基于小波回声状态网络的时间序列预测

Prediction of Time Series Based on Wavelet Echo State Network

摘要: 为了更好的对具有多尺度特性的时间序列进行预测，运用小波分析方法与回声状态网络模型相结合来创建小波回声状态网络预测模型。利用小波方法对原始时间序列进行处理，获得不同层上的细节部分序列和概貌部分序列，根据不同层上的序列特性分别创建与之相匹配的回声状态网络模型从而得到各层预测数据，将各层预测数据进行拟合，最终得到原始时间序列的预测结果。通过对某国国民生产总值的仿真研究表明，该模型能够很好的拟合时间序列的发展趋势，预测精度较高。

Abstract: In order to forecast time series with multi-scale characteristics better, the echo state network model is combined with wavelet analysis method to create a new prediction model which is wavelet echo state network. The original time series are processed with the Mallat algorithm and Dau-bechies wavelet based on wavelet multi-scale analysis theory that can respectively get the details of the different layers and overview part of sequences. Then different echo state network forecasting models are respectively created based on the different sequence characteristics to forecast and eventually add the forecast data of detail sequences and overview sequence to obtain the original time series prediction result. The simulation examples of forecast model for a country gross national product show that this model can well fit the development trend of time series and the forecast accuracy is higher.