融合知识图谱与检索增强生成的水利抢险预案智能生成方法

Intelligent generation method for hydraulic engineering emergency rescue plan fusing knowledge graph and retrieval-augmented generation

摘要:

针对水利工程险情应急过程中存在的历史案例查阅效率低、通用大语言模型专业知识“幻觉”严重，以及静态预案难以动态耦合实时雨、水、工情等问题，构建水利专业知识图谱与面向检索增强生成的文档向量库，提出协同混合检索算法，以精准召回专家知识；构建多模态上下文融合模块，通过雨水情数据接口实时注入当前工程水位、降雨量等感知数据，驱动大模型生成兼具历史经验借鉴与实时工情适应性的处置预案。结果表明：在安徽省典型险情案例的方案生成测试中，Top-5 知识召回准确率达 94.2%，生成预案的专家评分较传统单一检索方法提升 38.5%。研究成果突破单一模态检索的局限性，实现险情应急方案从被动查阅到主动生成的转变，可为水利应急决策提供有效支持。

Abstract:

To address the challenges in emergency response for hydraulic engineering hazards—specifi cally the low effi ciency of historical case consultation，the severe domain knowledge “hallucinations” of general-purpose models， and the diffi culty in dynamically coupling static plans with real-time rainfall，water and work conditions—this study constructed a hydraulic domain knowledge graph and a document vector database for retrieval-augmented generation. A collaborative hybrid retrieval algorithm was proposed to accurately retrieve expert knowledge. Furthermore，a multimodal context fusion module was developed to inject real-time sensing data，such as water levels and rainfall， via hydrological data interfaces. This enabled the large language model to generate emergency response plans that integrated historical experience with adaptability to real-time engineering conditions. Experimental results demonstrated that in plan generation tests based on typical hazard cases in Anhui Province，the proposed framework achieved a Top-5 knowledge retrieval accuracy of 94.2%. Additionally，expert evaluation scores for the generated plans increased by 38.5% compared to traditional single-retrieval methods. These fi ndings overcome the limitations of single-modal retrieval and realize a paradigm shift in emergency planning from passive consultation to active generation，providing eff ective support for decision-making in water conservancy emergencies.