Research On Highway Traffic Flow Forecasting Method Based On Spatial-Temporal Graph Convolutional Network

Traffic flow forecasting plays an important role in the intelligent transportation system.For the management and decision-making of the transportation department,the accurate prediction of traffic flow can provide scientific and reasonable information support,which is beneficial to alleviate traffic congestion and ensure the normal travel of residents.At present,there are two problems in the research: ⅰ)the correlation feature information between traffic flow and road occupancy rate,traffic flow and speed in the traffic flow data is not fully considered,and the multi-scale temporal features are ignored;ⅱ)the static and dynamic spatial information extraction is not fully considered,and adaptive periodic feature extraction is ignored.To solve the above problems,this paper proposes two traffic flow prediction methods based on spatialtemporal graph convolutional network,as follows:In order to solve the problem that the existing algorithms do not consider the correlation feature information between flow and road occupancy rate,flow and speed,and the multi-scale temporal features are ignored,this paper proposes a new model: Double Branch SpatialTemporal Graph Convolutional Network(DAM＿STGCN).Firstly,according to the periodicity of traffic data,the model divides the traffic data into two temporal granularity inputs: near-term input and periodic input.Secondly,in each input branch,the correlation feature information is extracted by AGLU between traffic flow and road occupancy rate,traffic flow and speed.Then,the spatial-temporal context information is extracted by graph convolutional layer and multiscale temporal convolutional layer,and the near-term and periodic double branch prediction results are outputted by prediction convolutional layer.Finally,the prediction results are fused by gating module.In order to solve the problems that existing algorithms do not consider the static and dynamic spatial information extraction,and the adaptive periodic feature extraction are ignored,based on DAM＿STGCN,this paper proposes a new model: Adaptive Spatial-Temporal Dynamic Graph Convolutional Network(PTrans＿DGCN).The method uses the dynamic graph convolutional layer to extract dynamic spatial information,and uses the gated fusion layer for static and dynamic spatial information fusion.Compared with the multi-scale temporal convolutional layer,an improved P-Transformer layer is proposed to pay attention to the influence degree of traffic flow at different moments,and adaptively extract periodic feature.For the two methods proposed above,experiments are carried out on two open traffic data sets Pe MSD4 and Pe MSD8.The results show that: The MAE,RMSE and MAPE of DAM＿STGCN model are better than other benchmark models on the whole.The prediction accuracy of PTrans＿DGCN model is improved after considering static and dynamic spatial information and adaptive periodic feature extraction,and the results of the above three indexes are better than DAM＿STGCN model on the whole.