Solar Cell Crack Image Generation for Power Loss Prediction

Due to the extreme thermomechanical stresses imposed on photovoltaic (PV) modules during manufacture, shipping, installation, regular operation, and extreme weather, fracture of Si cells may be inevitable. As the industry moves thinner glass and larger format modules, and as the frequency of extreme weather events continues to increase, the problem of cell cracks is likely here to stay. Therefore, when cracks are observed, it is important to understand their immediate effects on PV performance. Furthermore, the associated power loss in the years of operation following the fracture event is even more difficult to forecast, as the evolution of cell crack properties is not well understood. Currently, measuring the extent of power loss due to cell cracks in fielded PV systems is costly and requires several steps such as locating modules with cracks via electroluminescence (EL) imaging, manually removing identified cracked modules, and flash current-voltage (IV) testing identified modules with cell cracks. Then those modules are either re-installed or replaced. This process is costly, so often when a large site has widespread cell cracking and associated power loss, there are issues of identifying extent of damage and liability. Therefore, methods to costeffectively identify the current health state of modules and future risk of power loss due to cell cracks are of great interest to the solar industry, particularly asset owners. The scope of the work presented here collects three elements to form a model chain for predicting power loss due to c-Si cell cracks from EL images. First, we modify the open-source deep learning model pv-vision to segment EL images for locations of cracks and other cell features such as busbars. Second, we have developed a Variational Autoencoder (VAE) to parameterize crack geometries and their spatial relationships to cell busbars and edges. Third, power loss prediction will be a feature of our fully-trained VAE. Training of this model will include associated EL and IV data from cracked cells. The final trained VAE model will be capable of co-generation of EL and single-diode model parameters for cracked cells, as well as prediction of power loss given a previously unseen EL image.