A Segment-Wise Method for Pseudo Periodic Time Series Prediction

In many applications, the data in time series appears highly periodic, but never exactly repeats itself. Such series are called pseudo periodic time series. The prediction of pseudo periodic time series is an important and non-trivial problem. Since the period interval is not fixed and unpredictable, errors will accumulate when traditional periodic methods are employed. Meanwhile, many time series contain a vast number of abnormal variations. These variations can neither be simply filtered out nor predicted by its neighboring points. Given that no specific method is available for pseudo periodic time series as of yet, the paper proposes a segment-wise method for the prediction of pseudo periodic time series with abnormal variations. Time series are segmented by the variation patterns of each period in the method. Only the segment corresponding to the target time series is chosen for prediction, which leads to the reduction of input variables. At the same time, the choice of the value highly correlated to the points-to-be-predicted enhances the prediction precision. Experimental results produced using data sets of China Mobile and bio-medical signals both prove the effectiveness of the segment-wise method in improving the prediction accuracy of the pseudo periodic time series.