Estimation and prediction of urban traffic flows in response to global pandemic using machine learning and foundation models

Li, Yue (2026) Estimation and prediction of urban traffic flows in response to global pandemic using machine learning and foundation models. PhD thesis, University of Glasgow.

Abstract

Urban traffic flow refers to the dynamic movement of vehicles within a city, playing an important role in supporting sustainable and efficient transportation systems. This research concerns several methodological and empirical gaps related to urban traffic analysis, particularly in the context of data quality, model generalisability, and disruptions caused by external events such as the global COVID-19 pandemic. The aim of this research is to improve the understanding and prediction of urban traffic flows by integrating highresolution sensor data, emerging urban indicators, machine learning and foundation models. Specifically, the research intends to achieve the following goals: (1) to develop a publicly available, long-term traffic flow dataset with high spatio-temporal granularity; (2) to explore how spatial distribution of built environment and socio-demographics influences traffic dynamics; and (3) to predict the temporal distribution of traffic flows under both normal conditions and disruptive events. The goals are achieved through three empirical studies conducted in Glasgow, UK.

To achieve the first objective, a high-resolution intra-city traffic dataset covering four consecutive years before, during, and after COVID-19 is constructed. A multi-step cleaning process is applied to remove poor-quality sensor records using spatial, temporal, and numerical filters. The filtered dataset is then validated through spatial and temporal analyses, including comparisons with government policy stringency measures during the pandemic. Results show that the dataset reliably captures daily, seasonal, and disruption related variations in traffic flow across road types and neighbourhoods.

To achieve the second objective, the study integrates traffic flow data developed in the first study with a range of urban elements, including road characteristics, socio-demographics, surrounding built environments (land use and nearby points of interest), and the emerging urban big data source such as Google Street View (GSV) imagery. Spatial econometric models are used to understand the relationship between traffic flows and urban indicators before, during, and after pandemic periods. The results reveal that higher traffic flows are more frequently observed in areas with more young and white dwellers, while lower flows are observed in natural green spaces. Major roads between cities and towns also show heavier traffic flows. Besides, the application of GSV images in this research has revealed the heterogeneous effects of green space on urban traffic flows, as the magnitudes of their effects vary by distance. We also detect that the spatial dependence between adjacent neighbourhoods among the traffic flows and associated urban parameters is variable during the four COVID-19 periods. With the influence of COVID-19, there has been a significant decrease in long-distance travel.

To achieve the third objective, two pre-trained foundation models, Lag-Llama and Chronos, are applied for zero-shot traffic flow prediction and we have compared their accuracy against traditional deep learning models. The results show that foundation models outperform deep learning models in traffic flow prediction under both normal conditions and disruptive events. Unlike deep learning models, which require large-scale historical data and extensive training time for each task, pre-trained foundation models can be directly applied to datasets with different data sizes, traffic dynamics, and context lengths. We also find that foundation models with longer context lengths and larger model sizes achieve higher prediction accuracy but require increased inference times. Selecting an appropriate foundation model is also crucial – models trained on a comprehensive dataset are more likely to achieve superior zero-shot performance, making them a practical and efficient choice for real-world traffic prediction applications.

Overall, this thesis contributes to the development of urban traffic research by introducing high-resolution traffic data, analysing the quantitative relationships between traffic patterns and urban elements, and demonstrating the potential of pre-trained foundation models for efficient and accurate traffic prediction using limited data. The findings can support urban planners and policymakers in making effective planning and resource allocation decisions in diverse and dynamically changing urban traffic environments.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: H Social Sciences > HE Transportation and Communications
• Colleges/Schools: College of Social Sciences > School of Social and Political Sciences
• Supervisor's Name: Zhao, Professor Qunshan and Wang, Dr. Mingshu
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-86073
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 25 Jun 2026 09:23
• Last Modified: 25 Jun 2026 09:25
• Thesis DOI: 10.5525/gla.thesis.86073
• URI: https://theses.gla.ac.uk/id/eprint/86073