Operational Real Time Intra-Day Solar Irradiance and Variability Forecasts based on Computer Vision of Processed Satellite Images Enhanced with Machine Learning Algorithms

The currently bearable amount of energy converted from fluctuating and intermittent renewable energy sources in the existing power grid has reached its limit in many areas of highly industrialized countries. This imposes a technological constraint limiting potential growth rates of energy converted from renewable sources penetrating the grid. Besides the expansion of storage systems, high-fidelity solar forecasting engines covering all temporal and spatial horizons of interest are capable of increasing the maximum feasible solar energy grid penetration. Recent advances in the field provided significant improvements on accuracy and temporal resolution for short term (intra-minute to several minutes ahead) and long term (multiple hours to several days) solar forecasts. However, operational forecasts with high forecasting skills for time horizons 30 minutes to 240 minutes ahead are still deficient.

This contribution seeks to present recent advances in this area based on multiple computer vision techniques such as Optical Flow, Particle Index Velocimetry and Difference Centroid Algorithm. These methods are applied to satellite imagery to test their capabilities in cloud movement detection and cloud tracking. The accuracy of these techniques is evaluated based on the performance of cloud movement and irradiance forecasts. To increase the skill of the forecasting engine, various other available inputs (e.g statistical cloud cover indices, prevalent local movement patterns and image features of cloud segmented images in proximity to the location of interest) are explored and optimized with different machine learning techniques. Generic Algorithms are applied to enable the efficient selection of input parameters and to reduce calculation costs.

Findings show that our approach is suitable to forecast the solar irradiance on multiple time horizons with significant optimization potential. Our progress and preliminary results are valuable for independent system operators, utility companies and system operators.