Classification of Scalogram Signatures for Power Quality Disturbances Using Transfer Learning

The electrical power systems have gone through a process of transformations that will remain characterized by a wide penetration of renewable sources, electronic devices, and computerization. In this context, Power Quality (PQ) is associated with several challenges for the sector, presenting new issues and new scenarios for old problems. Signal processing (SP) plays an essential role in PQ applications as a tool that helps measure, characterize, and visualize electrical grid disturbances. At the same time, artificial intelligence (AI) is becomming more and more useful to classification tasks regarding PQ disturbances . This work aims to employ a transfer learning methodology for PQ disturbances classification. Wavelet scalograms of the signal are created using CWT for feature extraction of time-frequency signatures. The 2-D images of this representation are used to train and test pre-trained CNN models’ performance. The work aims to contribute to PQ disturbances classification through innovative methods and assess the performance of different CNNs models that have a significant role in image classification. The performance of four network models is assessed: ResNet-18, VGG-19, Inception-v3, and ResNet-101. Discussion and consideration about the results provide evaluation of the methodology.