基于频域介电谱和多输出支持向量回归的变压器油纸绝缘状态评估

Transformer Oil-paper Insulation Condition Assessment Based on Frequency Domain Dielectric Spectroscopy and Multi-output Support Vector Regression

为分别评估变压器油纸绝缘的老化程度和水分含量,文中通过加速热老化试验和吸潮实验制备了不同老化程度下含水量不同的油浸绝缘纸试品,测试了试品的介损因数频谱和阻抗相位频谱,分析了老化、水分对油纸绝缘频域介电谱特性的影响。采用多输出支持向量回归算法(multi-output support vectorregression,M-SVR)逼近频域介电谱曲线与油纸绝缘老化和水分之间的非线性映射关系,据此预测油纸绝缘试品的老化程度和水分含量。研究表明,对于含水量在0.5～5.6%范围内、具有不同老化程度的试品,M-SVR对含水量的预测精度较高,而对老化程度的预测精度较低,且均高于RBF神经网络;根据油纸绝缘试品的聚类分析结果,发现与老化程度相关性较大的第1、2类试品,M-SVR对其老化程度的预测精度明显提高,含水量预测精度变化不大。

In order to respectively assess the aging degree and moisture content of transformer oil-paper insulation,this article prepares oil impregnated insulation paper samples with different moisture content and aging degree through accelerated thermal aging test and moisture absorption experiment,then tests the dielectric loss factor spectrum and the argument of geometric impedance spectrum of different samples,and analyzes the influence of moisture and aging to frequency domain dielectric spectroscopy.Finally,adopted multi-output support vector regression algorithm(M-SVR)to approximate the nonlinear relationship between frequency dielectric spectrum curve moisture content and aging degree,then accordingly predict the aging degree and moisture content of the samples.Research results show that for different aged oil-paper samples with moisture from 0.5 to 5.6%,the prediction accuracy of moisture content is relatively high,and the prediction accuracy for aging degree is not acceptable,but both of the accuracy is higher than RBF neural network.According to the clustering analysis result to all samples,the study finds that for the samples belonging to class 1 and 2 which is more correlative to aging,the aging prediction accuracy by M-SVR increase obviously,while the moisture prediction accuracy changed a little.