事件影响下的时间序列多尺度集成预测

Multiscale integrated prediction for time series under event influences

针对时间序列的非线性、非平稳和多尺度特征,考虑到事件对序列结构产生的影响,提出事件影响下的时间序列多尺度集成预测策略。首先,基于经验模态分解将原始序列分解成若干分量,从多个尺度展现序列的基本构成;随后,基于迭代累积平方和实现分量序列的变点检验,从多个尺度判别和获取事件对序列产生的结构性影响;然后,基于干预分析构建事件对不同分量序列的影响模型,据此剔除事件影响,获取净化序列;最后,运用基于粒子群优化的支持向量回归,建立单一尺度的序列预测模型,进而实现事件影响下的时间序列多尺度集成预测。实证分析表明:该策略能够精细辨识事件对序列的多尺度影响,有效建立序列总体及分量的预测模型,与传统方法相比,具有更强的事件辨识能力、自适应建模能力和更高的预测精度。

Considering the nonlinear,nonstationary,multiscale characteristics of a complex time series,as well as the structural influence of events,an integrated prediction strategy based on the multiscale series decomposition is proposed.Firstly,an empirical mode decomposition is used to decompose the original series at multiple scales,which is reduced to its original components.Secondly,an iterated cumulative sum of squares method is used to test the structural change points of the original series and its each component,and the structural influence of events on the series at multiple scales can be identified and obtained.Thirdly,based on the identification of structural change points,an intervention analysis method is used to study the time range and intensity change rule of the event influence at different scales,then the influence of events is eliminated and the purified time series is obtained.Finally,a support vector regression optimized by particle swarm optimization is used to predict the single scale series,the prediction model at each scale is established separately and the integrated prediction on the complex time series under the influence of events is realized.The empirical analysis on the cost series of warship maintenance spare parts shows that this method can effectively obtain the influence of events on the series from a multi-scale perspective and build the prediction model of different scale series adaptively.Compared with the existing methods,it has a stronger ability of event analysis and identification,adaptive modeling and a higher prediction precision.