SoilingEdge: PV Soiling Power Loss Estimation at the Edge Using Surveillance Cameras

Solar panels are exposed to various pollutants in outdoor environments, such as dust, sediment, and bird excrement, which can cause the power generated by the panels to drop by up to 50%. To accurately estimate the power generated by photovoltaic (PV) systems, it is necessary to take into account th...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on sustainable energy 2024-01, Vol.15 (1), p.556-566
Main Authors: Zhang, Wenjie, Archana, Vaidheeswaran, Gandhi, Oktoviano, Rodriguez-Gallegos, Carlos D., Quan, Hao, Yang, Dazhi, Tan, Chin-Woo, Chung, C. Y., Srinivasan, Dipti
Format: Article
Language:English
Subjects:
Cameras
Comparative studies
Computational modeling
Deep learning
Edge computing
Field programmable gate arrays
Image edge detection
Kernel
MobileNet
Photovoltaic cells
Photovoltaics
Platforms
power loss
Shape
soiling
Soils
Solar energy
Solar panel
Solar panels
Surveillance
Tensors
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Solar panels are exposed to various pollutants in outdoor environments, such as dust, sediment, and bird excrement, which can cause the power generated by the panels to drop by up to 50%. To accurately estimate the power generated by photovoltaic (PV) systems, it is necessary to take into account the effects of soiling on the panels. In this article, we propose a deep learning approach that uses edge devices such as micro-controllers to estimate the power loss due to soiling based on images captured by surveillance cameras. The proposed model, called SoilingEdge, is based on MobileNet and has been implemented on several platforms, including CPUs, edge tensor processing units (EdgeTPUs), field programmable gate arrays (FPGAs), and vision processing units (VPUs). We present a comparative study of the performance of SoilingEdge on these different platforms and show that FPGA offers a balanced compromise between cost and inference performance, while VPU is the most cost-effective option, but has a slower execution time. Furthermore, the article also investigates the interpretability of the model through the visualization of attention maps, thus providing useful insights for researchers and engineers in understanding the model behaviors.
ISSN:1949-3029
1949-3037
DOI:10.1109/TSTE.2023.3320690