Safe reinforcement learning-based energy management for fuel cell hybrid electric aircraft with longevity considerations

Abstract

Fuel Cell Hybrid Electric Aircraft (FCHEA) represent a promising solution for decarbonizing short-to medium-range aviation. However, the hybrid-electric architecture introduces increased control complexity and poses challenges in ensuring component longevity and operational safety. Although reinforcement learning (RL)-based energy management strategies (EMS) have been explored in ground vehicle application, they often prioritize fuel efficiency while neglecting component degradation and safety-critical constraints, both of which are vital for the reliability of electric aviation. This study presents a Longevity-Conscious Safe Energy Management Strategy (LC-SEMS) to minimize operational and degradation-related costs over long-term use, while ensuring the satisfaction of multi-type constraint. The strategy is implemented within a multidisciplinary simulation framework that integrates propulsion, aerodynamics, hybrid powertrain, and flight dynamics models for mission-level evaluation. The EMS problem is formulated as a Constrained Markov Decision Process (CMDP) incorporating physical, cumulative, and instantaneous constraints. Instantaneous safety is enforced via an adaptive shielding mechanism that leverages a pretrained transition model to detect potential constraint violations and applies minimal corrective actions without interfering with policy learning. The proposed strategy is validated on a simulated FCHEA retrofitted from the NASA X-57 Maxwell, achieving fast convergence and strict constraint adherence across multi-mission scenarios. It achieves a 26.96 % reduction in depreciation cost compared to baseline RL-based EMS, with a minimal 4.21 % performance gap relative to the globally optimal Dynamic Programming (DP) benchmark, demonstrating its adaptability and robustness under uncertain and unseen mission scenarios.