Data-driven Fault Diagnosis And Prediction For Large-scale Power Transformer

Power transformer is one of widely distributed, complex and expensive equipment in the power system. It undertakes the heavy task of voltage conversion and power transmission, and its safety state plays a great effect on the stability and security level of 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.In view of the relationship between dissolved gases with the fault reasons and the characteristics of transformer malfunctions, this thesis focuses the research on new methods of fault diagnosis and predietion based on the data-driven technique. These methods presented are as follows:Fault detection model of cost-sensitive kernel principal component analysis based on modeling samples’purification; Fault diagnosis method based on reconstruction-based contribution and grey relation entropy; Fault diagnosis model based on dual-space feature extraction algorithm; Prediction model based on mutual information and multiple-kernel support vector regression. All methods are applied to the fault diagnosis and prediction of power transformer, and achieve good performances. The major innovation research achievements include:(1) In order to handle the invalidation problem of fault detection when the modeling samples is impure, and traditional kernel principal component analysis (KPCA) doesn’t consider the misclassiflcation cost, insensitive to fault condition, a new cost-sensitive kernel principal component analysis (CSKPCA) of transformer fault detection model based on purified modeling samples is proposed. First, a new purification algorithm is proposed to purify the modeling samples based on KPCA, then feature sample extraction method is adopted to process the modeling samples for solving the invalidation problem of fault monitoring model and the calculation problem of the kernel matrix; Second, cost-sensitive mechanism is introduced into kernel principal component analysis, the squared prediction error (SPE) threshold adjusting method is designed to get the threshold aimed to minimize misclassification cost; And the chaos particle swarm optimization (CPSO) algorithm is adopted to optimize the kernel parameters of kernel principal component analysis. Compared with cost-sensitive neural network the proposed method can reduce misclassification cost effectively with high fault sensitivity and diagnosis accuracy.(2) In order to extract the fault characteristic of dissolved gas analysis (DGA) data, a new transformer fault diagnosis method based on reconstruction-based contribution (RBC) and grey relation entropy is proposed. The proposed method is based on the idea of fault reconstruction when principal component analysis (PCA) model based on dissolved gases is set up, the fault characteristic information of dissolved gases are extracted by calculating RBCs and normalization. In order to overcome shortcomings of original grey relation analysis, such as partial relation and information losing, the proposed method adopt grey relation entropy (GRE) to fault diagnosis. Experimental results show that the proposed method has very good fault distinguishing ability and enhance fault diagnosis accuracy.(3) It is more attractive to relative content of feature gases as regarding to transformer fault diagnosis based on DGA. Considering the RBCs’comparability and the dynamic value range of each feature gas variable, feature information of dissolved gases are extracted by calculating relative RBCs and normalization. Then, transformer fault diagnosis model is set up based on relative RBCs and synthesis grey relation entropy. Experimental results show that compared with the features extracted by RBC-GRE and GRE, the proposed method increases the separability of feature data set and performs better classification accuracy.(4) A new transformer fault multilayer diagnosis model based on PCA and kernel independent component analysis (KICA) dual-space feature extraction algorithm is proposed to overcome the limits of single subspace. Firstly, DGA test sample was projected to PCA subspace, Taking advantage of stronger robustness, higher precision and less dependence of modeling accuracy on kernel function, multiple-kernel support vector machine (MKSVM) was used as classifier to predict the class label. The test sample was pre-classified as difficult one or easy one according to comparison of prediction result with the threshold which was obtained by kernel density estimation method. The class label of easy one was identified in the PCA subspace directly, as to the difficult, the test sample was re-projected to KICA subspace where another MKSVM was used to identify the class label. Thus, MKSVM of two class problem based on dual-space feature extraction algorithm was achieved. Finally, a multilayer diagnosis model was set up according to the fault characteristic of transformer. The experiment shows that mutual complementary dual-space feature extraction algorithm has a higher diagnosis rate, which proves its effectiveness and usefulness.(5) Considering the impact of noise in each dissolved gas variable, a dissolved gas-in-oil content prediction model based on improved standardization mutual information and multiple-kernel support vector regression machine (MKSVR) is proposed. Firstly, independent component analysis (ICA) was adopted to separate signal and noise of each dissolved gas-in-oil variable. Secondly, improved normalized mutual information feature selection method was proposed to select input variables considering the impact of noise, input variable selection results are consistently with mineral oil inferior thermal dynamics. At last, MKSVR was used to forecast dissolved gases content. Experiments show the proposed prediction model has a better prediction and generalization.