Deep Learning-Based Health Monitoring for Photovoltaic Systems

• Published in: IEEE journal of photovoltaics Vol. 15; no. 4; pp. 577 - 592
• Main Authors: Alnuaimi, Khaled, Al-Sumaiti, Ameena Saad, Alansari, Mohamad, Wang, Huai, Al Hosani, Khalifa Hassan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.07.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Anomalies; Anomaly detection; Artificial intelligence; Computer science; Deep learning; Deep learning (DL); Electrical engineering; Energy management systems; Environmental monitoring; Fault diagnosis; health monitoring; Inspection; Location awareness; Manuals; Monitoring; object detection; photovoltaic (PV); Photovoltaic cells; Plant monitoring; renewable energy; Renewable energy sources; Semantic segmentation; Solar panels; Unmanned aerial vehicles
• ISSN: 2156-3381; 2156-3403
• DOI: 10.1109/JPHOTOV.2025.3563887

The transition to renewable energy sources like photovoltaic (PV) systems is essential for societal progress, counteracting the adverse effects of fossil fuels. However, managing PV systems entails significant challenges and economic implications. PV fault occurrence necessitates swift detection and resolution, exacerbating financial burdens. Effective fault diagnosis relies heavily on data from PV plant monitoring and energy management systems. Historically, PV monitoring relied on manual inspections, but autonomous aerial vehicle (UAV) technology provides a more efficient and comprehensive solution, enhancing safety and offering detailed imagery, scalability, environmental monitoring, and advanced data analytics. This study utilizes deep learning (DL) approaches to monitor the health of the PV, focusing on analyzing UAV-captured scenes. Specifically, this article presents an end-to-end two-stage DL-based health monitoring framework that consists of semantic segmentation model, SegFormer, for isolating solar panels and object detection model, YOLOv8, for identifying anomalies within the PV modules. The proposed framework is validated and compared with state-of-the-art (SOTA) models on a three publicly available UAV-captured datasets. Results show improvements of 25.8% and 1.5% in solar panel segmentation, and 26.6% in solar panel anomaly detection compared with recent SOTA models.