Development of generalizable data-driven traffic state prediction solutions: a comprehensive approach

Author

• Md Mobasshir Rashid, University of Central FloridaFollow

ORCID

0000-0002-3253-8038

Keywords

generalizability, data-driven traffic prediction, transferability across contexts, knowledge graph, graph transformer

Abstract

Increased urbanization and population growth lead to more traffic in the transportation system, causing severe congestion. Particularly during disruptive events such as hurricanes or traffic incidents, the traffic flow in the network worsens significantly. As a result, there is a need for data-driven traffic prediction models to reduce congestion in real-time, and the prediction model should be generalizable to be applicable across different dynamic scenarios. Additionally, a robust and adaptable traffic prediction model can help traffic managers to take proactive actions to optimize traffic flow. However, the majority of data-driven traffic prediction models suffer from limited applicability as they are trained for a selected context with fixed datasets and require extensive retraining when applied to new contexts. To overcome such limitations, this dissertation addresses a comprehensive approach to increase the prediction model’s generalizability by addressing different generalizability-related aspects such as integration of novel dataset to incorporate dynamic demand information during hurricanes, semantic representation of a network, selection of optimal model architecture to reduce overfitting, and transferability of the prediction model across new contexts. The dissertation investigates the following problems. First, the dissertation conducts a case study by incorporating novel ‘Facebook Movement Data’ to capture dynamic demand patterns during hurricane evacuation to illustrate the usefulness of such generalizable data to enhance the model’s predictive accuracy. Second, a transportation knowledge graph incorporating semantic representation of different spatial entities of a network is constructed to increase traffic prediction model’s accuracy during incidents. The case study illustrates the use of network-related representation to enhance the interpretability of predicted results so that traffic management can apply them in real-world. Third, another case study is conducted where network-specific characteristics such as shortest path information between node pairs is integrated into a data-driven traffic assignment problem for predicting link-specific flows. The concept of incorporating shortest path information can be transferable to any new network, showcasing the cross-network adaptability of the proposed methodology. Finally, the dissertation develops a generalizable evacuation traffic prediction framework trained on historical hurricane data to predict future hurricanes’ evacuation traffic flow, minimizing the retraining process for each new hurricane to propose more robust and generalizable evacuation traffic management solutions. By addressing different aspects of data-driven generalizability, the dissertation develops a comprehensive approach of generalizable traffic prediction model to serve as an adaptable real-world decision-making tool for reducing congestion across diverse dynamic conditions.

Completion Date

2025

Semester

Summer

Committee Chair

Hasan, Samiul

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Civil, Environmental, and Construction Engineering

Format

PDF

Release Date

8-15-2026

Document Type

Dissertation

Campus Location

Orlando (Main) Campus

Subjects

Traffic flow--Forecasting; Traffic congestion--Forecasting; Urban transportation--Forecasting; Traffic flow--Research; Traffic congestion--Research

STARS Citation

Rashid, Md Mobasshir, "Development of generalizable data-driven traffic state prediction solutions: a comprehensive approach" (2025). Graduate Thesis and Dissertation post-2024. 567.
https://stars.library.ucf.edu/etd2024/567