Answers are grounded in this paper and cite their sources.
Demand response (DR) in renewable- and storage-rich integrated energy systems (IESs) must remain economically reliable and physically executable despite renewable and load uncertainty, volatile prices, and cyber disruptions affecting sensing and communication. Existing approaches commonly separate market scheduling, distributed control, cyber resilience, and feasibility enforcement, creating a gap between day-ahead commitments and real-time delivery. This paper proposes a unified market-to-execution framework based on large language model (LLM)-guided multi-agent deep reinforcement learning (MADRL). The framework coordinates electrical and thermal loads, renewable generation, battery and thermal storage, and multi-energy conversion resources through cyber- and uncertainty-aware observations, adversarial and domain-randomized training, and conditional value-at-risk regularization. The LLM is used only as a bounded supervisory layer that interprets structured operating conditions, prioritizes inter-agent coordination, and supports repair decisions; executable actions remain generated by the MADRL agents and are projected onto a feasible operating region defined by device, storage, comfort, network, and power-quality constraints. The framework is assessed under nominal, uncertainty-only, attack-only, and combined-stress regimes involving false-data injection, denial-of-service, and communication delay. Under combined stress, DR delivery remains within 93.8–96.4%, while operational violation rates remain within 1.2–2.3%. The results further show progressive rather than abrupt performance degradation, stronger economic and downside-risk performance than advanced learning-based baselines, and adaptive reallocation of flexibility toward more controllable storage-supported resources as stress increases. The main finding is that reliable market delivery under cyber-operational stress is achieved through the coordinated interaction of risk-aware distributed learning, bounded supervisory guidance, and feasibility-preserving execution rather than through any single control component.
No takes yet. Share an insight, caveat, or question.
Sepehrzad et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: