Nonlinear time series prediction by using RBF network

Abstract

This paper describes a numerical algorithm for short-term prediction of nonlinear time series by using time-delay embedding and radial basis function (RBF) neural networks. Unlike the existing RBF algorithms with centers preselected during training process and fixed during prediction process, the proposed method utilizes a simple selection algorithm to dynamically change the center positions, resulting in a local RBF model with time varying parameters. Analysis and methodology are detailed in the context of the Leuven competition. Results show that the proposed local dynamical RBF network performed remarkably well. © 2009 Springer Berlin Heidelberg.