Detecting Snow Layer on Solar Panels using Deep Learning

Ozturk O., Hangun B., Eyecioglu O.

10th IEEE International Conference on Renewable Energy Research and Applications, ICRERA 2021, İstanbul, Turkey, 26 - 29 September 2021, pp.434-438 identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/icrera52334.2021.9598700
  • City: İstanbul
  • Country: Turkey
  • Page Numbers: pp.434-438
  • Keywords: Convolutional neural networks, deep learning, solar panel defect detection, solar panels, transfer learning
  • Istanbul Gelisim University Affiliated: Yes


© 2021 IEEE.Renewable energy now plays a significant role in meeting rising energy demand while protecting the environment. Solar energy, generated by enormous solar panel farms, is a rapidly developing environmentally friendly technology. However, its efficiency degrades due to some factors. The climate is one of the most impactful factors that affect the electricity generation of a photovoltaic cell - especially countries with snowy climates face those downside effects. Hence, detection and removal of the snow layer on the solar panels are crucial. Firstly, most of the snow detection approaches are based on time series or momentary sensor data. Secondly, the removal of snow is based on surface coatings, heating, and mechanical clearing. Nowadays, vision-based solutions for detecting and removal of snow are trending. Since eliminating the human factor is a priority in physical labor, drones are suitable for vision-based operations. This paper presents a new deep learning-based approach that can be deployed on drones for detecting snowy conditions on solar panels using deep learning-based algorithms. As they are state-of-the-art neural networks in computer vision applications, ResNet-50, VGG-19, and InceptionV3 have been selected. In order to increase generalization in the training phase, we augmented the dataset using different image manipulation techniques. Our results show that we obtain 100%, 99%, and 91% F1-Score from InceptionV3, VGG-19, and ResNet-50 respectively.