Subway systems have become essential to modern life in big cities, as they ease traffic congestion by carrying a growing number of passengers every year, reducing carbon emissions in the process. As subway networks expand, operators face the continuing challenge of managing large and changing passenger flows, particularly during rush hour. Accurate forecasting of subway demand thus constitutes an important part of subway management.
To this end, several predictive tools have been developed. Most of these tools focus on predicting passenger volume at individual stations, failing to capture passenger origins, destinations, and travel durations, which prevents them from modeling the subway network as a whole. This is partly a data problem, since a rider’s destination is not recorded until they exit the subway system, which can occur tens of minutes after entering. On the other hand, it is also a modeling problem; many popular forecasting methods only consider limited temporal scales, such as passenger flows in the preceding hours. However, they do not explicitly explore passenger patterns across hours, days, and weeks, and often overlook inter-station correlations.
Against this backdrop, a research team led by Professor S. Joe Qin from Lingnan University in Hong Kong, China, set out to develop a new method that can accurately forecast both the destinations and travel times of subway passengers. Their study, published in Volume 13, Issue 7 of IEEE/CAA Journal of Automatica Sinica on August 3, 2026, introduces a multi-time-scale model that can make these predictions one day ahead.
The researchers’ approach is based on a matrix that they term the travel-time distribution and destinations (TDD) matrix. Instead of predicting only the total number of passengers leaving a station, the model estimates the proportion of passengers traveling from each origin station to different destinations, as well as their expected travel times.
To build a model that predicts the TDD matrix a day in advance, the researchers first analyzed years of real subway data from the Shenzhen subway system in China, using records from 2013 and 2019. They found rider behavior on a given day closely resembles the same day from the previous week, and weekdays closely resemble other weekdays. “ By quantifying multi-time-scale similarities across weeks, days, and time intervals, we unveil the underlying mechanisms of passengers' travel patterns, revealing predictable and repetitive mobility behavior ,” explains Prof. Qin.
Built on these patterns, the model uses data from past days and past weeks to forecast tomorrow’s travel patterns for each station. Once combined with real-time rider counts entering a station, the model estimates how many passengers will exit where and when. When tested against several widely used forecasting methods like ARIMA/SARIMA and neural network-based approaches, the new model outperformed them for the large majority of stations studied, cutting prediction error by 10% on average. “ The TDD model exhibits a significant advantage over the benchmark methods for stations that exhibit a high level of regularity, such as common commuter working and residential centers. These stations generate high volumes of passengers especially during rush hours and therefore it is important to obtain accurate predictions for them, ” remarks Prof. Qin.
Overall, the findings suggest that subway operators could perform TDD-based forecasting to better plan train schedules and manage congestion at stations proactively rather than reactively. Because the model relies on data that are already being collected through existing fare card systems, it could be adopted without the need for new devices or infrastructure, paving the way for more efficient public transport.
Reference
Title of original paper: Multi-Time-Scale Modeling for Day-Ahead Forecasting of Passenger Travel-Time and Destinations Distribution
Journal: IEEE/CAA Journal of Automatica Sinica
DOI: https://doi.org/10.1109/JAS.2026.126170
About Lingnan University
Lingnan University (LU) is a leading public tertiary institution in Hong Kong, China, recognized globally as a premier research-intensive liberal arts university in the digital era. Founded in Guangzhou in 1888 as the Christian College in China, the university was re-established in Hong Kong in 1967 and is now one of the eight public universities funded by the University Grants Committee (UGC). Combining the best of Chinese and Western academic traditions under its motto “Education for Service,” Lingnan University provides a whole-person education. Lingnan University enjoys high international recognition, including ranking first in the world for “Quality Education” in the Times Higher Education (THE) Impact Rankings 2025.
Website: https://www.ln.edu.hk/
About Professor S. Joe Qin from Lingnan University
Dr. S. Joe Qin obtained B.S. and M.S. degrees in automatic control from Tsinghua University, China, and a Ph.D. degree in chemical engineering from the University of Maryland at College Park, USA. He is currently Wai Kee Kau Chair Professor of data science and President of Lingnan University in Hong Kong, China. His research interests include data science and analytics, process monitoring, model predictive control, system identification, and smart cities. He received the 2022 IEEE Control Systems Society Transition to Practice Award, the 2022 CAST Computing Award by AIChE, the U.S. NSF CAREER Award, and the NSF-China Outstanding Young Investigator Award.
Funding information
This work was supported in part by a Shenzhen-Hong Kong-Macau Science and Technology Project Category C (9240086), a grant from ITF-Guangdong-Hong Kong Technology Cooperation Funding Scheme (GHP/145/20), and a Collaborative Research Fund by RGC of Hong Kong (C1143-20G).
IEEE/CAA Journal of Automatica Sinica
Computational simulation/modeling
Not applicable
Multi-Time-Scale Modeling for Day-Ahead Forecasting of Passenger Travel-Time and Destinations Distribution
3-Aug-2026