Dual-branch graph neural network with spatio-temporal fusion for traffic flow forecasting

Reliable traffic flow forecasting is fundamental to intelligent transportation systems. Existing spatio-temporal models often emphasise either localised variation or long-range dependency, rely on spatial relations that respond too slowly to traffic changes, and fuse multi-branch features in a relatively shallow way. To address these issues, this paper proposes a dual-branch spatio-temporal fusion graph neural network (DB-STFGNN). The first branch captures localised coupled dynamics through multi-scale temporal convolution and a hybrid diffusion dynamic graph convolution (HDGCN) module with iterative interaction updates. The second branch employs spatio-temporal multi-head attention to model broader dependencies across nodes and time steps. To strengthen collaboration between the two branches, the authors further introduce a bidirectional cross-attention fusion module (CAFM), which performs cross-branch interaction and feature refinement before prediction. A non-autoregressive prediction head then maps the fused representation to future traffic flow values. Experiments on real-world benchmarks, including PeMS03, PeMS04, PeMS07 and PeMS08, show that DB-STFGNN achieves competitive performance under mean absolute error, root mean square error and mean absolute percentage error compared with traditional spatio-temporal models and mainstream graph-based baselines.