Metro Passenger Flow Prediction with GNN: Station Attribute Influence Analysis

Metro passenger flow prediction is vital for optimizing scheduling, enhancing passenger experience, ensuring safety, and supporting urban transportation planning and economic development. Differences in station attributes lead to variations in how neighboring stations influence each other, a factor often overlooked in existing research. To address this, we propose a hybrid model, Graph Convolutional Station-Aware Long Short-Term Memory (GCSA-LSTM), which integrates spatial and temporal features with station-specific attributes. The model begins by classifying stations based on their passenger flow characteristics to tailor predictions to station types. It then uses a Graph Convolutional Network (GCN) to construct a metro network structure, dynamically adjusting the influence weights of neighboring stations to capture spatial dependencies. Finally, an LSTM model processes multi-source data to incorporate temporal dynamics, enabling accurate and robust passenger flow predictions. The model's effectiveness is tested on real-world datasets, outperforming traditional machine learning methods. Ablation experiments further confirm the value of its modular design. By leveraging spatial and temporal dynamics and stationspecific attributes, the GCSA-LSTM model enhances prediction accuracy empowering metro operators to make data-driven decisions that improve efficiency.