Intelligent dust monitoring and cleaning optimization on photovoltaic panels

Abstract

Dust deposition on photovoltaic (PV) panels is a significant operational issue, often leading to power losses exceeding 15–30% in regions with high airborne particle concentrations. Although numerous studies have investigated either visual detection of dust or analytical estimation of performance loss, most approaches focus on a single task and provide limited practical insight for real-time maintenance. This work introduces a dual-task deep learning framework that simultaneously classifies dust severity and predicts the corresponding power loss from panel images. Five recent architectures vision transformer (ViT), swin transformer, GhostNet, DenseNet, and MobileNetV2 are employed as backbone feature extractors, with extracted embeddings processed by a multi-head multi-layer perceptron (MLP) combining shared representation learning with separate classification and regression outputs. The system is trained and evaluated on a real-world dataset of PV panels, and performance is assessed using accuracy and mean absolute error. DenseNet achieves the highest accuracy (94%) and lowest prediction error, while lightweight convolutional neural network (CNN) backbones demonstrate the best balance between precision and computational efficiency. By integrating hybrid processing and dual predictive capability, the proposed method offers a more comprehensive and deployable solution for automated PV monitoring compared to existing single-output approaches.