A strategy for time series prediction using Segment Growing Neural Gas

Abstract

Segment Growing Neural Gas (Segment-GNG) has been recently proposed as a new spatiotemporal quantization method for time series. Unlike traditional quantization algorithms that are prototype-based, Segment-GNG uses segments as basic units of quantization. In this paper we extend the Segment-GNG model in order to deal with time series prediction. First Segment-GNG makes a quantization of the trajectories in the state-space representation of the time series. Then a local prediction model is associated with each segment, which allows us to make predictions. The proposed model is tested with the Mackey-Glass and Lorenz chaotic time series in one-step ahead prediction tasks. The results obtained are competitive with the best results published in the literature.