基于小波分解的AR-SVR一类非平稳时间序列预测

AR-SVR Forecasting of no-stationary time series with tendency based wavelet decomposition

本文提出了一种基于小波分解的均值具有趋向性的非平稳时间序列预测方法。方法首先利用具有平移不变性的小波分解,分离非平稳时间序列中的长期趋势和高频信息(短期行为),然后对高频信息构建自回归AR模型,而对于趋势则利用支撑向量回归(SVR)进行拟合,最后将各模型的预测结果进行叠加,从而得到原始序列的预测值。这样保证了充分拟合长期趋势的同时,避免了短期行为造成的过拟合。最后的实验结果表明本文提出的这类非平稳时间序列预测方法是有效的。

A non-stationary time series with tendency forecasting method based on translation invariant wavelet decomposition is proposed. Non-stationary time series with tendency are decomposed into tendency (smooth signal) and stochastic components (high frequency signal). The high frequency stochastic components are forecasted with auto-regression models, and the tendency with support vector regression, the addition of the forecasting result of tendency and that of stochastic components constitutes the forecasting result of the original time series. The forecasting method avoids the high frequency signal to be over-fitted, and better fits the tendency of the non-stationary time series so better forecasting performance can be obtained. Experiments show that the forecasting method proposed is effective.