Snow loss modeling for solar modules using image processing and deep learning

Abstract

Energy loss resulting from partial or full snow coverage on solar modules, such as photovoltaics (PV), poses serious challenges to the efficiency of renewable sources in cold climates. This study introduces a new method to quantify the impact of snow on installed PV panels using image processing and deep learning (DL) techniques. A set of 44 PV panels in four different arrays were examined in the experiment through per-pixel analysis that was correlated with various performance indicators. Results of two different PV selection methods, Direct Selection (DS) and Perspective Transformation (PT), were compared, with an acceptable deviation that ranges from 0.1% to 7.35% of one panel area. Energy production was computed where the estimated lowest average was 0.0368 kWh/panel with DS and 0.0395 kWh/panel with PT, due to high uncovered PV ratio with an average value of 84.1%. The validation with alternative methods resulted with an average deviation of 1.21%. Neural networks proved to offer strong correlation between input/output pairs with an average R value of 0.9. Additional uncertainty analysis showed that the error for area coverage detection increased due to higher snow deposition rates from 1.8% to 3.7%. Overall, remote monitoring of PV panels can be predicted to enhance energy production performance, while using DL-aided digital photography to track accumulation.