Neural modelling of chemical plant using MLP and B-spline networks

This paper demonstrates the potential of the B-spline neural network (BSNN) for the modelling of nonlinear processes. The performance of this paradigm is compared with the multi-layer perceptron (MLP), for the modelling of experimental and industrial case-studies. The experimental pH neutralisation plant of the University of California at Santa Barbara (UCSB) is modelled using both networks. This case-study demonstrates the higher performance of the B-spline network, for rapid on-line model adaptation. The second case-study focuses on the prediction of viscosity for an industrial polymerisation reactor. Predictive models of viscosity are developed, based on both networks to predict over and hence remove the measurement time-delay introduced at the viscometer. A novel disturbance modelling approach is also developed here, and demonstrated to perform excellently in on-line tests.