LLM and DRL Collaborative Optimization for Flexible Resource Scheduling in Integrated Energy Systems

Increasing renewable energy penetration imposes stringent requirements on flexible resource scheduling in Integrated Energy Systems (IES), yet conventional methods face significant challenges in computational efficiency and modeling complexity under multi-energy coupling, high-dimensional uncertainties, and multi-constraint operations. Here we propose a novel scheduling strategy that synergistically integrates Large Language Models (LLM) and Deep Reinforcement Learning (DRL) to address these challenges. Specifically, LLM leverages few-shot learning for intelligent reasoning of Battery Energy Storage System decisions with a three-layer Safety-Guard ensuring constraint satisfaction, while DRL optimizes Combined Heat and Power units and Gas Boilers through enhanced state space incorporating executed battery actions. This hierarchical collaboration effectively reduces action dimensionality while maintaining responsiveness to dynamic disturbances. Comprehensive simulations demonstrate substantial improvements in battery responsiveness to time-of-use tariffs and renewable fluctuations, achieving significant reductions in operating costs and cycling intensity. Multi-scenario tests validate robust performance under high-intensity perturbations and fault injections with zero constraint violations. This approach provides a promising pathway for intelligent multi-energy coordinated scheduling in IES.