Industrial concern regarding the use of 'black box' models has underst
andably limited their routine application. In an attempt to overcome s
ome of these concerns the technique of hybrid modelling (Thompson, M.
L. and Kramer, M. A., AIChE Journal, 1991, 40, 1328-1340) has become a
popular alternative. It has been applied to many different processes
and has shown a marked improvement on the now more traditional data-ba
sed models such as artificial neural networks (Psichogios, D. C. and U
ngar, L. H., AIChE Journal, 1992, 38, 1499-1511) and NARMAX (Billings,
S. A., Gray J. O. and Owens, D. H., Nonlinear Systems Design. Peter P
eregrinus Ltd, 1984) structures. Hybrid models consist of a 'black box
' model combined with a model of predefined structure (often mechanist
ic/first principles model) thereby gaining the advantages of both mode
lling techniques, namely the non-linear capabilities of a generally st
ructured neural network and the extrapolation capabilities of a more r
igid model. In this paper the ability of two different hybrid modellin
g techniques, combining an artificial neural network with a model of a
ppropriate structure, to predict loss of head of pressure profiles in
a water treatment plant, are compared. Data were obtained at 15-min in
tervals from on-line plant sensors. The input variables used were inco
ming water flowrate, raw water turbidity, supernatant return flowrate,
pH of water entering the filters and solids build up on the filters.
Results show the serial structure to be more robust allowing extrapola
tion outside the range of the training data set as compared to the par
allel arrangement yielding a 20% reduction in RMS validation error. (C
) 1997 Elsevier Science Limited.