Avaliação computacional de pontos quentes em painéis fotovoltaicos: uma estratégia de machine learning

Abstract

The growing representation of photovoltaic energy in the Brazilian energy matrix leads to the need for a more efficient maintenance strategy in detecting hot spots in solar modules. The traditional method for identifying these flaws is poorly scalable and dependent on human interpretation of thermal images, which highlights the low scalability of the process. The use of computational processing, through deep machine learning, of standardized thermographs is a scalable solution to this problem, in addition to presenting significant efficiency in identifying hot spots. The experimental results obtained with the created algorithm were 96%, 87%, 89% and 90% for the respective convolutional methods MobileNetV2, MobileNetV2, VGG16 and ResNet50, based on training and testing with a set of 419 images.