Multi-task graph-based model for metro flow prediction under dynamic urban conditions

Abstract

Accurately predicting metro commuter flows under changing urban conditions is essential for guiding infrastructure investments and service planning. However, existing methods show limited adaptability to evolving urban conditions. To address this, we propose an adaptive graph sharing embedding cascade interaction network (AGSECIN), which establishes a dynamic mapping relationship between changing urban conditions and commuter flows, enabling accurate predictions of metro inflows, outflows, and origin-destination (OD) flows simultaneously. A graph attention network is built on the long-term graph to capture the spatiotemporal evolving patterns of urban conditions. Then, an adaptive supply–demand sharing embedding network is designed to model the interaction between origin supply and destination demand. Finally, an adaptive feature interaction layer is developed to uncover the complex high-order relations among passenger flows and urban conditions. Experimental results on real-world Beijing datasets demonstrate the superior performance of AGSECIN, compared to contemporary models. Ablation experiments confirm the robustness of our model.