Dual-temporal inflow–outflow dependency modeling for short-term metro outflow prediction

Recent advances in deep learning have substantially improved short-term metro passenger flow prediction. However, existing approaches often inadequately model the dependency of outflow on inflow and typically rely on predefined station correlation graphs, which limits modeling flexibility and representational capacity. To address these issues, this study decomposes the influence of inflow on outflow into short-term and long-term temporal components and proposes a dual-temporal inflow-outflow dependency model (DTIOD). DTIOD adopts an asymmetric feature extraction scheme to encode inflow and outflow sequences according to their distinct roles in forecasting. Instead of using predefined station correlation graphs or explicit spatial modules, the model employs a dual-branch cross-attention mechanism to capture inflow-outflow dependencies across multiple temporal scales, thereby enabling implicit learning of spatial correlations. In addition, sample-level origin-destination (OD) matrices are incorporated as additive attention biases to embed prior inter-station relationships and guide attention allocation. The outflow features are adaptively fused with the long-term and short-term inflow effect representations through learnable weights, and final predictions are generated by a fully connected layer. Experiments on the Hangzhou metro dataset show that DTIOD reduces RMSE (root mean squared error), MAE (mean absolute error), and WMAPE (weighted mean absolute percentage error) by 10.75%, 11.60%, and 6.84%, respectively, compared with the strongest baseline, while completing training within 70 seconds. These results demonstrate that DTIOD achieves a favorable balance between predictive accuracy and computational efficiency, indicating its practical applicability.