Abstract
As solar energy integration grows, managing its variability due to cloud cover is critical for grid stability. Accurate short-term solar forecasting relies on ground-based sky cameras to predict cloud-Sun interactions. However, traditional Sun tracking methods often require precise camera calibration or fail when the Sun is obscured. This paper presents a robust, data-driven Sun tracking algorithm that utilizes ridge polynomial regression to model the Sun's trajectory directly in pixel space. By leveraging temporal consistency across multiple days, the method maintains tracking even during heavy cloud cover. The algorithm achieves a mean absolute error (MAE) of less than 1% of the image size without requiring GPS data or complex calibration.
Introduction
The inherent variability of solar power poses significant challenges for grid operators. A single passing cloud can cause an 80% drop in power production within seconds. To mitigate this, operators use "fish-eye" sky cameras to forecast short-term irradiance. A fundamental requirement for these forecasts is knowing the Sun's exact position in the image.
Existing methods typically rely on solar position algorithms (requiring GPS and precise camera orientation) or image-based tracking that fails when clouds hide the Sun. This research addresses these gaps by proposing a calibration-free, image-only tracking framework that remains robust under adverse weather conditions.
Methods
The proposed framework consists of three primary stages: visibility classification, localization, and trajectory prediction.
1. Visibility Classification and Localization
The system first determines if the Sun is visible or hidden using a thresholding method on the blue channel of short-exposure images, where solar contrast is highest. When visible, the Sun's coordinates are identified by calculating the median position of saturated pixels and applying a Gaussian filter to isolate the true solar disk from lens flares.
2. Trajectory Prediction via Ridge Regression
The core innovation lies in treating the Sun's path as a temporally consistent multi-day regression problem. The algorithm uses Ridge polynomial regression to model the Sun’s $x$ and $y$ coordinates as a function of time (minutes and days). This dual-axis approach allows the model to learn from the Sun's position at the same minute across previous days, effectively "filling in the blanks" when the Sun is obscured by clouds. This ensures a smooth, physically realistic trajectory regardless of instantaneous image noise.
Results and Discussion
The algorithm was evaluated using a multi-year dataset from the SIRTA lab in France, featuring sky images at 2-minute intervals.
Performance Metrics
- Visibility Accuracy: The classifier reached 94% accuracy and 98% precision, with most errors occurring during low-Sun conditions (sunrise/sunset).
- Tracking Precision: The MAE for the Sun’s position was approximately 4.7 pixels, representing less than 0.91% of the total image size. This significantly outperforms previous image-based benchmarks, which reported errors around 3.71%.
- Robustness: The model demonstrated remarkable resilience, providing smooth trajectories even during winter months with limited data or periods with up to 15 consecutive days of missing observations.
Conclusion
This research provides a "plug-and-play" solution for Sun tracking in solar forecasting. By bypassing the need for camera calibration and leveraging temporal consistency, the algorithm offers a robust tool for researchers using diverse or uncalibrated sky-camera datasets. Future work will focus on improving outlier detection near the horizon and testing the model's performance on standard long-exposure imagery.
Summary updated