Scholarlink

<p>Dust accumulation on photovoltaic (PV) panels pose a significant challenge to maintaining optimal power generation. Despite advancements in deep learning models for detecting dust on PV panels, these solutions often demand substantial computational resources. This study proposes the SoPaD2Net model, which is an enhancement of the MobileNetV2 that integrates a random data augmentation layer to boost detection performance while reducing complexity through modified dropout and dense layers. Using a combination of two datasets—one public and one self-developed, totaling 2,781 images—the model achieved an accuracy of 97.08% ± 1.20% with only 2.92&#xa0;million parameters and 0.601 giga floating-point operations per second (GFLOPs), validated through five-fold cross-validation. Compared to baseline models like ResNet50, VGG16, EfficientNet, InceptionV3, SolNet, and the original MobileNetV2, this approach demonstrates both superior accuracy and computational efficiency. Statistical validation with a t-test at a 95% confidence level further confirmed its significant performance advantage. These attributes make SoPaD2Net highly suitable for real-time dust detection on low-power devices, contributing to a more sustainable and efficient solar energy solution.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SoPaD2Net: efficient dusty solar photovoltaic panels detection using random data augmentation and adapted MobileNet architecture

  • 期刊论文

  • 发表日期：

    2025年4月15日

查看全文
  • Van-Trung Nguyen,
  • Huan Tran-Cong,
  • Isobel Timothea French,
  • Batan Le,
  • Xuan-Vien Nguyen,
  • Duc-Hung Pham,
  • Nghi C. Tran

摘要

Dust accumulation on photovoltaic (PV) panels pose a significant challenge to maintaining optimal power generation. Despite advancements in deep learning models for detecting dust on PV panels, these solutions often demand substantial computational resources. This study proposes the SoPaD2Net model, which is an enhancement of the MobileNetV2 that integrates a random data augmentation layer to boost detection performance while reducing complexity through modified dropout and dense layers. Using a combination of two datasets—one public and one self-developed, totaling 2,781 images—the model achieved an accuracy of 97.08% ± 1.20% with only 2.92 million parameters and 0.601 giga floating-point operations per second (GFLOPs), validated through five-fold cross-validation. Compared to baseline models like ResNet50, VGG16, EfficientNet, InceptionV3, SolNet, and the original MobileNetV2, this approach demonstrates both superior accuracy and computational efficiency. Statistical validation with a t-test at a 95% confidence level further confirmed its significant performance advantage. These attributes make SoPaD2Net highly suitable for real-time dust detection on low-power devices, contributing to a more sustainable and efficient solar energy solution.