A Traffic Adapative Physics-informed Learning Control for Energy Savings of Connected and Automated Vehicles
Yunli Shao
TL;DR
This paper tackles energy-efficient control for connected and automated vehicles under nonlinear traffic dynamics by proposing a traffic-adaptive augmented state-space formulation and a physics-informed learning control (PILC) framework. The PILC combines the Bellman-value learning from dynamic programming with derivatives derived from Pontryagin’s Maximum Principle into a single loss, improving data efficiency and robustness. A key contribution is an augmented state representation that transcripts predicted trajectories of preceding vehicles via piecewise polynomials, enabling trajectory-aware control without altering the fundamental dynamics. Empirical results on a real-world, data-calibrated simulation show energy savings of about $9\%$ in unseen scenarios and comparable performance to model-based MPC, while dramatically reducing online computation.
Abstract
Model predictive control has emerged as an effective approach for real-time optimal control of connected and automated vehicles. However, nonlinear dynamics of vehicle and traffic systems make accurate modeling and real-time optimization challenging. Learning-based control offer a promising alternative, as they adapt to environment without requiring an explicit model. For learning control framework, an augmented state space system design is necessary since optimal control depends on both the ego vehicle's state and predicted states of other vehicles. This work develops a traffic adaptive augmented state space system that allows the control strategy to intelligently adapt to varying traffic conditions. This design ensures that while different vehicle trajectories alter initial conditions, the system dynamics remain independent of specific trajectories. Additionally, a physics-informed learning control framework is presented that combines value function from Bellman's equation with derivative of value functions from Pontryagin's Maximum Principle into a unified loss function. This method aims to reduce required training data and time while enhancing robustness and efficiency. The proposed control framework is applied to car-following scenarios in real-world data calibrated simulation environments. The results show that this learning control approach alleviates real-time computational requirements while achieving car-following behaviors comparable to model-based methods, resulting in 9% energy savings in scenarios not previously seen in training dataset.