Presentation Details
| Physics-Informed Deep Learning for Current-Voltage Curve prediction of Photovoltaic devices from Electroluminescence Images (yes) Brandon K Byford1, 2, Laura E Boucheron2, Norman R Jost1, Jennifer L Braid2. 1Sandia National Labs, Albuquerque, NM, USA.2New Mexico State University, Las Cruces, NM, USA |
Abstract
Luminescence imaging and deep learning offer fast characterization of of photovoltaic (PV) devices. A common approach is to take a generalized network like ResNet or ViT and transfer learn with a dataset for a specific problem like module characterization or defect detection. The transfer learning approach has shown promising results with accurate characteristics predicted by several different PV researchers, however, proper datasets for training a generalized network able to identify or characterize various types of damage are difficult to create. Module datasets needed to train networks for defect identification or module characterization in production environments are often lacking in defect representations. Specialized datasets involving mini-modules or single cells focused on singular defect types are numerous but lack the scope for module level network. We demonstrate a physics-informed network leveraging domain knowledge of the PV module electrical layout. The network is able to use data from PV cells as well as mini-modules, and full size modules. The proposed network starts by extracting individual cell features with a convolutional network. These cell features are fed into a transformer for current-voltage (IV) curve prediction. The transformer allows for the network to be trained and tested with data from cells, mini-modules, or even module-level electroluminescence (EL) images for IV curve prediction. The network performs similarly to a generalized transfer learned approach while retaining the adaptability for cell, mini-module, or module data.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.