Redes neurais dinâmicas para predição e modelagem não-linear de séries temporais

Abstract

In this work, dynamic neural networks are evaluated as non-linear models for efficient prediction of complex time series. Among the evaluated architectures are the FTDNN networks, Elman and NARX. The predictive power of these networks are tested in prediction task a step ahead and multiple-steps-forward. To this end, the following time series are used: Series Laser chaotic Mackey-Glass chaotic series, and network traffic series of computers with self similar characteristics. The use of NARX network prediction time series is a contribution of this thesis. This network has a recurrent neural architecture originally used to identify input-output nonlinear systems. The input NARX network is formed by two sliding windows (sliding window time), one slipping over the other input signal and which slides on the output signal. When applied to chaotic time series prediction, the NARX network is usually designed as an autoregressive nonlinear model (NAR), eliminating the output delay window. In this paper, we propose a simple strategy, but effective to allow the network NARX fully explore the input time slots and output in order to improve its predictive ability. The results show that the proposed approach outperforms the performance presented by predictors based on FTDNN and Elman networks.