Graph convolutional dynamic recurrent network with attention for traffic forecasting

Traffic forecasting is a typical spatio-temporal graph modeling problem, which has become one of the key technical issues in modern intelligent transportation systems. However, existing methods cannot capture the long-range spatial and temporal characteristics very well because of the complexity and heterogeneity of the traffic flows. In this paper, a new deep learning framework called Graph Convolutional Dynamic Recurrent Network with Attention (GCDRNA) is proposed to predict the traffic state in the traffic network. GCDRNA mainly consists of two components, which are Graph Convolutional with Attention (GCA) block and Dynamic GRU with Attention (DGRUA) block. GCA block can capture both global and local spatial correlations of the traffic flows by k-hop GC, similarity GC and spatial attention modules. DGRUA block captures the long-term temporal correlation of the traffic flows by Dynamic GRU (DGRU) and Node Attention Unit (NAU) modules. Experimental results show that GCDRNA achieves the best prediction performance compared with other baseline models on two public real-world traffic datasets.