A CNN-LSTM model using elliptical constraints for temporally consistent sun position estimation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Renewable Energy Systems

Background:

• Accurate sun position estimation is crucial for solar power, weather forecasting, and augmented reality.
• Existing image-based methods struggle with cloud cover due to single-image reliance.
• Partial or complete sun occlusion significantly degrades performance of current systems.

Purpose of the Study:

• To develop a robust deep learning system for precise sun position estimation.
• To overcome limitations of single-image methods, especially under occluded conditions.
• To enhance the reliability of solar energy and augmented reality applications.

Main Methods:

• A deep learning model leveraging spatial, temporal, and geometric features.
• ResNet-based convolutional networks for spatial feature extraction from image sequences.
• LSTM layers to process temporal brightness changes and elliptical constraints for natural sun path adherence.

Main Results:

• Achieved an R² Score of 0.98 on diverse datasets (Sirta, Laval, custom).
• Outperformed previous approaches by at least 0.1 in accuracy.
• Demonstrated capability in estimating sun position even when occluded.

Conclusions:

• The proposed deep learning approach offers superior accuracy and robustness for sun position estimation.
• The system effectively handles partial and complete sun occlusion.
• Enables a simplified, camera-only sky imaging system, replacing complex sensor arrays.