Solar Panel Dust Detection: An Efficacy Analysis and Optimization using Deep Learning based Techniques

Monitoring and cleaning solar panels is an essential task in countries where output power loss due to dust accumulation is among the highest rates. Therefore, it is crucial to develop an optimal procedure for monitoring and cleaning these panels to increase module efficiency, reduce maintenance costs, and reduce the use of resources. To the best of our knowledge, this is the first research to perform a comprehensive performance analysis using seven pre-trained Convolutional Neural Networks (CNNs) such as VGG16, VGG19, InceptionNet, MobileNet, MobileNetv2, Xception, and ResNet50 CNNs to detect dust in solar panel images. The problem is formulated as binary classification, as the dataset used in this research includes clean and dusty solar panel images. The study conducted a comprehensive performance analysis of the solar panel's dust classification task. The study also fine-tunes the hyperparameters to enhance the performance of the best-performing model during implementation, thereby further enhancing its performance. Numerous performance evaluation metrics have been implemented, including accuracy, precision, recall, F1 score, and AUC score. MobileNet can classify clean and dusty solar panels with initial accuracy (83.6%) After the hyperparameter tuning using learning rate and batch size, the model achieved 86.7% accuracy, which has been improved as compared to the initial performance.