Abstract

This study presents a novel algorithm to evaluate the power losses in photovoltaic (PV) modules that are exposed to naturally accumulated soiling. The proposed algorithm uses red, green, and blue (RGB) images of the PV modules captured in the visible spectrum and measures features such as the average color intensity and standard deviation from every color channel. These features are then ranked based on the module performance by using statistical tools, and the most relevant set of features is selected. The selected features are used as an input, and the module power generation is used as the label to train an artificial neural network that outputs a continuous value that represents the instantaneous performance of the evaluated module when compared with a clean module operating under the same environmental conditions. The proposed method is effective for images captured within an irradiance ranging from 700 to 1000 (W/m(2)), with the predictions achieving an R-2 score of 0.96 and RMSE of 0.74% for power losses ranging from 0% to 20% in natural soiling. This percentage corresponds to a clean condition, where irradiance is not used as an input, making it a cheap and reliable solution to monitor soiling conditions. This is the first work of its kind to demonstrate a correlation between the extracted color features of RGB images and module performance under outdoor conditions for the studied dataset. The proposed algorithm presents less accurate results when it is tested in modules exposed to non-homogenous soiling, suggesting that the proposed methodology might be effective only for naturally accumulated soiling.