Neural network modelling of nonlinear systems for controller design

Nonlinear control design usually requires the system dynamics to be in some canonical form. Neural networks appear a promising tool to generate such canonical forms. The paper addresses some fundamental neural network modelling issues which must be considered in order to generate satisfactory models. Two ways of comparing a nonlinear model with its NN model are proposed, namely static comparison and dynamic comparison. Three extension concepts are proposed. Two of them, extension over time and extension over initial conditions, are discussed. These are closely related to the application of NN models in controller design. Pertinent examples are given.