Because wind resources are highly changeable in time, knowledge on the skills of the commonly used numerical models used in wind energy to forecast the variability is required. It is common knowledge that the atmosphere is inherently chaotic at both small and the larger time scales with some intermediate regime where some predictability has been demonstrated. The chaotic nature at the smaller time scales is predominantly caused by turbulence and the larger scales by the limit in predictability as the accuracy degrades in time.
Here we investigate, based on one year of observations carried out with a wind lidar from 126 to 626-m height at the FINO3 research platform in the North Sea, the ability of the Advanced Research version of the Weather Research and Forecasting model (WRF) to simulate the changes in the observations ahead of time (lead time). The WRF model was nudged towards the Global Forecast System (GFS), a weather forecast model developed by the National Centers for Environmental prediction (NCEP). Forecasts are available every 6 hours on a 0.5 x 0.5 grid. Forecasts of 10 min output were performed every 6 hours up to a forecast horizon of 8 days.
The variability of the change of the wind-speed and direction for lead times of 10 min to 8 days will be reported, thus dealing both with the lack of predictability caused at small lead times caused by turbulence as well as the limit in predictability caused by the non-linearity of the Navier-Stokes equations. Taking a correlation coefficient of 0.6 as lower limit for skills in the simulations corresponds to a lead time lead time of about 4 hours for both wind speed and direction for small lead times (turbulence limited). This value is larger than typically found over land - being about 2 hours. The difference is likely related to the marine conditions of the measurement site, with minor daily variation of the atmospheric conditions but the variability is to a larger degree controlled by the prevailing synoptic conditions.
For larger lead times when the predictability is limited by the non-linearity of the Navier-Stokes equations, the correlation coefficient of 0,6 was found at about 6 days for the wind speed and somewhat smaller – about 4 days for the wind direction. Thus the window of predictability of the WRF/GFS forecasts is found to be in the interval 4 hours up to 6 days (wind speed) and 4 days (wind direction). The predictive skills are found to be a function of height, at a height of 626m it is better than at 126m.
The persistence approach, which is a commonly used benchmark, will be presented for comparison and completeness.