S-RGAT: A hybrid relational graph attention network for interregional railway passenger flow prediction

Accurately modeling interregional railway passenger flows is important for understanding regional mobility demand and supporting the planning and management of railway systems. However, existing studies lack a railway-specific framework that simultaneously accounts for spillover demand from nearby areas due to limited rail coverage and integrates diverse semantic features that capture indirect interregional dependencies, limiting predictive accuracy. To address this gap, we propose the Semantic-enhanced Relational Graph Attention Network (S-RGAT), a two-stage model which integrates local spatial message passing with global semantic enhancement for interregional railway passenger flow modeling. In the first stage, we perform attention-based message passing on a multi-relational graph constructed from rail and road travel times to capture localized spatial dependencies. In the second stage, a multilayer perceptron decoder integrates the learned node embeddings with origin–destination level semantic features, while a soft gating mechanism conditionally regulates their use across OD contexts. Experiments conducted on China’s Yangtze River Delta demonstrate that S-RGAT outperforms baseline models in both predictive accuracy and structural consistency. Ablation experiments and attention footprint analysis further highlight the importance of incorporating road accessibility and areas without rail service in railway passenger flow prediction. Overall, S-RGAT provides not only improved model performance, but also a more interpretable and geographically grounded framework for understanding interregional railway mobility.