9B.3
Kalman filter, analog and wavelet postprocessing in the NCAR-Xcel operational wind-energy forecasting system

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Thursday, 27 January 2011: 4:00 PM
Kalman filter, analog and wavelet postprocessing in the NCAR-Xcel operational wind-energy forecasting system
4C-2 (Washington State Convention Center)
Luca Delle Monache, NCAR, Boulder, CO; and A. Fournier, T. M. Hopson, Y. Liu, B. Mahoney, G. Roux, and T. Warner

Two novel postprocessing methods are evaluated for wind predictions for energy applications. The first method is based on a Kalman filter (KF) bias correction algorithm that is run in analog space rather than in time (ANKF). The second method is based purely on the analog concept, and it is the weighted average of the observations that verified when the 10 best analog forecasts were issued (AN). The analog of a forecast for a given location and time is defined as a past prediction that matches selected features of the current forecast. Both AN and ANKF are able to predict drastic changes in forecast error (e.g., associated with rapid weather regime changes), a feature lacking in KF and standard model-output-statistic algorithms.

The raw wind observations collected at the nacelle-height of each wind turbine is noisier than standard surface observations because of the wake effects of upwind turbines as well as turbulence induced by the turbine rotors. For this reason a wavelet procedure is implemented to de-noise the 15-min raw wind farm observations as well as the model predictions produced with the NCAR RTFDDA (Real Time Four-Dimensional Data Assimilation and forecasting system) run with three nested-grid domains with 3.3 km grid spacing for the finest mesh domain.

Results from tests with different configurations and combinations of the ANKF and AN methods with the wavelet de-noising procedure will be presented. The different postprocessing strategies will be tested over multiple wind farms with data from a time period of several months up to one year. Preliminary results show that the new methods improve raw wind predictions at most wind farms by 10 to 30%. The corrected predictions exhibit lower systematic and random errors as measured by bias and centered-root-mean-square-error, respectively. The latter, underline the ability of the new methods to improve the predictive skill of the raw prediction. The application of ANKF and AN combined with wavelet filtering results also in wind predictions with a stronger monotone association with the observations (as measured by rank correlation).