Research On Traffic Flow Forecasting Method Of Traffic Stations Based On Spatio-temporal Multi-graph Convolutional Network

Traffic flow forecasting is an important part of the intelligent transportation system.Accurate traffic flow forecasting can provide government and related departments information and decision support for the reasonable allocation of transportation resources,reasonable planning of traffic control,and improvement of the travel experience of passengers.However,due to the comprehensive influence of various factors such as weather,time period,holidays,geographical location,surrounding events,etc.,the traffic flow is highly non-linear,complex and random,which make it hard to predict accurately.In this paper,the traffic flow prediction problem of the traffic stations in real life is analyzed and studied,and two types of spatio-temporal multi-graph convolutional network models considering various relationships among the stations are proposed to predict the flow of stations.First of all,aiming at the different relationships among traffic stations under multiple perspectives,a spatio-temporal multi-graph convolutional network model based on the mutual attention mechanism(STMGCN-MA)is proposed for the traffic flow prediction of traffic stations.In the model,the neighbor graph and OD(Origin-Destination)graph among the stations are constructed from the two perspectives of the geographical location relationship and the OD relationship among the stations,which are used to model the traffic similarity and traffic dependency among the stations,respectively.On this basis,the spatial feature representation of the station under two perspectives is learned by the graph convolution technology on the two graph structures.Further,for the high-order correlation between the two feature representations,an information enhancement unit based on the mutual attention mechanism is proposed to enhance the feature representations under two perspectives.Afterwards,the time-dimensional convolution unit is used to learn the spatio-temporal features of the station under two perspectives,and finally the traffic prediction is performed through the fully connected layer and the fusion layer.Subsequently,considering that the traffic relationship between the stations could change dynamically with the changes of weather,time,traffic conditions and other factors,a graph structure with dynamic edge features is constructed and a dynamic graph convolution unit is proposed,after that a spatio-temporal multi-graph network learning model based on dynamic graph convolution(STMGN-DGC)is proposed based on the STMGCN-MA in this paper.Specifically,the unit learns a dynamic mask matrix for graph convolution based on dynamic edge features,which changes dynamically as the edge features change at different times,and adjusts the weights of graph convolution feature aggregation,so as to use the dynamic correlation between stations in different time intervals to learn a better feature representation for each station.The experimental results on two real traffic stations datasets reveal that the STMGCN-MA and STMGN-DGC model can effectively predict the flow of traffic stations and the prediction accuracy is better than other methods of the same type.Especially when the prediction step is long,the advantage in prediction accuracy of the models is more obvious.The STMGN-DGC model has higher prediction accuracy than the STMGCN-MA model due to the dynamic nature of the relationship between the stations.At the same time,the STMGN-DGC model requires more training time and hardware resources than the STMGCN-MA model due to its higher complexity and more parameters.