基于策略可靠性评估的燃料电池汽车DRL-ECMS能量管理方法研究

Deep reinforcement learning-based energy management strategies （EMSs） often suffer from reduced reliability and lack of interpretability under unfamiliar driving conditions， which can negatively impact optimization performance. To address this issue， taking fuel cell vehicles as the research object， an energy management method based on policy reliability assessment is proposed. An offline training framework that integrates intelligent learning with model-based optimization is developed， where the soft actor-critic （SAC） algorithm is employed to dynamically determine the optimal equivalent factor for the equivalent consumption minimization strategy （ECMS）. A policy reliability assessment mechanism is designed， utilizing an ensemble policy network model to quantitatively evaluate the reliability of EMS decisions. Besides， based on the predefined reliability thresholds， the equivalent factor is corrected online to ensure the interpretability of safety-critical action. The experimental simulation results show that compared to standard SAC and conventional adaptive ECMS， the proposed SAC-ECMS training framework improves fuel economy by 4.32% and 7.82%， respectively. Under unfamiliar test cycles， the reliability assessment mechanism further enhances fuel economy by 2.87% while demonstrating real-time computational performance.