Anmerkungen:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Beschreibung:
The performance of a PV string can be assessed by extracting quantitative information from the electroluminescence images of each of the PV modules included in the string. In this work, we propose a method to predict PV module IV curves from electroluminescence (EL) images using a deep learning algorithm. The proposed method consists of creating eleven deep learning models that predict ten points on the IV curve, including ISC, Impp, Vmpp and VOC. We use 376 samples (EL images and corresponding IV curves) for training the deep learning models, 114 samples for validating the models, and 92 samples for testing the models, whereas 84 samples were measured in the lab and 8 samples were measured in the field. The dominant fault present in the PV modules used in this work is cell cracks resulting in disconnected areas. Results show that the PV module IV curves are predicted with a mean absolute error for all points below 5 W. This indicates that the deep learning models are able to find a relationship between the inactive areas present in the PV module EL image and the respective PV module IV curve. The module power that were used in this research experiment range from 210 W to 120 W. To compare field measurements to deep learning predictions, we corrected Pmpp predictions to values at Standard Testing Conditions, and at the same conditions we measured modules with the flasher. The results from the field experiment showed that the Pmpp prediction error of the high power modules in relation to the ground truth Pmpp values was less than 7 % or 16 W. While the low power modules with a high number of inactive cell areas presented a Pmpp prediction error of 15 % or 20 W.