Research On Traffic Speed Prediction Based On Spatial-Temporal Graph Attention Networks

With the acceleration of urbanization,the number of urban motor vehicles shows an increasing trend year by year,which brings a heavy burden to the traffic system.To solve this problem,researchers have deeply integrated new technologies such as big data,Internet,artificial intelligence,sensors and other new technologies with the transportation industry,and proposed intelligent transportation systems.Traffic speed prediction,as a key component of intelligent transportation systems,is of great significance for providing valuable traffic management,planning and control information.Accurate traffic speed prediction can help travelers avoid travel during congested periods,and can also help management departments carry out accurate traffic light control.Existing traffic speed prediction research methods try to use graph neural networks to model the spatial correlation of traffic,but most of these methods define the road network structure in a static distance graph way,only considering the static topological features of traffic,which ignores the complex correlation and dynamic attributes of traffic network structure,resulting in unsatisfactory prediction results.In addition,the data missing problem caused by sensor failure further reduces the prediction accuracy of existing methods.Aiming at the above problems,this thesis explores the traffic speed prediction methods in two situations:complete traffic data and missing data,proposes corresponding spatialtemporal graph attention models,and builds a traffic speed analysis and prediction system.The main contributions of this thesis are as follows:(1)In the case of complete traffic data,this thesis proposes a traffic speed prediction model based on spatial-temporal dynamic graph networks,which can effectively capture the deep static and dynamic features of traffic network.This method uses the Jensen-Shannon divergence value between node speed distributions to characterize the non-geographic correlation between nodes,and proposes a new dynamic graph attention mechanism,which can dynamically adjust the weights between nodes while retaining the information of past time slices,thus reflecting the dynamic correlation between nodes.Finally,it is verified on two real-world datasets that the model has better prediction accuracy than comparative methods.(2)For the situation where there are a large number of missing values in traffic speed data,this thesis proposes a traffic speed prediction model based on adversarial spatial-temporal graph attention networks.This model integrates data imputation and speed prediction seamlessly into a unified framework through multi-task learning to reduce the error accumulation caused by data imputation.This model uses spatial-temporal attention network to capture spatial-temporal correlation,and designs an adversarial training framework to enhance the robustness of traffic embedding and reduce the impact of missing data.The experimental results show that the model has a significant advantage in prediction accuracy.(3)This thesis implements a traffic speed analysis and prediction system,which provides users with a visual execution interface,which is convenient for users to perform traffic data analysis,model training,prediction analysis and other functions,reducing the difficulty of urban managers to perform traffic speed prediction,and providing a practical tool for urban traffic management.This thesis aims at the traffic speed prediction problem,and proposes two traffic speed prediction models based on spatial-temporal graph attention network,which significantly improves the prediction accuracy of existing prediction models.Combined with the traffic speed analysis and prediction system implemented in this thesis,these two models can be effectively used for urban traffic congestion problems,promote the construction of urban intelligent transportation systems,and have important research significance and practical value.