Research On Time Series Forecasting Based On Multivariate Analysis

The fluctuation of time series is affected by many factors.There are many irrelevant factors and redundant factors,among those interfering factors.These unnecessary factors increase the complexity of building the model,also affect the accuracy of the prediction.Besides,there are other data sharing the same data type,which will affect the target sequence for predicting.Among those interfering data,the similar one contains same fluctuation information,indicating that there are a high degree of mutual influences between them.Hence,the control amount of these similarity data has important reference value for accurate prediction.Therefore,it is necessary to evaluate various influencing factors and screen the important factors,while predicting the target sequence.At the same time,it is necessary to use the control amount of similarity data to improve the accuracy of the prediction model.From what mentioned all above,the factor data and the control amount of similarity data are studied,a deep learning prediction model is constructed for the time series.The main research contents are as follows:The predicted target sequence is affected by many factors.Aiming at the problem that too many irrelevant factors and redundant factors are reducing the prediction accuracy,the influencing factors of the target sequence are analyzed in this article,and these unimportant factor sequences are filtered out.A new deep learning prediction model is proposed,RF-CNN-GRUA,based on random forest importance analysis.Firstly,the random forest algorithm is used to evaluate the importance to select the sequence of factors with higher importance.Then,the CNN and GRU is used to extract the spatial features of data and the characteristics of data timing changes respectively.Moreover,the attention mechanism is introduced to construct a GRUA model,GRU with Attention Mechanism.Finally,the two prediction results are trained and predicted again through the fully connected layer.The RF-CNN-GRUA model fully considers the impact of important factors,and takes advantage of the spatial feature advantage of the CNN model and the time series change feature advantage of the GRUA model.The simulation experiment results show that this model has higher prediction accuracy than other single models.The predicted target sequence is affected by other data of the same type.The fluctuation of similarity data implies important relevant information.A new deep learning prediction model based on similar sequence control amount is proposed,DTWGPC-Seq2 Seq.In this model,the DTW algorithm is used to find sequences similar to the target sequence in the same type of data.For the similar sequences,the generalized predictive control model,GPC,is used to obtain the control value of similar sequences,which will represent the fluctuation information of the time series.The Seq2 Seq prediction model composed of GRU network is used for prediction.The experimental results show that the prediction accuracy obtained by using the analyzed data is higher.In summary,the predictive model that combines important factor analysis and similarity data analysis is proposed in this article.In this model,the Random Forest algorithm is used for importance analysis,the DTW algorithm is used for similarity analysis.Besides,the generalized predictive model,GPC,is used to obtain the control amount of similarity sequence.And,two kinds of data are formed into a new data set,and the Seq2 Seq deep learning model is used for prediction.The simulation experiment results show that these two kinds of data have higher prediction accuracy,under the method that given them the environmental factors of the predicted data and the influence of the same type of similar data.