Discrimination between internal faults and inrush currents in power transformers based on the discriminative-feature-focused CNN
Anyang He, Zaibin Jiao, Zongbo Li, Yonghui Liang
Abstract
Overcoming the drawback of the incorrect action of differential protection induced by the inrush current is a crucial issue for transformer protection. This paper proposes a deep learning protection scheme based on a discriminative-feature-focused convolutional neural network to improve the protection accuracy. Considering that internal faults and inrush currents can be reliably distinguished based on the unsaturated part of the differential current, the three-module network and specialized loss function are designed to guide the proposed network to extract the discriminative features of the unsaturated part of the differential current. By focusing on discriminative features with invariance and robustness, the proposed method presents high accuracy and remedies the problem of the limited generalization ability of artificial-intelligence-based protection in engineering applications. Test results on various datasets show that the proposed scheme can accurately identify transformer states with good robustness to noise and generalization to complex scenes, such as CT saturation, overexcitation, and field data. A time-consuming test deployed on hardware shows that this algorithm meets the protection time requirements. The comparison with previous methods confirms the superior reliability and generalization ability of the proposed scheme.