The Research Of Chaotic Time Series Prediction Methods Based On Modified ESN

Chaotic time series prediction is highly important and challenging jobs in chaotictime series analyst, and has always been the hot research topic. First, this paper makessystematical analysis on chaotic theory and summarizes some common chaotic timeseries prediction method based on the intensive research of phase space reconstruction.According to the characteristics of modeling data, chaotic time series predictionmethods are separated into chaotic global prediction method, local prediction methodand adaptive prediction method, and it also expounds each prediction methodsystematically. This paper puts the emphasis on the predominant global predictionmethod and deeply investigates one of the neutral network prediction methods.Second, the simulation results of variety of prediction methods show that recurrentneural network has better performance on chaotic time prediction. Based on this, bystudying the structure and learning principles of a new recurrent neural networknamed Echo State Network (ESN), the superiority in chaotic time series prediction ispointed out by building a simulation model. Last, this paper proposes a modifiedmethod to solve the problem of selecting parameters of ESN in chaotic time seriesprediction, and develop correlative simulation software.The ESN not only has great significance to theory research, but also has highlypractical value. Among all the problems during the ESN research, how to create andtrain Dynamic Reservoir (DR) for specific problems become one of the mostimportant problems, and the DR parameter’s selection becomes more crucial intraining Reservoir, which has great influence on performance of ESN in chaotic timeseries prediction. So far as, there is no exact way to create relevant optimal trainingparameters to solve different problems, mostly depending on personal experience ortrial error to setting parameters, which brings heavy task to practical applications.Therefore, this paper introduces Differential Evolution algorithm (DE) which withstrong global optimizing ability, and proposes a chaotic time series prediction modelcombined Differential Evolution algorithm with Echo State Network. With thisDE-ESN model, to train the inputted sample sequences to find suitable datacharacteristic training parameters of ESN and then do chaotic time series predictionby using the gotten ideal parameters.Simulation experimental results show the method based on Differential Evolutionalgorithm to optimize the ESN Reservoir parameters could adapt to different datacharacteristics to find the relative optimal parameters to build ideal prediction modelwhich improves the accuracy of chaotic time series prediction effectively.