Spatial Temporal Period Transformer Networks for Traffic Forecasting

As a crucial component of intelligent transportation system, traffic forecasting still faces challenges in achieving timely and accurate results. In this paper, we propose a novel approach called Spatial-Temporal Period Transformer Networks (STPTN) to improve the accuracy of traffic forecasting by incorporating both dynamic graph spatial dependencies and temporal periodic dependencies. We utilized Graph Attention Networks (GAT) and multi-head attention mechanism to simulate the real-world relations of traffic sequences and better extract their dependency from the static and dynamic perspectives, respectively. By extracting spatial and temporal dependencies in parallel, it not only enhances learning efficiency, but prevents interference between spatial and temporal dependencies. Experimental results demonstrate that STPTN has achieved state-of-the-art results in the accuracy of traffic forecasting compared to several existing methods.