Global and Local Information Aware Traffic Speed Prediction
Facing the increasingly severe traffic congestion problem,the intelligent transportation system has been rapidly developed and widely used,and the traffic speed prediction,a cornerstone task,has attracted much attention.In re-cent years,deep learning has been widely used in the research of traffic speed prediction,and the research direction has also shifted from modeling time correlation to considering complex spatiotemporal correlation.The graph neural network fits the graph structure data of the traffic network and has become the mainstream method for modeling spatial correlation.To date,most research works have noted the importance of modeling dynamic spatial correlations in the task of traffic speed predic-tion.However,predefined or adaptive matrices for spatial feature learning are essentially static,and are not sufficient to match the complex and dynamic characteristics of spatial correlations.Moreover,through the analysis of multiple real traf-fic speed datasets,we find that the local fluctuations of inter-node dependencies are more dynamic than the global influence of the traffic network,which indicates that the spatial correlation can be derived from the global and local angles.Therefore,we propose an end-to-end global and local aware dynamic graph neural network model for traffic speed prediction.The traf-fic speed flow is first decomposed into static components and dynamic components by the self-decomposition layer,and then the dynamic graph generation module constructs a real-time dynamic graph for the dynamic components to match their dynamics.With the constructed dynamic graph and the input predefined graph,we model higher-order representations of these two classes of spatial correlations through graph convolution operations.Besides,we use causal convolution in the temporal module to capture temporal correlations in traffic data.Finally,residual connections are used to aggregate spatio-temporal correlations and feed to the output layer for final speed prediction.Experimental results on two highway datasets and one urban road dataset show that our proposed model outperforms state-of-the art models in terms of MAE and RMSE.
万方数据
维普
