基于多参量LSTM模型的变压器油纸绝缘老化诊断方法  

Diagnosis method of oil-paper insulation aging in transformer based on multi-parameter LSTM model

作　　者：鄢仁武[1] 高硕勋 宋微浪 林穿 钟伦贵[1] 罗家满 YAN Renwu;GAO Shuoxun;SONG Weilang;LIN Chuan;ZHONG Lungui;LUO Jiaman(Fujian Colleges and Universities Engineering Research Center of Smart Grid Simulation&Analysis and Integrated Control,Fujian University of Technology,Fuzhou 350118,China;State Grid Fujian Electric Power Company Limited Maintenance Branch Company,Fuzhou 350013,China)

机构地区：[1]福建工程学院智能电网仿真分析与综合控制福建省高校工程研究中心,福建福州350118 [2]国网福建省电力有限公司检修分公司,福建福州350013

出　　处：《武汉大学学报（工学版）》2021年第6期541-550,共10页Engineering Journal of Wuhan University

基　　金：福建省自然科学基金项目(编号:2018H0003,2017J01731);福建省教育厅科技项目(编号:JT180339,JAT171096);福建工程学院基金项目(编号:GY-Z11002)。

摘　　要：针对单一特征参量进行油纸绝缘老化诊断容易造成诊断结果偏差和片面性的问题,提出了一种基于多参量长短时记忆(long short-term memory,LSTM)深度学习模型的油纸绝缘老化诊断方法。首先分析变压器油中糠醛与聚合度的相关性,根据糠醛含量和聚合度的良好线性关系建立油纸绝缘老化分级策略;其次深入分析多个时域特征量与油中糠醛的关系,利用现场获取的变压器绝缘测试数据,建立油中糠醛含量与时域特征参量组成的5维评价体系,并对模型进行优化,构建多参量LSTM神经网络模型;最后通过现场实例验证该评价体系的准确性和可靠性。该方法将深度学习与故障诊断方法相结合,可为变压器绝缘老化诊断提供一种新的思路。Most of the oil-paper insulation aging diagnosis is based on single characteristic parameter analysis,which is easy to cause deviations and one-sided results in the diagnosis.This paper proposed a diagnosis method of oil-paper insulation aging based on long short-term memory(LSTM)deep learning model to solve the problem.Firstly,the correlation between furfural content and degree of polymerization in transformer oil was analyzed,and an oil-paper insulation aging grading strategy was established based on the good linear relationship between furfural and polymerization degree.Secondly,the relationship between multiple timedomain characteristic parameters and furfural in oil was deeply analyzed.A 5-dimensional evaluation system composed of furfural concentration and time-domain characteristic parameters was established by using the transformer insulation test data obtained in the field,and the model was optimized to build a multi-parameter LSTM neural network model.Finally,the accuracy and reliability of the evaluation system were verified by field examples.The method proposed in this paper combines deep learning with fault diagnosis methods to provide a new idea for transformer insulation aging diagnosis.

关 键 词：油纸绝缘 老化诊断 LSTM 时域特征量 糠醛含量 

分 类 号：TM83[电气工程—高电压与绝缘技术]