A Dual Deep Neural Network Approach for Discriminating Internal Faults and Inrush Currents in Transformers Under CT Saturation

The dependability and security of transformer differential protection can be compromised by current transformer (CT) saturation, which distorts current measurements and may lead to the maloperation of protective relays. This paper proposes a dual deep neural network (dual-DNN) approach to accurately distinguish internal faults from simultaneous inrush currents, even under CT saturation. The proposed architecture consists of two stacked denoising autoencoder (SDAE)-based networks: a compensation DNN (Com-DNN) and a discrimination DNN (Dis-DNN). The Com-DNN mitigates the adverse effects of CT saturation by learning to recognize and correct distortions in the current signals, while the Dis-DNN discriminates between inrush currents and internal faults. The dual-DNN approach was evaluated using a 23-kV power transformer under various scenarios, including simultaneous inrush and fault events. Simulation results demonstrate that the proposed approach delivers robust performance and stability under challenging conditions. A comparative analysis with a conventional artificial neural network (ANN) and three other existing methods confirms that the dual-DNN approach offers superior dependability and security. Furthermore, hardware implementation validates its feasibility and effectiveness for real-world deployment.