A multi-scale photovoltaic (PV) panel dust accumulation simulation dataset based on physical consistency modeling

The lack of high-quality image datasets with quantifiable labels hinders the development of deep learning models for photovoltaic (PV) panel dust detection. To address this, this paper proposes a novel physical consistency modeling approach to generate a large-scale synthetic dataset with precise dust concentration labels. A particle cluster generation model is developed to simulate multiscale, heterogeneous dust aggregation governed by lognormal distributions, while an adaptive error feedback mechanism ensures accurate concentration estimation. Furthermore, an optical attenuation model based on the Beer-Lambert law and nonlinear hue, saturation, and value (HSV) color mapping is employed to ensure visual realism. Multi-dimensional evaluations demonstrate that the synthetic images achieve a histogram similarity exceeding 0.86, an entropy similarity above 0.94, and a comprehensive similarity score over 0.82 compared to real-world ground truth. These results significantly outperform conventional mask-based and generative adversarial networks (GANs) techniques, providing a reliable data source for training advanced dust detection algorithms.