Authors: Lueder von Bremen, Constantin Junk
ForWind – Center for Wind Energy Research
Institute of Physics, University of Oldenburg,
Ammerländer Heerstraße 136,
D-26129, Oldenburg, Germany.
Tel: +49 (0) 441 798 5071
Fax: +49 (0) 441 798 5099Lueder 'dot' von 'dot' Bremen 'at' forwind.de
Highlight results
- 25 % improvement of probabilistic score
-
50 % improvement of RMSE with ensemble mean forecast
Meteorological Ensembles
The uncertainty in Numerical Weather Prediction (NWP) can be quantified by Ensemble Forecasting [1]. At the European Centre for Medium-Range Weather Forecasts (ECMWF) 50 members of the Ensemble Prediction System (EPS) are computed with slightly different initial conditions. The different trajectories of the forecasts can be used to estimate the probability of certain events or to quantify in general if the uncertainty of the forecast is high or low.
Since years wind power forecasters have been correcting the effect of atmospheric stability on winds in hub height. Without stability correction hub height winds are underestimated (overestimated) in stable (non-stable) atmospheric conditions when logarithmically extrapolated from surface winds (10 m height). However, deterministic forecasts provide the required data for the stability correction. This is not the case for ensemble prediction systems due to data volume contraints. Within the SafeWind project ECMWF introduced 100 m winds as a new product tailored for the wind energy industry in the analysis, deterministic and ensemble forecasting suite at 26 January 2010 .
Benefits from 100 m ensemble winds
Wind power forecasts are evaluated for Germany interpolating 10 m and 100 m winds to hub height for each grid point, respectively. A regional power curve fromTradeWind is used for the conversion from wind to power [2]. Wind power forecasts with 100 m winds are clearly superior to 10 m winds [3]:
- decreased ensemble mean forecast error for Germany up to 50%
- RMSE forecast error <8 % for forecast Day+3 using ensemble mean
- strong improvement in ensemble spread and reliability
- up to 25 % improvement in proba-bilistic skill score CRPSS
Fig. 1: RMSE of wind power forecast normalized with installed capacity for Germany utilizing 100 m (black) and 10 m bias corrected (red) ensemble winds (black). The ensemble mean (solid line) clearly outperforms the deterministic forecast (dashed) at forecast Day+3. 10 m ensemble mean forecast without bias correction (green) and ensemble spread (dotted) is also shown for Feb 2010 to Apr 2011.
Evaluation of probabilistic wind power forecast for Germany
Wind power forecasts have been computed for each of the four control zones in Germany individually. Forecasts utilizing 10 m winds require an ex-post bias correction that is dependent on the time of the day as very large biases occur due to dismissed thermal stability effects on wind speed in hub height (Fig. 1, green and red curve). The Talagrand Rank Histogram for 10 m winds is not improved by the simple bias correction (Fig. 2, left).
Bias correction is not needed for 100 m wind power forecasts. Compared to 10 m winds the Talagrand Rank Histogram looks very much improved, i.e. the spread of 100 m wind power forecast members is considerably better.
Fig. 2: Talagrand Rank Histogram for German wind power at forecast Day+3. Bias corrected 10 m ensemble winds (left) and 100 m ensemble winds (right).
The criteria for a skillful probabilistic forecast (reliability, sharpness, resolution) are combined in the Continuous Ranked Probability Score. The comparison with a reference probabilistic forecast system leads to the Continuous Ranked Probability Skill Score (CRPSS). The reference system is outperformed when CRPSS is larger than zero.
The superiority of 100 m over 10m ensemble winds decreases with increasing forecast step (Fig. 3). The improvement for the control zone of 50Hertz is slightly smaller than for Germany.
Fig. 3: Improvement in skill (CRPSS) for an ensem-ble wind power prediction system for Germany (full line) and 50Hertz (dashed line) utilizing 100 m ensemble winds. The reference ensemble system uses 10 m ensemble winds. The time period is Feb 2010 to Apr 2011.
Evaluation of probabilistic wind power forecast for French wind farm
The impact of ensemble model level winds on wind power forecasts for a French wind farm was analysed with respect to the usage of 10 m ensemble winds. The wind power forecasts were computed with a Neural Network.
It is found that forecasts obtained from model levels closest to hub height (78 m) and the linear interpolation between adjacent levels have the best probabilistic skill (Fig. 4). In general, the improvements in CRPSS are less compared to Germany. It will be shown later if 100 m winds have the highest probabilistic relevance for large regions (e.g. Germany).
The calibration of 10 m winds [4] does not improve the CRPSS substantially for a single site. However, it is expected that the calibration of raw model level winds or 100 m winds will improve probabilistic scores in future studies.
Fig. 4: Improvement in skill (CRPSS) for an ensem-ble wind power prediction system for an EDF wind park utilizing ensemble winds from different heights: 10 m calibrated (black), model level 61 (35 m, blue), model level 60 (67 m, red), model level 59 (110 m, turquoise) and linear interpolation between model level 61 and 59 (green). The reference ensemble system uses 10 m uncalibrated ensemble winds. The time period is Jan-Oct 2008.
Bibliography
[1] Leutbecher, M., T. Palmer (2008), Ensemble forecasting, Journal of Computational Physics, 227(7), 3515–3539.
[2] McLean JR (2008) Equivalent Wind Power Curves. Deliverable 2.4 of the TradeWind Project
[3] Deliverable 5.10 of the SafeWind Project
[4] Pinson P, 2012: Adaptive calibration of (u,v)-wind ensemble forecasts. Q.J.R. Meteorol. Soc., doi:10.1002/qj.1873.
