Posters

Abstract

The sub-seasonal to seasonal forecasts (S2S) provided by ECMWF through the Copernicus Climate Change Service (C3S) predict the evolution of Earth’s components with long timescales such as El Niño and the North Atlantic Oscillation. At Meteodyn, we combine the S2S predictions and reanalysis data to build our site-specific seasonal forecast model. With state-of-the-art machine learning algorithms for time-series and statistical analysis, our model is able to deliver monthly yields for the next 6 months and yearly productions of a wind farm, updated on a monthly basis. This long- term prevision of wind farm production can be used to mitigate financial risk for investors and to assist long-term budgeting and operational planning for operators. The non-trivial problem of the prediction of long-term wind speed variations poses several challenges to our approach. Firstly, downscaling is required to deal with coarse spatial and temporal resolutions of S2S forecast data, and wind speed inputs need to be extrapolated to turbine hub heights. Secondly, wind energy yield forecasts based on wind speed forecasts are required as a final deliverable to the user. Last but not the least, since forecast error increases with longer horizons, some estimation of uncertainty would be beneficial for decision-making purposes.   The aim of this work is to validate our approach by running a benchmark case on wind farms operated by TotalEnergies, who has provided us with their consolidated historical wind farm production datasets. We start with the prediction of wind speed anomalies regarding a climatology reference and then that of wind farm yields. Finally, our benchmark uses data from wind farms located in different regions to compare the effects of seasonality under different climate zones and terrain complexity. For each wind farm, differences between reanalysis and forecast values are analyzed with metrics assessing the quality and reliability of our model.