Application Study Of Support Vector Machine In Transformer Condition Assessment

The transformer is a major apparatus in the power system and the traction power supply system.Failure of a transformer will definitely harm the safety and stability of the power system and the electrified railway.Transformer condition assessment is the foundation of the transformer condition based maintenance.Support vector machine(SVM) based on statistical learning theory(SLT) accomplishes the structural risk minimization principle,avoids being trapped into local minima and has fine performance to limited samples.It solves the difficulty that fault intelligent diagnosis system faces,which is terrible lack of typical fault samples,and is becoming gradually one of the powerful tools in intelligent diagnosis.Although the application research of SVM in diagnosis has obtained certain achievement,how to perform the novel method effectively in transformer condition assessment is a great challenge.Therefore,the application of SVM to the transformer intelligent fault diagnosis,fault forecasting and running condition assessment are developed in this thesis.The main contributions of this dissertation are as follows:1.The intelligent diagnosis of transformer based on gases dissolved in transformer oil is studied with the support vector classification.The method for conversion from two-class to multi-class is analyzed.Two kinds of multi-class algorithms are studied, i.e.the one-against-one method and the M-ary method.An improved M-ary SVM algorithm is proposed.The test results prove the effectiveness and superiority of the improved algorithm.Furthermore,the selection of kernel function parameters is discussed.The parameters are optimized using improved genetic algorithm.Compared with general genetic algorithm,the optimum can be found accurately in a wide range using the proposed method and the value can be used to diagnose the transformer effectively.2.The prediction of gases dissolved in transformer oil is investigated using support vector regression.The least square support vector machine(LS-SVM) is introduced into the concentration prediction of dissolved gases.The adaptability of the LS-SVM model and the situation of the parameters varied with the data for building the model are analyzed.In order to get optimized grey model,the multi-variable grey model is recommended to the concentration prediction of dissolved gases in transformer oil,an improved discrete grey model and an improved multi-variable grey model with higher accuracy are proposed.Moreover,a combined forecasting model is put forward using the LS-SVM combination.In the combined model the advantages of the optimized grey model and the LS-SVM model are integrated.The effectiveness and superiority of the mentioned models have been verified with the results of the actual case.3.The running condition assessment of transformer is realized with support vector machine.Aiming at the status that fuzzy synthetic evaluation is a main method for transformer running condition assessment,a fuzzy synthetic SVM model based on scores is proposed.The results of the fuzzy synthetic evaluation are the inputs of the SVM,the actual states of the transformer are the outputs.The examples show that,the running condition assessment for transformer using the new model is more accurate.In order to reduce the human factors,eliminate the misleading features and retain the genuine information,some condition assessment models for transformer based on component analysis and SVM are proposed.Four methods for extracting features are studied,viz.PCA,KPCA,ICA,and PCA+ICA.From the results it is concluded that the feature data after extracted lead to better separation,which suggests that data preprocessing in practical running condition assessment is favorable to the realization of classification algorithm,KPCA is an effectively feasible data preprocessing method for the transformer running condition assessment.In order to improve the accuracy and mend the effect of evaluation,an improved KPCA+SVM model is proposed.Mixtures of kernels and parallel optimized strategy are used.