%0 Conference Proceedings
%T Monitoring of PV Modules and Hotspot Detection Using Convolution Neural Network Based Approach
%+ Amrita Vishwa Vidyapeetham
%A Sandeep, B.
%A Saiteja Reddy, D.
%A Aswin, R.
%A Mahalakshmi, R.
%< avec comité de lecture
%@ 978-3-031-16363-0
%( IFIP Advances in Information and Communication Technology
%B 5th International Conference on Computational Intelligence in Data Science (ICCIDS)
%C Virtual, India
%Y Lekshmi Kalinathan
%Y Priyadharsini R.
%Y Madheswari Kanmani
%Y Manisha S.
%I Springer International Publishing
%3 Computational Intelligence in Data Science
%V AICT-654
%P 311-323
%8 2022-03-24
%D 2022
%R 10.1007/978-3-031-16364-7_24
%K Convolution neural network
%K PV modules
%K Detection of hotspots
%K CNN
%K TensorFlow
%K Solar panels
%Z Computer Science [cs]Conference papers
%X The use of solar photovoltaic systems in green energy harvesting has increased greatly in the last few years. Fossil fuels reaching the end are also growing rapidly at the same rate. Despite the fact that solar energy is renewable and more efficient, it still needs regular Inspection and maintenance for maximizing solar modules’ lifetime, reducing energy leakage, and protecting the environment. Our research proposes the use of infrared radiation (IR) cameras and convolution neural networks as an efficient way for detecting and categorizing anomaly solar modules. The IR cameras were able to detect the temperature distribution on the solar modules remotely, and the convolution neural networks correctly predicted the anomaly modules and classified the anomaly types based on those predictions. A convolution neural network, based on a VGG-based neural network approach, was proposed in this study to accurately predict and classify anomalous solar modules from IR images. The proposed approach was trained using IR images of solar modules with 5000 images of generated solar panel images. The experimental results indicated that the proposed model can correctly predict an anomaly module by 99% on average. Since it can be costly and time-consuming to collect real images containing hotspots, the model is trained with generated images rather than real images. A generated image can be used more efficiently and can also have custom features added to it. In the prediction process, the real image is processed and it is sent to the model to determine bounding boxes. It provides a more accurate prediction than direct use of the real image. Here we have used CNN custom model and TensorFlow libraries.
%G English
%Z TC 12
%2 https://inria.hal.science/hal-04381277/document
%2 https://inria.hal.science/hal-04381277/file/526570_1_En_24_Chapter.pdf
%L hal-04381277
%U https://inria.hal.science/hal-04381277
%~ IFIP-LNCS
%~ IFIP
%~ IFIP-AICT
%~ IFIP-TC
%~ IFIP-TC12
%~ IFIP-ICCIDS
%~ IFIP-AICT-654