Research And Application Of Traffic Flow Prediction Method Based On Graph Attention Network

With the development of urbanization in China,the number of cars is increasing rapidly.As a result,urban traffic becomes more congested and the environment deteriorates,which causes huge economic losses to society.Traffic flow prediction has become one of the important methods to solve urban traffic problems.Due to the complex pattern of traffic flow changing rapidly in time and space,it is difficult for existing prediction methods to accurately capture the spatio-temporal coupling characteristics of traffic data.Therefore,based on the research and analysis of urban road traffic data,this thesis proposes an optimized traffic flow prediction model and designs a traffic flow prediction visualization management system.The main contributions of this thesis are as follows:(1)This thesis proposes a traffic flow prediction model GA2T(Graph Attention Network and Transformer)based on Graph Attention Network(GAT)and Transformer.In order to solve the problem that Graph Convolution Network(GCN)can only statically fuse the features of road network nodes,GA2 T uses GAT to dynamically aggregate spatial features to model the spatial dependence of traffic flow.Aiming at the poor long-term prediction effect of traffic flow,the highly parallelized self-attention mechanism in Transformer is used to adaptively capture long-term dependencies from long sequences.For the problem that the attention mechanism is not sensitive to the location information of traffic data,a spatio-temporal code is embedded in the feature vector of the data,so that the model can recognize its spatio-temporal order.Due to the high complexity of the existing model structure,this thesis proposes a nonautoregressive fusion encoder-decoder framework,which embeds the decoder in the original Transformer into the encoder,and reduces the number of parameters to half of the original.GA2 T is evaluated on two traffic datasets,METR-LA and PEMS-BAY.The experimental results show that GA2 T achieves better prediction performance than the comparison model.Meanwhile,the necessity of each module in the model is evaluated by ablation experiments.(2)Based on the above research work,this thesis designs and builds a traffic flow prediction visualization management system.The system consists of two parts: a traffic flow prediction visualization platform and a traffic flow data management system.The traffic flow prediction visualization platform mainly displays the traffic flow data in real time in the form of a large screen.It includes various forms of chart presentation such as maps,bar charts,carousel charts,ring charts,etc.As the background support of the visualization platform,the traffic flow data management system has many strong functions.It can not only independently select prediction models and parameters,create new prediction tasks and visualize the results,but also maintain and manage information such as system users,road sensors,historical and forecast data of traffic flow and historical forecast records.The research in this thesis runs through an integrated process from the proposed model to practical application.The proposed model GA2 T improves the accuracy of traffic flow prediction and makes more timely and accurate prediction.Based on the proposed model,a traffic flow prediction visualization management system is built to assist traffic managers to formulate prevention plans and guide travelers to plan more reasonable paths.Therefore,it can alleviate the imbalance of time and space resources of the road network and improve the road traffic capacity.