Research on Feature Reduction Method for Partial Discharge Pattern Recognition of Converter Transformer

Abstract

The effective identification of the partial discharge fault type of the converter transformer can significantly improve the power supply reliability of the HVDC transmission system. In the partial discharge pattern recognition, the feature extraction will directly affect the ability of the classifier to identify the fault type. So, it is a key problem to extract features that can reflect the essential characteristics of faults by adopting appropriate feature reduction methods currently. Therefore, in this paper, the feature reduction method is deeply studied, and the application results of several representative methods in partial discharge feature reduction are compared. Based on a large number of existing partial discharge test data, firstly, the fingerprint features of the partial discharge are extracted to construct the feature space. Then, three representative feature reduction methods are used to optimize and reduce dimension of feature space, including kernel principal component analysis(KPCA), rough set theory(RST) and correlation coefficient matrix respectively. Finally, the feature space with reduced dimensionality is taken as the input of the back propagation (BP) neural network. The advantages and disadvantages of the three different Feature reduction methods are comprehensively considered in terms of convergence time and recognition accuracy. This paper compares the commonly used feature reduction methods in the process of feature extraction, and combines it with BP neural network for partial discharge pattern recognition. It can be concluded from this paper that these methods can be applied to the feature reduction of the feature space of the partial discharge. Applying these methods to existing experimental data, we can find that the combination of feature space reduction method and BP neural network can significantly improve the accuracy of partial discharge pattern recognition.