Using Artifical Neural Networks and the Rapid Refresh Model for Wind Energy Forecasting

The goal of this research is to use machine learning to improve wind energy forecasting applications. The forecasting includes both Annual Energy Production (AEP) of a single wind turbine and short-term (1-18 hours ahead) of an 86 wind farm array. The basis of wind energy forecasting is the wind turbine power curve. The power curve is a powerful tool but does not take into account atmospheric effects. The study used Artificial Neural Networks (ANNs) to create an improved power curve that uses wind speed, turbulence intensity, and density which capture atmospheric effects. The ANNs were validated with previous research and reduced the peak Mean Average Error (MAE) of power by 40%. The reduction in power estimation improved the energy production from 5% to 1% error.
Similar ANNs were developed for the wind farm. The data from the wind farm include hour averages from a nearby meteoroidal station and wind speed from each nacelle. In addition, data from the Rapid Refresh (RAP) model was used as inputs into the ANN. Using RAP assimilation data, in addition to the available data, improved the ANN models instantaneous MEA of power by 8%. The RAP forecasted data also reduced the hour ahead MEA of power by 52%.