Reinforcement Learning for Vehicle Air Conditioning Systems: Dynamic Optimization of Energy Efficiency and Occupant Thermal Comfort

To address the challenges of high en-ergy consumption and insufficient temperature con-trol precision in electric vehicle air conditioning (AC) systems, this study proposes a reinforcement learning control method based on Recurrent Proximal Policy Optimization (Recurrent PPO). A dynamic thermal balance model for the cabin is established, incorpo-rating heat conduction, radiation, occupant metabolic heat, and ventilation loads. The control framework is formulated as a Markov Decision Process (MDP) with continuous action space (AC cooling capacity) and multidimensional state information, including ambient temperature, solar radiation, and energy consumption. The Recurrent PPO algorithm integrates LSTM net-works (Long Short-Term Memory) to capture temporal features of historical states, combined with policy clipping and generalized advantage estimation (GAE), to achieve multi-objective optimization for AC systems. Simulation results demonstrate that compared to Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) algorithms, Recurrent PPO reduces the av-erage absolute temperature error by 17.6%-65.2% and energy consumption by 0.35%-0.51 % under high-temperature cooling conditions. Additionally, it exhibits strong robustness in scenarios with unknown passenger numbers and initial temperature fluctuations. This research validates the potential of reinforcement learning (RL) in enhancing energy efficiency and thermal comfort for electric vehicle AC systems, providing novel insights for intelligent AC system design.