Short-term wind power ramp forecasting using statistical and machine-learning techniques and off-site observations

To integrate variable power generation sources, such as wind energy, into the electric grid economically and reliably, the power output needs to be predicted on short time scales from minutes to hours ahead. In particular, the rapid changes in wind power generation need to be anticipated to allow utilities and grid operators to appropriately manage reserve electric capacity and the transmission system. State-of-the-art operational methods for short-term wind power forecasting are rooted in both auto-regressive statistical models and supervised machine learning techniques (e.g., neural networks) using predictors from both on-site data and nearby (off-site) meteorological observations. Numerical weather prediction output is sometimes used directly as further input variables to the statistical models, or it is blended with the statistical model output with weights that increase with the forecast horizon. Adaptive methods for predictor selection are employed using information theoretic techniques. Since model performance varies with the weather regime and the power generation characteristics, further accuracy can be obtained using adaptive regime-switching methodologies. Typical skill scores for the one-hour forecast horizon range from 5% to 25% relative to the root-mean-squared error of a persistence benchmark forecast.