Residual adaptation with contextual information and structural priors for traffic flow forecasting

Accurate multi-step traffic flow forecasting is important for intelligent transportation systems but remains challenging due to complex temporal dynamics, external contexts, and spatial dependencies. Existing spatio-temporal Transformer models have achieved strong performance in traffic forecasting, but heterogeneous contextual information and complementary structural priors are still insufficiently utilized. To address this issue, this study proposes MSIFA-PDFormer, a context-structural residual adaptation framework for traffic flow forecasting. The proposed method extends the PDFormer backbone by introducing two complementary adaptation modules. The Horizon-wise Late Context Adapter incorporates weather and calendar variables at the prediction stage to perform horizon-aware contextual correction, while the Multi-Graph Structural Prior Adapter adaptively fuses road topology, Dynamic Time Warping (DTW)-based temporal similarity, and self-loop priors to refine hidden spatial representations. Across five independent runs on MTD-300, MSIFA-PDFormer achieved an MAE of 51.60 ± 0.21, an RMSE of 106.83 ± 0.28, and a MAPE of 17.23% ± 0.19%. Compared with PDFormer, these results represent average reductions of 3.82% in MAE, 3.09% in RMSE, and 4.70% in MAPE. Further five-run evaluations on PeMS04-W and PeMS08-W showed consistent average improvements across different traffic networks. These results indicate that context-structural residual adaptation can effectively improve PDFormer-based traffic flow forecasting by jointly exploiting external contextual information and multi-graph structural priors.