• wind power forecasting;
  • probabilistic;
  • uncertainty;
  • persistence;
  • quantile regression


We discuss some ways of formulating quantile-type models for forecasting variations in wind power in the short term (within a few hours). Such models predict quantiles of the conditional distribution of the wind power available at some future time using information presently available. A natural reference for models of this kind is a ‘probabilistic-persistence’ quantile forecast whose only input is the present wind power. Using data from some New Zealand wind farms, we find that more complex quantile models can readily improve on probabilistic persistence in resolution but not in sharpness. The most valuable model inputs, apart from the present power, are found to be real-time air pressure measurements and a power total-variation indicator. Copyright © 2010 John Wiley & Sons, Ltd.