Machine Aided Estimation of Solar Cell Crack Caused Power Loss from Electroluminescence Images

In this work, we present and validate a machine learning approach for estimating power loss in monocrystalline silicon PV cells due to cell cracks. The primary focus is selection and evaluation of robust and accurate prediction-features derived from electroluminescence (EL) images of solar cells. To achieve this, we have experimentally created training and validation EL and I-V measurements datasets of cells with two to up to five busbars, with different extents of cell crack degradation. We have evaluated simple features derived from the cells' luminescence histogram, which are easy to compute; these are the cell inactive area, the mean, standard deviation (STD), skewness, and kurtosis. Learning models are then trained against cell I-V measurements and validated. This cell-level approach has the advantage of being scalable so that it could be used to estimate the power loss of a full PV module based on a module EL image that is segmented in cell images. Similarly, this bottom-up approach could facilitate the development of more generalizable machine algorithms, where just one model can be used to simulate power no matter the size of the module, number, layout, or type of the cracked cells. As a baseline, a linear regression approach was attempted, which yielded a R2 score of 0.85 and a maximum error of 0.26. These results were subsequently improved by applying a 2nd degree polynomial linear regression and a custom non-linear regression to 0.869 of R2 and 0.17 of maximum error. The model selection has been chosen from a statistical exploratory analysis of the dataset, where it has been demonstrated that mono-crystalline cells of with different number of busbars behave differently in terms of histogram features correlations with the output power, and that adding a classifier variable as the busbar number as a predictor it is possible to account for the non-linearity of each cell type and model it accordingly.