Study On Power Transformer Fault Diagnosis Based On Ensemble Learning

Power transformer is the core of energy conversion and transmission in power grid, and it is the key hub equipment in the first line of grid security defense. Breakdowns of power transformer may cause huge losses like blackouts and damage to other equipment assets, or may even bring serious social influence. Therefore, potential fault diagnosis of the power transformer has important theoretical and practical significance for guiding the transformer’s maintenance of operation and condition, and for preventing and reducing the probability of failure. Based on the analysis of the primary failure mode of power transformer and transformer oil-paper insulation decomposition mechanism, by using dissolved gases in transformer oil as characteristic value, this paper builds least squares support vector machine(LS-SVM) model which are on the base of sample weight form intelligent learning perspective, and builds cloud membership space fault diagnosis model form the perspective of study and statistical probability, also analyses and optimizes these algorithms from the point of sample weights. The main contents of this paper are as follows:1) A new least squares support vector machine diagnosis model based on sample weight is established. This model puts sample importance as an influence factor of the punishment of the LS-SVM optimization problem, making the original optimal hyperplane of SVM raise form number to amount of information. This model not only can recognize the importance of samples or the samples carrying more information to prevent the loss of important sample and fault classification,but also can improve the rate of correct identification of the important sample under the condition of keeping accuracy of fault diagnosis, and provides a base for ensemble learning.2) A new transformer fault diagnosis model based on cloud relationship space is established from knowledge learning and mathematical statistics perspective. This model can get the fault rules of the sample, and store these rules into the cloud combination as knowledge. Membership algorithm can use the knowledge to diagnosis. The cooperation of this model and intelligent algorithm can achieve accurate fault diagnosis of transformer.3) A new combination diagnosis model composed by sample weight LS-SVM is established, the sample at their segment boundaries can be corrected by the combination hyperplane. In case of the knowledge covering each other, the weights is set to sample knowledge on the foundation of cloud relationship space model. A dynamic combination diagnosis model is composed by different diagnosis models by the way of Ensemble learning.