Publication

Abstract :

Power quality disturbances (PQDs) are significant irregularities in electrical power systems that can negatively impact system stability and delicate equipment. To ensure the re liability of the power system, PQDs must be classified accurately and effectively. This work investigates the application of Vision Transformer (ViT) based architectures for classifying PQDs, which are preprocessed into spectral images. The performance of the proposed model is evaluated and compared against the convolutional neural networks (CNNs). The three vision transformer variants considered in this work include the Basic ViT model, the Swin Transformer model, and the Data-Efficient Image Transformer (DeiT). Among these three variants, DeiT employs a knowledge distillation strategy, achieving accuracies of 99.33% and 99.38%, and offers qualitative interpretability through attention maps. Through comparative analysis with CNN architectures, such as EfficientNet and DenseNet, this paper highlights the potential of ViT-based models for PQD classification and proposes future extensions toward real-world validation and improved teacher-student frameworks.