Estimating PV Soiling Loss Using Panel Images and a Feature-Based Regression Model

Abstract

Solar energy from solar photovoltaics (PV) has become a rapidly growing sustainable energy source around the world. However, maintaining PV system efficiency remains a challenging problem. In desert regions, soiling is one of the most significant environmental factors that can cause PV system loss. In our early work, a PV soiling loss estimation method based on a single-image feature and in-lab testing was developed. In this study, we extend our previous work by incorporating various image features in a machine-learning regression model to predict PV soiling loss. The new model is trained and tested using PV performance data and RAW panel images collected in the field over several months, covering real-time soiling loss levels up to about 28%. There are 479 RAW images with 21 unique soiling loss levels, which were taken under different camera settings. The results show that the new method can reliably predict the soiling loss when the images are taken under similar settings as the training data (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i>-squared value of 0.98 and normalized RMSE is 0.01 for the training dataset).