基于大语言模型的水库调度知识图谱智能构建

冯仲恺; 林腾; 牛文静; 肖洋; 杨涛; 唐洪武

文章摘要

冯仲恺,林腾,牛文静,肖洋,杨涛,唐洪武.基于大语言模型的水库调度知识图谱智能构建[J].水利学报,2025,56(12):1556-1569

基于大语言模型的水库调度知识图谱智能构建

Intelligent construction of reservoir operation knowledge graphs based on large language models

投稿时间：2025-04-28 修订日期：2025-12-24

DOI：10.13243/j.cnki.slxb.20250261

中文关键词: 水库调度知识图谱大语言模型知识抽取知识驱动

英文关键词: reservoir operation knowledge graph large language model knowledge extraction knowledge-driven

基金项目:国家自然科学基金项目（52379009，52441901，U2240209）；江苏省自然科学基金优秀青年基金项目（BK20240189）；北京江河水利发展基金会—水利青年科技英才资助项目（JHYC202310）；水灾害防御全国重点实验室自主研究项目（5240152E2）；江苏省科技智库计划项目（JSKX0225047）

作者	单位
冯仲恺	河海大学水灾害防御全国重点实验室, 江苏南京 210098 河海大学水文水资源学院, 江苏南京 210098 河海大学洪涝灾害风险预警与防控应急管理部重点实验室, 江苏南京 210098
林腾	河海大学水灾害防御全国重点实验室, 江苏南京 210098 河海大学水文水资源学院, 江苏南京 210098 河海大学洪涝灾害风险预警与防控应急管理部重点实验室, 江苏南京 210098
牛文静	长江水利委员会长江水文局, 湖北武汉 430010
肖洋	苏州科技大学环境科学与工程学院, 江苏苏州 215009
杨涛	河海大学水灾害防御全国重点实验室, 江苏南京 210098 河海大学水文水资源学院, 江苏南京 210098 河海大学洪涝灾害风险预警与防控应急管理部重点实验室, 江苏南京 210098
唐洪武	河海大学水利水电学院, 江苏南京 210098

摘要点击次数: 27

全文下载次数: 22

中文摘要:

受气候变化与人类活动双重影响，水库调度面临多源异构数据激增、知识关联性弱化等挑战，传统知识管理模式难以支撑精细化决策需求。针对水库调度知识碎片化、语义关联缺失等问题，本文提出融合大语言模型（LLM）与深度学习技术的知识图谱智能构建方法。首先，构建覆盖水文要素、工程属性、调度方法、约束条件及优化模型的多维知识体系，设计面向非结构化知识的知识抽取框架，采用动态编码文本语义特征，通过BertBiLSTM-CRF模型实现水库调度实体边界识别，结合注意力机制提升调度方法、调度模型等专业实体抽取效果；其次，提出基于语义角色标注与依存句法分析的关系抽取策略，建立水库调度知识冲突消解规则，解决跨文献实体对齐难题。基于国内核心期刊文献等资料构建的水库调度知识图谱包含1590个实体和922组关系，实体识别准确率、召回率分别达97.38%和97.96%，F₁值较传统BiLSTM-CRF与BiLSTM-CNN模型分别提升13.29%与13.02%。应用表明，知识图谱可充分展现水库调度知识的拓扑关联，支持调度知识推理、知识问答等应用，可为流域水库群智能调度提供可扩展的知识中枢，其构建范式对水利数字孪生体系建设具有理论借鉴意义。

英文摘要:

Under the dual influence of climate change and human activities，reservoir operation faces challenges such as the proliferation of multi-source heterogeneous data and the weakening of knowledge correlations，making traditional knowledge management models inadequate for supporting refined decision-making. To address the issues of knowledge fragmentation and lack of semantic associations in reservoir operation，this paper proposes an intelligent method for constructing knowledge graphs by integrating Large Language Models（LLMs）and deep learning techniques. Firstly，a multi-dimensional knowledge system covering hydrological elements，engineering attributes，operation methods，constraints，and optimization models is constructed. A knowledge extraction framework for unstructured texts is designed，which employs dynamic encoding of textual semantic features. The Bert-BiLSTM-CRF model is utilized to identify entity boundaries in reservoir operation texts，and an attention mechanism is incorporated to enhance the extraction of specialized entities such as operation methods and models. Secondly，a relation extraction strategy based on semantic role labeling and dependency parsing is proposed. Conflict resolution rules for reservoir operation knowledge are established to address the challenge of cross-literature entity alignment. A reservoir operation knowledge graph，built using materials from core domestic journal literature and other sources，contains 1，590 entities and 922 relation groups. The entity recognition accuracy and recall rates reach 97.38% and 97.96%，respectively，with the F₁ score improving by 13.29% and 13.02% compared to the traditional BiLSTM-CRF and BiLSTMCNN models. Application results demonstrate that the knowledge graph effectively reveals the topological associations within reservoir operation knowledge，supporting applications such as knowledge reasoning and question answering. It can serve as a scalable knowledge hub for the intelligent operation of river basin reservoir groups. The construction paradigm offers theoretical reference for the development of digital twin systems in water resources.

查看全文查看/发表评论下载PDF阅读器

关闭