Multi-Graph Inductive Representation Learning for Large-Scale Urban Rail Demand Prediction under Disruptions

With the expansion of cities over time, Urban Rail Transit (URT) networks have also grown significantly. Accurate demand prediction plays a crucial role in supporting planning, scheduling, fleet management, and other operational decisions. This study proposes an Origin-Destination (OD) demand prediction model called Multi-Graph Inductive Representation Learning (mGraphSAGE) for large-scale URT networks under operational uncertainties. The proposed model represents each OD pair as a node in multiple graphs that capture distinct spatial and temporal correlations, thereby enhancing the spatial learning capability of graph-based methods while maintaining scalability. Moreover, operational uncertainties such as train delays and cancellations are explicitly incorporated as model inputs to improve robustness under real-world disruptions. The model is validated on three network scales of the Copenhagen URT system. Experimental results show that mGraphSAGE outperforms both conventional graph-based and machine learning baselines, achieving up to a 5% reduction in RMSE across network scales. The consistent improvement demonstrates the model’s enhanced spatial representation and robustness under operational uncertainties, confirming its suitability for large-scale and disrupted URT environments. • We propose mGraphSAGE, a novel multi-graph inductive representation learning model for OD demand prediction in Urban Rail Transit (URT) networks. • The model represents each OD pair as a node and simultaneously learns from multiple graphs capturing spatial and temporal correlations. • Operational disruptions such as train delays and cancellations are incorporated as input features to improve prediction performance. • The approach is validated on three different scales of the Copenhagen URT network, demonstrating scalability and adaptability.