Short-term metro passenger flow forecasting via decoupled spatio-temporal modeling with a diffusion-convergence graph and Swin-Enhanced LSTM

Short‑term metro passenger flow forecasting—predicting station‑level inflow and outflow volumes is challenging, mainly due to the complexity of multimodal spatial couplings and the difficulty of modeling long‑range temporal dependencies while retaining local variations. Therefore, this paper proposes an Integrated Spatio-Temporal LSTM (IST-LSTM). IST-LSTM adopts a dual-branch architecture in which a diffusion branch explicitly models OD-driven propagation and a convergence branch captures structural aggregation among stations; these branches are coupled through a Diffusion and Convergence Graph (DCG) that fuses structural links, inter-station distances, and behavior-driven OD diffusion into a unified spatial representation. To strengthen temporal representation, we refine the LSTM gating mechanism by integrating a Swin Transformer module, enabling the model to better represent long-range dependencies while retaining sensitivity to local, fine-grained temporal patterns. To evaluate model generalization under dense and heterogeneous urban conditions, we construct the NJMetro dataset from Nanjing AFC smart-card transaction data, which reflects inner-ring high-density travel dynamics and pronounced spatio-temporal heterogeneity; this dataset is used alongside the public HZMetro dataset. IST-LSTM reduces RMSE by 9.19% on NJMetro relative to GCN-SBULSTM and achieves an average MAE improvement of 9.62% over Graph WaveNet on HZMetro and 8.68% versus LSTM on NJMetro. On HZMetro, it also achieves competitive RMSE performance relative to PB-GRU, with notable advantages at 15- and 30-minute horizons. Overall, the results indicate that the proposed design more effectively captures the coupled spatial and temporal mechanisms underlying metro passenger flow.