ANFIS Synthesis by Hyperplane Clustering for Time Series Prediction

Abstract

Time series prediction can be considered as a function approximation problem whose inputs are determined by using past samples of the sequence to be predicted. ANFIS networks are neural models particularly suited to the solution of such problems, which usually require a data driven estimation technique. In this context, clustering procedures represent a straightforward approach to the synthesis of ANFIS networks. A novel use of clustering, working in the joint input-output data space, is proposed in the paper. It is intended to improve the ANFIS approximation accuracy by directly estimating the hyperplanes associated with the consequent parts of Sugeno first-order rules. Simulation tests and comparisons with other prediction techniques are discussed to validate the proposed synthesis approach. In particular, we consider the prediction of environmental data sequences, which are often characterized by a chaotic behaviour.