Copyright @ IFAC Adaptive Systems in Control and
Signal Processing, Glasgow, Scotland, UK, 1998
TWO-DEGREE-OF-FREEDOM STABLE GPC DESIGN
Zdzislaw Kowalczuk and Piotr Suchomski
Technical University ofGdGllsk, Department of Automatic Control (W.E. T.l.),
Narutowicza 11/ 12, P.G.Box 612, 80-952 Gdansk, Poland; e-mail kova@pg. gda . pl
Abstract: The problem of discrete-time generalised predictive control of plants
described by discrete-time polynomial models, which can have either minimal or non-
minimal characteristics, is deliberated. Design specifications for robust GPC control are
presented and a stable two-degree-of-freedom formulation of the GPC control strategy is
proposed in which nominal stability, nominal performance and robust stability
properties are obtained by properly choosing the control and observation horizons and
the observer polynomial. Copyright © 19981FAC
Keywords: system design, polynomials, predictive control, adaptive and robust control.
l. INTRODUCTION
The Generalised Predictive Control approach of
Clarke et aJ. (1987) has attracted a considerable
attention in the control literature. In that
methodology the controlled plant is described by the
following discrete-time model of the CARIMA type
(1)
where {u(t)} and {y(t)} are the input and output
signals, {v(t)} is a zero-mean white-noise, q -I is the
backward shift operator, and 6. = 1- q -I. The
polynomials A(q-I), B(q-I) and C(q-I) are of
degree N A ' NB and N c N A + I, respectively. In
the above model an incremental-control channel and
a disturbance channel can be distinguished
(2)
where A(q-I) = M(q-I). Let nb indicate a pure
transportation delay in the control channel, and let
nB = nb + I . It is also assumed that NB N A + I
that confines the dimension of the minimal state
space representation of this channel to N A + I .
207
1.1 Basic GPC control/er design
Future controls 6.u(t) = [6.u(t)... 6.u(t + Nu-I) r '
where Nu denotes the control horizon, is sought
after via minimisation of the quadratic cost function
J(6.u(t)) =
N2 Nu
= +i) -8y(t +i)f +A.L 6.u(t +i _1)2,
(3)
;=1
where, for i = I, ... , N
2
, e(t + i) = e(t + ilt) = lie(t) is
a sequence scaled by the control error
e(t) = wet) - yet) (within Nu a constant set point
w(t +i) = wet) is assumed), the 'i 's are coefficients
of the step response of anticipation filter (AF), and
8y(t + i) = y(t + i) - y(t) denotes a trajectory of the
predicted incremental plant output, while y(t +i)
stands for the minimum-variance i-step ahead
predictor of the plant output. For the model (1) the
predictor can be represented in the following form
.v(t +i) = H,(q - I)6.u(t+i -1)+ y(t+ilt), (4)
where, for i=I , ... ,N
2
, H,(q -I) =ho+· ·-+h;_lq-(i-I) ,
and y(t + il t) denotes the free component that can be
obtained by employing the following formula