NONLINEAR SYSTEM IDENTIFICATION AND PREDICTIVE CONTROL USING ORTHONORMAL BASIS FUNCTION (OBF)-NEURAL NETWORKS MODEL

Nonlinear empirical models in general, and neural networks models in particular,
are normally developed using data from limited region of experimentation. In
practice, time and cost are the two main factors that restrict the complete coverage of
the whole input space. As a result, these models tend to be performing poorly in
regions beyond the original operating conditions (also known as extrapolation
regions). In practice however, the underlying conditions of every process plant are
continually changing and extrapolation is completely inevitable.
In this thesis, a nonlinear system identification framework using linear-plus-neural
networks model is developed to address the widely acknowledged extrapolation
limitations inherent in conventional neural networks models. The framework is
established by integrating a linear Orthonormal Basis Filter (OBF) model and a
nonlinear multi-layer perceptron neural networks (NN) model in a parallel structure.
The overall nonlinear model is then taken as the sum of these two models. A parallel
OBF-NN model is therefore obtained.