A Graph Deep Learning Model for Station Ridership Prediction in Expanding Metro Networks

Abstract

Due to their reliability, efficiency, and environmental friendliness, metro systems have become a crucial solution to transportation challenges associated with urbanization. Many countries have constructed or expanded their metro networks over the past decades. During the planning stage, accurately predicting station ridership post-expansion, particularly for new stations, is essential to enhance the effectiveness of infrastructure investments. However, station-level metro ridership prediction under expansion scenarios (MRP-E) has not been thoroughly explored, as most advanced models currently focus on short-term predictions. MRP-E presents significant challenges due to the absence of historical data for newly built stations and the dynamic, complex spatiotemporal relationships between stations during expansion phases. In this study, we propose a Metro-specific Multi-Graph Attention Network model (Metro-MGAT) to address these issues. Our model leverages multi-sourced urban context data and network topology information to generate station features. Multi-relation graphs are constructed to capture the spatial correlations between stations, and an attention mechanism is employed to facilitate graph encoding. The model has been evaluated through realistic experiments using multi-year metro ridership data from Shanghai, China. The results validate the superior performance of our approach compared to existing methods, particularly in predicting ridership at new stations.