Evaluating passenger flow variations across an urban rail network induced by new lines using spatio-temporal transfer learning method

Abstract

In many major cities worldwide, the expansion and construction of new rail transit lines are actively pursued to alleviate operational pressures on existing networks. Evaluating the impacts of new lines on existing ones, particularly through network-wide Origin–Destination (OD) passenger flow forecasting that accounts for newly constructed lines, is crucial for efficient line planning and network operations. However, OD flow prediction faces significant challenges due to the absence of historical passenger flow data for new lines and the changes they introduce to overall passenger volumes and distribution. This study presents a transfer learning-based hypergraph approach to represent OD flow data, tackling computational challenges in megacities with hundreds of urban rail stations. In this model, OD pairs serve as vertices, while their spatiotemporal similarities are captured by hyperedges. Spatial features are extracted from geographical data, while temporal features of existing OD pairs are learned from historical passenger flows. For new OD pairs lacking historical data, transfer learning infers temporal features from spatially similar pairs. These spatiotemporal similarities are then used to construct the hypergraph. Then, the hypergraph convolution is applied to extract high-order spatiotemporal features from the proposed hypergraph model, enabling the prediction of OD flow changes in the expanded urban rail transit network. A logit-based passenger assignment model is adopted to estimate how passengers redistribute across the network in response to the introduction of new lines. The effectiveness and accuracy of the proposed method are validated using real-world data from the Shanghai urban rail network. Results demonstrate that the modeling framework enables detailed analysis of station- and section-level passenger flow changes across the entire network. The integrated prediction–assignment framework presented in this study offers a novel and practical tool to support data-driven planning and operational decision-making in large urban rail systems.