Abstract
Automated traffic signal performance measures (ATSPM) are increasingly used to collect and evaluate traffic data from signalized intersections. ATSPM provide invaluable insights into both current and historical traffic signal performance. With their adoption by over 39 state and local transportation agencies, ATSPM play a crucial role in traffic operation and maintenance. ATSPM are continually evolving, introducing new performance measures and applications tailored to the needs of various transportation agencies. However, despite their potential, several challenges require attention and resolution. The first challenge involves the underutilization of ATSPM data due to its massive volume, which can overwhelm traditional identification and prioritization methods. Second, data anomalies frequently occur in ATSPM datasets, creating a significant issue that can undermine data reliability and accuracy. Last, discrepancies often arise between the traffic volumes recorded by ATSPM and the actual traffic patterns observed at intersections, impairing the decision-making capabilities of transportation agencies. This dissertation presents a combination of statistical and machine learning approaches to address these challenges. First, a method to summarize intersection- and corridor-level performance using ATSPM data is introduced, facilitating the prioritization of further analysis. Performance measures, including platoon ratio, split failures, arrivals on green, and red-light violations, are employed to establish threshold values. Through the application of k-means cluster analysis and expert input, scores are generated for each intersection and corridor, providing valuable information for decision-making and future studies. Second, a method to automatically detect data anomalies in ATSPM datasets is developed. This method identifies anomalies when the z-score, calculated using moving averages and standard deviations, exceeds a predefined threshold. It enhances the usability of ATSPM data by empowering engineers, planners, and other users to address and rectify anomalies, leading to more reliable and accurate data analysis. Third, a Long Short-Term Memory (LSTM) model, enhanced by additional features such as weather conditions, crash data, and holidays, is incorporated to predict traffic volumes and overcome data anomalies in ATSPM datasets. A comparative analysis involving two statistical methods and three other machine learning algorithms reveals that the LSTM model outperforms in predicting traffic volume, as evidenced by lower root mean square error and mean absolute percentage error. The application of the LSTM model allows transportation agencies to enhance the quality or usefulness of their ATSPM data, facilitating more informed decision-making in traffic operations. The methods presented in this dissertation contribute to the robustness and usability of ATSPM datasets, benefiting traffic engineers, planners, and all ATSPM users in their future analyses. By addressing the challenges associated with ATSPM data, this dissertation enables more informed transportation planning, improved traffic operations, and enhanced efficiency in managing signalized intersections. The findings emphasize the importance of fully utilizing ATSPM data, paving the way for further advancements in the field and promoting more effective transportation management strategies.
Degree
PhD
College and Department
Ira A. Fulton College of Engineering; Civil and Environmental Engineering
Rights
https://lib.byu.edu/about/copyright/
BYU ScholarsArchive Citation
Wang, Bangyu, "Evaluating and Advancing Automated Traffic Signal Performance Measures: Statistical and Machine Learning Approaches" (2023). Theses and Dissertations. 10504.
https://scholarsarchive.byu.edu/etd/10504
Date Submitted
2023-08-16
Document Type
Dissertation
Handle
http://hdl.lib.byu.edu/1877/etd13342
Keywords
ATSPM, big data, k-means cluster analysis, moving average and standard deviation, LSTM, traffic signal, traffic volume prediction, machine learning algorithms
Language
english