Abstract:

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.

Key words: time series, event influence, empirical mode decomposition, multiscale analysis, structural change point, integrated prediction