DMGF-STAN: Spatio-Temporal Attention Network and Dynamic Multi-Graph Fusion for Traffic Flow Prediction

The development of ITS is vital for decreasing traffic congestion and improving traffic scheduling procedures. Traffic prediction is a fundamental component of the development of ITS. Even though a lot of research has been done on modeling intricate spatiotemporal correlations in order to make accurate predictions, traditional methods primarily use predefined graph structures for feature extraction, which leaves out important correlations in the data and leads to limited prediction accuracy. The objective of the DMGF-STAN that we have recently created is to recognize both explicit and latent connections between time and space in traffic flow data that are subjected to various types of alterations. Our framework introduces a dynamic multi-graph expert selection module (DMGE) that combines a multi-graph information aggregation component with a sparse gating network to effectively model complex spatial dependencies. The Dynamic Multi-graph Gating (DMGG) module subsequently integrates global and local spatial feature extraction units-specifically the Adaptive Global Similarity Graph Convolution (AGS-GConv) module and Local Spatio-Temporal Attention Graph Convolution (LSTA-GConv) module-through integration of their outputs via dynamic gating fusion mechanisms. These processed features are then coupled with GRU-based codecs for comprehensive spatio-temporal feature learning, ultimately enabling future traffic state prediction. Our model outperforms the most advanced benchmark approaches in terms of prediction accuracy, according to comparative experiments conducted on real-world traffic datasets. The proposed framework can provide urban traffic management centers with short-term congestion forecasts and support dynamic signal cycle adjustments to reduce average delay during peak hours.