Knowledge-based convolutional neural networks for transformer protection

Deep learning based transformer protection has attracted increasing attention. However, its poor generalization abilities hinder the application of deep learning in the power system owing to the limited training samples. In order to improve its generalization abilities, this paper proposes a knowledge- based convolutional neural network (CNN) for the transformer protection. In general, the power experts can reliably discriminate between faulty transformers and healthy transformers only through the unsaturated parts of equivalent magnetization curve (voltage of magnetizing branch-differential current curve) but deep learning intends to focus on the combined features of saturated and unsaturated parts. Inspired by the identification process of power experts, CNN adopted a specially designed loss function in this paper which is used to identify the running states of power transformers. Specifically, the presented Restrictive Weight Sparsity substitutes a special regularization term for the common L1 regularization. The presented Adaptive Sample Weight Adjustment endows the softmax loss of each sample with the optimizable weight the softmax loss of each sample with the optimizable weights to increase the impact of more-difficult-to-identify cases on the training process. With the modified loss function, the knowledge is abstractly introduced into the training process of CNN so as to successfully imitate the identification process of power experts. Accordingly, the proposed knowledge- based CNN will pay more attention to the unsaturated parts of equivalent magnetization curve even if only limited samples are included in the training process. The results of simulations and dynamic model experiments reveal that the knowledge-based CNN exhibits an improved generalization ability and the knowledge-based deep learning algorithm is a promising research direction.