Image Cognition-based Power Transformer Protection Scheme Using Convolutional Neural Network

A novel equivalent magnetization hysteresis (V-I curve)-based transformer protection scheme using convolutional neural network (CNN) is proposed in this paper. The characteristics of equivalent magnetization hysteresis demonstrating the relationship between excitation branch voltage and differential current can describe figuratively whether the power transformer is faulty. Due to the generality of V-I curves, the performance of artificial intelligence (AI)-based transformer protection methods can be improved. In this paper, the images of V-I curves are used as the input of CNN to identify the faults. The training samples are generated with typical 2-winding transformers through PSCAD simulations. And the test samples which are composed of field test data and PSCAD simulation data are obtained with different transformer parameters from that of training samples. With much work of adjusting CNN parameters, a CNN structure with better performance is decided. The results show that the generalization ability of proposed method is very well and the image cognition-based methods are free from the different sampling frequency. Despite there are still a lot of works for further improving the performance, the proposed method has shown a good application prospects.