基于数据增强和深度学习的水电站告警事件诊断  

Diagnosis method of hydropower alarm events based on data augmentation and deep learning

作　　者：孙国强[1] 章逸舟 唐杰阳 唐凡 卫志农[1] 臧海祥[1] 杨东 SUN Guoqiang;ZHANG Yizhou;TANG Jieyang;TANG Fan;WEI Zhinong;ZANG Haixiang;YANG Dong(College of Energy and Electrical Engineering,Hohai University,Nanjing 211100,China;Yalong River Hydropower Development Co.,Ltd.,Chengdu 610051,China)

机构地区：[1]河海大学能源与电气学院,江苏南京211100 [2]雅砻江流域水电开发有限公司,四川成都610051

出　　处：《电力自动化设备》2023年第8期88-95,共8页Electric Power Automation Equipment

基　　金：雅砻江流域水电开发有限公司科技项目(0023-20XJ0017)。

摘　　要：针对水电告警事件传统诊断方法存在效率低下、准确率不足等缺陷,设计了一种融合先验知识的数据增强方法和基于双向简单循环单元网络的层级注意力深度学习框架。针对水电告警规则不完善的问题,采用隐含狄利克雷分布-序列推理增强模型构建告警信号与告警特征间的映射机制;结合该水电告警先验知识提出改进隐含狄利克雷分布方法增强样本数据,最终由层级注意力模型学习样本特征并输出诊断结果。测试算例为某水电集控中心的实际告警数据,测试结果表明,所提方法可在低资源训练环境下实现快速和高准确率的水电告警事件诊断。Aiming at the shortcomings of traditional diagnosis methods of hydropower alarm events,such as low efficiency and low accuracy,a data augmentation method combining prior knowledge and a hierarchical attention deep learning framework based on bidirectional simple recurrent units++(Bi-SRU++)are designed.Aiming at the problem of imperfect hydropower alarm rules,the latent Dirichlet allocation-enhanced sequential inference model(LDA-ESIM)is used to construct the mapping mechanism between warning signals and warning features.Then,combined with the prior knowledge of hydropower alarm rules,an improved LDA method is proposed to augment the sample data.The hierarchical attention model learns the sample features and outputs the diagnosis results.The test example is actual alarm data of a hydropower centralized control center.The test results show that the proposed method can realize rapid diagnosis of hydropower alarm events with high accuracy in low resource training environment.

关 键 词：水电站告警事件 文本数据增强 注意力机制 深度学习 先验知识 

分 类 号：TM73[电气工程—电力系统及自动化] TM622