随着人工智能技术的发展，大语言模型（LLM）在电力系统运维中的应用潜力巨大，但其在处理电力操作票时存在的输出不确定性高、可靠性难以量化以及领域适应性不足等问题，严重制约了其在安全关键场景下的实际应用。为解决上述挑战，实现电力机器人对操作票的精准理解与可靠执行，本文旨在提出一种融合大语言模型语义理解能力与贝叶斯推理不确定性量化优势的智能系统。本文构建了基于LLM-贝叶斯级联架构的智能理解执行系统。首先，设计了“理解-执行-反思”（UAR）三模块协同架构，通过标准化的信息流闭环实现领域适应性。其次，建立多维度置信度评估算法，从一致性检查、注意力模式分析和历史准确率统计三个维度对LLM输出进行校验。最后，构建涵盖分解、语义和策略三层次的贝叶斯不确定性量化模型，并据此设计置信度驱动的动态执行策略（自主、监督、逐步确认、专家介入）。在包含800条真实电力操作票（涵盖220kV及500kV变电站）的数据集上进行实验验证。结果表明，该系统的理解准确率达到95.8%，执行成功率为92.4%，风险识别率为94.5%。与现有纯大模型（如Qwen2.5-72B、DeepSeek-R1）及传统专家系统相比，理解准确率分别提升了6.6%以上，执行成功率提升了10.7%以上。LLM-贝叶斯级联架构通过概率建模有效弥补了大语言模型的“黑盒”缺陷，显著提升了电力操作票处理的安全性与可靠性，为电力系统智能运维提供了可行的技术支撑。

With the rapid development of artificial intelligence, Large Language Models (LLMs) show great potential in power system operation and maintenance. However, challenges such as high output uncertainty, difficulty in quantifying reliability, and insufficient domain adaptability severely restrict their application in safety-critical scenarios like power operation ticket processing. To address these challenges and achieve precise understanding and reliable execution of operation tickets by power robots, this paper proposes an intelligent system that fuses the semantic understanding capabilities of LLMs with the uncertainty quantification advantages of Bayesian inference. An intelligent understanding and execution system based on an LLM-Bayesian cascade architecture is constructed. First, a Understanding-Action-Reflection (UAR) three-module collaborative architecture is designed to enhance domain adaptability through a closed-loop information flow. Second, a multi-dimensional confidence evaluation algorithm is established to verify LLM outputs from three dimensions: consistency checking, attention pattern analysis, and historical accuracy statistics. Finally, a three-level Bayesian uncertainty quantification model covering decomposition, semantic, and strategic levels is built to support confidence-driven dynamic execution strategies (autonomous, supervised, step-by-step confirmation, and expert intervention). Experiments were conducted on a dataset of 800 real power operation tickets (covering 220kV and 500kV substations). The results demonstrate that the system achieves 95.8% understanding accuracy, 92.4% execution success rate, and 94.5% risk identification rate. Compared with existing pure LLM methods (such as Qwen2.5-72B, DeepSeek-R1) and traditional expert systems, the understanding accuracy and execution success rate are improved by over 6.6% and 10.7%, respectively. The LLM-Bayesian cascade architecture effectively compensates for the "black box" defects of LLMs through probabilistic modeling, significantly improving the safety and reliability of power operation ticket processing, and providing technical support for intelligent operation and maintenance of power systems.