Abstract
Energy storage systems (ESSs) are becoming critical flexibility resources in renewable-rich integrated energy systems (IESs), where demand response (DR) must support renewable energy source (RES) variability, network feasibility, and market-oriented operation. However, ESS-supported DR faces a practical but underexplored inconsistency: the flexibility physically delivered by storage assets may differ from the delivery observed, measured, and settled by the market when uncertainty and cyber-communication disturbances occur together. RES forecast errors, load and price deviations, false data injection, denial-of-service events, packet loss, and latency can jointly distort ESS dispatch, metered DR delivery, imbalance exposure, and settlement reliability. To address this problem, this paper proposes a storage-centric, uncertainty-aware, and cyber-resilient DR framework for RES-rich IESs. The proposed model connects ESS dynamics, multi-carrier operational constraints, DR commitments, cyber-degraded observability, metered delivery, and risk-aware settlement within a unified operational chain. A large language model-guided hierarchical multi-agent deep reinforcement learning architecture coordinates distributed ESS-backed DR decisions, while a constrained action repair and projection layer converts raw control actions into feasible dispatch commands satisfying storage, network, comfort, ramping, and power-quality limits before execution. The framework is evaluated on the IEEE 69-bus system, the Caracas 141-bus feeder, and the IEEE 342-node system under nominal, uncertainty-only, attack-only, and combined uncertainty-cyber regimes. In the combined regime, the proposed method maintains 93.8–96.4% DR delivery with only 1.2–2.3% violation rates and consistently outperforms the benchmark methods in profit, constraint satisfaction, and CVaR0.95, confirming its ability to preserve feasible and economically reliable ESS-supported DR under degraded real-time conditions.