A Multi-graph Fusion Attention Network for Traffic Flow Forecasting

Traffic flow forecasting constitutes a fundamental task in Intelligent Transportation Systems (ITSs). Although existing methods capture spatiotemporal dependencies from traffic data, most rely solely on a single graph structure, making it difficult to jointly model long-term topological priors and short-term dynamics. To this end, this paper proposes a Multi-Graph Fusion Attention Network (MGFAN) for traffic flow forecasting, including a Periodic Attention Feature Selection Module (PAFSM) and a Multi-Graph Fusion Attention Gated Recurrent Unit (MGFA-GRU). Specifically, PAFSM adopts attention mechanisms to extract periodic patterns across different temporal scales. Moreover, MGFA-GRU integrates a multi-graph fusion module that dynamically constructs adjacency matrices from periodic features and historical data, where adaptive fusion weights are utilized to jointly model static and dynamic graphs to preserve stable topological priors while exploring dynamic variations. Additionally, a dual attention mechanism is combined with the graph convolution network to model long-term dependencies and multi-scale spatial representations. Extensive experiments across four real-world datasets demonstrate that MGFAN outperforms comparable baselines.