Traffic Prediction Based On Dynamic Spatial-temporal Graph Convolutional Neural Network

Traffic prediction is gradually becoming a core component of intelligent city building.Accurate and reasonable traffic prediction not only helps traffic management agencies to make important decisions and adjustments,but also improves the efficiency of pedestrian travel.With the increasing demand for improved intelligent transport systems,the need for accurate and efficient prediction of future traffic conditions has become a pressing issue.However,this task is challenging due to the fact that traffic data is continuously influenced by dynamic stochastic factors as well as spatially hidden behaviour.Traditional traffic prediction models and current research have progressed to some extent,but there is still a general focus on modelling relationships between node pairs and between node history information,neglecting the analysis of the nature of the nodes themselves and the mining of potential features,leading to performance bottlenecks.Therefore,this paper chooses to take traffic prediction as the research background and proposes a series of effective prediction models based on graph convolutional neural networks for different characteristics of traffic data.(1)To have the ability to capture dynamic traffic information,a dynamic spatial-temporal graph convolutional network prediction model is proposed.A dynamic graph generation module was designed to capture geographical proximity and spatial heterogeneity information between nodes,adaptively fusing the two types of information and generating a new dynamic graph as the time step increases.In the temporal dimension,a graph convolution recurrent module was constructed to capture local temporal dependence on the basis of merging spatial relationships.Experimental results on both datasets demonstrate the effectiveness of the model,with predictive performance outperforming the vast majority of baseline models.(2)A dynamic graph generation module was designed to capture geographical proximity information and spatial heterogeneity information between nodes,adaptively fuse the two types of information with the time step of superposition and generate a new dynamic graph.In the temporal dimension,a graph convolution recurrent module is constructed to capture local temporal dependencies based on merging spatial relationships.Experimental results on both datasets demonstrate the effectiveness of the model,with predictive performance outperforming the vast majority of baseline models.(3)To effectively learn from traffic data with multidimensional temporal correlation and complex spatial correlation,a multi-graph network prediction model based on graph convolution is proposed.A novel spatial-temporal graph convolution module is designed to fuse information about the traffic network from multiple perspectives such as static distance of nodes,dynamic attributes and historical connections.The model as a whole uses a new decoder architecture to resolve future traffic conditions along both temporal and spatial dimensions.Experimental validation on two real traffic datasets demonstrates the feasibility and effectiveness of the model.In summary,this paper focuses on traffic prediction as the learning context and proposes three different prediction models based on graph convolutional networks,which also capture the spatial-temporal dependence of traffic data while deeply mining the potential features of the data in order to improve traffic prediction performance.