Traffic Flow Forecasting Method Research Based On Spatial-Temporal Graph Neural Networks

Intelligent traffic control and guidance system is one of the core components of intelligent traffic system.Excellent traffic control system can ensure efficient and safe operation of traffic.Traffic flow prediction plays an important role in various traffic control systems.The accuracy of prediction determines the effectiveness of traffic control and induction strategies.Nowadays,the traffic data that can be collected are increasingly abundant,and we have entered the era of traffic big data.How to mine and utilize abundant traffic big data for more accurate and timely traffic flow prediction,help managers to make better traffic control plans,and provide support for traffic congestion management is an important topic.In order to mine the complex spatio-temporal correlation of traffic flow data on the traffic network and accurately and timely predict traffic flow parameters,this paper studies the traffic network as a graph to model the spatio-temporal relationship,and constructs three spatio-temporal convolution network models to extract the complex spatio-temporal correlation of traffic flow data.The main contents of this paper are as follows:(1)Analyzing the research progress of graph neural network.we focus on the most important breakthrough in graph neural network research,graph convolutional neural network,and its theoretical knowledge background and representative methods.Then we introduces the related theory and research of graph attention mechanism GAT,and finally introduce specific examples of spatio-temporal graph network application.All of these provide guidance for the establishment of traffic flow prediction model based on spatio-temporal graph neural network.(2)By studying the limitations of applying convolutional neural network(CNN)in the traffic network to capture spatial relations,following the latest development of graph neural network,using the graph convolutional neural networks to model the traffic road network relationship,learning the traffic as a graph and using the bidirectional LSTM network to model the time relationship,a traffic flow prediction method GCBLSTM combining graph convolution network GCN and LSTM is proposed,which can realize the prediction of graph-structured spatio-temporal data.We test the performance of the proposed model on two real-world traffic speed data sets.(3)By studying the structure and principle of time convolutional neural network and its advantages of processing sequence data,a traffic flow prediction method STGCN combining graph convolutional network GCN and time convolutional network is proposed.STGCN uses the dilated causal convolution layer to extract the temporal correlation of traffic flow data,and the graph convolution layer to extract the spatial correlation of nodes,stacking the spatio-temporal convolution modules,which can capture the time-dependence in a long range and deal with the graph-structured spatio-temporal prediction problem.We use the traffic flow data collected by the California Department of Transportation Performance Measurement System PeMS for training and testing.(4)By analyzing the limitations of the graph convolutional network GCN in the application of traffic prediction,the graph attention network GAT is introduced.By analyzing the structure and principle of the graph attention network and the gated recurrent unit GRU network,a graph attention network GAGRU combining GAT and GRU networks for traffic flow prediction is proposed.GAGRU uses the GRU network to extract the time correlation of traffic flow data,and uses the graph attention mechanism to obtain an adaptive adjacency matrix at each time step to capture the spatial correlation of the road network,which can realize the prediction of the temporal and spatial sequence of the graph structure.We use the traffic flow data collected by the PeMS system for example verification.