Accurate solar energy prediction is vital for its seamless integration with existing power grids. In this paper, we propose and test a novel hybrid Numerical Weather Prediction (NWP)-Machine Learning (ML) approach that uses a broad range of atmospheric variables from Global Forecast System (GFS) operational forecasts to accurately predict the Global Horizontal Irradiance (GHI) up to one day in advance for a given location. We investigate three ML based scenarios, where one scenario utilizes GHI observations from ground weather stations, whereas the other two utilize GHI observations from the METEOSAT satellite and ERA5 reanalysis respectively. The results demonstrate that the proposed approach effectively predicts 1-day ahead GHI with errors under 10%. Moreover, it was found that ML based scenarios that explicitly used station data for model training performed the best yielding a reduction in MAE of ∼40% over the best baseline. On the other hand, for scenarios that do not explicitly use station data for ML model training, the METEOSAT based scenario performed the best yielding up to 27% reduction of MAE over the best baseline. It is thus postulated that the use of METEOSAT based GHI estimates can be a promising alternative to station data for ML-based GHI prediction, especially in locations where station data does not exist or is not easily available.
A Hybrid Machine Learning - Numerical Weather Prediction Approach for Day Ahead Solar Irradiance Prediction
31.07.2024
548467 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
AN UNSTRUCTURED CFD APPROACH FOR NUMERICAL WEATHER PREDICTION
British Library Conference Proceedings | 2010
|Limited Area Numerical Weather Prediction
Deutsches Zentrum für Luft- und Raumfahrt (DLR) | 2012
|