Browse by author
Lookup NU author(s): Dr Wenxian YangORCiD
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND).
With the continuously increasing application of photovoltaic (PV) panels, how to effectively manage these valuable facilities has become an issue of concern. To date, some methods have been developed to meet this purpose. However, to date, a satisfactory solution has not been achieved for managing large-scale solar PV power plants. To address this issue, a new PV panel condition monitoring and fault diagnosis technique is developed in this paper. The new technique uses a U-Net neural network and a classifier in combination to intelligently analyse the PV panel’s infrared thermal images taken by drones or other kinds of remote operating systems. The novelties of this research include: (1) a U-net neural network is developed and trained to carry out image segmentation, thereby significantly improving the efficiency of image processing; (2) the condition monitoring and fault diagnosis is based on the contour features in the ‘mask’ of true colour infrared images. As compared to the true colour images, their mask images have little interference information, the reliability and accuracy of condition monitoring and fault diagnosis based on mask images can be guaranteed to a large extent. In the research, 295 infrared images were taken first from the PV panels in different health states, and then their ‘masks’ were manually created using the software LabelMe. Secondly, enlarge the number of image samples using three image expansion methods (i.e. mirroring left and right, flipping up and down, and cropping and then zooming in) to establish the image sample database for training and testing the U-net neural network. Thirdly, use 1852 infrared images and their mask images stored in the database to train a U-Net neural network until the trained U-net neural network can create mask images as accurately as the software LabelMe does. The trained neural network will guarantee that the achieved segmentation accuracy can be as high as 95.2%. Finally, four potential criteria (i.e., Contour area, Perimeter, Aspect ratio and Ratio of contour area to the area of contour outer rectangle) are proposed to characterise the contour of mask images and then based on the calculation results of these four criteria, different types of PV panel faults are diagnosed with the aid of three kinds of classifiers. The classifiers are Decision tree, K-Nearest Neighbours algorithm (KNN), and Support-vector machine (SVM). The research results have shown that the combined use of a well-trained U-Net neural network and Decision tree can diagnose the PV panel faults with 99.8% accuracy. Therefore, it may arguably provide a promising intelligent tool for condition monitoring the PV panels.
Author(s): Wang X, Yang W, Yang B, Wei K, Ma Y, Zhang D
Publication type: Article
Publication status: Published
Journal: Energy Reports
Year: 2022
Volume: 8
Pages: 5005-5015
Print publication date: 01/11/2022
Online publication date: 09/04/2022
Acceptance date: 24/03/2022
Date deposited: 24/03/2022
ISSN (electronic): 2352-4847
Publisher: Elsevier
URL: https://doi.org/10.1016/j.egyr.2022.03.173
DOI: 10.1016/j.egyr.2022.03.173
Altmetrics provided by Altmetric
