NowCasting: Accurate and Precise Short-Term Wind Power Prediction using Hyperlocal Wind Forecasts

To increase wind power integration, it is essential for electric utilities to accurately predict how much power wind turbines will generate. While purely autoregressive (and nonlinear autoregressive) approaches to prediction using historical data perform well for immediate future (10 to 30 minutes ahead) horizons, their accuracy dramatically deteriorates for farther time horizons. Predicting generation up to 5-6 hours ahead is essential for scheduling multi-tier generation systems that have varying dynamic response. We propose a method that augments autoregressive approaches with exogenous inputs from hyperlocal wind speed forecasts to improve the prediction accuracy and precision beyond 30 min ahead. Our approach reduces the mean absolute error to 2.11%--14.25% for predictions made 10 min to 6 hours ahead. Importantly, it also reduces the uncertainty associated with the predictions by over 15% in comparison with approaches presented in related work.