Wind energy is generated intermittently due to dramatic fluctuations in the availability of the fuel (near-surface wind speed). The predictability of these significant changes are of paramount importance to electric system operators, wind-plant managers, and energy traders.

As with all predictions of future events, no forecast is perfect. Therefore, accounting for the uncertainty in wind energy forecasts is both necessary and appropriate. This paper describes a comprehensive approach toward the uncertainty estimation of day-ahead (forecast horizons of 30-56 hours) wind energy forecasts at several North American wind energy projects.

The wind speed forecast uncertainty is estimated at each wind project using multi-model ensembles of both global and mesoscale numerical weather forecasts. The uncertainty estimates are calibrated with Bayesian model averaging (BMA). The BMA statistical model uses a mixture of gamma probability densities to handle wind speed forecast errors. Several novel training methodologies for BMA are implemented to address both forecast timing errors and regime dependence.

The conversion of energy from kinetic to electric is a non-linear, non-deterministic function of the wind speed and the mechanical characteristics of the wind turbines. Therefore, additional forecast uncertainty must be included due to the power conversion process. Empirical deviations of actual generated energy from the wind turbine rating curve are used to estimate this source of wind energy forecast uncertainty.

Probabilistic wind energy forecasts are produced by random sampling from both the BMA wind speed forecast and the empirical "delta-power" probability distributions. The calibration and sharpness of the resulting prediction intervals are examined.