Research On Traffic Flow Prediction Based On Graph Neural Networks

In recent years,the rapid development of transportation system has not only enriched people’s travel options,but also brought them closer to each other in terms of geographic distance.At the same time,traffic congestion has become increasingly serious,which not only increases people’s commuting costs,but also leads to public-safety incidents.In order to alleviate the influence of traffic congestion,accurate traffic prediction is of great importance.Traffic flow prediction aims to analyze the historical traffic flow patterns and thus deduce the future traffic evolution.In this thesis,we focus on the development of traffic flow prediction technology.We aim to analyze the advantages and shortcomings of existing methods and provide effective solutions.Existing traffic flow forecasting methods can be divided into region-level and road-level according to the research objects.Region-level traffic flow prediction methods divide cities into multiple regions and usually predict the traffic flow variation within each region.Existing methods try to model the spatial dependencies between regions with the help of graph neural networks.However,most of them define the correlation strength between regions in the form of static graphs,ignoring the spatial-temporal dynamics of traffic flow.To this end,this thesis proposes a spatial-temporal attention mechanism that captures the varying spatial dependencies between regions at different time.Also,we perform multi-granularity temporal modeling of traffic flow sequences to reduce redundant input information.Furthermore,to improve the accuracy of traffic flow prediction,a multi-task prediction model is proposed in this thesis.The model effectively enhances the prediction performance of the main task through a multi-task collaboration mechanism.In addition,the method also designs a dual attention mechanism to further enhance the spatial representation of the graph convolutional network.On the other hand,road-level traffic flow prediction studies road segments and usually focuses on the travel speed of each road segment within a certain time interval.Existing methods model different turning lanes in a crude way,ignoring the latent traffic patterns embedded in the turning types.In this thesis,we propose a spatial-temporal graph attention network that can effectively perceive the heterogeneous attributes of different turning lanes.We also introduce prior knowledge to make the model more effective in capturing the spatial dependencies between road segments.We conduct extensive experiments on real-world datasets and compare our method with existing mainstream traffic flow prediction methods.The experimental results show that our proposed model has an advantage in prediction accuracy and outperforms the compared methods in both region-level and road-level traffic flow prediction tasks.Our proposed method also shows more stable performance in multi-step prediction experiments.