Research On Traffic Flow Prediction Based On Spatio-temporal Graph Multi-feature Fusion Of Variational Autoencoder

With the development of the logistics industry,the logistics industry’s dependence on transportation is increasing.Smooth transportation can effectively improve transportation efficiency.Timely and accurate traffic flow forecasts provide logistics practitioners with great opportunities and convenient for downstream tasks such as transportation route planning and scheduling.At present,traffic flow forecasting is mainly focused on using the historical time series information of observations to learn sequence features.Such forecasting will cause the lack of information and lead to poor accuracy of the forecast results.Although some studies use the spatio-temporal information,but they are not very accurate.The features information extraction of the traffic road network is not perfect,which also has an impact on the traffic flow prediction results.Based on the above background and problems,this paper proposes a variational multi-autoencoder traffic flow prediction model based on multi-feature fusion of spatio-temporal graphs.The specific research content is as follows:Firstly,this paper adopts the variational multi-encoder-decoder framework to model the multi-features information such as traffic map and flow,and find the hidden variables that plays a key decisive role.Both multi-encoders and decoders are composed of multiple neural networks,which are used to construct the sum of probability distributions.The traffic flow time series of the input model is divided into two parts,one is to directly input the sequence to the maximum feature encoder in order to preserve the original features of the data,and the other input of the spatio-temporal graph feature encoder to extract the features of the data spatiotemporal graph.The two parts perform coding learning together and input to the decoder for output prediction.The proposed variational multi-self-encoding framework can effectively learn the distribution characteristics of time series.Then,in the construction of spatio-temporal map feature encoder,in order to solve the problem of insufficient spatiotemporal information feature extraction by traditional traffic flow prediction methods,a method based on spatio-temporal map convolution multi-feature fusion is proposed to extract spatio-temporal map dependent features of data.In order to extract spatial features through spatial map convolution and temporal residual convolution to extract temporal features,the leading matrix is used to construct the domain structure of the traffic road network map.At the same time,in order to enhance the prediction effect,the spatio-temporal map features including periodic features are fused.Finally,this paper verifies the feasibility of the model on the public data set through the analysis of the training process and results,and the analysis of hidden variables.Hereafter,three evaluation indicators are selected,and the proposed traffic flow prediction model is compared and analyzed with five baseline models of ARIMA,GAT,T-GCN,Graph Wave Net,and DCRNN.The experimental results show that the proposed traffic flow prediction model can be gain an advantage in evaluation indicators.