Physics-Guided Machine Learning for Freeway Traffic Modelling and Control

Innovative solutions that combine predictive modeling with improved traffic control strategies are needed to address the persistent problem of traffic congestion in urban transportation networks. This dissertation introduces novel methods for improving highway traffic con dition prediction, optimization, and control by integrating machine learning methods with well-established theories of traffic flow. Physics-Regularized Gaussian Process (PR-GP) and Physics-Informed Long Short-Term Memory (PI-LSTM) network are two traffic state prediction models developed by the research. Especially in cases with multiple classes of vehicles, these models enhance prediction accuracy and interpretability by incorporating domain knowledge from traffic flow theory. As a regularization component, the PR-GP incor porates error terms from the METANET model; to guarantee physically consistent forecasts, the PI-LSTM incorporates traffic dynamics equations into training. The dissertation expands upon these models to present two control strategies to improve predictive modeling-based real-time traffic management. By anticipating traffic densities and modifying control inputs accordingly, the Localized Predictive ALINEA Controller provides proactive ramp metering. It is enhanced with traffic predictions from the Physics-Informed LSTM model. This is where the Integrated Predictive ALINEA Controller comes in; it takes the localized method and applies it to a multi-ramp setup, coordinating metering rates across several highway access points in real time. This approach improves traffic efficiency on the freeway as a whole and stops the spread of congestion by taking spatial and temporal dependencies into account. Freeway operations may be optimized in dynamic situations with the help of ramp metering and predictive traffic control, which are both based on physics and machine learning. Notable difficulties in traffic state estimation, interpretability, and control optimization are tackled by this work’s contributions. Freeway traffic management solutions that are resilient, scalable, and adaptive are provided by this research, which advances the state of the art in intelligent transportation systems by combining physics-based and data-driven approaches.