The use of advanced signal processing and deep learning for pattern recognition in integrated metrics of quality performance: a smart grid application

Abstract

Power quality (PQ) is not a new theme, but it should not be neglected in any way, as its performance parameters will reveal problems in the adequacy between the consumer equipment and the electrical grid. With the ongoing transformations in electrical power systems, characterized by the high penetration of renewable energy sources, the massive insertion of components based on power electronics in the network, and the decentralization of generation, these issues are becoming increasingly important. In Smart Grids, solutions are sought for more advanced solutions to solve PQ disturbances problems. Advanced signal processing plays an essential role in dealing with the network and supporting various applications within this context and Artificial Intelligence (AI), which has gained significant prominence to feed applications with innovative solutions in several areas. This research investigates the use of advanced signal processing and Deep Learning techniques for pattern recognition and classification of signals with PQ disorders. For this purpose, the Continuous Wavelet Transform with a filter bank is used to generate 2-D images with the time-frequency representation from signals with voltage disturbances. The work aims to use Convolutional Neural Networks (CNN) to classify this data according to the images’ distortion. In this implementation of AI, specific stages of design, training, validation, and testing were carried out for a model elaborated by the case file and a knowledge transfer technique with the pre-trained networks SqueezeNet, GoogleNet, and ResNet-50. The work was developed in the MATLAB/Simulink software, all signal processing stages, CNN design, simulation, and the investigated data generation. All steps have their objectives fulfilled, culminating in the excellent execution and development of the research. The results sought high precision for CNN de Scratch and ResNet-50 in classify the test set. The other two models obtained not-so-high accuracy, and the results are consistent when compared with different methodologies. Considerations about the results were pointed out. Finally, some conclusions were established and a philosophical reflection on the role of AI and advanced signal processing in electrical power systems.