Noisy Chaotic Time Series Prediction Based on Wavelet Echo State Network

As a research focus of intelligence algorithm, the prediction of classic noiseless chaotic time series has a great development in recent years. However, the existing prediction models cannot get good performance for real-world chaotic time series because of the interference of noise components. In order to take full advantage of the property of real-world chaotic time series, the paper proposes a novel prediction model based on wavelet transform and echo state network (WESN). The basic idea of WESN is that firstly wavelet decomposition is used to separate the chaotic dynamics component and noise components, then the gotten components can be predicted by echo state network (ESN) independently, and finally the prediction results of time series are obtained by assembling the prediction values of all components. By using real-world sunspot time series for verification, the prediction results show that the proposed model has higher prediction accuracy by comparing with the models of direct echo state network (DESN), and direct echo state network (DSVM).