Solar Irradiance Forecasting with Fisheye Images

Presenter Information

• Name: Vanessa Leguin
• Position: PhD student
• Affiliation: EDF (French Electricity Company) and CNARM, supervised by Professor Nicolato
• Event: CVPR Omnicv Workshop

Introduction

• Solar photovoltaic power generation is a promising renewable energy source.
• Challenges exist due to the intermittent nature of solar energy, necessitating accurate forecasting for grid integration.

Tools for Solar Energy Forecasting

• Different tools are used based on forecasting horizons:
  • Short-term (Intraday): Use of satellite images and local weather data.
  • Very Short-term: Fisheye images to extrapolate cloud motion.

Meteorological Campaign

• Location: La Reno Island
• Data Collection:
  • Fisheye images captured every 10 seconds.
  • Pyranometers measured solar irradiance components.
• Dataset contains millions of images labeled by solar irradiance.

Machine Learning Approach

• Objective: Forecast solar irradiance using only camera data (cost-effective compared to high-grade pyranometers).
• Forecasting Steps:
  1. Calibration of fisheye cameras using calibration patterns and the Ochem Calip toolbox.
  2. Projection of hemispherical images onto a plane.
  3. Compute optical flow between images to determine cloud motion.
  4. Future images are generated using segmentation methods and machine learning techniques.

Deep Learning and Cloud Motion Forecasting

• Challenges: Extrapolating cloud motion is complex.
• Proposed solution: Incorporate physical knowledge into deep neural networks for better regularization.
• Model Presented: Feed Net, a deep video prediction model using learned partial differential equations.
  • Architecture: Two branches (physical and residual dynamics) that operate in latent space.
  • Mechanism:
    • Normal physical prediction followed by correction with input images (data assimilation).

Contribution: Fitness Jewel Model

• Improved version of Feed Net with specific encoders and decoders.
• Context vector summarizes previous images to predict future solar irradiance values.
• Results:
  • Better performance in predicting solar irradiance and images compared to baseline models (e.g., STM, Thread RNN, and Feed Net).
• Qualitative Results:
  • Accurately predicts rapid variations in irradiance.
  • Demonstrated challenging cloud motion predictions.

Future Directions

• More specific physical models for atmospheric phenomena.
• Extension to probabilistic forecasting for uncertainty estimation to assist grid managers.
• Explore cooperation between networks of fisheye cameras over small territories.

Conclusion

• Emphasized the importance of integrating physical insights with machine learning for enhanced solar irradiance forecasting.