A neural network approach to improving identification of nonlinear polynomial models

This paper presents a neural network approach to improving the identification of nonlinear polynomial model. The idea is to realize the identification in two steps. In the first step, a quasi-ARX neural network is first used to approximate the system under study; then a reasonable number of important monomial terms are selected as candidates, by introducing an importance index based on a Taylor expansion of the identified quasi-ARX neural network; In the second step, Genetic algorithm (GA) is applied to the selected important terms to further determine a set of significant terms to include in the polynomial model. In this way, the whole identification algorithm is implemented very efficiently. Numerical simulations are carried out to demonstrate the effectiveness of the proposed two-step identification method.