Deep Reinforcement Learning-Based Multi-Source Energy Management for Hypersonic Vehicle Mode Transition

作者：Yiqun Wang, Xingjian Jin, Lei Yu, Si Gao, Jian Lu, Jiecheng Fu · 年份：2025 · DOI：10.1109/aaac66612.2025.11427777 · 研究领域：Traffic control and management、Model Reduction and Neural Networks、Plasma and Flow Control in Aerodynamics

During the mode transition phase of hypersonic vehicles, significant changes in the propulsion system's state lead to instability in the electric energy supply, necessitating the development of a multi-source, adaptive, and high-energy-efficiency power generation solution. This paper proposes a multi-source energy management strategy based on DRL (Deep Reinforcement Learning) and the DQMR-DDPG (Double Q-learning with Multi Replay Buffer for DDPG) algorithm, aimed at minimizing fuel consumption during power generation while meeting system performance and safety constraints. The improved DQMR-DDPG algorithm is proposed to solve the optimal multi-source energy allocation strategy. By introducing innovations such as state-partitioned multi-stage experience replay, uncertainty-weighted dual Q-network fusion, and adaptive gradient soft update strategies, the algorithm effectively addresses issues such as poor strategy adaptability, unstable value estimation, and inaccurate convergence under strong nonlinear changes in system efficiency. Finally, simulations validate the effectiveness of the proposed multi-source energy management strategy.