MRR-YOLO: an instance segmentation technique for ground-based cloud images.

Area of Science:

• Artificial Intelligence
• Renewable Energy Systems
• Computer Vision

Background:

• Cloud cover significantly impacts solar energy generation efficiency and system longevity.
• Accurate cloud detection is crucial for optimizing solar power plant operations and reducing costs.

Purpose of the Study:

• To propose and evaluate the MRR-YOLO model for enhanced cloud segmentation in solar energy applications.
• To investigate the effectiveness of specific modules (MSDA, RCS-OSA, RFAConv) within the MRR-YOLO framework.

Main Methods:

• Deep learning and instance segmentation techniques were employed.
• The MRR-YOLO model incorporated MSDA, RCS-OSA, and RFAConv modules for cloud feature extraction.
• Performance was evaluated using metrics like Precision (PB, PM), Recall (RB, RM), and mean Average Precision (mAP50B, mAP50M).

Main Results:

• The MRR-YOLO model demonstrated superior cloud detection compared to semantic segmentation, particularly for varied cloud shapes.
• The RFAConv module proved effective for cloud feature extraction, while MSDA maintained model efficiency.
• The model achieved high performance metrics: PB 79.2%, RB 66%, mAP50B 74.7% for detection and PM 79.3%, RM 64.8%, mAP50M 73% for segmentation.

Conclusions:

• The MRR-YOLO model offers improved detection performance and applicability for cloud segmentation in solar energy contexts.
• Validation across multiple datasets and real-world images confirms the model's robustness and accuracy.
• The study highlights the potential of advanced deep learning models for optimizing solar energy system management.