基于模拟退火法与多层感知机的变压器故障诊断模型及其泛化性能研究  

Fault Diagnosis Model of Transformer and Its Generalization Performance Based on Simulated Annealing and Multilayer Perceptron

作　　者：高超 王志武 冯玉辉 杜预[1] 宋兵 高二亚 李乾 饶召伟 邹国平[2] 杨仕友[2] GAO Chao;WANG Zhiwu;FENG Yuhui;DU Yu;SONG Bing;GAO Erya;LI Qian;RAO Shaowei;ZOU Guoping;YANG Shiyou(Suzhou Nuclear Power Research Institute,Jiangsu Suzhou 215004,China;College of Electrical Engineering,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]苏州热工研究院,江苏苏州215004 [2]浙江大学电气工程学院,杭州310027

出　　处：《高压电器》2024年第11期77-85,共9页High Voltage Apparatus

基　　金：中广核智能核电战略专项(R-2020SZEM21TF)。

摘　　要：为诊断电力变压器内部的潜伏性故障,以溶解气体分析(DGA)数据为特征量,提出了一种基于多层感知机(MLP)的变压器故障诊断模型。以实际运行变压器的故障数据为学习样本,利用模拟退火法实现多层感知机内部节点之间的连接权重优化。以不同特征组合作为MLP的输入,对比、分析了MLP诊断故障类型的正确率;研究了MLP拓扑结构、参数正则化等对诊断模型泛化性能的影响。使用训练数据以外的变压器故障数据测试学习完成的诊断模型,获得较高的测试准确率。For diagnosing potential internal fault of power transformer,a kind of fault diagnosis model of transformer based on a multilayer perceptron(MLP)is proposed with the dissolved gas analysis(DGA)as the characteristic quantity.The fault data of transformer in actual operation is taken as the learning sample,the simulated annealing method is used to achieve connection weight optimization between the internal nodes of the MLP.The different characteristic combination is taken as the input of the MLP,the accuracy rate of the MLP diagnosis fault type is compared and analyzed.The influence of topology of the MLP and the regularization of parameters on the generalization performance of the diagnosis model is studied.The data of diagnosis model other than the training data is used to test the learned diagnosis model,and a higher diagnosis accuracy is obtained.

关 键 词：人工神经网络 多层感知机 模拟退火 DGA 故障诊断 

分 类 号：TM41[电气工程—电器]