Comparative study of pre-trained CNN models for multiclass fault detection in solar panels

Abstract

Solar power as renewable energy can be an alternative to fossil-based power where it is carbonless, environmentally friendly, and combats climate change. In solar panel systems, manual assessment by personnel is time-demanding and prone to human error, necessitating automated solutions. The implementation of smart systems that can automatically detect objects that hinder the solar panel from receiving solar energy can be very helpful in reducing the potential threat of decreasing performance in power generation. This study proposes a convolutional neural networks (CNN)-based image classification approach to automatically identify common solar panel conditions using visual data. The dataset used was a public dataset titled “Solar Panel Images: Clean and Faulty Images”, obtained from Kaggle, containing six classes for multiclass classification. The most effective pre-trained CNN-based deep learning model for uncovering issues in solar panels was inception-V3, achieving an overall accuracy of 91% and the highest F1-score in four categories: clean, electrical damage, physical damage, and snow coverage. These outcomes confirm the potential of implementing deep learning image classification for enhancing solar panel maintenance and monitoring systems in real-world applications.