Discrimination Between Faults and Inrush Currents Using a Feature Transfer Strategy-Based CNN for Power Transformers

Transfer learning can enhance artificial intelligences (AIs) ability in identifying faults and inrush current for the transformer differential protection. However, due to the high complexity and scarcity of operating data, conventional transfer methods are not applicable. Additionally, significant differences in differential currents among transformers hinder the AI’s generalization ability. To address these issues, a novel transfer strategy for convolutional neural networks (CNNs) is introduced in this article. This strategy guides the CNN to focus on the unsaturated segments of the differential current fault components (FCDCs), which reliably indicate the presence of a fault. It defines the images of the unsaturated segments as the source domain which is used to train a CNN called source-CNN (SCNN). The original images, containing both unsaturated and saturated segments, are defined as the target domain and are used to train another CNN called target-CNN (TCNN). The proposed transfer strategy helps TCNN focus on the unsaturated segments by minimizing the feature differences between TCNN and SCNN. Specifically, adaptation layers are incorporated to compute adaptation losses in each convolutional layer between TCNN and SCNN. TCNN is then trained using a loss function that is the weighted sum of classification and adaptation losses. Additionally, a dropout strategy for the adaptation layers is proposed to mitigate overfitting. Finally, TCNN is employed to develop an inrush current blocking method for transformer differential protection, demonstrating enhanced generalization ability in dynamic simulation experiments.