Transformer Fault Diagnosis Based On Least Squares Support Vector Machines And D-S Evidence Theory

Power transformer is an important electric power equipment, its operation is directly related to the security and stability of overall power system. Due to the complexity and particularity of the operation environment, failure is inevitable for a long-term operation of a power transformer. With the social and economical progress today, because of the increasing requirements for the power quality, reliability and safety of the power system, it is necessary to study the new technology deeply for power transformer fault diagnosis.In all technologies for transformer fault diagnosis based on DGA(Dissolved Gas Analysis), the traditional three-ratio method is replaced by some intelligent methods due to the shortcomings of owning fuzzy ratio boundaries. Some transformer fault diagnosis technologies based on AI(Artificial Intelligence) are increasingly popular, but the fault diagnosis effects are less than satisfactory owing to some deficiencies existing in these fault diagnosis technologies themselves. The fusion of many intelligent methods for transformer fault diagnosis can remedy their deficiencies, and become the research trends. Therefore, this thesis proposes a new transformer fault diagnosis method combining LS-SVM(least squares support vector machine) and D-S evidence theory.Firstly, the thesis studies the transformer fault diagnosis based on DGA, and discusses the advantages and the disadvantages about the early ratio method. Because of these disadvantages, the necessity of artificial Intelligence fault diagnosis for transformer is presented. After comparing several commonly used artificial intelligence algorithms, the thesis puts forwards the transformer fault diagnosis method which combines with LS-SVM and D-S evidence theory. The posterior probability that is output by the multiclass LS-SVM models becomes the basic probability assignment of D-S evidence theory, then D-S combination rule and D-S decision rule are used to diagnose transformer fault. In the process of establishing the multiclass LS-SVM models, the thesis studies the classification methods of multiclass LS-SVMs, the selection methods of kernel function and the outputs of the posterior probability, etc. In order to obtain the optimal LS-SVM’s parameters, the thesis establishes the bayesian framework for LS-SVM. In this framework, the parameter is optimized by the method that the posterior probability of the parameter is maximized. In the end, D-S evidence theory is used to complete power transformer fault diagnosis, the basic probability assignment of D-S evidence theory is calculated by the posterior probability which is output by the multiclass LS-SVM models, the advantage of this method is that the way of achieving the basic probability assignment is objective. Based on the above theory, the model of po wer transformer fault diagnosis is set up by using DGA data which is collected form the transformer fault, then the thesis completes fault diagnosis using the model and gets the accurate results. In order to validate the performance of the model, the three ratio method transformer fault diagnosis model and the single LS-SVM transformer fault diagnosis model are set up as the references. According to compare the results of the three models, the thesis proves that the model combining LS-SVM and D-S evidence theory is better than the other two.Through the study of this thesis, the method combining LS-SVM and D-S evidence theory is useful, has a high accuracy, and the same time, this method provides a reference for the other systems that are similar with transformer system.