A two-stage algorithm for structure identification of polynomial NARX models

Abstract

This paper deals with the problem of structure identification for polynomial NARX models in long-term prediction, based on the minimization of the simulation error. The effect of the sampling time on structure selection is analyzed first, comparing the identification approach with classical prediction error minimization (PEM) methods. A two-stage identification algorithm is then proposed, to cope with the high computational load inherent in simulation-based approaches. The first stage performs a coarse identification of the model structure considering oversampled input-output data, while in the second stage the structure is iteratively refined considering a decimated version of the data.