Reservoir neural state reconstruction and chaotic time series prediction

Abstract

ESN(Echo state network) is a new type of recurrent neural network, which is based on the “reservoir”. ESN has been proved to be significantly efficient to deal with some chaotic time series prediction tasks. This paper makes an analysis of the current iterative prediction method based on “reservoir”, and points out some problems in theory and the obstacles in application. And then, a direct prediction method is proposed, which relates the prediction origin and prediction horizon directly. The direct method does not close the loop in the process of prediction, so there are no such problems as instability and error accumulation. The simulation results show how the reservoir property changes when the loop is closed in the iterative prediction, and then demonstrate the feasibility of the direct prediction method in application to the Mackey-Glass benchmark prediction problem.