Dynamic Memory Derived Graph learning for Spatio-Temporal Metro Flow Prediction*

Accurate prediction of urban metro ridership flows is crucial for effective transportation operation management and optimized individual travel planning. Addressing the complex spatio-temporal dynamics data and topology information within the metro system, we introduce a novel spatio-temporal graph structure learning model: Memory Derive Graph Convolutional Recursive Network (MDGCRN). Specifically, We implement this method to Graph Convolutional Recursive Network(GCRN) to predict the urban metro station flow. First, we propose a Spatio-Temporal Memory Node Library module to capture similarity and correlation at different stations. We then incorporate a Graph Memory Learner, guided by the Node Library, into the GCRN’s encoder-decoder architecture. This allows the model to embrace the intricate spatio-temporal data heterogeneity and non-stationarity inherent in both intra-station and inter-station metro ridership patterns. Finally, we propose a transformer-based mid-term relationship prediction model for metro flow prediction. We conducted a comprehensive evaluation based on metro flow data from Beijing and Shanghai, the advantages of our method in predicting metro ridership flow at the station level has been well demonstrated through experiments with more than 11 reference lines. The source code of the paper is available at https://github.com/chenw396/MDGCRN.git.