SAC 전이학습 기법을 적용한 SDV 에너지 효율 연구 = SDV Energy Efficiency Study Using SAC Transfer Learning Techniques

다국어 초록 (Multilingual Abstract)

The need for energy efficiency improvement is increasing due to the widespread use of electric vehicles and the development of autonomous driving technology. Existing rule-based control methods have limitations in energy optimization in complex road environments. In this study, we used Soft Actor-Critic (SAC) reinforcement learning to optimize electric vehicle energy efficiency through transfer learning techniques. We transferred a pre-trained LunarLander model to the electric vehicle environment using an adapter network with a 6-stage progressive learning strategy. The simulation was conducted using real traffic data from Seoul's Gangnam-daero and Hyundai IONIQ5 specifications, achieving 22.7% energy efficiency improvement compared to cruise mode baseline. Through a 28-dimensional extended state space and multi-objective optimization reward function, we comprehensively considered traffic congestion, weather conditions, and road gradients. Comparison with existing studies validated the effectiveness of the transfer learning approach. The results of this study are expected to contribute to the development of autonomous electric vehicles or software-defined vehicle (SDV) energy management systems.