An adaptive transfer learning framework for real-time detection and prediction of anomalous returning passenger flow in railway stations
Yanchun Shao, Enze Liu, Zhiyuan Lin, Shuguang Zhan, S.C. Wong · Transportmetrica A Transport Science · 2026
Successive train delays and cancellations often trigger unexpected passenger responses, particularly in the form of returning passenger flow (RPF), where passengers abandon their trips and return from railway stations. This increased passenger flow generates an anomalous surge in urban transit demand, thereby exacerbating the effects of railway disruptions on broader transit systems. However, when observed through certain data sources (e.g. mobile dataset), these passenger flow patterns tend to be infrequent, irregular, and data insufficient, presenting a challenge for effective model training for prediction. To address this, this paper proposes a transfer learning framework, integrating an online adaptive fine-tuning mechanism, for the real-time detection and prediction of anomalous RPF. A lightweight deep learning model, comprising a convolutional neural network-Transformer hybrid architecture synergized with dynamic XGBoost regularisation, is designed to strike a balance between computational time and prediction accuracy. The framework is extensively validated through a suite of experiments on a real-world railway disruption in China, using a large-scale mobile dataset from two heavy-load railway stations. The results show that, compared to state-of-the-art models, the proposed framework reduces the root mean square error by 36% and the mean absolute error by 39%, while also decreasing computational time by 31%. Thus, the proposed framework demonstrates potential for real-time RPF prediction during railway disruptions.