A deep graph convolution spatial-temporal attention learning model for traffic flow prediction

Abstract As an effective means to solve the challenges of online measurement of difficult-to-measure variables in complicated traffic processes, deep learning–based traffic flow prediction has emerged as one of the primary research objectives in the field of soft measurement with its strong data feature extraction ability. However, most of the existing traffic flow prediction methods follow a stacked structure of spatial-temporal blocks to capture traffic flow features, which encounter limitations in solving the over-smoothing problem caused by the deep-stacked graph convolution network (GCN), as well as the interaction effects endured by the spatial-temporal attention learning based on cascaded structure. In this paper, we adopt a sequential structure to connect the spatial module, the temporal module, and the attention module to achieve the spatial-temporal correlations learning. Along this line, we propose a deep graph convolution spatial-temporal attention learning network, which considers both large range spatial dependence and global joint spatial-temporal correlation, to predict traffic flows. In particular, a deep stacked GCN module is adopted in our model to capture multiscale spatial features of the traffic flow data, which are constructed by leveraging a combination of residual networks, jump connections and multilayer perceptrons. Afterward, the temporal features are learned using the gated temporal convolution. Finally, a spatial-temporal attention mechanism is introduced to simultaneously capture dynamic global correlations in both spatial and temporal dimensions. Our experimental results, based on four real-world traffic flow datasets called PeMSD3, PeMSD4, PeMSD7, and PeMSD8, reveal that the prediction accuracy of the proposed model outperforms other baseline models. Moreover, it can efficiently alleviate over-smoothing effects while maintaining manageable computational overhead.