Energy supervisory control (ESC) for electric flying vehicles is challenged by strong air–ground mode coupling and stringent safety constraints. Conventional reinforcement learning–based ESC typically requires extensive training and are difficult to bootstrap from scratch, which often leads to inefficient exploration and delayed satisfaction of safety constraints. To address this issue, we propose a large language model–enhanced curriculum learning warm start for soft actor–critic (CL-LSAC), which incorporates expert prior knowledge from large language models. The proposed curriculum prioritizes safety constraints before fuel efficiency, guiding policy learning through a highly non-convex reward landscape. Evaluated on a high-fidelity urban air mobility mission, CL-LSAC significantly suppresses the control saturation observed in unguided baselines. Compared with static reward weighting, our method improves state-of-charge regulation accuracy by 49% and accelerates convergence by more than 15% via LLM-guided initialization, yielding data-efficient and physically consistent policies that demonstrate high real-time execution feasibility on hardware-in-the-loop platforms.
