Spatio–temporal graph hierarchical learning framework for metro passenger flow prediction across stations and lines

Accurate prediction of metro passenger flow is crucial for the public and metro managers as it can provide decision support. Previous research has predominantly focused on predicting passenger flow at individual stations and lines, often encountering challenges in simultaneously predicting both aspects. Furthermore, some studies that mine spatio–temporal data from metro networks have tended to remain at a global level and have not deeply explored individual stations. In this study, we propose a hybrid prediction framework using spatio-temporal graph neural networks to accurately predict inter-station and inter-line passenger flows while also considering the overall network dynamics. This approach not only captures global information but also emphasizes the importance of precise predictions for individual stations. By utilizing spatio-temporal graph convolutional networks, we derive the global spatio–temporal information to construct a feature flow. Then, by employing the proposed Local Feature Extraction Module, we perform an initial prediction to obtain the prediction value of each individual station, thereby completing the first stage of feature extraction and model training. Furthermore, we establish a new hierarchical prediction module to generate line-level passenger flow predictions while correcting station-level prediction errors in the first stage. Four experiments based on real data from the Hangzhou and Shanghai metro systems demonstrate that our framework outperforms all baseline models, highlighting its outstanding performance and versatility.