A novel vision transformer-based power quality disturbance classification method

Abstract

Accurate and automated classification of power quality disturbances (PQDs) is essential for ensuring the reliability and stability of smart power systems. This study introduces a novel classification framework that combines a Vision Transformer (ViT) model with an innovative signal-to-image transformation technique, which directly reshapes 1D time series (T.S.) signals into 32 × 32 grayscale images, thereby eliminating complex preprocessing steps such as feature extraction or wavelet transforms. Unlike traditional approaches that rely on handcrafted features or spectrogram-based methods, this lightweight conversion preserves temporal characteristics while enabling efficient end-to-end learning through ViT’s attention mechanism. The model was evaluated on a comprehensive dataset comprising 21 distinct PQD classes, systematically generated under real-world conditions (20–50 dB noise levels and ± 0.5 Hz frequency deviations) using two Arbitrary Waveform Generators (AWGs). The proposed system achieved state-of-the-art performance with 99.23 % classification accuracy and an exceptionally fast inference time of 3.58 ms per sample, demonstrating both precision and suitability for real-time applications. Remarkably, the architecture maintained robust performance across all noise levels, confirming its strong generalization capability. Despite using an image-based approach, the method’s computational efficiency, achieved through optimized patch processing and compact input dimensions, makes it deployable in resource-constrained embedded systems. These findings position the framework as a practical foundation for next-generation PQD monitoring systems. The study advances the field by: (1) introducing the first ViT-based solution for raw PQD signal classification, (2) establishing a new benchmark in processing efficiency (3.58 ms runtime), and (3) demonstrating unprecedented robustness to both noise and frequency variations. Overall, this work provides a scalable, accurate, and hardware-friendly solution for intelligent power quality management, showcasing the untapped potential of transformer architectures in T.S. industrial applications.

Topics & Concepts