Improvements in short-term solar energy forecasting
The field of view of the developed imaging system is approximately 2 miles, which yields in up to 10 min prediction horizons at ~ 15 miles per hour cloud movements. After some geometrical corrections of the cloud images optical flow algorithms are applied, which effectively remove the "background" in the image by distinguishing between moving objects (i.e, clouds) and stationary objects (e.g, trees etc.). Cross-correlation between consecutive cloud images yields the direction and speed of the cloud, which in combination with the sun's position and the location of the solar panels is used to predict the solar energy. Although the cloud recognition algorithm is (so far) binary we adjusted the level of solar energy using a modification function F, which depends on time of the day and is based on past data for a "clear" and "cloudy" days (for a given date of the year). Further adjustments (i.e., level of shading) are made continuously using the actual measurements of the solar panels, which are available a couple of minutes later after the prediction was made.
In this paper we demonstrate how this approach was used to make continuously short-term forecasts (<10 min ahead) for a small solar installation. The average forecast error ε was estimated by ε=∫|PM(t)-PF(t)|dt / ∫PM(t)dt with PF(t) and PM(t) as the forecasted and produced/ measured power, respectively yielding 10% for a 6 min forecasts and slightly less for a 3 minutes forecasts, which is one of the highest reported to date.
Acknowledgement: The work is partially funded by the DOE SunShot program.