Improved near surface wind speed predictions using Gaussian process regression combined with numerical weather predictions and observed meteorological data

This study presents a hybrid numerical weather prediction model (NWP) and a Gaussian process regression (GPR) model for near surface wind speed prediction up to 72 h ahead using data partitioned on atmospheric stability class to improve model performance. NWP wind speed data from the UK meteorological office was corrected using a GPR model, where the data was subdivided using the atmospheric stability class calculated using the Pasquill-Gifford-Turner method based on observations at the time of prediction. The method was validated using data from 15 UK MIDAS (Met office Integrated Data Archive System) sites with a 9 month training and 3 month test period. Results are also shown for hub height wind speed prediction at one turbine. Additionally, power output is predicted for this turbine by translating hub height wind speed to power using a turbine power curve. While various forecasting methods exist, limited methods consider the impact of atmospheric stability on prediction accuracy. Therefore the method presented in this study gives a new way to improve wind speed predictions. Outputs show the GPR model improves forecast accuracy over the original NWP data, and consideration of atmospheric stability further reduces prediction errors. Comparing predicted power output to measured output reveals power predictions are also enhanced, demonstrating the potential of this novel wind speed prediction technique.