Differential Fault Diagnosis in Power Transformers Under Limited Data Condition via Few-Shot Learning

Power transformers require fast and reliable protection to prevent catastrophic failures. However, collecting labeled fault data in real-world conditions is difficult, as transformers cannot operate under faulted states. This challenge is further complicated by imbalanced datasets and the need for retraining when deploying models on new, unseen transformers, which hinder the effectiveness of artificial intelligence based methods. To address these limitations, we propose Few-DP, a few-shot learning-based differential protection method that operates effectively under limited and imbalanced data. Few-DP uses metric learning and a Siamese convolutional neural network architecture to extract fault features from raw current signals and measure similarity between samples, enabling fault classification with fewer than five examples per class. The model is validated on a simulated 400-kV grid in power systems computer aided design (PSCAD)/electro-magnetic transients in dc systems (EMTDC) and through experimental testing. Results show that Few-DP is robust to noise, data scarcity, saturation, and transformer variability, outperforming state-of-the-art methods in both accuracy and generalization. All simulation models and code are available on GitHub to support future research and development.