Research On Time Series Prediction Method Based On Deep Learning

Multivariate time series(MTS)is a set of observed values of multiple univariate according to the chronological order of time,which is generally composed of more than one Univariate Time Series(UTS).With the wide generation and practical application of MTS data,the prediction of MTS has gradually become one of the more and more concerned research topics.With the development of Deep Learning(DL)technology,the performance of Time Series(TS)prediction model based on DL is getting better and more widely used.The existing models of MTS prediction methods generally use attention mechanism and Recurrent Neural Network(RNN)to extract MTS complex spatiotemporal features,but these methods have insufficient ability to capture spatial dependencies between MTS variables.Graph Convolutional Network(GCN)has strong ability to extract spatial features of complex data.Therefore,the following main research work is carried out around the MTS prediction method based on Graph Neural Network(GNN)in this paper.(1)A MTS prediction model Dual Self-attention Graph Convolutional Network(DSAGCN)of fusion GCN,attention mechanism and Convolutional Neural Network(CNN)is proposed.DSAGCN uses GCN to extract spatial dependencies between variables,and extracts the temporal dependence to take the MTS variables values using CNN and attention mechanisms,and then integrates the captured spatial and temporal dependencies,with MTS prediction from the fused spatio-temporal features.Experimental situations on the ExchangeRate MTS dataset show that the proposed DSAGCN out performs selected several state-of-theart baseline models.(2)A network framework model Gated Recurrent Unit Graph Convolutional Network(GRUGCN)that fuses GCN with Gated Recurrent Unit(GRU)is designed.The model extracts the temporal features of the GRU network and Multilayer Perceptron,then merges with the spatial-dependent features extracted by GCN,and the resulting spatial-temporal fusion features perform MTS prediction.Experimental results show that GRUGCN out performs performance on the Exchange-Rate MTS dataset over the selected several current state-of-theart baseline models.Under the condition of ensuring MTS prediction accuracy,the model GRUGCN has fewer parameters and is easier to train than DSAGCN.