Using MLP to locate transformer winding fault based on digital twin

Abstract

There is no doubt that transformer plays a fundamental role in power system. At the same time, transformer winding fault diagnosis is an important topic. Many works put the most emphasis on the identification of fault type and degree, while ignoring the fault location. However, fault location is an urgent problem to be solved, which is worth studying and discussing. The contribution of this paper lies in the location of Disk space variation (DSV) fault. The introduction of digital twin can solve the problem of insufficient fault cases, and pave the way for the intellectualization of fault diagnosis. In this paper, the digital twin of transformer winding is established based on double ladder network, in which the distributed parameters are calculated by finite element method. Frequency response analysis (FRA) is one of the most widely accepted methods for transformer winding mechanical deformation fault diagnosis. Aiming at the interpretation code of FRA, this paper disproves the view that phase information is useful. Then, by extracting the mathematical index of FRA, multi-layer perceptron (MLP) is trained and DSV fault location is realized. In addition, the popular support vector machine is also compared with the MLP model in this paper, which further highlights the advantages of MLP. The proposed method is verified by an actual transformer, and the results are satisfactory.