A computationally light data-driven alternative to cloud-motion prediction for PV forecasting
No Thumbnail Available
Author
Carrillo, Rafael E.
Alet, Pierre-Jean
Müller, Stephan
Remund, Jan
Abstract
In this paper we compare a state-of-the-art cloud-motion tracking solution against a data-driven solution for photovoltaic (PV) power production to provide insights into their performance drivers. The solutions are MeteoTest’s CloudMove, based on cloud-motion tracking from satellite images, and CSEM’s data-driven solution based on graph machine learning forecasting models that use only past PV production data from a network of PV systems. The study compared the two approaches in terms of accuracy in different scenarios for a forecasting horizon of up to six hours ahead, with a resolution of 15 minutes, on a dataset of 21 days and 18 locations in Switzerland. The days and locations were selected as a representative sample of the whole range of different conditions in terms of seasons, weather conditions, terrain, and distance to other instrumented sites. Over the whole benchmarking set, CSEM’s data-driven technique yields a lower normalized root-mean-square error than the cloud-motion tracking method for forecasting horizons above one hour, though the error spread is larger with CSEM’s data-driven solution. Regarding computational load, the data-driven methods can accelerate the computation of forecasts by a factor 100, after an initial training, thus offering a viable alternative to satellite-based cloud-tracking methods.
Publication Reference
Proceedings of the World Conference on Photovoltaic Energy Conversion (WCPEC-8)
Year
2022-09