An automated data analysis pipeline is developed to preprocess electroluminescence (EL) module images, and parse the images into individual cells to be used as an input for machine learning algorithms. The dataset used in the study includes EL images of three 60 cell modules from each of five commercial brands at six steps of damp heat exposure, from 500 to 3000 h. Preprocessing of the original raw EL images includes lens distortion correction, filtering, thresholding, convex hull, regression fitting, and perspective transformation to produce planar indexed module and single cell images. Parsing of PV cells from each of the preprocessed 90 EL module images gives us 5400 cell images, which are function of module brand and damp heat exposure step. From the dataset, two unique degradation categories ("cracked" and "corroded") were observed, while cells that did not degrade were classified as "good." For supervised machine learning modeling, cell images were sorted into these three classes yielding 3550 images. A training and testing framework with 80:20 sampling ratio was generated using stratified sampling. Three machine learning algorithms (support vector machine, Random Forest, and convolutional neural network) were trained and tuned independently on the training set and then given the test set to predict the scores for each of the three models. Five-fold cross validation was done on training set to tune hyper-parameters of the models. Model prediction scores showed that convolutional neural network outperforms support vector machine and Random Forest for supervised PV cell classification.
The authors demonstrate the feasibility of quantifying cell-level performance heterogeneity from module-level I–V curves by determining conditions of bypass diode turn-on. Analysis of these curves falls outside of typical diode-based models of photovoltaic (PV) performance. The authors show that this approach can leverage statistical and machine learning techniques for broad application to massive datasets, and combine those insights with simulations and laboratory-based experiments to provide useful information into the metastability of the interfaces of a PV cell. The authors find good agreement between the experimentally determined curves and the simulated curves, which guide the variable selection in the massive dataset collected from sites in Cleveland, OH, USA, the Negev Desert, Israel, Isla Gran Canaria, Spain, and Mount Zugspitze, Germany.
We present a description of an electroluminescence (EL) apparatus, easily sourced from commercially available components, with a quantitative image processing platform that demonstrates feasibility for the widespread utility of EL imaging as a characterization tool. We validated our system using a Gage R&R analysis to find a variance contribution by the measurement system of 80.56%, which is typically unacceptable, but through quantitative image processing and development of correction factors a variance contribution by the measurement system of 2.41% was obtained. We further validated the system by quantifying the signal-to-noise ratio (SNR) and found values consistent with other systems published in the literature, at SNR values of 10-100, albeit at exposure times of greater than 1 s compared to 10 ms for other systems. This SNR value range is acceptable for image feature recognition, providing the opportunity for widespread data acquisition and large scale data analytics of photovoltaics.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.