【作者】 刘帆；

【导师】 唐炬；

【作者基本信息】 重庆大学 ， 电气工程， 2013， 博士

【摘要】 现有的局部放电检测和识别方法均存在各种不足，难以满足气体绝缘组合电器（Gas Insulated Switchgear，GIS）局部放电在线监测的要求。脉冲电流法、超声波法和光测法抗干扰能力较弱，测量精度不高，且超声波法和光测法都不能进行放电量的标定，超高频法抗电磁干扰能力强，但其模式识别的性能还需继续提高才能满足现场应用的要求，其放电量标定问题也未完全解决。国内外大量研究表明，SF6气体在局部放电下会发生分解并与GIS气室中的微量水分和氧气发生反应，生成一系列化学物质，通过检测GIS气室中的SF6分解组分即可判断是否有局部放电发生，且不同类型绝缘缺陷引起的局部放电，其导致SF6分解产生的组分在含量、产气速率、含量比值等各方面会有所差异，可以通过检测SF6分解组分对放电故障类型进行识别。此外SF6气体的分解总量与放电量之间存在联系，通过检测SF6气体的分解速率及放电重复率可以大致判断放电量水平。这种局部放电检测和识别方法本质上是一种化学方法，可以避免现场强烈的电磁干扰和噪声干扰，具有较高的理论研究意义和良好的应用前景。但选取何种分解组分作为特征分解组分，如何利用组分分析来诊断故障类别，如何消除各种因素对识别的影响，以及如何利用分解速率来确定放电量，目前还没有得到解决。本文综合考虑分解组分的稳定性、易检测性以及所表征的物理意义等各方面，从SF6众多的分解组分中选取了SO2F2、SOF2、CF4和CO2作为特征分解组分，用于分析各种局部放电下SF6的分解特性。利用SF6局部放电分解实验装置，进行了包括金属突出物、自由导电微粒、绝缘子表面污秽和绝缘子气隙四种绝缘缺陷引起的局部放电下的SF6分解实验，在每种局部放电下都测量了四种特征分解组分的含量，分析了各种局部放电下特征分解组分含量以及特征组分含量比值的特点，发现各种局部放电下特征分解组分在含量、产气率以及含量比值等方面存在明显的差异，可以用SF6分解组分对局部放电类型进行识别。采用模糊C均值聚类作为分析方法，分别以四种特征分解组分含量和三组特征组分含量比值作为特征量，对四种局部放电下得到的SF6分解组分数据进行了聚类分析，结果表明以特征组分含量比值作为特征量时的聚类结果要优于以特征分解组分含量作为特征量时的结果，因此本文选择特征组分含量比值作为识别局部放电的特征量，所选取的三组特征组分含量比值为c(SO2F2)/c(SOF2)、c(CF4)/c(CO2)和c(CF4+CO2)/c(SO2F2+SOF2)，结合SF6在局部放电下的分解过程，分析了它们所对应的物理意义。利用决策树理论，以上述的特征量作为输入量，构建了用于局部放电识别的决策树，并取得了良好的识别效果。同时基于支持向量机理论，构建了用于识别不同局部放电的智能识别系统，采用粒子群优化理论对支持向量机的参数进行了优化，验证结果显示，该智能识别系统比决策树具有更好的识别效果。利用金属突出物缺陷，进行了不同微水和微氧含量下的SF6分解实验，得到了不同微水和微氧含量下的四种特征分解组分含量随时间变化的规律，分析发现，SOF2、SO2F2和CO2三种特征分解组分的含量都是随着水分和氧气含量的增加而增加的，CF4的含量几乎不随着水分和氧气含量的变化而变化。通过对不同微水和微氧含量下的c(SO2F2)/c(SOF2)、 c(CF4)/c(CO2)和c(CF4+CO2)/c(SO2F2+SOF2)三组特征组分含量比值的分析可以发现，它们的值都是随微水和微氧含量的增加而逐渐减小的。微水微氧含量对基于SF6分解组分的局部放电识别系统影响非常显著，当微水微氧含量变化时，所建立的支持向量机识别系统的识别率大幅下降，甚至完全不能识别。结合各种分解产物的生成过程，并利用化学动力学理论分析发现，SF6与O2和H2O反应生成SOF2和SO2F2的化学反应属于二级反应，有机绝缘材料和金属中所含有的C与O2和F原子反应生成CO2和CF4的化学反应也为二级反应。基于二级反应的反应物与生成物浓度之间的数学关系，推导得出三组特征组分含量比值与气室中初始O2和H2O含量的关系可以统一写成幂函数的形式，结合各种O2和H2O含量下特征组分含量比值的数据，采用最小二乘拟合的方法，得到了O2和H2O含量对特征组分含量比值影响规律的数学模型。根据数学模型，提出了通过校正来消除微水微氧影响的思想，并建立了微水和微氧对特征组分含量比值影响的校正方法，给出了相应的校正公式，可以将一种水分和氧气含量下得到的特征组分含量比值数据校正到另外一种氧气和水分含量下。利用构建的局部放电支持向量机识别系统，对比了数据在校正前和校正后的识别正确率，结果显示，通过所提出的微水微氧校正方法校正后，识别正确率大幅提升，表明所提出的微水微氧校正方法具有良好的校正效果。更多还原

【Abstract】 The existing partial discharge detecting methods for gas insulated switchgear (GIS)have some kinds of disadvantages. For pulse current method, ultrasonic method, andoptic detecting method, the problems of anti-jamming capability, detection accuracy,and costing make them not suitable for accurate on-line partial discharge monitoring.Though ultrahigh frequency method has lots of advantages, its pattern recognitioncapability has a far way to meet the industry requirement. Also discharge capabilitycalibration has not been solved yet for ultrasonic method, optic detecting method, andultrahigh frequency method. Therefore, the partial discharge detecting methods usednow can not completely meet the requirement of on-ling partial discharge monitoring.Research shows that partial discharge can decompose SF6. The decomposedproducts continue to react with moisture and oxygen in GIS and lots of chemicals aregenerated. Therefore, partial discharge in GIS can be discovered by detecting SF6decomposition products. As the SF6decomposition products generated by differentkinds of partial discharge are different in concentration, generation rate, andconcentration ratio, it is also possible to recognize partial discharge category by SF6decomposition products. It is essentially a kind of chemical method to detect partialdischarge by SF6decomposition products, which has the capability to resist the strongelectromagnetic and noise interference in field. So this method is significant in boththeory and application. But lots of problems have not been solved yet for this method,such as which decomposition product should be chose as the feature decompositionproduct, how to use the decomposition product to recognize the partial dischargecategory, severity and development trend, and what is the specific criteria.Based on stability, how easy to detect, and physical significance, four kinds of SF6decomposition products, namely, SO2F2、SOF2、CF4and CO2, were chose as thefeature decomposition products to analyze the SF6decomposition characteristic underdifferent kinds of partial discharge. The SF6decomposition experiments were conductedon the SF6decomposition device under four kinds of partial discharge including metalprotrusion, free conductive particle, contamination on insulator surface, and gap oninsulator-conductor interface. The concentration of the four kinds of featuredecomposition products were measured under each experiment. The characteristic of thefeature decomposition products concentration and feature concentration ratio was analyzed. Also the decomposition characteristics of SF6under each kind of partialdischarge were compared. It is found that the decomposition products are obviouslydifferent in concentration, generation rate, and concentration ratio under the four kindsof partial discharge. It is feasible to use SF6decomposition products to recognize thepartial discharge category.The concentration data of the feature decomposition products from the four kindsof partial discharge were clustered be fuzzy C-means clustering. And the concentrationand concentration ratio were used as the feature parameter respectively. The comparisonof the results shows that the performance using the concentration ratio as the featureparameter is much better than that of the concentration. So the concentration ratio waschose as the feature parameter for partial discharge recognition in this paper. Threeconcentration ratios were employed, namely, c(SO2F2)/c(SOF2), c(CF4)/c(CO2), andc(CF4+CO2)/c(SO2F2+SOF2). The physical significance of the three concentration ratioswas also analyzed. A decision tree was established for partial discharge recognition bydecision tree theory, using the three concentration ratios as the feature parameters. Thetest result shows that the decision tree has a good performance on partial dischargerecognition. Meanwhile, a smart recognition system was established for partialdischarge recognition, based on support vector machine theory. The parameters of thesupport vector machine were optimized by particle swarm optimization theory. The testresult shows that this smart system has a better performance than that of decision tree.The SF6decomposition experiments were conducted under different moisture andoxygen contents. Partial discharge was generated by the protrusion defect. And theconcentration of the four feature decomposition products was measured in eachexperiment. Through analysis of the data, it is found that the concentration of SOF2,SO2F2, and CO2increases with the initial content of moisture and oxygen. And theconcentration of CF4barely changes with the content of moisture and oxygen. After that,the characteristic of the three concentration ratios was also analyzed. It is shown that thevalue of the three concentration ratios all decreases with the content of moisture andoxygen. The content of moisture and oxygen has a strong influence on the partialdischarge recognition system. The recognition accuracy decreases dramatically whenthe content of moisture and oxygen changes.The chemical reactions generating the four feature decomposition products are allsecond order reactions by chemical kinetics theory. Based on the mathematical relationof reactant and resultant, it is found that the relation of the three concentration ratios and the initial content of moisture and oxygen in the gas chamber can be expressed bypower function. Then the mathematical model of influence rule of moisture and oxygencontent on the feature concentration ratios was developed by least square fitting methodusing the concentration data of feature concentration ratios under different moisture andoxygen content. According to the mathematical model, the calibration method ofinfluence of moisture and oxygen on the feature concentration ratios was proposed,which can calibrate the value of the feature concentration ratios from one moisture andoxygen content to the other content. The recognition accuracy before calibration wascompared with that of after calibration using the partial discharge recognition systembased on support vector machine. The result shows that the recognition accuracyincreases significantly when the input data are calibrated by the calibration method,which means the calibration method has a good performance.更多还原