关键词: 非平稳时间序列, 小波变换, 自回归移动平均模型, BP神经网络

Abstract: According to the theory of wavelet analysis, a non-stationary time series forecasting method which is based on wavelet is put forward. Through the wavelet decomposition and single reconstruction, the original non-stationary time series is decomposed into a layer of approximation coefficients and several layers of detail coefficients. In the next step, each layer of coefficients is used to model and forecast, using the Auto-Regressive and Moving Average（ARMA） model once, and the BP neural network model once. After integrating layers of coefficients, the predictive value of the original time series is obtained. The result of the experiment, in which the network traffic data of internet nodes and daily maximum temperature data is used to model and forecast, demonstrates good accuracy of the method mentioned above. And it also shows that the prediction accuracy and curve fitting of the model using the BP neural network are better, which means that this model can be applied to the analysis and forecasting of non-stationary time series.

Key words: non-stationary time series, wavelet transform, wavelet analysis, Auto-Regressive and Moving Average（ARMA） model, BP neural network