On Model Structure Selection for Nonparametric Prediction Methods

Abstract

In this paper we continue to explore identification of nonlinear systems using the previously proposed concept of model-on-demand. The idea is to estimate the process dynamics locally and on-line using process data stored in a database, and has in earlier contributions proven to be capable to produce results comparable to (or better than) other nonlinear black-box approaches. The modeling part of the method is based on local polynomial modeling ideas. This has several implications on the choice of model structure, which is discussed at length in the paper. It is concluded that the NARX structure should be considered as the default choice in the local polynomial context. Furthermore, it is shown that the predictions in some situations can be enhanced by tuning other parameters that are special for the nonparametric case. The usefulness of the method is illustrated in numerical simulations. For the chosen application it is shown that the prediction errors are in order of magnitude directly comparable to more established modeling tools such as artificial neural nets.