Origin-destination flow generation for metro network expansion using spatiotemporal gated graph neural networks

Abstract

• Tackles OD flow generation for cold-start pairs in metro network expansion. • Constructs dynamic localized graphs to capture evolving metro connectivity. • Develops an adaptive gate to modulate spatial-temporal feature fusion. • Achieves strong performance and generalization across scenarios. Rapid urbanization drives continuous expansion of metro networks to meet rising mobility demand. Due to the large investment, such expansion requires careful planning, including detailed assessment of potential Origin-Destination (OD) passenger flows. However, forecasting OD flows becomes particularly challenging in evolving metro networks; new stations and lines introduce OD pairs without historical ridership data, while travel patterns shift as passengers adapt to network changes. Traditional gravity-based models rely on oversimplified assumptions and overlook complex network dependencies within metro systems. Meanwhile, recent deep learning methods designed for regional-scale mobility often struggle to capture the fine-grained dynamics and structural evolution unique to metro networks. To address these limitations, we propose a Spatiotemporal Gated Graph Neural Network (STG-GNN) to address the OD flow generation for metro network expansion problem. STG-GNN integrates diverse urban contextual data and dynamically evolving metro network structures. Spatial dependencies are modeled using graph attention networks built on both transfer-aware travel time and geographic distance, while temporal patterns are captured via Gated Recurrent Units (GRUs). A novel gated fusion module adaptively combines spatial and temporal outputs, weighting their contributions based on OD pair characteristics. Additionally, an age-aware weighted loss function is used to reflect the maturation process of new OD pairs. We validate STG-GNN with extensive experiments using multi-year metro ridership data from Shenzhen, China. Results show that STG-GNN consistently outperforms existing state-of-the-art models, improving Common Part of Commute (CPC) by 13.6% and reducing RMSE and MAE by 11.3% and 4.7% for new OD pairs. The proposed model is general and can be adapted to other evolving transportation networks. The source code and a synthetic dataset are made publicly available for transparency and reproducibility.