Abstract: Using the wind power observation and training data from the 2014 Global Energy Forecasting Competition (GEFCom2014),a study on the application of the XGBoost machine learning method in wind power forecasting was conducted.To assess the potential influence of variable distribution on machine learning models,wind power forecasting models were developed using zonal wind and meridional wind,as well as wind speed and wind direction as feature variables.Forecast experiments show that although the distributions of actual observed wind power and numerically forecasted wind speed generally follow the wind power curve,the high degree of dispersion is a major factor contributing to the uncertainty in wind power forecasts.While the correlation between zonal/meridional winds and wind power is not high,models trained with these features using XGBoost still achieve good forecasting performance comparable to those built directly with wind speed and wind direction.The forecast results of the models tend to underestimate wind power peaks and overestimate low power values,which may be attributed to errors in numerically forecasted wind speed.

Key words: Wind power, Machine learning, XGBoost, Wind energy forecasting