Multi-agent collaborative optimization framework for integrated energy system scheduling

Integrated energy system (IES) scheduling optimization involves multi-energy coupling, multidevice coordination, and complex constraints. Conventional approaches rely on experts to manually formulate optimization models, resulting in high technical barriers, lengthy development cycles, and limited transferability. To address requirement uncertainty, attention drift in long contexts, and high debugging costs when applying large language models (LLMs) to IES scheduling optimization, this study proposes the MASEO, a multi-agent system for energy optimization. MASEO decomposes the optimization workflow into four sequential stages—information collection, mathematical modeling, code implementation, and execution verification—each handled by a dedicated LLM agent. A structured checklist is introduced to standardize information collection, while an error-classification-based traceable repair mechanism routes errors to the responsible agent for targeted correction. Case studies on a Beijing data center IES and a German community IES show that MASEO achieves results highly consistent with expert benchmarks, with an annualized cost deviation of approximately 1.4% in the Beijing case. Compared with a single-agent baseline, modeling accuracy improves by approximately 35%, while the code execution pass rate increases from 63% to 94%. Ablation, multi-backbone, and sensitivity experiments further validate the effectiveness of the core mechanisms and the applicability of the framework. These results indicate that MASEO can reduce reliance on specialized modeling expertise and provide a practical pathway for the reliable application of LLMs to energy system optimization.