Research On Highway Traffic Forecasting Based On Spatial-Temporal Graph Convolutional Networks

In recent years,the surge in the number of cars in China leads to increasingly prominent contradictions between booming demand on road resource and the limited road carrying capacity.Traffic flow is one of the main parameters to reflect the running state of a road.If the traffic flow situations on roads can be timely and accurately grasped and predicted,the vehicles can be effectively channeled in advance,and the running capacity and efficiency of the road network can be improved,which is of great significance to traffic managers,operators and participants.Traffic flow data are highly nonlinear and complex,so it is a great challenge to accurately predict them.Most of the existing prediction methods cannot directly use the traffic network structure to capture the dynamic spatial-temporal correlations of data.Aiming at solving this problem,this thesis applies graph convolutional network and spatial-temporal attention mechanism to predict traffic flow.The main research contents and results are as follows:First of all,there are many inherent patterns of traffic data in the time dimension.In this thesis,the input data is divided into multiple components,which are used to separately capture the recent,daily and weekly periodic patterns of data.This multi-component partitioning method can fully extract the time dimension information of data and reduce the overhead caused by processing irrelevant historical information.At the same time,considering that the traffic network is a kind of graph structure data,graph convolutional network is introduced to mine the spatial features of traffic data,and a spatial-temporal convolution module is designed.A multi-component spatial-temporal graph convolutional network(MCSTGCN)model is proposed,which can effectively extract the features of data in both time and space dimensions.Then,a more ingenious spatial-temporal module is designed,which consists of spatial-temporal attention mechanism and spatial-temporal graph convolutional network.The spatial-temporal attention mechanism is responsible for adjusting the degree of interaction between spatial-temporal nodes in different scenes.The spatial-temporal graph convolutional network extracts the spatial-temporal features of data based on the adjusted results of the spatial-temporal attention mechanism.Each spatial-temporal module is trained by residual learning.Thus,a spatial-temporal graph convolutional network model based on spatial-temporal attention mechanism(ASTGCN)is proposed,which can mine the dynamic spatial-temporal patterns of data more effectively.Extensive experiments are carried out on the open dataset PeMS from California Freeway.The experimental results show that the prediction effect of the MCSTGCN model is obviously better than the existing traffic flow prediction methods.The ASTGCN model takes more training time than MCSTGCN,but its prediction accuracy is further improved.This demonstrates that the models proposed in this thesis can effectively mine the dynamic spatial-temporal patterns of traffic flow data and make accurate prediction.