Using Reservoir Computing for Forecasting Time Series: Brazilian Case Study

This paper presents a Brazilian case study of forecasting a wind speed time series with reservoir computing (RC). RC is a research area, in which an untrained recurrent network of nodes is used for the recognition of temporal patters. In RC only the weights of the connections in a linear output layer are trained. This reduces the complexity of recurrent neural networks (RNN) training to simple linear regression. In this work we used echo state network (ESN) to create the case study and compare the results with Multilayer Perceptron Networks and persistence method. Our case study concerns forecasting the wind speed, which is fundamental information in the operation planning for electrical wind power systems. The results showed that the RC performed significantly better than multilayer perceptron networks or persistence method, even though it presents a significantly simpler and faster, training algorithm.