A Generalized Classification Framework for Power Quality Disturbances Based on Synchrosqueezed Wavelet Transform and Convolutional Neural Networks

Power quality disturbances (PQDs) pose a significant threat to the reliability, efficiency, and security of electrical power systems and ought to be detected and classified. This allows for appropriate corrective action to be taken based on the power quality event, which is essential for effectively mitigating the impact of PQDs. This article presents a new framework for the accurate detection and classification of PQDs in electrical power systems using synchrosqueezed wavelet transform (SWT) and convolutional neural networks (CNNs) based on the EfficientNetB0 architecture. First, the SWT is utilized to transform the signals into time–frequency matrices, which are then visualized as 2-D time–frequency contour images. The obtained time–frequency representations (TFRs) based on SWT have been compared with short-time Fourier transform and pseudo Wigner–Ville distribution. Then, time–frequency contour images are used for training, validation, and testing of the CNNs to detect and classify 15 different types of single and multiple PQDs. The effectiveness of the proposed methodology is demonstrated by testing on unseen data: 1) synthetic PQDs corrupted by pink noise (PN) and 2) PQDs generated using power system simulations. The results obtained from the proposed scheme are compared with continuous wavelet transform (CWT)-based time–frequency analysis and machine learning approaches. Furthermore, visual explanations for the correct and incorrect classifications made by the CNNs have been obtained through gradient-weighted class activation maps (Grad-CAM).