首页|期刊导航|电力信息与通信技术|调度推理知识驱动的电力调度大模型训练微调框架

调度推理知识驱动的电力调度大模型训练微调框架OA

A Training and Fine-tuning Framework for Power Grid Dispatching Large Language Models Driven by Dispatching Reasoning Knowledge

中文摘要英文摘要

为解决新型电力系统中电力调度问答所面临的领域推理数据稀缺及模型领域适配性不足等问题,文章提出一种基于电力调度推理知识驱动的大模型训练微调框架.首先,提出规则-生成相结合的电力调度问答数据合成方法,通过规则引导与生成蒸馏的策略,构建覆盖关键调度场景的高质量领域推理数据集;其次,设计面向多类调度需求的多专家智能体奖励模型,实现对模型生成内容的多维度综合评估;最后,采用结合监督微调与强化学习的两阶段训练策略,基于主流基座模型完成电力调度领域的深度适配.实验结果表明,本方法在专用测试集上的核心评估指标表现优异,显著优于当前主流闭源及开源模型,有效提升了电力调度问答系统的逻辑准确性与场景实用性,为电网安全稳定运行提供了关键技术支持.

To address the problems of scarce domain reasoning data and insufficient model domain adaptability faced by power dispatching question answering in new-type power systems,this paper proposes a large model training and fine-tuning framework driven by power dispatching reasoning knowledge.First,a rule-generation combined data synthesis method for power dispatching question answering is proposed.By means of rule-guided and generative distillation strategies,a high-quality domain reasoning dataset covering key dispatching scenarios is constructed.Second,a multi-expert agent reward model oriented to multiple types of dispatching demands is designed to realize multi-dimensional comprehensive evaluation of model-generated content.Finally,a two-stage training strategy combining supervised fine-tuning and reinforcement learning is adopted to achieve in-depth domain adaptation for power dispatching based on mainstream foundation models.Experimental results show that the proposed method achieves excellent performance in core evaluation metrics on the dedicated test set,significantly outperforming current mainstream closed-source and open-source models.It effectively improves the logical accuracy and scenario practicality of the power dispatching question answering system,providing key technical support for the safe and stable operation of power grids.

姚铁锤;赵如意;张文思;常珂;柴博;王轶申

中国电力科学研究院有限公司,北京市 海淀区 100192中国电力科学研究院有限公司,北京市 海淀区 100192中国电力科学研究院有限公司,北京市 海淀区 100192中国电力科学研究院有限公司,北京市 海淀区 100192中国电力科学研究院有限公司,北京市 海淀区 100192中国电力科学研究院有限公司,北京市 海淀区 100192

信息技术与安全科学

电力调度知识问答推理驱动训练两阶段微调

power grid dispatching knowledge QAreasoning-driven trainingtwo-stage fine-tuning

《电力信息与通信技术》 2026 (8)

36-44,9

国家电网有限公司总部科技项目"面向电网调度智能辅助的大模型关键技术研究"(5108-202455044A-1-1-ZN).

10.16543/j.2095-641x.electric.power.ict.2026.08.04

评论