Spatio-temporal Dynamic Graph Convolutional Probability Sparse Attention Networks for Traffic Flow Forecasting

Accurate traffic flow forecasting plays a vital role in the effective management and control of intelligent transportation systems. However, existing forecasting methods face constraints due to insufficient information sharing and a limited understanding of global temporal relationships, impeding a thorough analysis of temporal and spatial features. To address this issue, we propose a novel approach called ST-DGCPSA (spatio-temporal dynamic graph probabilistic sparse attention networks) for traffic flow forecasting. This framework serves as a general forecasting solution, effectively extracting the temporal and spatial characteristics of traffic flow. First, a temporal graph convolution block is constructed, treating historical time slots as graph nodes and employing graph convolution to capture flexible global temporal dependencies. Next, a dynamic graph constructor is designed to explore spatial correlations between nodes and dynamic temporal dependencies across different time points, fully extracting dynamic spatio-temporal relationships. Finally, the spatio-temporal convolution block enhances the model's ability to capture spatial correlations, temporal dependencies, and multi-scale modeling through the probabilistic sparse attention mechanism, leading to improved forecasting accuracy. The experimental results on four real datasets demonstrate the superiority of ST-DGCPSA over existing techniques.