Research On Urban Traffic Flow Predicition Based On Dynamic Spatial-temporal Graph Convolution Networks

Traffic congestion has become the main source of negative impacts on urban economy and environment,and predicting future traffic flow is one of the most effective methods to alleviate traffic congestion.However,accurate urban traffic flow prediction is still challenging.On the one hand,due to the complex topological structure of urban traffic network and the dynamic impact of different events in the urban,there are multi levels of spatial dependencies in the traffic network,and most existing methods often neglect to extract the spatial dependencies of different levels.On the other hand,because there are different commuting modes in the urban traffic network,these different traffic modes make the flow between stations in the road network change with a variety of different rules,and the various of traffic modes makes the flow change more dynamic.Finally,the traffic flow at a certain time slice may be affected by traffic flow in different historical periods.Existing methods often predict future traffic from the perspective of a single historical period without considering the impact of different historical periods on traffic,resulting in poor prediction results.This paper,in response to the three issues discussed,undertakes an investigation into urban traffic flow forecasting and offers the following contributions:(1)Aiming at the multi-level spatial dependence of traffic flow,this paper proposes a Multi-step Coupled Graph Convolution Neural network traffic prediction model(MCGCN)based on temporal attention mechanism.The model mainly includes two modules: Multi-step coupled Graph Convolution Network module,Gated Recurrent coupled graph convolution module and Multi-step temporal attention module.The multi-step coupling graph convolution module uses the multi-step coupled updating mechanism to dynamically update the adjacency matrix of each convolution layer,so as to capture the spatial dependencies of different levels.The gated recurrent coupled graph convolution module combines the coupled graph convolution with the gated recurrent unit to extract the temporal and spatial characteristics of traffic flow simultaneously.Finally,the multi-step temporal attention module is used to filter useful information from the historical time steps,so as to capture the global temporal dependencies of traffic flow.(2)Aiming at the influence of various traffic modes on traffic flow,this paper proposes a Multi-mode Dynamic Residual Graph Convolution Network for traffic flow prediction(MDRGCN),which mainly includes three modules: multi-mode dynamic residual graph convolution module,gated recurrent dynamic graph convolution module and dynamic residual module.The multi-mode dynamic graph convolution module captures the spatial dependence of different traffic modes in the traffic network by learning two different relationship matrices,and dynamically fuses the spatial characteristics of different modes.The gated recurrent dynamic graph convolution module combines the dynamic graph convolution with the gated recurrent unit to achieve the combination of temporal and spatial characteristics of traffic flow.The dynamic residual fusion module can dynamically combine the historical traffic data with the prediction output of the encoder-decoder module,and update the final prediction result combining with the historical traffic flow.(3)In view of the impact of multiple historical periods traffic flow on traffic flow prediction,this paper proposes a Multi-View Dynamic Graph Convolution neural Network for traffic flow prediction(MVDGCN).MVDGCN mainly includes two modules: multi-view encoder-decoder module and dynamic fusion module.Firstly,three traffic flow data with different historical periods are constructed from the historical traffic data,and three encoder-decoder structures are constructed to extract the spatial-temporal dependencies of traffic flow from different perspectives.The dynamic fusion module can dynamically fuse the traffic characteristics extracted from the encoder-decoder in different historical periods,so as to obtain more spatial-temporal dependencies.This paper conducts experiments on two real traffic flow data sets in New York City,NYCTaxi and NYCBike,the United States.The experimental results show that MCGCN model,MDRGCN model and MVDGCN model can extract more spatio-temporal dependencies of traffic flow,and all evaluation criteria are better than other baseline models.