Neural network automates the identification of dirty solar panels

Over the past two decades, solar panels have exploded in popularity, and they now produce more than two terawatt-hours of energy worldwide. However, when photovoltaic modules get dirty, their energy efficiency drops and hot spot effects arise that can accelerate a device’s degradation.

To improve the monitoring of solar panel cleanliness, Li et al. developed a neural network that can automate the detection of dusty photovoltaic modules. The network, named Photovoltaic Dust Perception UNet (PDP-UNet), was trained on an artificial dataset the researchers generated with an optical blending model. The dataset included three common types of dirt — red soil, urban dust, and sand — as each has unique particles sizes, colors, and densities that affect light transmittance in different ways.

As a multi-task perception method, PDP-UNET classifies dust type, identifies dusty regions on the photovoltaic module, and determines transmittance values. Using a shared “Residual-SE” encoder, the model can perform these three simultaneous tasks efficiently. This allows for the quantification of dust accumulation at fine-grained spatial resolution, which can be used to assess the cleanliness of a solar panel.

“We are most excited about the model’s performance on real-world photovoltaic images,” said author Yu Shen. “Even though it was trained on synthetic data, it successfully identifies dust patterns in practical inspection scenarios.”

The researchers’ results showed the network consistently outperforms mainstream baselines in multiple metrics. With pixel-level loss analysis, the authors hope the network can enable cleaning prioritization and be integrated into automated inspection pipelines using drones or fixed-angle cameras.

Source: “A transmittance-driven multi-task perception method for dust region analysis in photovoltaic images,” by Chunrong Li, Zhaoyun Li, Yiming Liu, Xinyi Chen, Yu Shen, Kanjian Zhang, and Haikun Wei, Journal of Renewable and Sustainable Energy (2026). The article can be accessed at https://doi.org/10.1063/5.0301351 .