Intelligent Protection for Power Transformer Using Convolutional Neural Network Integrated into Features Transferring Strategy

Abstract

Multi-feature fusion strategy based on artificial intelligence (AI) can theoretically improve the reliability of transformer protection but its poor generalization ability hinders the application in engineering. Based on the images of the equivalent magnetization curve (voltage of magnetizing branch-differential current), this paper presented a novel feature transferring strategy of convolutional neural network (CNN) to improve the generalization ability. In generally, according to the professional knowledge, the power experts can reliably identify the running states only through the unsaturated parts of equivalent magnetization curve. Meanwhile, the saturated parts have little effects on the identification process. Therefore, this paper tried to imitate the identification process of power experts for a reliable transformer protection. Specifically, equivalent magnetization curves whose saturated parts were removed were defined as source domain and the original ones were target domain. The source and target domains were used to train two CNNs with the same structures, named as SDCNN and TDCNN respectively. Adaptation layers were embedded between the convolutional layers of SDCNN and TDCNN to calculate the adaptation loss, namely feature differences. TDCNN which aimed at minimizing the weighted sum of adaption loss and classification loss will pay more attentions to the unsaturated parts. Finally, an optimal TDCNN was used to construct an intelligent protection of power transformer. In order to decrease the over-fitting of TDCNN, a dropout strategy of adaption layers was further proposed to decrease the coupling relationships. The results of PSCAD simulations and dynamic model experiments reveal that the feature transferring strategy effectively improves the generalization ability of AI based transformer protection.