Abstract
<jats:p>This paper addresses a scientific and applied problem of ensuring the energy resilience of critical medical infrastructure operating under conditions of extreme resource scarcity and instability of external power supply caused by deliberate destruction of the energy system. Unlike conventional Building Energy Management Systems (BEMS), which are primarily focused on economic efficiency in stable environments, the proposed approach shifts the focus toward maximizing the autonomous operation time of life-support equipment.The methodological foundation of the study is the development of a hybrid decentralized model for adaptive energy resource prioritization implemented within an Edge–Fog Computing architecture. This architectural design ensures full system autonomy: TinyML models deployed at the Edge level provide real-time anomaly detection, while strategic decision-making is performed by a Deep Reinforcement Learning (DRL) agent at the Fog node level, without reliance on cloud services. This eliminates a critical single point of failure and significantly enhances data privacy.The scientific novelty of the work lies in the introduction of an ethically oriented reward function for a Deep Q-Network (DQN) agent, which mathematically formalizes the prioritization of the bioethical principles of Beneficence and Non-Maleficence through the integration of risk categories defined by the NFPA 99 standard. The model employs a hard prioritization mechanism in which the unconditional supply of critical loads prevails over patient comfort and operational costs, while also accounting for physical system constraints, such as minimum battery charge levels, by incorporating elements of Constrained Reinforcement Learning.The proposed model was validated in the simulation environment HospitalEnergyEnv using the “Blackout-48” scenario. Comparative simulation results demonstrated that the proposed DQN agent maintained 100% operational availability of critical medical equipment throughout 48 hours of autonomous operation by proactively switching to a deep energy-saving mode. In contrast, traditional Greedy and Rule-Based algorithms depleted available resources at the 32nd and 41st hours, respectively, resulting in emergency power loss in the intensive care unit.The practical significance of the obtained results lies in the development of a deployment-ready architecture for autonomous hospital energy systems capable of adaptively allocating limited resources during prolonged blackouts while ensuring patient safety in accordance with international standards.</jats:p>