Passenger Flow Prediction for Urban Rail Transit Transfer Stations Based on Multi-Scale Dual-Graph Spatial-Temporal Fusion Network

Abstract

To improve the accuracy of short-term passenger flow prediction at urban rail transit transfer stations, this paper proposes a Multi-Scale Dual-Graph Spatial-Temporal Fusion Network (MS-DSTGFN). First, AFC data and network topology are integrated to identify transfer events through travel-time decomposition and a two-stage path search method, thereby constructing inbound, outbound, and transfer passenger-flow time series. Then, a physical topology graph and a data-driven adaptive graph are jointly introduced into the forecasting model, and a dual-graph hybrid propagation mechanism is designed to capture both explicit and implicit spatial dependencies. On the temporal side, three encoders are built for recent, daily-periodic, and weekly-periodic sequences, and a cross-scale attention module is adopted to fuse multi-scale temporal patterns adaptively. Finally, a channel-aware decoder is used for differentiated joint prediction of the three passenger-flow types. Experiments on Hangzhou Metro AFC data show that MS-DSTGFN outperforms all baseline models. Compared with the best-performing baseline DCRNN, MAE, RMSE, and MAPE are reduced by 8.88%, 6.58%, and 10.06%, respectively. Ablation results further confirm the effectiveness of the dual-graph fusion, cross-scale fusion, and channel-aware decoding modules.