Predictive control of the outputs or states of a system relies on the
quality of the predictions over a certain time horizon into the future
. The conventional ways of obtaining these predictions suffice when th
e system characteristics allow either the development of a relatively
accurate model or the use of a relatively short time horizon. Such is
the case for many continuous processes, where the system dynamics rend
er short prediction horizons adequate, or where relatively simple, pos
sibly on-line adapted, low-order, input-output models yield accurate p
redictions through integration. However, conventional methods cannot y
ield reliable predictions for processes that are poorly modelled and r
equire long prediction horizons for good control performance. Batch pr
ocesses especially belong in this category, due to their high-order ti
me-varying characteristics. Their changing nonlinear behaviour allows
only inadequate state-space modelling from first principles, and the s
ignificant long-range future effect of their current manipulated input
dictates the use of long prediction horizons for satisfactory control
. This paper presents a reliable strategy for long-range prediction in
poorly-modelled systems. A state-space-based combination of two neura
l networks in series is employed to predict system outputs and unmeasu
red states over a long horizon into the future. A simulation example o
f a poorly modelled, exothermic, semi-batch reaction system demonstrat
es the excellent prediction capability of the new structure. The resul
ts establish that the new strategy solves a realistic prediction probl
em where the previously available alternatives fail.