A Short-term Traffic Flow Forecasting Model Based on Spatial-temporal Attention Neural Network

In recent years, with the enrichment of information collection methods, traffic flow data are no longer limited to a single section or a certain region, and often presented in the form of large-scale road networks. Traffic flow data of road networks are mainly spatial-temporal sequential data. This paper combines graph theory with prediction of temporal sequences to propose the Spatial-temporal Attention Neural Network (STAtt) from the perspective of spatial-temporal feature fusion, aiming at providing data support and reference for traffic guidance and path planning. The Graph Attention Network (GAT) is used to describe the directivity, difference and variability of the roads' interactions, which is embedded into the loop unit to replace the gate structure to achieve the spatial and temporal information fusion. At the same time, the Sequence-to-Sequence architecture and temporal attention mechanism are used to enlarge the receptive field of the model and reduce the cumulative error of multi-step prediction. The results of the experiment show that the model is superior to other models in the short-time prediction task within 1 hour.