面向电网两票的知识增强智能审核方法

A knowledge-enhanced intelligent auditing method for work and operation tickets in power grids

摘要:

电网工作票与操作票（简称“两票”）的审核质量直接关系电力安全生产，现有审核方式存在标准不统一、问题识别不充分、判定依据不足等问题。为此，提出了一种融合规则约束与检索增强生成（retrieval-augmented generation，RAG）的电网两票智能审核方法：对两票审核任务进行范式化建模，构建覆盖规程条款、票据模板、设备图谱、资质安措和历史案例的多源知识体系，设计“票据解析—审核项识别—多源知识召回—联合判定—缺陷定位—建议生成”的完整审核链路，并以规则优先与冲突仲裁机制兼顾审核的稳定性与适应性。在 10 000 条脱敏两票样本上的实验结果表明，该方法的票据级准确率、问题级 F1 和检索命中率 Hit@k 分别达到 0.876 7、0.779 4 和 0.719 2，均优于规则、大语言模型和检索增强各基线，验证了规则约束与多源知识融合对提升审核准确性与证据支撑能力的有效性。

Abstract:

The auditing quality of power grid work tickets and operation tickets (referred to as the “two tickets”) is directly related to the safety of power production, while existing auditing approaches suffer from inconsistent criteria, insufficient problem identifica-tion and inadequate decision basis. To address these problems, an intelligent two-ticket auditing method integrating rule con-straints and retrieval-augmented generation (RAG) was proposed. The two-ticket auditing task was firstly modeled in a para-digmatic manner, and a multi-source knowledge system covering regulation clauses, ticket templates, equipment graphs, qualifi-cation and safety-measure knowledge and historical cases was constructed. Then a complete auditing pipeline of “ticket parsing–audit item identification–multi-source knowledge retrieval–joint decision–defect localization–suggestion generation” was de-signed, in which a rule-priority and conflict-arbitration mechanism was adopted to balance the stability and adaptability of audit-ing. Experimental results on 10 000 desensitized two-ticket samples showed that the proposed method achieved a ticket-level accuracy of 0.876 7, an issue-level F1 of 0.779 4 and a retrieval hit rate (Hit@k) of 0.719 2, outperforming the rule-based, large language model and retrieval-augmented baselines. The results verify the effectiveness of integrating rule constraints and multi-source knowledge fusion in improving auditing accuracy and evidence support.