Date of Award

7-21-2025

Degree Type

Thesis

Degree Name

Master of Philosophy (MPHIL)

Discipline

Data Science

First Advisor

Prof. QIN Si Zhao Joe

Abstract

Accurate forecasting of metro passenger flow is vital for efficient urban transportation management and optimal resource allocation in modern cities. Traditional ARIMA-based models effectively capture regular, cyclical patterns but struggle with sudden, nonlinear fluctuations caused by random events such as weather disruptions, special events, or service interruptions. Moreover, existing research predominantly focuses on individual stations, overlooking the complex cross-station interactions inherent in networked metro systems where passenger flows are interconnected across the entire network.

To address these critical limitations, we propose the Three-Stage Latent Dynamics Forecasting (T-LDF) Framework, a novel approach that systematically integrates temporal decomposition, latent dynamics extraction, and cross-station interaction modeling. Initially, the observed passenger flow is decomposed into a periodic component, representing regular daily and weekly commuting patterns, and a non-periodic component, capturing irregular fluctuations and anomalies. Next, the Latent Variable AutoRegression (LaVAR) model extracts low-dimensional dynamic latent variables (DLVs) from the non-periodic component, effectively filtering noise while preserving essential dynamic information for model training. Finally, a supervised learning model with an innovative alignment mechanism accounts for travel delay effects in passenger movements across the metro network, facilitating accurate real-time forecasting. Comprehensive experiments conducted on the Shenzhen Metro passenger flow dataset demonstrate that T-LDF significantly outperforms traditional statistical models and existing single-station approaches, achieving remarkable Test-RMSE reductions of 27% on weekdays and 20.4% on weekends, while simultaneously enhancing model interpretability through explicit cross-station interaction modeling.

Keywords

Metro Passenger Flow, Latent Dynamics Extraction, Spatio-Temporal Interactions, Machine Learning

Language

English

Recommended Citation

Zhang, T. (2025). Three-stage latent dynamics forecasting (T-LDF) framework for Shenzhen Metro passenger flow prediction (Master's thesis, Lingnan University, Hong Kong). Retrieved from https://commons.ln.edu.hk/otd/261/

Share

COinS