Prediction of future behavior of complex set of data sets is a difficult task. Data mining is a potential technical way. For the real-time evolutionary data sets containing multiple time series and non time sequence, a method of integrating the sequence segmentation, clustering, and pattern matching is proposed, which combines the theme discovery and joint decision. In the whole method construction, the topic discovery prediction and joint decision prediction are fused into the early sequence segmentation and clustering. The sequences are stratified and segmented for forming standard pattern sets of each layer, using multi time granularity and multi span. Then, according to the standard pattern set, with the prediction strategy, the compound pattern with high stability extension behavior is used as the theme pattern. This can predict with online pattern matching. Finally, a distributed parallel computing architecture is used to implement the whole processing algorithm. Theoretical deduction and experimental data analysis show that the accuracy of the method is improved compared with the traditional time series prediction method.