Short-term Traffic Flow Prediction With Residual Graph Attention Network

Traffic flow prediction has been essential for traffic management and road network planning. However, the complex urban road network and the strong spatial-temporal correlation of traffic flow data make this problem difficult. Existing prediction methods cannot fully utilize the spatial-temporal correlations in traffic flow data. Therefore, we propose a deep learning model called ResGAT-ABiGRU which combines Residual Network (ResNet), Graph Attention Network (GAT), Attention Mechanism, and the Bidirectional Gated Recurrent Unit (BiGRU). Firstly, GAT is used to capture the spatial correlations of traffic flow data, and then the time characteristics are extracted by Bidirectional GRU. Secondly, the ResNet module stacks multiple GAT layers and designs the attention mechanism to assign weights for different flow sequences to further capture spatial relations. Finally, we obtain the output through the fully connected layers. Validation traffic data from California, USA, is used for verification. The results show that the ResGAT-ABiGRU model proposed in this paper has higher prediction accuracy. Compared the model's performance with the Gated Recurrent Unit (GRU) baseline model, and the root means square error (RMSE) is reduced by 22.75%, and compared to the T-gcn model, the root mean square error is reduced by 3.29%.