SOLAR PANEL CONTAMINATION CLASSIFICATION USING MACHINE LEARNING

Abstract

This thesis introduces an automated methodology for solar panel inspection using deep learning techniques to enable faster and accurate inspections. Traditional inspection techniques frequently include manual inspections, which can be demanding of labor and vulnerable to human error. A dataset of slightly over 4,000 high-resolution solar panel photos has been collected under diverse environmental conditions. Each image is annotated by its contamination type (dust, snow, water droplets, or clean) and the associated electrical measurements (voltage and current). Contamination on solar panels, such as dust or snow, may block sunlight and reduce the panel’s effectiveness by limiting the light that reaches the photovoltaic cells. The main goals are to precisely categorize the type of contamination present on each panel image and to predict important electrical characteristics. Various convolutional neural network (CNN) designs were evaluated such as VGG, ResNet, EfficientNetB2, and ConvNeXt Tiny. Initially, several models were created for classification and regression tasks to examine the influence of network depth and complexity on performance. Based on these observations, multi-head models were developed to provide simultaneous learning of both tasks within a singular backbone. During testing, EfficientNetB2 achieved the highest classification accuracy at 97% on unseen test images, but ConvNeXt Tiny resulted in the most accurate regression predictions for voltage and current, achieving the highest R² among the assessed architectures. In conclusion, this research introduces the potential of advanced deep learning models to offer a more consistent and objective alternative to manual panel inspections, thereby offering information about potential power losses and surface contamination. The proposed method provides practical data for maintenance making decisions by incorporating regression and classification capabilities. The software prototype that has been created also demonstrates the ability of an intuitive interface to assist individuals in the uploading of images, immediately receiving contamination labels, and the estimation of electrical parameters.