• Alessandrini, S., , Sperati S. , , and Pinson P. , 2012: The influence of the new ECMWF Ensemble Prediction System resolution on wind power forecast accuracy and uncertainty estimation. Adv. Sci. Res., 8, 143147, doi:10.5194/asr-8-143-2012.

    • Search Google Scholar
    • Export Citation
  • Alessandrini, S., , Sperati S. , , and Pinson P. , 2013: A comparison between the ECMWF and COSMO ensemble prediction systems applied to short-term wind power forecasting on real data. Appl. Energy, 107, 271280, doi:10.1016/j.apenergy.2013.02.041.

    • Search Google Scholar
    • Export Citation
  • Bremnes, J. B., 2004: Probabilistic wind power forecasts using local quantile regression. Wind Energy, 7, 4754, doi:10.1002/we.107.

  • Brett, A. C., , and Tuller S. E. , 1991: The autocorrelation of hourly wind speed observations. J. Appl. Meteor., 30, 823833, doi:10.1175/1520-0450(1991)030<0823:TAOHWS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Drechsel, S., , Mayr G. , , Messner J. , , and Stauffer R. , 2012: Wind speeds at heights crucial for wind energy: Measurements and verification of forecasts. J. Appl. Meteor. Climatol., 51, 16021617, doi:10.1175/JAMC-D-11-0247.1.

    • Search Google Scholar
    • Export Citation
  • Giebel, G., , Brownsword R. , , Kariniotakis G. , , Denhard M. , , and Draxl C. , 2011: The state-of-the-art in short-term prediction of wind power: A literature overview. ANEMOS Tech. Rep., 109 pp.

  • Gisinger, S., , Mayr G. J. , , Messner J. W. , , and Stauffer R. , 2013: Spatial and temporal variation of wind power at hub height over Europe. Nonlinear Processes Geophys., 20, 305310, doi:10.5194/npg-20-305-2013.

    • Search Google Scholar
    • Export Citation
  • Glahn, H. R., , and Lowry D. A. , 1972: The use of model output statistics (MOS) in objective weather forecasting. J. Appl. Meteor., 11, 12031211, doi:10.1175/1520-0450(1972)011<1203:TUOMOS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Hagedorn, R., , Hamill T. M. , , and Whitaker J. S. , 2008: Probabilistic forecast calibration using ECMWF and GFS ensemble reforecasts. Part I: Two-meter temperatures. Mon. Wea. Rev., 136, 26082619, doi:10.1175/2007MWR2410.1.

    • Search Google Scholar
    • Export Citation
  • Hagedorn, R., , Buizza R. , , Hamill T. M. , , Leutbecher M. , , and Palmer T. N. , 2012: Comparing TIGGE multimodel forecasts with reforecast-calibrated ECMWF ensemble forecasts. Quart. J. Roy. Meteor. Soc., 138, 18141827, doi:10.1002/qj.1895.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Whitaker J. S. , , and Wei X. , 2004: Ensemble reforecasting: Improving medium-range forecast skill using retrospective forecasts. Mon. Wea. Rev., 132, 14341447, doi:10.1175/1520-0493(2004)132<1434:ERIMFS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Hagedorn R. , , and Whitaker J. S. , 2008: Probabilistic forecast calibration using ECMWF and GFS ensemble reforecasts. Part II: Precipitation. Mon. Wea. Rev., 136, 26202632, doi:10.1175/2007MWR2411.1.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Bates G. T. , , Whitaker J. S. , , Murray D. R. , , Fiorino M. , , Galarneau T. J. , , Zhu Y. , , and Lapenta W. , 2013: NOAA’s Second-Generation Global Medium-Range Ensemble Reforecast Dataset. Bull. Amer. Meteor. Soc., 94, 15531565, doi:10.1175/BAMS-D-12-00014.1.

    • Search Google Scholar
    • Export Citation
  • Messner, J. W., , Zeileis A. , , Broecker J. , , and Mayr G. J. , 2014: Probabilistic wind power forecasts with an inverse power curve transformation and censored regression. Wind Energy, 17, 17531766, doi:10.1002/we.1666.

    • Search Google Scholar
    • Export Citation
  • Miller, M., , Buizza R. , , Haseler J. , , Hortal M. , , Janssen P. , , and Untch A. , 2010: Increased resolution in the ECMWF deterministic and ensemble prediction systems. ECMWF Newsletter, No. 124, ECMWF, Reading, United Kingdom, 10–16.

  • Müller, M., 2011: Effects of model resolution and statistical postprocessing on shelter temperature and wind forecasts. J. Appl. Meteor. Climatol., 50, 16271636, doi:10.1175/2011JAMC2615.1.

    • Search Google Scholar
    • Export Citation
  • Obersteiner, C., , and Bremen L. V. , 2009: Influence of market rules on the economic value of wind power: An Austrian case study. Int. J. Environ. Pollut., 39, 112, doi:10.1504/IJEP.2009.027146.

    • Search Google Scholar
    • Export Citation
  • Palmer, T., , Buizza R. , , and Doblas-Reyes F. , 2009: Stochastic parametrization and model uncertainty. ECMWF Tech. Memo. 598, 42 pp.

  • Pineda, I., , and Wilkes J. , 2015: Wind in power: 2014 European statistics. European Wind Energy Association Tech. Rep., 12 pp. [Available online at http://www.ewea.org/fileadmin/files/library/publications/statistics/EWEA-Annual-Statistics-2014.pdf.]

  • Pinson, P., , Chevallier C. , , and Kariniotakis G. N. , 2007: Trading wind generation from short-term probabilistic forecasts of wind power. IEEE Trans. Power Syst., 22, 11481156, doi:10.1109/TPWRS.2007.901117.

    • Search Google Scholar
    • Export Citation
  • Tobin, J., 1958: Estimation of relationships for limited dependent variables. Econometrica, 26, 24–36, doi:10.2307/1907382.

  • Wilks, D. S., , and Hamill T. M. , 2007: Comparison of ensemble-MOS methods using GFS reforecasts. Mon. Wea. Rev., 135, 23792390, doi:10.1175/MWR3402.1.

    • Search Google Scholar
    • Export Citation
  • ZAMG, 2014: ALARO–ALADIN (in German). Zentralanstalt für Meteorologie und Geodynamik. [Available online at http://www.zamg.ac.at/cms/de/forschung/wetter/alaro.]

  • Zugno, M., , Jónsson T. , , and Pinson P. , 2013: Trading wind energy on the basis of probabilistic forecasts both of wind generation and of market quantities. Wind Energy, 16, 909926, doi:10.1002/we.1531.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Turbine locations, area predicted by ALARO, and different model grid boxes.

  • View in gallery

    SSRMSE of the power production forecasts from different NWP models relative to ECMWF-DET aggregated over turbines (left) 1–3 and (right) 4–7. The horizontal lines mark the median and the boxes the interquartile ranges of the 24 × 3 (for turbines 1–3) or 24 × 4 (for turbines 4–7) values at lead times between 24 and 47 h. For the aggregated turbines, the whiskers show the most extreme values that are <1.5 times the length of the box away from the box, and empty circles are plotted for values that are outside the whiskers.

  • View in gallery

    As in Fig. 2, but for the financial loss instead of RMSE.

  • View in gallery

    RMSE of power production forecast (%) of nominal power at turbine 6 for different NWP models and varying training data lengths, averaged over lead times between 24 and 47 h.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 153 153 7
PDF Downloads 99 99 8

Predicting Wind Power with Reforecasts

View More View Less
  • 1 Institute of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innsbruck, Austria
  • | 2 Department of Statistics and Institute of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innsbruck, Austria
© Get Permissions
Full access

Abstract

Energy traders and decision-makers need accurate wind power forecasts. For this purpose, numerical weather predictions (NWPs) are often statistically postprocessed to correct systematic errors. This requires a dataset of past forecasts and observations that is often limited by frequent NWP model enhancements that change the statistical model properties. Reforecasts that recompute past forecasts with a recent model provide considerably longer datasets but usually have weaker setups than operational models. This study tests the reforecasts from the National Oceanic and Atmospheric Administration (NOAA) and the European Centre for Medium-Range Weather Forecasts (ECMWF) for wind power predictions. The NOAA reforecast clearly performs worse than the ECMWF reforecast, the operational ECMWF deterministic and ensemble forecasts, and a limited-area model of the Austrian weather service [Zentralanstalt für Meteorologie und Geodynamik (ZAMG)]. On the contrary, the ECMWF reforecast has, of all tested models, the smallest squared errors and one of the highest financial values in an energy market.

Denotes Open Access content.

Corresponding author address: Markus Dabernig, Institute of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innrain 52, 6020 Innsbruck, Austria. E-mail: markus.dabernig@uibk.ac.at

Abstract

Energy traders and decision-makers need accurate wind power forecasts. For this purpose, numerical weather predictions (NWPs) are often statistically postprocessed to correct systematic errors. This requires a dataset of past forecasts and observations that is often limited by frequent NWP model enhancements that change the statistical model properties. Reforecasts that recompute past forecasts with a recent model provide considerably longer datasets but usually have weaker setups than operational models. This study tests the reforecasts from the National Oceanic and Atmospheric Administration (NOAA) and the European Centre for Medium-Range Weather Forecasts (ECMWF) for wind power predictions. The NOAA reforecast clearly performs worse than the ECMWF reforecast, the operational ECMWF deterministic and ensemble forecasts, and a limited-area model of the Austrian weather service [Zentralanstalt für Meteorologie und Geodynamik (ZAMG)]. On the contrary, the ECMWF reforecast has, of all tested models, the smallest squared errors and one of the highest financial values in an energy market.

Denotes Open Access content.

Corresponding author address: Markus Dabernig, Institute of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innrain 52, 6020 Innsbruck, Austria. E-mail: markus.dabernig@uibk.ac.at

1. Introduction

Over the past decades, wind energy has become an important power source. In 2014, wind energy accounted for 14.1% of the installed capacity in Europe (Pineda and Wilkes 2015). However, the volatility of wind and therefore the fluctuating energy supply complicates an integration of wind power into the power grid. To keep the energy demand and supply balanced, accurate wind power forecasts are important for energy traders, producers, and distributors. Usually, wind power forecasts for lead times longer than 6 h are based on numerical weather prediction (NWP) models (Giebel et al. 2011). NWP models differ widely in their model formulations, space and time resolutions, and parameterizations. Global models provide global forecasts with a limited spatial resolution. Limited-area models use the output of global models as boundary conditions and usually have higher resolutions. Furthermore, to estimate forecast errors arising from errors in the initial state or parameterization, many weather services also provide an ensemble of forecasts, with perturbed initial conditions and/or different model formulations. However, the ensemble size is not always large enough to cover the range of these errors. Additionally, a representativeness error between the state in an NWP grid box and a wind turbine remains. Fortunately, systematic errors can be corrected with statistical postprocessing, such as model output statistics (MOS; Glahn and Lowry 1972). Statistical relationships between NWP forecasts and corresponding observations are formulated from past data and then used to correct future forecasts. To get stable estimates of these relationships, the dataset of past forecasts and observations should be as long as possible. However, frequent model enhancements like increased resolution or improved parameterizations usually change the statistical properties of NWP model output. This limits the usable length of the dataset of past forecasts (Hamill et al. 2004) for commonly used statistical models.

Reforecasts (or hindcasts) overcome this problem (Hamill et al. 2013; Hagedorn et al. 2008). They are ensemble forecasts of past dates that are recomputed with an NWP model that is the same for all of these dates. To reduce computational costs, the ensemble sizes are decreased or only specific dates are recomputed.

In this study, we test whether wind power forecasts can benefit from these longer datasets compared to higher-resolution deterministic forecasts or larger ensemble sizes of operational global ensemble forecasts. For this purpose, short-term (24–47 h) wind power forecasts for several wind parks in Austria, Germany, and the Czech Republic are made with an MOS method tailored to wind power predictions (Messner et al. 2014). As NWP input into MOS, we compare the European Centre for Medium-Range Weather Forecasts (ECMWF) deterministic and ensemble models, the ALADIN–Applications of Research to Operations at Mesoscale (AROME) combined limited-area model (ALARO) of the Austrian weather service [Zentralanstalt für Meteorologie und Geodynamik (ZAMG)], the second reforecast generation of the North American Global Ensemble Forecast System (GEFS), and the ECMWF Ensemble Prediction System hindcasts.

In section 2, we describe the turbine data, the NWP models, and the MOS method. Section 3 presents the results, which are summarized and discussed in section 4.

2. Data and methods

This section first describes the wind power data that are used to test the different NWP models. Subsequently, we describe these NWP models and present an MOS model for removing systematic NWP model errors.

a. Wind power data

To compare the different NWP models, we used data from seven wind parks in Austria, Germany, and the Czech Republic where for each wind park only data from one turbine were available. Figure 1 shows the turbine locations. These turbines have nominal powers between 600 kW and 2 MW, hub heights between 63 and 100 m, and were constructed between the years 2000 and 2007. Turbine data record lengths through 2013 are between 6 and 13 yr (Table 1). To compare turbines with different nominal powers, we use percentages of the nominal power instead of absolute power values.

Fig. 1.
Fig. 1.

Turbine locations, area predicted by ALARO, and different model grid boxes.

Citation: Weather and Forecasting 30, 6; 10.1175/WAF-D-15-0095.1

Table 1.

Date of construction and availability of wind power measurements for each turbine.

Table 1.

b. NWP models

This section describes the setup of the NWP models during the investigation period in detail. We investigate four different NWP model approaches: global deterministic (ECMWF-DET), limited-area deterministic (ALARO), global ensemble prediction system (ECMWF-EPS), and global ensemble hindcast and reforecast (ECMWF-HC and GEFS-RF, respectively).

We use three model chains of the ECMWF Integrated Forecast System (IFS). Because of a major model update in January 2010 (Miller et al. 2010) that changed its statistical properties (Gisinger et al. 2013), we only used data compiled after this date. ECMWF-DET had a spatial resolution of ~16 km (T1279) and 91 vertical levels (Miller et al. 2010). ECMWF-EPS ran at a coarser resolution of 62 vertical levels and ~32 km (T639) horizontally (Miller et al. 2010). It produced one control run that otherwise had the same setup and initial conditions as ECMWF-DET, and 50 perturbed (initial state and model formulation) members (Palmer et al. 2009). Once a week the ECMWF computed ECMWF-EPS forecasts for a specific date (current date + 2 weeks) of the past 20 yr: the so-called ECMWF hindcasts (ECMWF-HC). For example on 2 January 2014, ECMWF-EPS forecasts are initialized with reanalyses of 16 January 2013, 16 January 2012, …, 16 January 1994. The model setup was essentially the same as for ECMWF-EPS, but to save computational costs only four perturbations were computed in addition to the control run. Because the hindcast was computed only once a week, the dataset contains gaps. However, because the model has not changed significantly since January 2010, we were able to collect hindcasts over three years. Thus, we could partly fill these gaps so that our dataset always contained 3 days in a row separated by gaps of 4 days; for example, the gap between hindcasts from 2 to 9 February 2012 (i.e., from 2 to 9 February during 1992–2012) can partly be filled with hindcasts from 10 February 2011 (10 February during 1991–2011) and 11 February 2010 (11 February during 1990–2010).

As a limited-area model we used ALARO, which is operated by ZAMG. ALARO takes its initial and boundary conditions from the ECMWF deterministic model. ALARO computed forecasts for a smaller area, as shown in Fig. 1, with a relatively high horizontal resolution of 4.8 km. However, with only 60 vertical levels (ZAMG 2014), the vertical resolution is smaller than that of ECMWF-DET. As a result of frequent model updates, only data since 2011 were used. Only three of the seven turbines (1–3) are located within the region covered by ALARO. GEFS is the counterpart to the ECMWF-EPS. The National Oceanic and Atmospheric Administration (NOAA) recently computed an extensive reforecast dataset (Hamill et al. 2013) with GEFS for every single day in the past 28 yr. However, its horizontal resolution (~40 km; T254) is relatively coarse and the number of vertical levels (42) is smaller compared to ECMWF-EPS and hindcasts. The ensemble contains 1 control and 10 perturbed forecasts. Table 2 summarizes the characteristics of the different models. Additional forecasts with only the control run (ECMWF-EPS cr, ECMWF-HC cr, and GEFS-RF cr) were used to investigate the influence of ensemble forecasts compared to a single member.

Table 2.

Summary of NWP model characteristics.

Table 2.

For trading purposes, mainly wind power forecasts for the following day are demanded (see the appendix). Thus, we used forecasts with lead times between 24 and 47 h, initialized at 0000 UTC. Since power is usually sold in hourly units, we linearly interpolated the 3-hourly forecasts of ECMWF and GEFS to hourly values. Furthermore, the forecasts were bilinearly interpolated from neighboring grid points to the turbine locations. The direct model output of wind speed forecasts for 80 or 100 m above ground level was used directly since Drechsel et al. (2012) could not find significant benefits from interpolating from NWP model levels to the turbine hub height. For the ensemble forecasts, we used the ensemble mean and standard deviation as input for the statistical postprocessing. Furthermore, we evaluated the benefit of an EPS over a single forecast by comparing it with its control run.

c. Statistical methods

Statistical postprocessing of the NWP model output reduces its systematic errors. It is used here to provide a fair (from an end user’s perspective) comparison of the NWP models. In this study, we use a relatively simple MOS method that is particularly tailored to wind power forecasting (Messner et al. 2014). A specific problem of wind power forecasting is that it is not directly predicted by NWP models so that usually wind speed predictions have to be used as the main MOS input. Unfortunately, the relationship between wind speed and power is strongly nonlinear (Messner et al. 2014) so that linear regression models that are originally used for MOS (Glahn and Lowry 1972) cannot be applied directly. However the wind–power relationship is usually known (with the power curve given by the turbine manufacturer) and can be used to transform wind power observations into wind speed measurements; that is, the wind turbine can be seen as a huge anemometer. These transformed measurements should then relate almost linearly to the NWP wind speed forecasts so that systematic errors of wind speed forecasts can be eliminated by linear MOS models (i.e., linear regression):
e1
where is the transformed power production, is the numerical wind speed forecast (depending on the numerical model it is the deterministic, control, or ensemble mean forecast), and are estimated coefficients, and is the remaining error. Once the coefficients are estimated, Eq. (1) can be used to derive a corrected numerical forecast for , which can then be transformed back into wind power predictions by using the turbine power curves.
An additional problem is that turbines need a certain cut-in wind speed to start rotating and are down-regulated for wind speeds above a certain nominal wind speed. This means that the true wind speed υ can only be observed between the cut-in and nominal wind speeds. Observations outside this range are mapped on the range limits:
e2
To overcome this problem, Messner et al. (2014) proposed to use censored (also known as tobit) regression (Tobin 1958), which considers this unobserved (censored) wind speed when estimating the coefficients and .

Note that turbines usually have to be shut down at very high wind speeds, but since this happens very rarely in our dataset, we did not consider this additional problem. However, principally an extension of the censored regression model would be possible.

Several studies have shown that probabilistic information about the expected forecast uncertainty is beneficial for trading or managing wind power (Pinson et al. 2007; Zugno et al. 2013). The standard censored regression can already be used to provide censored Gaussian predictive distributions but with a constant variance that does not depend on the current weather situation. To include flow-dependent uncertainty information provided by the ensemble variance , the censored regression model can be extended to consider conditional heteroscedasticity with an additional regression equation for the variance:
e3
where and are additional model coefficients and the logarithm is used to assure positive values of σ. A detailed description of this model and a comparison with other wind power MOS approaches can be found in Messner et al. (2014).

3. Results

This section compares the performance of the statistically postprocessed NWP models. For this purpose, the MOS model was trained on data derived before 1 March 2012 and evaluated on data recorded between 1 March 2012 and 1 March 2013. The beginning of the training dataset varies among turbines and NWP models and is limited by the availability of either observations or NWP data (see Tables 1 and 2). Separate MOS models are estimated for each turbine and lead time (24, 25, 26, … , 47 h).

As objective performance measures, we use the root-mean-square error (RMSE) and financial loss, respectively. For every turbine and lead time, an individual RMSE and financial loss was calculated. Figure 2 shows the RMSE for the different NWP models. To ease the comparison of the different models, the RMSE is shown as the skill score (SS):
e4
where is the RMSE of a reference forecast, which is the ECMWF deterministic model in our case. Turbines 1–3 are in more complex topography, and turbines 4–7 are in simpler topography, and they are thus combined into two groups. Thus, the groups contain 3 × 24 and 4 × 24 RMSE values for each model to form the boxes, respectively. Note that although the dates of the training datasets vary, the dates in the test datasets are the same for all turbines. Furthermore, location-dependent error differences are removed by taking skill scores so that the results of different locations are comparable and can be combined.
Fig. 2.
Fig. 2.

SSRMSE of the power production forecasts from different NWP models relative to ECMWF-DET aggregated over turbines (left) 1–3 and (right) 4–7. The horizontal lines mark the median and the boxes the interquartile ranges of the 24 × 3 (for turbines 1–3) or 24 × 4 (for turbines 4–7) values at lead times between 24 and 47 h. For the aggregated turbines, the whiskers show the most extreme values that are <1.5 times the length of the box away from the box, and empty circles are plotted for values that are outside the whiskers.

Citation: Weather and Forecasting 30, 6; 10.1175/WAF-D-15-0095.1

ALARO is only shown in Fig. 2 (left) because turbines 4–7 are located outside its model domain. The ECMWF models outperform the others, followed by ALARO, and then GEFS-RF. Of the three ECMWF configurations, the hindcasts (ECMWF-HC) top the others, especially at turbines 4–7. ECMWF-DET is on par, however, with turbines 1–3, presumably because it represents the relatively complex topography of this region better. For all ensemble forecasts, using ensemble mean and standard deviation slightly improves the RMSE compared to using only the control run.

The RMSE does not directly measure the monetary value that a forecast has for a wind park operator. For example, probabilistic forecast information that should be used for trading wind power (Pinson et al. 2007; Zugno et al. 2013) is not evaluated by the RMSE. As an alternative measure, we therefore use the revenue in the Austrian power market. This metric also measures the probabilistic performance since predictive quantiles are used instead of expected values. The appendix contains a detailed description of the Austrian power market rules. To get a negatively oriented (smaller is better) score similar to the RMSE, we use the financial loss caused by imperfect forecasts instead of revenue. The results for the financial loss skill score in Fig. 3 do not differ drastically from those of RMSE (Fig. 2): the advantage of the ECMWF models over GEFS-RF is less pronounced and GEFS-RF performs almost as well as ALARO. Contrary to the RMSE, however, ECMWF-EPS performs better than does ECMWF-HC. Since the score for the five-member hindcast ensemble is similar to that of the control run, we hypothesize that its ensemble size is too small to sufficiently represent the forecast uncertainty.

Fig. 3.
Fig. 3.

As in Fig. 2, but for the financial loss instead of RMSE.

Citation: Weather and Forecasting 30, 6; 10.1175/WAF-D-15-0095.1

Note that the financial loss only evaluates the performance of the 0.46-quantile forecast. We also evaluated quantile scores (see the appendix for details) of other quantiles from 0.1 to 0.9 and found very similar results (not shown).

The main advantage of the hindcast and reforecast (ECMWF-HC and GEFS-RF, respectively) is that long datasets can be used to train the MOS system. ECMWF-HC uses the same model configuration as ECMWF-EPS but with training data gaps and only 5 instead of 51 members. Nevertheless, because of the longer dataset, its performance is comparable and even better across simple topography (Figs. 2 and 3). Figure 4 isolates the effects of the training data length more thoroughly. It shows the RMSEs of selected NWP models for different training data lengths at turbine 6, which has the longest observational time series. As expected, using longer datasets improves the RMSEs of both NWP models. However, while GEFS-RF improves strongly, especially with more than 2 yr of training data, ECMWF-HC improves only slightly when the training dataset becomes longer than 1 yr. This is especially surprising since ECMWF-HC contains less data per year than does GEFS-RF. However, since the ECMWF-HC data are less autocorrelated, the information content per day is higher.

Fig. 4.
Fig. 4.

RMSE of power production forecast (%) of nominal power at turbine 6 for different NWP models and varying training data lengths, averaged over lead times between 24 and 47 h.

Citation: Weather and Forecasting 30, 6; 10.1175/WAF-D-15-0095.1

4. Discussion and conclusions

This study tested the usability of the ECMWF and GEFS reforecast datasets for wind power forecasts. For this purpose, these reforecasts and operational global, limited-area, and ensemble forecasts were tested as input for a statistical postprocessing model to predict wind power for lead times between 24 and 47 h. Overall, ECMWF showed the best root-mean-square error (RMSE) of all our tested NWP models, confirming the results of Hagedorn et al. (2012). A direct comparison with similar operational models to the ECMWF showed that the long training dataset more than compensates for the coarser resolution compared to the operational deterministic model and the smaller ensemble size compared to the operational ensemble forecast. However, the probabilistic performance measured by the revenue in a power market is slightly worse for the ECMWF reforecast than for the operational ECMWF ensemble.

The Global Ensemble Forecast System reforecast does not have the skill of the ECWMF models. This is in accordance with previous studies (e.g., Hagedorn et al. 2008; Hamill et al. 2008), who found the ECMWF Ensemble Prediction System to be superior to the Global Ensemble Forecast System for precipitation and 2-m temperature.

The skill of the high-resolution limited-area model ALARO is also lower than the skill of the ECMWF models. Several factors could be responsible for this result. One possible reason could be the relatively small training dataset; another might be the coarser vertical resolution. Additionally, although the higher spatial resolution produces more realistic forecasts, RMSE as a verification measure punishes small timing and location errors more harshly. These results agree with those of Müller (2011) (from 40 to 12 to 3 km), Alessandrini et al. (2012) (from 25 to 15 km), and Alessandrini et al. (2013) (from 60 to 32 km), who showed that the spatial model resolution has only minor effects on the performance of postprocessed forecasts.

The ECMWF reforecasts were only computed once a week for the last 20 yr. Thus, the dataset contains gaps of 6 days. However, since the forecast errors are assumed to be autocorrelated, the loss of information is limited (Brett and Tuller 1991). Furthermore, we gathered 3 yr of data and the gaps can partly be filled with reforecasts computed in past years. However, we found that only 1 yr of reforecast training data is enough to reach a skill in the postprocessed ECMWF reforecasts that is almost equal to that of the deterministic ECMWF forecasts, which indicates that this computationally efficient reforecast approach is reasonable.

One advantage of long training datasets is that statistical models can find more complex relationships (e.g., nonlinear effects or multiple input variables). However, in this study we used only a relatively simple linear MOS model with only wind speed forecasts as input. Thus, we speculate that reforecasts could be even more advantageous with more complex statistical models.

The reforecast datasets contain 20 or 28 yr of forecasts. However, our training datasets are limited by the observation data from the turbines that were constructed between 2000 and 2007. Wilks and Hamill (2007), for example, suggest that for turbines or weather stations with longer data records the reforecast datasets might be even more useful.

Acknowledgments

The authors thank the three reviewers for their helpful suggestions to improve the paper, WEB Windenergie AG for providing the wind turbine data, ZAMG for providing the Alaro data and access to ECMWF data, and NOAA and Tom Hamill for making the GEFS reforecasts available. This study was supported by the Austrian Science Fund (FWF; L615-N10).

APPENDIX

Austrian Power Market

The revenue of a wind park operator when trading power in day-ahead markets is a practical measure of the benefits from wind power forecasts. Thus, we simulate participation in the Austrian power market (APCS; Obersteiner and Bremen 2009). In the APCS day-ahead market, operators sell their expected hourly energy supply 1 day in advance for a certain spot price . If the energy y that they actually deliver is less, they have to buy the missing energy for a price with . If they can deliver more energy than they expected, they can sell it for a price with . The prices and depend on the current demand and supply. If the demand outstrips supply, then and . If the demand is greater than the supply, then and . Thus, the total revenue R can be calculated as
ea1
It is clear from these market rules that the revenue is maximized with a perfect energy supply forecast . For forecasts with known uncertainty, Bremnes (2004) showed that the expected income can be maximized by predicting the quantile of the predictive distribution where , , and are the expected prices, respectively.
Zugno et al. (2013) proposed to use time-series models to predict the expected prices. This approach cannot be used here because in the APCS clearing prices are published only once a month. Therefore, we used the mean prices over the past year as a quantile prediction with a resulting quantile of 0.46. To get a negatively oriented score (smaller is better), we use the financial loss caused by forecast errors:
ea2
It can be shown (Messner et al. 2014) that for a constant price combination the financial loss up to a constant factor is equivalent to the quantile score:
ea3
where p is the respective quantile probability. We also tested different quantile probabilities between 0.1 and 0.9, but since the results were very similar to the financial loss, they are not shown.

REFERENCES

  • Alessandrini, S., , Sperati S. , , and Pinson P. , 2012: The influence of the new ECMWF Ensemble Prediction System resolution on wind power forecast accuracy and uncertainty estimation. Adv. Sci. Res., 8, 143147, doi:10.5194/asr-8-143-2012.

    • Search Google Scholar
    • Export Citation
  • Alessandrini, S., , Sperati S. , , and Pinson P. , 2013: A comparison between the ECMWF and COSMO ensemble prediction systems applied to short-term wind power forecasting on real data. Appl. Energy, 107, 271280, doi:10.1016/j.apenergy.2013.02.041.

    • Search Google Scholar
    • Export Citation
  • Bremnes, J. B., 2004: Probabilistic wind power forecasts using local quantile regression. Wind Energy, 7, 4754, doi:10.1002/we.107.

  • Brett, A. C., , and Tuller S. E. , 1991: The autocorrelation of hourly wind speed observations. J. Appl. Meteor., 30, 823833, doi:10.1175/1520-0450(1991)030<0823:TAOHWS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Drechsel, S., , Mayr G. , , Messner J. , , and Stauffer R. , 2012: Wind speeds at heights crucial for wind energy: Measurements and verification of forecasts. J. Appl. Meteor. Climatol., 51, 16021617, doi:10.1175/JAMC-D-11-0247.1.

    • Search Google Scholar
    • Export Citation
  • Giebel, G., , Brownsword R. , , Kariniotakis G. , , Denhard M. , , and Draxl C. , 2011: The state-of-the-art in short-term prediction of wind power: A literature overview. ANEMOS Tech. Rep., 109 pp.

  • Gisinger, S., , Mayr G. J. , , Messner J. W. , , and Stauffer R. , 2013: Spatial and temporal variation of wind power at hub height over Europe. Nonlinear Processes Geophys., 20, 305310, doi:10.5194/npg-20-305-2013.

    • Search Google Scholar
    • Export Citation
  • Glahn, H. R., , and Lowry D. A. , 1972: The use of model output statistics (MOS) in objective weather forecasting. J. Appl. Meteor., 11, 12031211, doi:10.1175/1520-0450(1972)011<1203:TUOMOS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Hagedorn, R., , Hamill T. M. , , and Whitaker J. S. , 2008: Probabilistic forecast calibration using ECMWF and GFS ensemble reforecasts. Part I: Two-meter temperatures. Mon. Wea. Rev., 136, 26082619, doi:10.1175/2007MWR2410.1.

    • Search Google Scholar
    • Export Citation
  • Hagedorn, R., , Buizza R. , , Hamill T. M. , , Leutbecher M. , , and Palmer T. N. , 2012: Comparing TIGGE multimodel forecasts with reforecast-calibrated ECMWF ensemble forecasts. Quart. J. Roy. Meteor. Soc., 138, 18141827, doi:10.1002/qj.1895.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Whitaker J. S. , , and Wei X. , 2004: Ensemble reforecasting: Improving medium-range forecast skill using retrospective forecasts. Mon. Wea. Rev., 132, 14341447, doi:10.1175/1520-0493(2004)132<1434:ERIMFS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Hagedorn R. , , and Whitaker J. S. , 2008: Probabilistic forecast calibration using ECMWF and GFS ensemble reforecasts. Part II: Precipitation. Mon. Wea. Rev., 136, 26202632, doi:10.1175/2007MWR2411.1.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., , Bates G. T. , , Whitaker J. S. , , Murray D. R. , , Fiorino M. , , Galarneau T. J. , , Zhu Y. , , and Lapenta W. , 2013: NOAA’s Second-Generation Global Medium-Range Ensemble Reforecast Dataset. Bull. Amer. Meteor. Soc., 94, 15531565, doi:10.1175/BAMS-D-12-00014.1.

    • Search Google Scholar
    • Export Citation
  • Messner, J. W., , Zeileis A. , , Broecker J. , , and Mayr G. J. , 2014: Probabilistic wind power forecasts with an inverse power curve transformation and censored regression. Wind Energy, 17, 17531766, doi:10.1002/we.1666.

    • Search Google Scholar
    • Export Citation
  • Miller, M., , Buizza R. , , Haseler J. , , Hortal M. , , Janssen P. , , and Untch A. , 2010: Increased resolution in the ECMWF deterministic and ensemble prediction systems. ECMWF Newsletter, No. 124, ECMWF, Reading, United Kingdom, 10–16.

  • Müller, M., 2011: Effects of model resolution and statistical postprocessing on shelter temperature and wind forecasts. J. Appl. Meteor. Climatol., 50, 16271636, doi:10.1175/2011JAMC2615.1.

    • Search Google Scholar
    • Export Citation
  • Obersteiner, C., , and Bremen L. V. , 2009: Influence of market rules on the economic value of wind power: An Austrian case study. Int. J. Environ. Pollut., 39, 112, doi:10.1504/IJEP.2009.027146.

    • Search Google Scholar
    • Export Citation
  • Palmer, T., , Buizza R. , , and Doblas-Reyes F. , 2009: Stochastic parametrization and model uncertainty. ECMWF Tech. Memo. 598, 42 pp.

  • Pineda, I., , and Wilkes J. , 2015: Wind in power: 2014 European statistics. European Wind Energy Association Tech. Rep., 12 pp. [Available online at http://www.ewea.org/fileadmin/files/library/publications/statistics/EWEA-Annual-Statistics-2014.pdf.]

  • Pinson, P., , Chevallier C. , , and Kariniotakis G. N. , 2007: Trading wind generation from short-term probabilistic forecasts of wind power. IEEE Trans. Power Syst., 22, 11481156, doi:10.1109/TPWRS.2007.901117.

    • Search Google Scholar
    • Export Citation
  • Tobin, J., 1958: Estimation of relationships for limited dependent variables. Econometrica, 26, 24–36, doi:10.2307/1907382.

  • Wilks, D. S., , and Hamill T. M. , 2007: Comparison of ensemble-MOS methods using GFS reforecasts. Mon. Wea. Rev., 135, 23792390, doi:10.1175/MWR3402.1.

    • Search Google Scholar
    • Export Citation
  • ZAMG, 2014: ALARO–ALADIN (in German). Zentralanstalt für Meteorologie und Geodynamik. [Available online at http://www.zamg.ac.at/cms/de/forschung/wetter/alaro.]

  • Zugno, M., , Jónsson T. , , and Pinson P. , 2013: Trading wind energy on the basis of probabilistic forecasts both of wind generation and of market quantities. Wind Energy, 16, 909926, doi:10.1002/we.1531.

    • Search Google Scholar
    • Export Citation
Save