Improving multi-step series prediction with recurrent neural modelling

Abstract

Multi-step prediction is a difficult task that has been attracted increasing the interest in recent years. It tries to achieve predictions several steps ahead into the future starting from information al time k. This paper is focused on the development of nonlinear neural models with the purpose of building long-term or multi-step time series prediction schemes. In these contexts, the most popular neural models are based on the traditional feedforward neural network. However, these kinds of models may present some problems when a long-term prediction problem is formulated. In this paper, a neural model based on a partially recurrent neural network is proposed as an alternative. For the new model, a learning phase with the purpose of long-term prediction is imposed, which allows to obtain better predictions of time series in the future. This recurrent neural model has been applied to the logistic time series with the aim to predict the dynamic behaviour Di the series in the future. Mode1s based on feedforward neural networks have been also used and compared against the proposed model.