Gerenciamento da qualidade da energia elétrica em smart grids baseado em técnicas de soft computing

• Autor(a) principal: Moraes, Anderson Luis de
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/14390
• topic: Aprendizado de máquinas; Classificação de distúrbios; Compressão de dados; Detecção de distúrbios; Processamento de sinais; Qualidade da energia elétrica; Machine learning; Disturbances classification; Data compression; Disturbances detection; Signal processing; Power quality; ENGENHARIAS::ENGENHARIA ELETRICA::SISTEMAS ELETRICOS DE POTENCIA

The increasing use of nonlinear loads (mainly those based on power electronics), the integration of renewable sources (such as wind and photovoltaic), atmospheric discharges, starting of motors and driving large load blocks generate disturbances that affect the Power Quality, i.e., the energy delivered to consumers. In the Smart Grids context, the distribution utilities seek ways to monitor the Power Quality, so that disturbances can be detected by smart meters and the resulting data should be compressed to ensure an efficient exchange of data packets. In this sense, it is expected that, after unpacking the data, the signals will be recovered with few information losses and can be classified to assist in the utilities’ decision making. Therefore, this work proposes a framework based on the edge and cloud computing technologies, where the processes of detection/segmentation, compression and classification of power quality disturbances will be properly performed. To analyze the performance of this framework, a synthetic database with 15 disturbance classes (simple and combined) was generated. Thus, detection of disturbances was performed by a Decision Tree capable of identifying 94.71% of the disturbance windows. Next, the disturbances detected were submitted to a treatment stage in order to guarantee a more efficient segmentation of the signals. The resulting windows of disturbances were then compressed using a Wavelet Transform, considering filters from the Daubechies family, in which it was possible to reduce the data packets to a compression rate greater than 3.6. Through data unpacking, a low information loss was observed. Finally, there was a transformation of the temporal signals in Recurrence Plots, Gramian Angular Summation Field and Gramian Angular Difference Field in order to identify the voltage signal patterns through a set of Convolutional Neural Networks. In this context, the proposed approach allows to obtain an average accuracy above 94%. Thus, the results of this research will contribute to advance the state-of-the-art in Power Quality automatic signal processing.