Time Series Data Stream Forecasting Based On Online-VAR

The last few years, data size increases at a very high speed every day, in sensor field and electronic information field. The scientist can extract a lot of useful information from a large number of data and use the information to determine lots of questions. How to find valuable information from the outpouring, not stable data stream, is a question that scientists need to develop and find.This article explore time series data stream forecasting methods. Online-VAR method produces based on sliding window technology and EMD method which study non-linear, non-steady data. Put forward a common prediction module which is fit for time series data streams prediction in all application fields. How to turn the sequential, vast data into only once processing of data is key problems in time series data stream prediction. Concerning about the continuity of data stream, this paper uses sliding window technique for modeling original data stream, made applications only concern with the data changes in sliding window in a recent time, sliding widow was realized by for loop in this paper; use VAR method to analyze the data in sliding window and forecast. In this paper, EMD method had been used with the idea of parallel processing, when there is too much data in sliding window, then the data needs to be divided into segmentation and disposed. First use EMD method decomposes every piece of data, put IMF component of every piece of data together to form joint IMF and put remaining signal together to form joint remaining signal. Then use EMD to decompose joint IMF, put remaining signal from joint IMF with joint remaining signal together, go on this course until the final joint remaining signal can’t be decomposed using EMD. This method is used to enhance rate of algorithm. Online-VAR method is a new method with high accuracy, high speed and suits online, self-adaption forecasting.Finally, Online-VAR method was used to analyze real rolling bearing time series data stream, and prove this method has a good effect in data series data stream forecasting by emulational experiments.