Research And Application Of Time Series Prediction Method Based On Autoregressive Model

In the era of big data,the amount of data is increasing rapidly,and the time series is also changing from unitary time series to multivariate time series.Both unary time series model and multivariate time series models based on Autoregression(AR)have some limitations.Therefore,this thesis studies and improves the classical time series model for its problems in prediction.The ARFIMA model is a further extension of the existing unary AR time series model,which mainly explores the long memory of the series.Long memory parameters are calculated by R/S analysis method,but R/S analysis method has strict requirements on the number of time series,which must meet the amount of data with a large number of common factors.Otherwise it cannot describe the long memory process of sequence well.In addition,the model parameter estimation method often contains assumptions,resulting in poor initial parameter values.In this thesis,R/S analysis method based on sliding window method and improved fractional order PSO algorithm are proposed.The improved R/S analysis method can be applied to any number of data sequences and can effectively avoid the inaccurate calculation of long memory parameters.Fractional order PSO algorithm has fast convergence speed and better searching ability.Experimental results show that the ARFIMA model established by this method has higher prediction accuracy.Vector autoregression(VAR)model is used for multivariate time series modeling,but it model only considers the correlation sequence,but does not consider the influence of residual sequence on itself.In this thesis,an improved IVAR model of VAR is proposed to solve this problem,and the lag value of residual sequence is integrated into the expression.At the same time,the model can only capture the linear relationship,and the rest of the nonlinear relationship data is lost.Therefore,this thesis proposes a hybrid prediction model based on linear and nonlinear(IVAR-CNN-AGRU).Firstly,IVAR model is used to linearly fit the multivariate time series,and then the residual series is obtained.Then,CNN was used to extract spatial features of data,and GRU was used to extract temporal variation features of data,and an attention mechanism was introduced into GRU to construct the AGRU model.Then,the residual sequence prediction results are output through the full connection layer.The end result is the accumulation of linear and residual sequences.The IVAR-CNN-AGRU model fully integrates the convenience of IVAR model in dealing with linear relations,and makes use of the spatial characteristic advantages of CNN model and the temporal variation characteristic advantages of AGRU model,thus making up for the shortcomings of IVAR model in nonlinear.Experimental results show that this model has higher prediction accuracy than other models.