Nonlinear time series modeling and prediction sing RBF network with improved clustering algorithm

Modeling and prediction of time series are important problems in various fields. Radial basis function (RBF) networks are able to approximate any continuous nonlinear function with any accuracy and have been applied successively to nonlinear time series modeling and prediction. One crucial problem for training the RBF network is that the number and locations of the centers in the hidden layer should be selected properly, or the network will perform badly. In this paper, an improved clustering algorithm is proposed, which can set an optimal centers configuration for the RBF network. Simulations results show that the improved clustering algorithm outperforms the previous clustering method for clustering analysis, and the RBF network trained with it achieves good generalization performance for nonlinear time series modeling and prediction.