Intelligent deep learning-based dual-task approach for robust power quality event classification

Ensuring the stability and reliability of modern power systems relies on accurately detecting and classifying power quality (PQ) disturbances. This study presents an advanced automated recognition approach capable of identifying and classifying 14 distinct PQ event types, including voltage sags, swells, transients, interruptions, harmonics, and flicker. The methodology integrates the tunable-Q wavelet transform (TQWT) for signal decomposition, optimizing Q-factor parameters to extract precise features, and morphological component analysis (MCA) with the Split Augmented Lagrangian Shrinkage Algorithm (SALSA) for effective component separation. A novel dual-task deep learning model is developed, incorporating a dynamic nonstationary redundancy factor, r(t,f), to enhance the localization of signal components across time and frequency domains. Simulation evaluation demonstrates that the proposed model outperforms conventional machine learning and dense networks, achieving superior accuracy (99.7%), faster convergence (by 30%), and reduced computational cost (by 25%). These findings underscore the model's efficacy in real-time PQ monitoring, contributing to improved reliability and stability in evolving power systems.