Dynamic Spatial-Temporal Graph Attention Graph Convolutional Networks for Short-Term Traffic Flow Forecasting

The application of graph convolutional networks in short-term traffic speed forecasting of road networks has effectively improved the prediction accuracy. The key point of this method is to construct the Laplacian matrix through extracting spatial features among nodes of the road network. However, most available methods mainly rely on the spatial distance among nodes to construct Laplacian matrix, then optimize the Laplacian matrix by other methods, which limits the wide application of the model. In this paper, we propose a dynamic spatial-temporal graph attention graph convolutional networks (GAGCN) method to improve the applicability of the model. The Laplacian matrix in this model is constructed directly by the dependencies among the nodes hidden in the traffic data which are identified by the graph attention network, and can be dynamically adjusted over time. The information of spatial distance among nodes and human intervention are not required in the process. Experimental results of two real-world datasets from the Caltrans Performance Measurement System (PeMS) show that both the applicability and prediction accuracy of the proposed model have been significantly improved.