Application of Machine Learning in Spatiotemporal Analysis of Traffic Flow Characteristics: Classification, Prediction, and Representation

Traffic data, regardless of its collection method—detector-based, probe-based, or crowdsourced approaches—fundamentally constitutes spatiotemporal information. This intrinsic characteristic underscores the complexity and dynamic variability of traffic flow across space and time. This dissertation explores the innovative application of machine learning techniques to the spatiotemporal analysis of traffic flow characteristics, offering a novel perspective on managing and understanding traffic dynamics. The dissertation consists of three interrelated chapters that collectively aim to enhance the accuracy and reliability of traffic data analysis, focusing on both truck volume estimation and short-term traffic volume forecasting under various conditions, alongside a study on representing network traffic condition.

The dissertation begins with a chapter introducing a novel approach that combines quantile regression and Light Gradient Boosting Machine to improve the accuracy of truck volume estimates from single loop detectors. This method significantly enhances the precision of truck volume estimations, providing a solid foundation for freight management and planning from a microscopic perspective. The second chapter proposes and reviews a new training and testing framework in predicting short-term traffic volume. This investigation highlights the strengths and limitations of existing research, proposing improvements that could lead to more reliable traffic forecasts. The final chapter presents a case study on the application of matrix decomposition to infer and represent traffic conditions across a traffic network by focusing on a group of key detectors, which contributes to better traffic management and operations.

Overall, this dissertation contributes to the field of data science and traffic engineering by leveraging machine learning for the spatiotemporal analysis of traffic flow, offering new frameworks for spatiotemporal analysis including truck estimation, volume prediction, and network condition representation. Its findings have significant implications for improving traffic management and operation with advanced data-driven methods, particularly in the face of growing urbanization and the increasing complexity of transportation networks.