Dual-layer energy management strategy combining model predictive control and deep reinforcement learning for fuel cell-battery hybrid power systems

Proton exchange membrane fuel cell (PEMFC) and lithium-ion battery (LIB) hybrid systems are critical for developing long-endurance unmanned aerial vehicles (UAV). However, frequent load fluctuations accelerate component degradation, and conventional standalone energy management strategies (EMSs) cannot simultaneously balance real‑time power allocation, hydrogen‑economy objectives, and system durability. To address convergence instability in standalone deep reinforcement learning (DRL) approaches and the model dependence in predictive control, this study proposes a novel health‑aware dual‑layer MPC‑TD3 EMS, which embeds PEMFC and LIB degradation models within an equivalent hydrogen consumption (EHC) metric. The upper-layer model predictive control (MPC) establishes a finite-horizon power baseline, while the lower-layer twin delayed deep deterministic policy gradient (TD3) agent executes real-time online corrections. Simulation results demonstrate that this hierarchical approach effectively constrains the action space, improving the cumulative reward by 10.53% compared to standalone MPC. The strategy efficiently mitigates PEMFC performance degradation and prevents LIB over-discharge. Compared with standard MPC and equivalent consumption minimization strategy (ECMS), the proposed scheme reduces total lifecycle‑oriented EHC by 6.47% and 28.99%, respectively. This dual-layer framework offers a robust solution for the collaborative optimization of operational economy and multi-source durability in UAV hybrid power systems.