Ensemble Learning of Numerical Weather Prediction for Improved Wind Ramp Forecasting

Numerical weather predictions used for wind power forecasting might not be updated in a timely manner in practice, due to its high computational complexity and complicated postprocessing. Thus, the accuracy of wind power forecasts could be significantly compromised especially during wind ramp events. This paper presents an innovative method for improving regional wind power ramp forecasting through ensemble learning of numerical weather prediction models, by using real-time weather measurements as the supervisory data. The numerical weather prediction models are combined to minimize the discrepancy between the forecast values and the real-time measurements in the trend of wind ramps, and the weights of the linear combination are calculated through gradient boosting. The proposed method is non-intrusive and could be efficiently carried out online. The proposed method is evaluated on historical ERCOT wind power ramp events, and compared with existing ensemble aggregation method using simple averaging. The results reveal the effectiveness of the proposed method for improving wind power forecasting during wind ramp events.