Application Research Of Oil-immersed Transformer Fault Diagnosis Based On Intelligent Algorithms

Oil-immersed transformer is a core equipment of power system. The safe operation oftransformer is related to the normal operation in all fields of national economy and the safetyof life and property. It also greatly affect the running cost of power system. Using electricaltest methods to diagnose transformer not only need outage, but also are difficult to detectsome local faults. However, it doesn’t need power failure to find transformer’s interior faultsby analyzing the components of gases dissolved in transformer oil. This method can findlatent fault of transformer effectively and quickly, so the deficiencies of electrical testmethods are remedied.The ratio method based on three ratio is a conventional method of transformer faults byDGA widely used at home and abroad. The ratio method exist some faultiness when itdiagnose faults of uncertain and complex transformer, such as "encoding boundary tooabsolute" and "code missing".With the development of intelligent information processingmethod, all kinds of intelligent algorithm have been used in transformer fault diagnosis.Support vector machine (SVM) and bayesian network have been used for application researchof transformer fault diagnosis in this paper.Support vector machine (SVM) has achieved good results in the application oftransformer fault diagnosis, but nowadays there is no generally accepted effective theory toselect kernel function and parameters. This paper adopt methods of grid search, GA and PSOfor SVM parameters optimization in the sense of cross validation. After getting a originaloptimized parameters, the parameters of final best are obtained by parameters fine-tuning inthe experiment. It is found that the RBF kernel is superior than others in experimentalcomparison of dealing with nonlinear input. Although fault diagnosis has achieved goodresults via choosing optimization of kernel function parameters svm, Directly using SVMtakes a long time in convergence of training set modeling when the transformer fault data arevery large and redundant, the classification accuracy is low. Due to this issue, SVM andk-means are proposed in fault diagnosis of transformer. Under the accuracy of model isguaranteed, training set of transformer faults are pre-selected through using k-meansclustering algorithm. To choose effectively numbers of support vectors are regard as inputfeature vectors of SVM classifier. Specific instances are adopted to verify the validity of modein the weka platform.This paper aim at the phenomenon of data loss in practice, Bayesian network ispresented in diagnosis of transformer fault.It is found that the method of entropy discretization is the best through comparing different discrete methods. Finally, two specificexamples are used to verify the effective diagnosis of Bayesian network method.