Abstract

This article presents an artificial neural network tool able to quantify the power loss due to soiling and partial shading effects of solar photovoltaic modules in the field, which may play a key factor on an optimal operation and maintenance of PV systems. The proposed approach uses visible spectrum RGB images of multiple solar panels and environmental data to predict each module's performance individually. The algorithm consists of three main stages. The first step is segmentation, which takes the image input and identifies every module present in the scene using Region Based Convolutional Neural Networks (RCNN) and supervised learning. In the second step, each of these regions is resized and reshaped to achieve a homogeneous format. The final step uses the processed regions and environmental data to predict the performance of each module, categorizing power loss according to a percentile classification. This step uses a convolutional neural network (CNN) designed specifically for this task. When compared to state-of-the-art computer vision architectures, the proposed approach achieved similar results with a significant reduction in computational cost. Preliminary experiments show that the classifier has an accuracy of over 73% when power loss predictions are divided into 8 percentiles ranging from 0 to 100%, where most of the errors originate from minimal differences between the actual and predicted percentiles.