LLM-powered distributed optimal scheduling for industrial heat-electricity micro-grids

Flexible loads, such as thermal loads and electric vehicles (EVs), act as Virtual Energy Storage Sources (VESS) in microgrid scheduling to enhance economic performance via load shifting. Due to centralized optimization methods face scalability and privacy limitations in flexible loads scheduling, distributed methods like Alternating Direction Method of Multipliers (ADMM) provide scalable alternatives. However, performance of ADMM is highly sensitive to the penalty parameter, which must be tailored to the problem's intrinsic characteristics. This dependency necessitates complex and labor-intensive manual tuning. To address this issue, this paper proposed a novel parameter tuning mechanism that leverages large language models (LLM). Under a domain-specific crafted prompt, LLM dynamically refines the penalty parameter to improve convergence efficiency and adaptability, reducing the reliance on expert knowledge and manual intervention. Moreover, a high-resolution temperature constraint is established through Predicted Mean Vote (PMV) method to ensure user thermal comfort during scheduling interval. Simulation results demonstrate: 1) Compared to fixed-parameter and heuristic-based tuning methods, the LLM-assisted parameter tuning method accelerates ADMM convergence by up to 50% (with average improvements exceeding 30%), 2) the method's effectiveness is validated across diverse LLMs, and its high performance is critically dependent on trial-and-error search process, and 3) the integrated indoor temperature constraint ensures thermal comfort compliance while reducing microgrid operational costs through VESS-enabled load shifting and price optimization.