A sliding window-based multi-stage clustering and probabilistic forecasting approach for large multivariate time series data

Abstract

ABSTRACTTime series data analysis, such as temporal pattern recognition and trend forecasting, plays an increasingly significant part in temporal data statistics and analytics. Yet challenges still exist in the efficiency of pattern extracting and trend prediction for large multivariate time series. The paper proposes a multi-stage clustering approach towards multivariate time series by using dynamic sliding time windows. The segmented multivariate time series are clustered separately in each time window to product first-stage clustering centres, and which are used to generate second-stage clustering results involving all time windows. The method can simplify large scale multivariate time series mining problems through multi-stage clustering on multiple sliding time windows thus achieve improved efficiency. Based on the clustering outcomes, a correlation rules mining method is given to discover frequent patterns in the time series and generate self-correlation rules. Then, the paper presents a probabilistic forecasting model that leverages the extracted rules to make short-term predictions. Finally, experiments are presented to show the efficiency and effectiveness of the proposed approach.