login

Search documents

Browse topics

Document details

 
Title:
 
Deep Learning Based Image Feature Extraction for Predicting Climate Related Degradation of PV Modules
 
Author(s):
 
L. Neumaier, J. Scherer, C. Schwarzlmüller, B. Kubicek, F. Mödritscher, C. Hirschl
 
Keywords:
 
Accelerated Aging, Image Processing, Electroluminescence Imaging, Machine Learning, Lifetime Prediction, Deep Learning
 
Topic:
 
PV Systems – Modelling, Design, Operation and Performance
Subtopic: Operation, Performance and Maintenance of PV Systems
Event: 38th European Photovoltaic Solar Energy Conference and Exhibition
Session: 5CV.2.11
 
Pages:
 
1207 - 1211
ISBN: 3-936338-78-7
Paper DOI: 10.4229/EUPVSEC20212021-5CV.2.11
 
Price:
 
 
0,00 EUR
 
Document(s): paper, poster
 

Abstract/Summary:


To avoid long time failures of PV power plants, reliability testing and modeling of novel photovoltaic (PV) material combinations has gained importance within the last years. In this work, conducted within the research project ADVANCE!, we present an image based deep learning approach to extract degradation-relevant information from aged PV modules. The basis of the work is a comprehensive database, established within the project INFINITY, which includes extensive time-resolved measurement and characterization data of aged PV modules, that have been subjected to precisely defined accelerated aging scenarios. Within this work, electroluminescence images serve as input for the development and application of a deep learning approach to extract degradation parameters. For defect classification and localization, we use a Mask R-CNN architecture, which allows instance-level segmentation and classification in parallel. The extracted information can be used for subsequent statistical analysis, where additional information is available, to make predictions about the operational lifetime of a module under different climatic conditions. Although the amount of data is comparatively small and the distribution of the defect classes within the dataset is not uniform, the available results show the potential of our approach.