Abstract:

Accurate prediction of Origin-Destination (OD) passenger flows is essential for improving metro system efficiency and service quality. While prior research has mainly focused on forecasting inflow and outflow at individual stations, OD flow prediction across the entire metro network remains underexplored due to several challenges: 1) the high temporal variability and intricate spatial dependencies of OD flows, 2) the influence of various external factors, and 3) the sparsity and incompleteness of OD data. To address these issues, we propose an Adaptive Feature Fusion Network (AFFN) that (a) adaptively integrates spatial relationships from multiple knowledge-based graphs, including latent inter-station correlations, and (b) effectively captures periodic flow patterns by learning the dynamic impact of external variables. To further enhance prediction accuracy under sparse data conditions, we introduce a multi-task AFFN framework that jointly predicts station-level inflow and outflow as auxiliary tasks to support OD matrix estimation. Extensive experiments on real-world metro datasets from Nanjing and Xi’an, China, demonstrate that both AFFN and its multi-task variant significantly outperform state-of-the-art baselines across multiple evaluation metrics, highlighting the robustness and effectiveness of our approach in complex urban transit environments.

Description:

.