Neuro-fuzzy networks-dynamic modeling and performance enhancement incorporating process knowledge

A study of the mapping capabilities of neuro-fuzzy networks in relation to conventional neural nets is carried out. Two representative systems, a time series model and an actual chemical process, are studied to analyze the ability of the empirical structure to capture complex nonlinear dynamics. The superiority of the neuro-fuzzy network in terms of its mapping ability is demonstrated. Performance enhancement of the empirical model is sought through incorporation of process knowledge into the identification procedure. The importance of appropriate choice of identification experiments and their role in model enhancement is highlighted through simulation studies. A nonlinear model predictive control scheme employing the neuro-fuzzy models is designed. The utility of this scheme in terms of its wide range of applicability is discussed