Attention-based spatial–temporal adaptive dual-graph convolutional network for traffic flow forecasting

Accurate traffic flow forecasting (TFF) is a prerequisite for urban traffic control and guidance, which has become the key to avoiding traffic congestion and improving traffic management in intelligent transportation systems. To precisely characterize the spatial structure of road networks and discover temporal and spatial characteristics, we propose an attention-based spatial–temporal adaptive dual-graph convolutional network (ASTA-DGCN) for TFF in this paper. Specifically, we employ a spatial–temporal attention module to explore the hidden temporal correlation information of traffic data and the implicit influence of weights among road network nodes and to further capture the dynamic influence of different spatial–temporal positions on the current spatial–temporal position. Then, we utilize an adaptive graph modeling module to automatically extract the one-way relationship between variables and integrate external knowledge into the module. The FastDTW algorithm is exploited to measure the similarity of road network nodes, and the non-Euclidean pairwise association between regions is encoded into graphs to discover the hidden temporal pattern similarity effectively. Furthermore, temporal and spatial correlations are explicitly modeled using dual-graph convolution and sequential convolution based on the obtained graphs to mine the spatial–temporal patterns in dynamic traffic flow, and the final prediction result is produced based on the weighted fusion of the output values of the recent, daily, and weekly components. Finally, the ASTA-DGCN algorithm is successfully applied to TFF on two real-world traffic datasets. The experimental results indicate that our ASTA-DGCN algorithm outperforms ARIMA, VAR, FNN, GCN, GAT, GWNet, STGCN, ASTGCN, and STSGCN.