general case of CARI MA model s wher e t he di st ur bances
are non- whi t e t he pr e di c t i ons f ( t + i) are onl y opt i mal as ympt ot -
ically as t hey have pol es at t he zeros of t he C( q- 1) and t he effect
of initial condi t i ons is onl y el i mi nat ed as t i me t ends t o infinity.
The cl osed- l oop pol i cy requi res an opt i mal Ka l ma n filter wi t h
t i me-varyi ng filter gains, see Schweppe (1973); t he par amet er s
of t he C- pol ynomi al are onl y t he gai ns of t he st eady- st at e
Kal man filter once conver gence is at t ai ned. Thus t he cl osed-
l oop and t he open- l oop- f eedback policies exami ned bel ow are
different in t he case of col our ed noise.
A.1. RELATI ON OF GP C AND OP E N- L OOP -
F EEDBACK- OP TI MAL CONTROL
In t hi s st rat egy it is assumed t hat t he fut ure cont r ol si gnal s
are i ndependent of fut ure measur ement s (i.e. all will be per f or med
in open- l oop) and we cal cul at e t he set fi such t hat it mi ni mi zes
t he cost J~. The first cont r ol in t he sequence is appl i ed and at
t he next sampl e t he cal cul at i ons are repeat ed. The cost becomes:
J l = E{iiT( Gr G + 2I)fi + 2/ i r ( GT( f - W) + Gre)
+ f r f + ere + WTW + 2fr(e _ w) -- 2eTw}.
Because of t he as s umpt i on above, E{/iTGTe} = 0; not e al so t hat
t he cost is onl y affected by t he first t wo t erms. By put t i ng t he
first deri vat i ve of t he cost equal t o zer o one obt ai ns:
fi = ( Gr G + 21)- IGT(w - t).
A.2. RELATI ON OF GP C AND CLOS ED- LOOP -
F EEDBACK- OP TI MAL CONTROL
Choos e t he set ii such t hat t he cost J l is mi ni mi zed subj ect
to t he condi t i on t hat Au(i) is onl y dependent on dat a up t o t i me
i. Thi s is t he st at ement of causality. In or der to find t he set ii
assume t hat by some means {Au(t), Au(t + 1) . . . . . Au(t + N - 2)}
is available.
Jl = E{(Au(t) . . . . .
Au(t + N - 1))(GrG + ~I)(Au(t) . . . . .
A u ( t + N - 1) "r
+ 2 i l r ( Gr ( f - w))
+ 2(Au(t) . . . . . Au(t + N - 1))Gre}
+ ext r a t er ms i ndependent of mi ni mi zat i on.
Not e t hat in eval uat i ng Au(t + N - 1) t he last row Gr e is
g o ~ e i ~ ( t + N - i ) and E { A u ( t + N - 1) ( t +N) } = 0 by t he
as s umpt i on of part i al i nf or mat i on. Ther ef or e Au(t + N - 1) can
be cal cul at ed assumi ng ~(t + N) = 0. Fr om general st at e- space
consi der at i ons it is evi dent t hat when mi ni mi zi ng a quadr at i c
cost, assumi ng a l i near pl ant model wi t h addi t i ve noise, t he
resul t i ng cont r ol l er is a l i near funct i on of t he avai l abl e measure-
ment s (i.e. u(t) = - k ( t ) wher e k is t he feedback gai n and ~(t)
is t he st at e est i mat e from t he appr opr i at e Ka l ma n filter).
Au(t + N - 1) = linear funct i on(Au(t + N - 2),
A u ( t + N - 3 ) . . . . . ~ ( t + N- 1 ) ,
~(t + N - 2) . . . . . y(t ), y(t - 1),...).
148 D. W. CLARKE et al .
In c a l c u l a t i n g A u ( t + N 2) t he f o l l o wi n g e x p e c t a t i o n s ar e zer o:
E I A u l t + N 2)~(t + N) ] - 0,
E f L A u ( t + N 2)~{t + N 1) ~, - 0.
I n o r d e r t o wo r k o u t E I A u { t + N - 1)Au(t + N - 2)I, t he pa r t
of t he e x p e c t a t i o n a s s o c i a t e d wi t h {(t + N - 1) i s set t o zer o as
t he c o n t r o l s i gna l A u ( t + N 2) c a n n o t be a f unc t i on of
5 , ( t + N - 1 ) . As A u ( t + N - 1 ) is a l i n e a r f unc t i on of
( t + N- I) t hi s i mpl i e s t h a t as f ar as c a l c u l a t i o n of
A u ( t + N - 2) i s c o n c e r n e d , A u ( t + N - 1) c o u l d be c a l c u l a t e d
by a s s u mi n g t h a t {(t + N - 1) a n d {(t + N) wer e zer o. Si mi l a r l y,
t he s a me a r g u me n t a p p l i e s f or t he r est of c o n t r o l s i gna l s a t e a c h
s t e p i a s s u mi n g E { A u ( t + i ) { ( t + j ) t _ 0 f o r ) > i. Th i s i mpl i e s
t h a t e a c h c o n t r o l s i gna l A u ( t + i ) i s t he i t h c o n t r o l i n t he
s e q u e n c e a s s u mi n g t h a t (t + j ) = 0 f or j > i. He n c e f or Au(t)
t he s o l u t i o n a mo u n t s t o c a l c u l a t i n g t he f i r st c o n t r o l s i g n a l
s e t t i n g {(t + j ) f or j > 0 t o zer o t he s a me as t h a t c o n s i d e r e d
i n pa r t I f or a n Op e n - L o o p - F e e d b a c k - Op t i ma l c ont r ol l e r . No t e
t h a t i n t he cas e of N U < N 2 t he mi n i mi z a t i o n i s p e r f o r me d
s ubj e c t t o t he c o n s t r a i n t t h a t t he l a s t N 2 - N U + 1 c o n t r o l
s i g n a l s a r e al l equal . Co n s i d e r a t i o n of c a u s a l i t y of t he c o n t r o l l e r
i mp l i e s t h a t A u ( t + N U - 1) i s t he fi rst free s i gna l t h a t c a n be
c a l c u l a t e d as a f unc t i on of d a t a a v a i l a b l e u p t o t i me t + N U - 1
i n t he s e que nc e a n d t he a r g u me n t a b o v e f ol l ows a c c or di ngl y.
The r e f or e , t he GP C mi n i mi z a t i o n ( OL F O) i s e q u i v a l e n t t o t he
c l o s e d - l o o p pol i c y f or mo d e l s of r e gr e s s i on t ype.
AP P E NDI X B. AN E XAMP L E F OR A F I RS T OR DE R
P R OC E S S
Th i s a p p e n d i x e x a mi n e s t he r e l a t i o n of N 2 a n d t he c l os e d-
l o o p pol e p o s i t i o n f or a s i mp l e e x a mp l e . Co n s i d e r a n o n mi n i m-
ur n- pha s e f i r s t - or de r pr oc e s s ( f i r s t - or de r + f r a c t i ona l de a d- t i me ) :
(1 + a l q - 1 ) y ( t } = ( b o + b l q - l l u ( t - I)
1 1 - - 0 . g q - l ) y { t ) = ( I + 2 q - l ) u l t - - I ) .
B. I. E AND F P AR AME T E R S
F r o m t he r e c ur r e nc e r e l a t i o n s h i p s i n t he ma i n t ext i t f ol l ows
t hat :
e i = 1 - - a l e i _ G.lio - 1 -- J l G . t i l = e i a a
eo = l : [ 1. O] ; f l o = (1 - aO: [ 1. 9] ; . l ; ~ = a , : [ - 0 . 9 ]
e 1 = 1 - a l : [ l . 9 ] ; J 2 o = 1 - al{1 - a 0 : [ 2 . 7 1 ] ;
J21 = a d l - a 0 : [ - 1.71]
e 2 = 1 - a d l - a l ) : [ 2. 71] ;
f3o = 1 a l ( l -- at ( 1 -- a 0 ) : [ 3 . 4 3 9 ]
.131 = a d 1 - al ( 1 - a 0 ) : [ - 2 . 4 3 9 ] -
B.2. C ONT R OL L E R P AR AME T E R S
go = bo: [ 1. 0] ; g t 1 = b~: [ 2. 0]
g l = bo{l - a l ) + b l : [ 3 . 9 ] ; g 2 2 = bt{1 - a 0 : [ 3 . 8 ]
g2 = b0( l - al ( 1 - a~)) + bl (1 - a t ) : [ 6. 51] ;
g33 = bl ( 1 - a d l - a 0 ) : [ 5 . 4 2 ] .
As s u mi n g N U = 1, t he effect of N 2 o n t he c o n t r o l c a l c u l a t i o n
a n d t he c l o s e d - l o o p pol e p o s i t i o n i s n o w e x a mi n e d :
A u { t ) = ( ~ , g d w , - J i o y ( t - 1 ) - l i l y ( t - 2 ) )
- g , A u t t - l)y~ (g~)2.
Fo r a c o n t r o l l e r of t he f or m R l q ~ l A u { t ) = w ( t ) - S ( q ~l,~,t~
'SO = 'ff'~ g i . iO, L,~gt
. ~ 1 = Y . ~ , J i , . . ' Z g ,
2 ,
ro = Z ( g a . ' ~ . ~ ,
r , = ~ , , g l g l i , " ~ g i
a n d t he c l o s e d - l o o p p o l e - p o s i t i o n s a r e at t he r o o t s of
R A A + q - t S B .
N2 = I
Au(t) = [ w - 1.9y(t) + 0. 9y(t - 1} - 2Au( t - 1)]
R A A + q - I S B : : : > ( I + 2q- 1) .
Th i s i s t he mi n i mu m- p r o t o t y p e c o n t r o l l e r whi c h, b e c a u s e of t he
c a n c e l l a t i o n of t he n o n mi n i mu m- p h a s e zer o, i s uns t a bl e .
N 2 = 2 ( i . e . g r e a t e r t h a n deg(B))
Au(t) = [ w - 1.9y(t) + 0. 9y(t - 1) - 2Au(t - 1)
+ 3.91w - 2.71y~t} + 1.71y(t - 1)
- 3. 8Au(t - 1))]/16.21
o r ~
and:
Au(t ) = [ 4. 9w - 12. 469y(t) + 7. 569y(t - 1)
- 16. 82Au(t - 1)]/16. 21
A R A + q - I S B ~ ( I - 0. 09q- 1).
Th e c l o s e d - l o o p pol e i s wi t h i n t he u n i t ci r cl e a n d t her ef or e t he
c l o s e d - l o o p i s s t abl e. I n f act i t c a n e a s i l y be d e mo n s t r a t e d t h a t
f or al l f i r st o r d e r pr oc e s s e s N2 > 6 B ( q - 1 ) + k - 1 s t a bi l i z e s a n
/ k
o p e n - l o o p s t a b l e pl a nt , f or s i g n ( Z g i ] = s i g n ( D. C. g a i n ).
\ /
N2 = 3
and:
Au(t ) = [ 11. 41w - 34. 857y{t) + 23. 447y( t - 1)
- 52. 104Au(t - 1)]/58. 591
A R A + q I S B = * . ( 1 - 0 . 4 1 6 q - 1 ) .
Th e pol e i s a g a i n wi t h i n t he uni t ci rcl e. No t e t h a t t he c l os e d-
l o o p pol e i s t e n d i n g t o wa r d s t he o p e n - l o o p pol e as N 2 i nc r e a s e s
(see Pa r t II).