A Dual-Graph Spatio-Temporal Network for Urban Rail Transit Ridership Prediction

Abstract

To predict the multi-step urban rail transit passenger flow, we propose a dual-graph spatio-temporal network (DSGN). This network can perform station-level passenger flow prediction and is beneficial for real-time operation of the stations. Predicting multiple future time steps remains challenging due to complex spatial dependencies and nonlinear temporal dynamics. The DSGN designed in this paper adopts a decoupled architecture to separately handle the passenger flow signals and time covariates. It performs parallel convolution on the physical topology graph, applies adaptive convolution on the learned graph, and then fuses through gates. The model also incorporates dilated causal convolution and attention pooling to model temporal dynamics. To validate the model, we conducted tests on the rapid transit dataset of the Massachusetts Bay Transportation Authority (MBTA) (121 stations, 30-minute resolution). The results showed that DSGN achieved a Mean Absolute Error (MAE) of 34.67 and weighted average absolute percentage error (WMAPE) of 18.60%, outperforming the competing spatio-temporal baseline models, and having lower variance across seeds. To further explore the model, we also designed ablation experiments, which showed that dual-graph fusion, time attention pooling, and decoupled feature encoding can all improve the model performance, while replacing the learned node embeddings with demographic priors would reduce accuracy and training stability.