OPTIMAL CONTROL FOR SELF-DRIVING PATH PLANNING AND THERMAL MANAGEMENT

Abstract

In the first part of this dissertation, a novel local path planning approach for autonomous vehicles (AVs) is proposed, based on Lyapunov Optimization (LO). The performance of the developed LO-based planner is benchmarked against two established techniques: Model Predictive Control (MPC) and a sampling-based method. To support real-world validation, an AV platform was developed using a modified 2015 Nissan Leaf S. The vehicle was equipped with a drive-by-wire system, perception sensors, and onboard computation units. Data collection was carried out in Norman, Oklahoma, enabling a comparative evaluation of the three control algorithms under real traffic conditions, including both straight and curved urban roads. In order to reduce system cost and enhance deployment feasibility, a vision-only perception solution was adopted for object detection and bird’s-eye view coordinate generation. Using the collected data, each method—LO, MPC, and the sampling-based approach—was independently deployed to generate safe and smooth driving paths. The comparison focused on key performance metrics: path smoothness, safety, and computation time. Results show that the proposed LO method significantly outperforms the baseline approaches, achieving up to a 97.5% less computation time. Additionally, the LO approach was tested in the high-fidelity CARLA simulation environment, providing further validation prior to real-world deployment. It is worth noting that, during the initial phase of this research, MPC and the sampling-based method were directly compared through real-world experiments. Findings revealed that MPC consistently outperforms the sampling-based method, particularly in terms of speed and safety. In the second part of the dissertation, a new optimization strategy for Battery Thermal Management Systems (BTMS) in electric vehicles (EVs) is introduced. The proposed method employs a model-free deep reinforcement learning (RL) framework designed to operate under extreme fast charging (XFC) conditions. The optimization objective is to minimize both battery degradation and the power consumption of the BTMS. The thermal dynamics of the system are modeled in detail, accounting for the air-conditioning refrigerant loop and the indirect liquid cooling loop for battery thermal regulation. The proposed RL method is implemented on a real battery pack and compared against two benchmark strategies: MPC, representing optimal control, and Proportional-Integral-Derivative (PID) control, representing traditional tracking control. The results show that, when provided with an accurate system model, MPC can perform comparably to the proposed RL approach. However, the RL method demonstrates a significant advantage in computational efficiency—operating up to 48 times faster than MPC during testing. This performance gain is attributed to the model-free nature of the RL algorithm, which eliminates the need for complex system modeling. Moreover, the RL approach outperforms PID control in both key metrics: battery degradation is reduced by up to 1.05%, and BTMS power consumption is lowered by up to 43.68%.