Author ORCID Identifier
https://orcid.org/0000-0002-1638-952X
Date Available
5-1-2027
Year of Publication
2025
Document Type
Doctoral Dissertation
Degree Name
Doctor of Philosophy (PhD)
College
Engineering
Department/School/Program
Civil Engineering
Faculty
Mei Chen
Faculty
Mei Chen
Abstract
Efficient transportation systems are increasingly vital for fostering socio-economic progress and addressing the needs of a growing population. This study investigates critical elements of traffic speed distribution estimation, hourly volume estimation, and establishes a linkage between reliability and Level of Service (LOS). These elements are essential for assessing transportation performance and informing resource allocation strategies within transportation agencies.
Leveraging advancements in probe data quality and machine learning techniques, the study developed models for accurately estimating traffic speed distributions across various roadway types and times of day, including peak and non-peak hours. The XGBoostLSS model demonstrated superior performance, surpassing state-of-the-art models such as generalized random forests (GRF) and quantile regression forests (QRF). The analysis revealed valuable insights into traffic management, including a reaffirming the negative correlation between access point densities and speeds. It emphasized that effective strategies must extend beyond mere capacity expansion; peak capacity positively affected speeds only up to a certain threshold. Moreover, traffic volumes and truck percentages, particularly on freeways and multilane highways, significantly influence speed predictions, underscoring the need for tailored management approaches that consider specific traffic conditions and roadway characteristics.
The study further explored enhancements in hourly volume estimation through attention-based deep learning, leveraging the TabNet model to significantly outperform baseline methods including XGBoost and traditional artificial neural networks. By incorporating a diverse set of features – including hourly short-term counts from portable stations, probe vehicle data, weather conditions, and roadway characteristics – the model captured temporal and contextual variation more effectively than conventional approaches. The integration of short-term counts led to a notable reduction in RMSE and MAPE, underscoring their value in improving model accuracy. Additionally, the inclusion of spatial correlation features, such as the count of origin-destination pairs derived from network topology, contributed meaningfully to model performance, demonstrating the importance of accounting for spatial structure. Probe data further enhanced prediction accuracy, consistent with previous findings, contributing approximately 20% to overall model improvement. These results affirm the effectiveness of combining deep learning with multi-source data and network-aware features for accurate hourly traffic volume estimation.
Additionally, the research introduced a complementary perspective to the traditional LOS framework by integrating travel time reliability, establishing a quantitative link between density-based LOS and speed variability. Leveraging GPS-probe speed data from HERE Technologies and traffic flow data from continuous count stations across Kentucky, the study examined how speed variability – measured by both the standard deviation and coefficient of variation (COV) of speed – relates to traffic density. Focusing on urban freeways, where the full LOS spectrum from A to F was observed, the analysis employed a quantile regression model to capture this relationship and developed a lookup tool to estimate speed variability metrics across different density levels. This approach enables practitioners to assess how LOS targets influence the consistency of travel times, offering a practical tool for sketch planning applications in data-scarce environments. The study demonstrated that speed variability tends to be moderate under free-flow conditions and lowest under severe congestion, underscoring the value of incorporating reliability into performance assessments of proposed roadway improvements.
Ultimately, this study offers valuable insights and tools for transportation agencies, enhancing their capacity to evaluate and improve roadway performance. By addressing both current service quality and future infrastructure needs, the research supports effective resource allocation and promotes sustainable transportation practices. The findings have broader implications for transportation planning and policy, providing a framework for integrating reliability into performance assessments and facilitating the development of responsive traffic management strategies.
Digital Object Identifier (DOI)
https://doi.org/10.13023/etd.2025.113
Funding Information
This study was supported by the Kentucky Transportation Center through two grants: PL 42, Congestion and Travel Time Reliability, Performance Measures for SHIFT 2024 (April 15, 2022 – June 30, 2023), and OHS, Develop an Online Dashboard for Automatic ATR and Classification Data Processing and Demand Factors Update (October 1, 2023 – September 30, 2024).
Recommended Citation
Boasiako Antwi, Eugene, "Modeling Speed Distribution and Volumes for Evaluating Travel Time Reliability Benefits" (2025). Theses and Dissertations--Civil Engineering. 154.
https://uknowledge.uky.edu/ce_etds/154
