Dynamic–Static Graph Fusion Multi-Head Flow Attention Networks for Traffic Flow Forecasting

Traditional traffic flow forecasting methods still face challenges in capturing complex spatiotemporal correlations. Static graph convolutional networks are unable to capture spatiotemporal dynamics, while dynamic graphs can adaptively adjust spatial dependencies but often ignore the inherent static connectivity of traffic networks. To address these limitations, this paper proposes a Dynamic–Static Graph Fusion Multi-Head Flow Attention Network (DSGFMFAN). Specifically, an Information-Enhanced Gated Recurrent Unit (IE-GRU) is designed to more effectively capture temporal correlations. Meanwhile, a Dynamic–Static Graph Fusion Gating (DSGFG) mechanism is introduced to integrate dynamic and static graphs, enabling more comprehensive modeling of latent spatial dependencies. Furthermore, a Gated Multi-Head Flow Attention mechanism (G-MFA) is proposed, which replaces the conventional linear projection in multi-head attention with a dynamic–static graph fusion gating module to capture complex spatiotemporal interactions. In addition, flow attention is incorporated into the model, along with a source competition mechanism and a sink allocation mechanism, to efficiently capture critical information while alleviating the quadratic complexity caused by similarity computations in traditional attention mechanisms. Extensive experiments on four real-world traffic datasets demonstrate that DSGFMFAN significantly outperforms existing baseline methods in terms of prediction accuracy.