Short-Term Metro OD Passenger Flow Prediction Based on Origin–Destination Bidirectional Graph Convolution and Dynamic Gated Fusion

Accurate short-term origin–destination (OD) passenger flow prediction provides important demand information for metro operation. An OD flow is associated with both origin-side and destination-side spatial dependencies, but their relative predictive value may differ across OD pairs and operating periods. Separately extracting the two representations does not establish whether their fusion should vary with the current state. To investigate this problem, an origin–destination bidirectional graph convolution and dynamic gated fusion model, termed OD-BGCN-GRU, is proposed. Origin-GCN and Destination-GCN propagate information along the two dimensions of the OD matrix, respectively. A state-dependent gate adaptively combines the two spatial representations, while a gated recurrent unit captures temporal evolution. Experiments on Hangzhou Metro automatic fare collection data show that OD-BGCN-GRU achieves an MAE of 0.704±0.007 and an RMSE of 2.253±0.021 on active OD pairs, reducing MAE by 3.03% compared with ODMixer. Moreover, dynamic fusion further reduces MAE by 2.36% relative to OD-specific static fusion, supporting the predictive benefit of state-dependent coordination between origin-side and destination-side spatial information. These findings provide a useful modeling basis for fine-grained short-term metro OD demand forecasting and subsequent demand-oriented operational analysis.