【作者】 胡青；

【导师】 孙才新；

【作者基本信息】 重庆大学 ， 电气工程， 2010， 博士

【摘要】 电力变压器是电网中能量转换、传输的核心,是电网中最重要和最关键的电气设备之一,其运行状态直接影响系统的安全运行水平,变压器一旦发生事故,导致电网不能正常供电,将带来巨大的经济损失。变压器故障诊断和故障预测是确保变压器正常运行的基础,也是实施状态维修的基础。本文对变压器故障诊断和故障预测方法进行了深入研究,主要研究内容如下:①针对变压器故障诊断的特点,提出分层诊断模型,依据特征气体对各层诊断所提供的有效信息量大小,选择信息量大,冗余度小的特征量集合作为最优诊断特征量,建立分层诊断模型。②应用互信息理论实现变压器故障诊断的特征量选择,针对互信息理论计算方法中不合理的地方,改进了互信息的计算方法;针对互信息理论不适合用于衡量特征量与类变量之间互信息的问题,提出用卡方距离计算特征量与类变量之间的互信息,并修改卡方距离的计算公式,使之与互信息计算具有相同的数量级。用修正的卡方距离度量特征量提供的有效分类信息量,用互信息衡量各个特征量之间的冗余度,为各层故障诊断选择出最优的诊断特征量集合。③用模糊数学解释故障诊断模型中神经网络各层节点的作用,基于此提出了一种自动设计神经网络结构,初始化神经网络权重矩阵的方法。该方法首先按卡方距离的计算思想,分析特征子空间与各个故障类别之间的关系,定义了卡方关系,然后依据子空间与各个类别之间的卡方关系,初始化神经网络连接权重。④为了进一步提高故障诊断的正判率和抗干扰能力,将神经网络群方法应用于变压器故障诊断领域,分析了分层诊断的神经网络群结构,为了提高分类器之间的差异,提出用核主成分分析方法增加特征量,详细分析比较了经过核主成分分析,所得的核特征值、核特征量与分类之间的关系,实验结果表明,核特征值大小不能反映核特征量所提供的有效分类信息量的多少,依据修正的卡方距离和互信息大小选择合适的核特征量,建立分层神经网络群是有效的。⑤分析变压器油中溶解气体组分预测的特点,认为气体组分预测问题的实质是对气体组分变化情况的预测,基于此,提出用气体组分的差值序列建立预测模型,并用气体组分序列的偏差对差值序列进行预处理,消除气体组分大小差异带来的影响,使不同特征气体的差值序列幅值相近,便于后续建立预测模型。⑥由于变压器内部各特征气体组分之间存在互相影响,为每个特征气体单独建立预测模型增加了建模和预测的计算量,不能反映气体变化的实际情况,也会影响预测精度。针对这个问题,提出用模糊认知图方法建立变压器特征气体的统一预测模型,该方法能够从历史数据中学习系统内在的变化规律,在一定程度上实现了对变压器特征气体的长期预测,其短期预测和长期预测效果均达到了一定的预测精度。更多还原

【Abstract】 Power transformer is the crucial device of energy conversion and transmission in the power grid, it’s also the most important electric device, its in-service state directly affects the safety and stability of the power system. The failure of power transformer may cause huge economic loss. The transformer fault diagnosis and fault prediction is the basis of keeping it running normally and carrying out the condition based maintenance, which was studied in-depth in this paper. The main contribution of this thesis was the following.①the layered fault diagnosis model was proposed based on the analysis of the character of transformer fault diagnosis, the layered diagnosis model should be constructed with the optimal features which are selected according to the effective information amount. The optimal features are the feature subset which may provide the most classification information with the least redundancy.②application of mutual information to feature selection for transformer fault diagnosis, to overcome the drawbacks in the calculation method of mutual information, an improved calculation method was proposed. As mutual information is not suited to evaluate the mutual information between a feature and a class variable, mutual information was replaced with chi-square distance, which was modified to have the same order of magnitude as mutual information. The feature selection method proposed here used modified chi-square distance to evaluate the effective information provided by the feature, used mutual information to evaluate the redundancy degree between features, selected the optimal feature subset for each layer.③a fuzzy mathematic interpretation of neural networks was given, and based on this interpretation, a automatic design and initialization of neural networks was proposed. First this approach analyzed the relationships between feature subsapces and fault types, gave the definition of the chi-square relation, then constructed and initialized the neural network based on the chi-square relations.④to improve the performance and the anti-noise ability of transformer fault diagnosis, neural network ensemble method was applied, this paper analyzed the ensemble structure of layered fault diagnosis, applied kernel PCA to increase the number of features, compared the kernel eigen value, eigen feature and its classification information, the experiment results showed that the eigen value didn’t reflect the amount of effective information, the neural network ensemble was constructed by optimal kernel features which were selected according to modified chi-square distance and mutual information.⑤analyzed the characters of the prediction of DGA gas concentrations in transformer, the key issue of gas concentration prediction is the prediction of the change of gas concentrations, so the change series was used to construct prediction model, the change series was pre-processed with the deviation of gas concentration series, this pre-precession eliminated the influence to the prediction brought by the magnitude differences among concentrations of different gases,⑥As in a transformer, there are mutual affections among feature gases, so individually developing prediction models for each gas is not appropriate, fuzzy cognitive map method was applied to construct a universal prediction model for all feature gases in a transformer. This method could learn the behavior rules of a given system from its history data, to some extent this prediction model achieved the long tern prediction.更多还原