Abstract Real-time prediction of dynamic origin–destination (OD) passenger flows is essential for efficient passenger flow management in urban rail transit (URT) systems. Existing studies have primarily focused on commuting OD flows, which exhibit strong regularity and are supported by abundant data samples. In contrast, non-commuting OD flows—especially those generated by irregular passengers with limited historical data—are characterized by high stochasticity and data sparsity and have received relatively little attention, with existing studies often reporting unsatisfactory predictive performance. To address these challenges, this study proposes a novel real-time OD flow prediction framework for irregular non-commuting passengers through multi-source data fusion and feature extraction. Specifically, individual-level spatiotemporal behavioral features are extracted from metro AFC data using a density-based clustering algorithm. Land-use and geo-economic data are then integrated to characterize individual travel preferences and construct a multidimensional behavioral indicator system. Building upon these features, hierarchical clustering and machine learning models are employed to perform personalized destination prediction. Empirical experiments conducted on Nanjing Metro data demonstrate that the proposed framework substantially improves prediction accuracy for non-commuting passengers and provides new insights into dynamic OD modeling. The results highlight the strong applicability and potential of the method for real-time passenger flow prediction in complex urban rail systems.
Cui et al. (Mon,) studied this question.