Deep Learning for Dust Accumulation Analysis on Desert Solar Panels: A CNN-Transformer Approach

In the energy transition towards sustainability, photovoltaic power is increasingly valued for its eco-friendly and renewable attributes. Northern and northwestern China’s deserts, abundant in solar energy and vast in land, are prime areas for photovoltaic generation. High winds and dust storms in these regions cause dust buildup on PV panels, reducing efficiency and potentially harming the equipment. In water-scarce deserts, cleaning photovoltaic panels is particularly difficult. Therefore, developing robust and precise dust accumulation detection techniques is crucial to ensure the stable operation of PV power plants in deserts. To address this challenge, this study proposes a solution based on the convolutional neural network (CNN) structure UTran-ResNet50, a model capable of accurately recognizing images with various levels of dust coverage on photovoltaic panels. By combining ResNet50 and Transformer architectures, their strengths in global and local feature extraction are effectively exploited. In addition, the study incorporates the Attention U-Net module to reduce the recognition error and introduces the Positional Attention module to enhance the feature recognition capability. It is particularly noteworthy that the experimental validation conducted with a homemade dataset, despite its limited sample size, demonstrates the high efficacy of the proposed network, achieving an accuracy of 96.41%. It exemplifies the model’s robustness in dealing with small sample datasets, highlighting its superior performance in accurately detecting dust accumulation on photovoltaic panels.