Contribuições ao problema de predição recursiva de séries temporais univariadas usando redes neurais recorrentes

Abstract

In this thesis, we tackle the problem of recursive prediction of univariate time series, also known as long-term prediction, using recurrent neural networks. This type of problem often emerges from nonlinear dynamical systems modelling and prediction tasks, particularly from those producing signals of chaotic nature, where one can observe the presence of long-term temporal dependencies. In recursive prediction, differently from the one-step-ahead prediction task, predicted values are fed back to the input of the neural model, a feature that makes time series with long-term temporal dependencies more difficult to deal with due to the propagation of prediction errors. That being said, in order to handle the problem of recursive prediction of univariate time series, extensions of the neural NARX (Nonlinear AutoRegressive model with eXogenous inputs) model ar eintroduced in this thesis. These extensions result from attempts to embed into the NARX model different strategies to capture temporal information, either of short-term or long-term nature. Among such strategies, we highlight the following ones: (i) simultaneous prediction of several steps ahead, also known as MIMO (multi-input, multi-output model) prediction, (ii) prediction via dynamical random projections, as in the ESN (echo state network) model, (iii) prediction via static random projections, as in the ELM (extreme learning machine) network, and (iv) prediction via hybrid recurrent models based the NARX and ELMAN networks. Additionally, a novel methodology for the design (i.e. parameter selection) and performance comparison of the proposed models is also introduced in this model with the aim of evaluating them under similar conditions and to serve as reference for further studies. For this purpose, synthetic and real-world benchmarking time series are used. The obtained results suggest that the proposed neural models present themselves as efficient alternatives to the state of the art in recursive prediction of univariate time series using recurrent neural architectures.