Research On Nonlinear Time Series Prediction Based On Echo State Networks

Nonlinear time series widely exists in the engineering projects and scientificresearch. Accurate predictions can provide a theoretical and practical support forscientific decision-making. Recently, the recurrent neural networks gradually becomeone of the main research directions of the time series prediction. But the traditionalRNN method has many problems, such as the complex training algorithm, slowconvergence, easy to fall into local optimum and difficult to determine the networkstructure. Echo state networks (Echo State Network, the ESN) trains only parts of theweights, thus it overcomes the inherent problems of RNN. ESN gradually become oneof the main tools for time series prediction field. Nonlinear time series prediction basedon echo state networks become a focus of theoretical research.This paper considers two aspects to carry out methods research of nonlinear timeseries prediction: parameters selection and the improvement of the network modelstructure.In the problem of reservoir parameters, this paper introduces genetic algorithm tooptimize reservoir parameters, and improve the traditional genetic algorithm mutationoperation to get a strong global optimization capability. This paper proposes echo statenetworks prediction model based on reservoir optimization to build the optimalreservoir.In improving the echo state networks model structure, considering differentprediction demands and the ability of handle multi-output of the echo state networks,this paper combines the respective advantages of iterative prediction method and directprediction method, we propose a fusion model structure to achieve precise multi-stepprediction of actual noisy time series.