Research Of Time Series Prediction Model Based On BP Neural Networks

Time series analysis and prediction is a kind of time series mining technique,which is of great significance to scientific research,industrial production and business management.Scholars in the field of time series prediction analyze and forecast time series by constructing models.Nowadays,theres is two categories of time series forecasting models: traditional prediction model and prediction model based on machine learning.Traditional prediction models are restricted by their own mathematic structure and regression mechanism,result in imprecise prediction on complex time series;time series prediction models based on machine learning technology need training repeatedly sufficient time series,and neglect possibly small amount of key features in time series,So that it cannot capture placidly movement trend of time series.It is obvious that training process of models in a large amount of time series reduces the efficiency of prediction model building.To solve the above preblem,this thesis regards time series prediction models as the research object,consummated dimension reduction technique in time series,combined traditional time series prediction model and machine learning technology,and proposed novel time series prediction hybrid model.The work content is as follow:(1)Similarity measure algorithm of time series based on Binary-dividing SAX.In order to solve the problem of selecting segments of the parameters in SAX representation,this algorithm combined with sliding window method,introduced variance into segments dividing process and changed the way and order of sliding window.It can adaptively divide into segments according to the numerical distribution of time series.Accordingly,uncertainty of segments is solved and accuracy of SAX representation is improved.(2)Log periodic power law based on BP neural networks hybrid model.This model constructed time series trend model based on Log periodic power law,introduced BP neural networks that capture patterns of time series exiguous fluctuation,conducted training process in BP neural networks for residual series of the trend model,and combined the trend model and residual neural networks.It achieved more accurate time series prediction.(3)Log periodic power law based on SAX representation hybrid model.This model introduced SAX representation into Log periodic power law based on BP neural networks hybrid model,simplify training data of BP neural networks in hybrid model,and significantly improve the efficiency of constructing the hybrid model and hold accuracy of prediction.