A Deep Learning Framework for Rail Station Passenger Flow Prediction with Temporal Knowledge Graph Embedding

To enhance the dispatching efficiency and service capacity of urban rail transit systems during peak hours, this study proposes a short-term subway passenger flow prediction framework named TKG-ResLSTM-GCN, which integrates Temporal Knowledge Graphs (TKG), Long Short-Term Memory (LSTM) networks, and Graph Convolutional Networks (GCN). The proposed method constructs a temporal knowledge graph of rail passenger flow to mine the dynamic flow relationships among stations over time. Point of Interest (POI) data is introduced to enhance the semantic representation of stations, and dynamic embeddings are generated based on the knowledge graph. The model further combines the spatial modeling capability of GCN with similarity matrices and the temporal sequence modeling ability of LSTM, enabling accurate short-term passenger flow prediction for all stations in the subway network. Experiments demonstrate that the proposed model outperforms baseline prediction methods such as ARIMA and LSTM across multiple temporal granularities (10 min, 15 min, and 30 min), achieving the best results in terms of RMSE, MAE, and WMAPE. This confirms its effectiveness and superiority in capturing the spatiotemporal evolution patterns of metro networks.