Abstract

Wind speed measurements from one year from meteorological towers and wind turbines at heights between 20 and 250 m for various European sites are analyzed and are compared with operational short-term forecasts of the global ECMWF model. The measurement sites encompass a variety of terrain: offshore, coastal, flat, hilly, and mountainous regions, with low and high vegetation and also urban influences. The strongly differing site characteristics modulate the relative contribution of synoptic-scale and smaller-scale forcing to local wind conditions and thus the performance of the NWP model. The goal of this study was to determine the best-verifying model wind among various standard wind outputs and interpolation methods as well as to reveal its skill relative to the different site characteristics. Highest skill is reached by wind from a neighboring model level, as well as by linearly interpolated wind from neighboring model levels, whereas the frequently applied 10-m wind logarithmically extrapolated to higher elevations yields the largest errors. The logarithmically extrapolated 100-m model wind reaches the best compromise between availability and low cost for data even when the vertical resolution of the model changes. It is a good choice as input for further statistical postprocessing. The amplitude of measured, height-dependent diurnal variations is underestimated by the model. At low levels, the model wind speed is smaller than observed during the day and is higher during the night. At higher elevations, the opposite is the case.

1. Introduction

Near-surface wind fields are of particular interest for many reasons. They affect a wide spectrum of human activities including safety concerns (e.g., severe weather or traffic), health matters (e.g., air quality), and economic issues (e.g., wind energy). The use of wind energy for electricity production has experienced an enormous boom during recent decades. In 2009, wind turbines with a total capacity of more than 10 000 MW were installed in Europe and more than 37 000 MW were installed worldwide (Global Wind Energy Council 2010). This boom is accompanied by the corresponding demand for wind information in a layer up to approximately 200 m AGL. For simplicity only, this layer is called the wind energy layer (WEL) hereinafter. With respect to wind power economics, the demand extends from long-term wind observations for both regional wind resource assessment and optimal siting of wind turbines to high-quality forecasts for already-installed wind parks (Petersen et al. 1997).

Long-term observations of relatively dense spatial coverage (at least for industrial countries) mainly stem from standard synoptic measurements, at approximately 10 m AGL. Farther aloft, wind measurements are commonly obtained by radiosondes, remote sensing systems (Doppler radar, Doppler lidar, and wind profiler), and meteorological towers (MTs). These methods are comparatively expensive in operation and/or initial costs so that they are much more sparsely deployed. With the increasing number of wind turbines (TBs), an additional observational source has become available: quasi-routine observations at the top of the nacelle of some of the TBs. Although these measurements are downwind of the rotor blades, they can provide valuable information on WEL wind at heights around 100 m AGL, as will be shown here.

In the absence of WEL measurements, wind profiles are commonly estimated by extrapolation of measured near-surface wind either by using the power law (Peterson and Hennessey 1978) or by using stability-corrected logarithmic relations as described by the Monin–Obukhov similarity theory (Stull 2008, chapter 9) that require additional information on stability and surface roughness (Petersen et al. 1997). Despite their limitations, these methods are applied because of their relatively simple nature; see, for example, Elliott et al. (1986) ( power law) and Troen and Petersen (1989) (similarity theory), which are the standard references for WEL wind climatological descriptions in the United States and Europe, respectively. In extending the profiles to higher elevations, the varying stability has to be considered, as shown by, for example, Gryning et al. (2007), who used appropriate length scales for the different layers, and by Kelly and Gryning (2010), who suggested a probabilistic approach to the stability distributions.

Since 2010 the European Centre for Medium-Range Weather Forecasts (ECMWF) has met the demand for WEL wind information by providing 100-m wind in addition to the conventional model outputs, which “traditionally” are 10-m wind and wind at discrete model and pressure levels. Operational numerical weather prediction (NWP) models like that of the ECMWF not only offer deterministic and probabilistic predictions but also offer long-term analyses and forecast data for any arbitrary location in the world even without any nearby observations. A crucial role in the representation of WEL wind is played by the vertical and horizontal model resolutions. As shown by Rife et al. (2004), the required resolution cannot be given in absolute numbers, but rather it depends on the scale of the dominating wind-driving forces. Rife et al. compared measurements of 28 surface stations with the output of four NWP models differing in horizontal resolution by two orders of magnitude. Though the mean absolute error (MAE) hardly differed (clustering around 2 m s−1), the decomposition of observation time series into spectral components revealed that only high-resolution models (1.33-km horizontal resolution) can account for differences in mesoscale diurnal forcing. These differences are caused by small-scale site characteristics such as orography, land use, and vegetation. Increasing resolution has little impact for locations dominated either by synoptic-scale processes or by subdiurnal and thus unpredictable small scales. Using 10-m wind from the ECMWF Ensemble Prediction System (EPS), Pinson and Hagedorn (2012) quantified the model performance for short-range forecasts of 0–12 h by the root-mean-square error (RMSE) of ~2.2 m s−1, averaged over more than 600 stations. When model resolution is increased from about 50 km to a value of 33 km, the error decreased significantly for all verification measures applied, with, for example, an improvement in RMSE by 2%–3%. The largest deviations from observed 10-m wind were found in complex terrain and at coastal sites, that is, in regions of largely differing site characteristics at a small scale.

Petroliagis et al. (2010) investigated the skill of ECMWF deterministic and ensemble predictions not only at the surface but also at the typical TB height of 110 m AGL. They used a wide range of verification measures to compare forecasts of different lead times out to 5 days with model analyses. Introducing a “coefficient based on spread” (COBOS), a weighted combination of deterministic (high resolution) and EPS mean (lower resolution) components, proved to be superior to using only one of them at all studied lead times. The evaluation of the skill at a single grid point (see Fig. 15 in Petroliagis et al.) yields an RMSE of about 1 m s−1 at a lead time of 12 h, increasing to 1.7–2 m s−1 at a lead time of 60 h for the COBOS deterministic and EPS mean approach.

The common practice, when estimating model skill by comparing forecast with analysis, is denoted as “fair” by Pinson and Hagedorn (2012), because the temporal and spatial scales are consistent. They point out, however, that the “end user” actually is interested in the skill revealed by comparison with the “truth,” that is, the observation at a point location.

In this study, we take the perspective of the end users and analyze elevated WEL wind observations between 20 and 250 m AGL and compare them with short-term forecasts of various model wind products derived from standard model output and interpolation methods. NWP 10-m winds are still the main input in many applications such as, for example, wind power forecasts (Pinson and Hagedorn 2012), since “most if not all operational NWP wind products that can be ‘purchased’ for wind power prediction are at 10 m” (P. Pinson 2010, personal communication). Motta et al. (2005) show that the extrapolation of 10-m wind may result in large errors, however, especially in (very) stable conditions. Furthermore, the performance of the “best” model wind with respect to WEL winds depending on individual site characteristics is investigated and is related to their representation in the model.

This paper is organized as follows. The next section describes observational data from both meteorological towers and wind turbines as well as the model data with which they are compared. The evaluation of measurements and the verification results are detailed in sections 3 and 4, respectively. The findings are summarized and discussed in the last section.

2. Data

One year (2009) of WEL wind velocity data from various European regions is analyzed and is compared with short-term forecasts (0–9 h) of the deterministic output of the ECMWF model. The observations stem from both conventional MTs and measurements at TBs. The measurements, which generally have a temporal resolution of 10 min, are quality controlled according to Jiménez et al. (2010). The procedure includes the exceedance check for lower and upper thresholds, elimination of periods of constant data of longer than 60 min, and omission of data of excessive variability. Furthermore, data availability of at least 75% of the investigated period (2009) is required in this study. Seasonal variability is studied with data from 2007 to 2009, for which only a subset of the measurements is available.

The spatial resolutions of the observations and model data differ strongly. Measurements are at a point, whereas the model data represent the spatial mean of an individual grid box of roughly 25 km × 25 km × 50 m. The model wind is at a particular time step of approximately 10 min within the NWP simulation. The discrepancy in spatial resolution is partly compensated by invoking Taylor’s hypothesis and using measurements averaged over a whole hour at ±0.5 h of model verification time. In 1 h, air at a typical speed of about 7 m s−1 (see section 3) traverses the 25-km model box length. Because wind direction was not available for most TBs, the study concentrated on wind speed only.

a. Meteorological towers

Wind profiles of the following eight MTs are analyzed (Table 1): the four German research MTs at Garching, Hamburg, Karlsruhe, and Lindenberg; the tower at Cardington, Great Britain, and the research MT at Cabauw, Netherlands. Also analyzed were data from the MTs associated with two offshore wind parks: Fino 1, about 45 km north of the German island of Borkum (Dankert and Horstmann 2007; http://www.fino-offshore.com), and tower M7 at Horns Rev in the North Sea, about 6 km east of the wind farm and 10 km southwest of Blavands Huk at the Danish coast (Méchali et al. 2006). Winds are usually measured at several (up to eight) heights between 20 and 250 m AGL. The sites of the MTs are indicated by white plus signs in Fig. 1.

Table 1.

Station name, abbreviation (abbrev), latitude and longitude, and anemometer heights of MTs used in this study.

Station name, abbreviation (abbrev), latitude and longitude, and anemometer heights of MTs used in this study.
Station name, abbreviation (abbrev), latitude and longitude, and anemometer heights of MTs used in this study.
Fig. 1.

Overview map of TB (white circles) and MT (plus signs) sites with abbreviations for MT sites (Table 1): (left) topography of ECMWF-T799 with horizontal resolution of about 25 km and (right) topography of the Global 30 arc-s Elevation Dataset (GTOPO 30) DEM with 1-km horizontal resolution.

Fig. 1.

Overview map of TB (white circles) and MT (plus signs) sites with abbreviations for MT sites (Table 1): (left) topography of ECMWF-T799 with horizontal resolution of about 25 km and (right) topography of the Global 30 arc-s Elevation Dataset (GTOPO 30) DEM with 1-km horizontal resolution.

The onshore MTs greatly differ in their site characteristics, which influences the individual wind profile enormously. The Cabauw (Baas et al. 2009) and Lindenberg (Görsdorf et al. 2002) MTs are sited in relatively flat terrain with low vegetation. The area surrounding the MT at Hamburg is also flat but features rural and industrial development to the north, west, and southwest and a mixture of mostly low vegetation and sparsely populated hamlets elsewhere (Gryning et al. 2007). The tower at Garching, which is located about 6 km north-northeast of the border of Munich, consists of a 15-m mast on the roof of a building that is 47 m high, and so the total height is 62 m. The terrain is slightly hilly, with agricultural land, scattered forests, and some hamlets. The surroundings near the Cardington MT, which is roughly 2 km south of Bedford village, are in a slightly hilly area (Lapworth 2003). The Karlsruhe MT is located in the Rhine Valley between the complex terrain of the Vosges Mountains and the Black Forest in a broad-leaved forest (Wenzel et al. 1997). See section 2d for details on the terrain classifications such as “hilly.”

In terms of the influence of the mast and nearby buildings on the wind speed measurements, a trade-off between the quality and quantity of data was made. Therefore, only lee effects were accounted for, that is, the sectors of 300°–360° at Fino, 110°–160° at Horns Rev, 330°–030° at Hamburg owing to the mast, and 350°–040° at Cardington resulting from two huge hangars. For Lindenberg, Cabauw, and Karlsruhe, the data were provided having already been precontrolled. Because of the large distance between the Horns Rev wind farm and the MT, wake effects on the wind speed could be neglected (Frandsen et al. 2007).

b. Wind turbine data

In addition to the MT dataset, wind speed data from 17 wind turbines at different wind parks could be used. From the operator’s point of view, TB data are sensitive with respect to economic competition. As a consequence, only the positions of the total wind parks, and not those of individual turbines, were provided. Thus, detailed information on site characteristics like land use and vegetation height of the surroundings were inaccessible to the authors. In general, the TBs are surrounded by agricultural land and scattered forests for some distance. They are located in flat to moderately complex terrain in middle and northern Germany, in the southern Czech Republic, and in the eastern part of Austria, at altitudes between sea level and approximately 500 m MSL (Fig. 1). More details on the topography are given in section 2d. The wind anemometers of the TBs used for this investigation are installed at the top of the turbine-housing nacelle. The observation heights hobs range from 63 to 105 m AGL. Except for downtimes due to maintenance or economic reasons, the turbine runs for wind speeds between cut-in (here usually 3–5 m s−1) and cutoff (here usually 20–25 m s−1) velocities. During operations, the rotors automatically turn into the wind, which moves the anemometer to the downstream side of the rotor blades. Assuming for simplicity that the turbine extracts 50% (because of the Betz limit of 59% and a maximum effectiveness of the turbine of approximately 80%) of the advected kinetic energy (~power), wind speed behind the rotor is on average reduced by about 20%, since wind speed is proportional to the cube root of power. Since TBs measure only a rotor-modified wind speed, outcomes that are based on wind turbine data are considered only in a qualitative instead of a quantitative way.

c. Model wind data

The observations described above are compared with diverse wind products of the global, operational model of the ECMWF-T799 to determine the individual skill. Short-term forecasts with lead times of 0–9 h for the 0000 and 1200 UTC runs are used, depicting the diurnal cycle in steps of 3 h. In 2009, model resolution for deterministic forecasts was 0.25°, which approximately corresponds to 25 km. For details, see ECMWF (2008). The following wind products, which are bilinearly interpolated to the individual observation sites, are investigated:

  1. L91, L90, … , L85—winds of the lowest seven model levels at approximately 10, 34, 67, 111, 165, 229, and 305 m AGL;

  2. lin—model wind linearly interpolated from the nearest model levels to hobs;

  3. 100 m—model wind linearly interpolated from the nearest levels to 100 m AGL [the ECMWF has been directly providing this information since January of 2010 for both deterministic and probabilistic runs, and it has been increasingly used for wind power forecasts (Pinson and Hagedorn 2012)];

  4. 10 m—10-m wind [the ECMWF also provides a postprocessed wind that is designed to be compatible with surface synoptic observations (SYNOP), that is, an open-terrain wind at 10 m AGL, and this open-terrain wind is obtained by interpolating wind information from an elevated level that is less influenced by the underlying terrain to 10 m AGL using a roughness length typical for open terrain (ECMWF 2008, chapter 3.12.2); its performance is investigated since it is (still) one of the main inputs to wind power forecasting methods (P. Pinson 2010, personal communication)];

  5. ln10—model wind extrapolated from 10-m wind to hobs, using a logarithmic wind profile [e.g., Stull 2008, chapter 9; the roughness length z0 is obtained from the 2006 updated version of “ECOCLIMAP” (Champeaux et al. 2005) for the respective location] [although logarithmically extrapolated wind provides only realistic results for neutral stratification, this method is the simplest way to obtain wind information at higher elevations when only near-surface wind data are available; therefore, it is (still) a frequently used product (e.g., in wind power prediction; Giebel et al. 2011, p. 45)];

  6. ln100—model wind logarithmically extrapolated from 100-m wind to hobs using the same z0 as for ln10; and

  7. 10 ln100—model wind logarithmically interpolated between 10-m wind and 100-m wind to hobs.

d. Topography

The inadequate representation of actual topography and surface roughness by global models as a result of their relatively coarse resolution affects low-level wind forecasts. The complexity of the terrain and its representation in ECMWF-T799 was roughly estimated by the standard deviation of the orographic height as given by a digital elevation model (DEM) with horizontal resolution of 30″ [σ(t30)] and by the ECMWF [σ(ec)], respectively, in a radius of 0.3° around the individual observation sites. For analyzing the model performance as a function of the topography (section 4a), the terrain was grouped in four classes: 1) offshore, 2) flat with σ(t30) < 25 m, 3) hilly with 25 ≤ σ(t30) < 75 m, and 4) moderately complex terrain with σ(t30) ≥ 75 m. The results are summarized in Table 2. The standard deviations are close to zero for offshore locations and near coastal sites but increase to several decameters as revealed by σ(t30) = 113 m for Karlsruhe and σ(t30) = 188 m for one TB in moderately complex topography.

Table 2.

Standard deviation σ of topographical height (m) as represented by a DEM of 30″ resolution (t30) and by ECMWF-T799 (ec), within a radius of 0.3° around the wind towers (Hor, Fin, Cab, Lin, Ham, Gar, and KIT; see Table 1 for abbreviations) and TB (tb) sites. The fourth column indicates the surface roughness z0 (m) as given by the 2006 updated version of ECOCLIMAP (Champeaux et al. 2005) for the respective locations. The last column contains the terrain class according to section 2d.

Standard deviation σ of topographical height (m) as represented by a DEM of 30″ resolution (t30) and by ECMWF-T799 (ec), within a radius of 0.3° around the wind towers (Hor, Fin, Cab, Lin, Ham, Gar, and KIT; see Table 1 for abbreviations) and TB (tb) sites. The fourth column indicates the surface roughness z0 (m) as given by the 2006 updated version of ECOCLIMAP (Champeaux et al. 2005) for the respective locations. The last column contains the terrain class according to section 2d.
Standard deviation σ of topographical height (m) as represented by a DEM of 30″ resolution (t30) and by ECMWF-T799 (ec), within a radius of 0.3° around the wind towers (Hor, Fin, Cab, Lin, Ham, Gar, and KIT; see Table 1 for abbreviations) and TB (tb) sites. The fourth column indicates the surface roughness z0 (m) as given by the 2006 updated version of ECOCLIMAP (Champeaux et al. 2005) for the respective locations. The last column contains the terrain class according to section 2d.

3. Wind in the lowest 20–250 m of the troposphere

The analysis of mean WEL wind speed and its diurnal variations reveals the influence of topography, local site conditions, and stratification.

Wind observations

Mean observed wind speed spans a wide range in the WEL from roughly 2 to almost 10 m s−1 (Fig. 2). The complex topography and the tall vegetation of a broad-leaved forest at the Karlsruhe MT result in the lowest of all locations at each hobs. Speed increases nearly logarithmically from 2.2 m s−1 at 30 m AGL up to 5.1 m s−1 at 130 m AGL, which is 5–6 m s−1 lower than at the offshore MTs. The wind velocities of the very similar profiles at Lindenberg and Hamburg are 1.5 m s−1 higher than for Karlsruhe, followed by Cardington and Cabauw, which have higher wind values than those for Karlsruhe by approximately 1.8–3.8 m s−1. The profiles at Lindenberg and Hamburg are almost identical, with marginally higher values at Hamburg. It is assumed, however, that, at least up to hobs = 50 m, at Hamburg is dampened as a result of the rough urban surface conditions of this site.

Fig. 2.

Mean observed wind speed for 2009 data measured at MTs (circles; see Table 1 for abbreviations) and TBs (plus signs).

Fig. 2.

Mean observed wind speed for 2009 data measured at MTs (circles; see Table 1 for abbreviations) and TBs (plus signs).

The mean wind velocities of the TBs range from 4.7 to 7.3 m s−1. For all but two of the TBs, exceeds those of the profile of Lindenberg, which indicates a good choice of location for the examined TBs.

Wind speed distributions of TBs and MTs differ strikingly. At wind towers, the typical, single-peaked Weibull distribution is found. It is narrow for onshore MTs but is wide for the offshore MTs. Although a few TBs also reveal a Weibull-shaped distribution, the distributions of most TBs have a plateau or secondary maximum (Fig. 3). The presumable reason is the extraction of kinetic energy by the rotor blades before the wind reaches the anemometers, which are installed at the top of the nacelle downwind of the rotor. As long as wind speeds are lower than the cut-in wind speed of the turbine (typically 3–4 m s−1), no kinetic energy will be extracted by the turbine. The turning on of the rotor around the cut-in wind speed, however, results in a reduction of the wind speed downstream of the rotor blades and thus partly eliminates wind speeds at this range. Therefore, the TB data are only used for qualitative analysis, for which they can provide valuable information.

Fig. 3.

Empirical probability density functions of wind speed for three typical distribution shapes: Weibull-shaped distribution at the Lindenberg MT (60 m; gray solid line); flat and wide, quasi-symmetrical distribution at the offshore Horns Rev MT (70 m; gray dashed line); and double-peak distribution of a TB observation (63 m; black solid line).

Fig. 3.

Empirical probability density functions of wind speed for three typical distribution shapes: Weibull-shaped distribution at the Lindenberg MT (60 m; gray solid line); flat and wide, quasi-symmetrical distribution at the offshore Horns Rev MT (70 m; gray dashed line); and double-peak distribution of a TB observation (63 m; black solid line).

The mean diurnal wind variations mainly depend on the site-specific mean thermal cycle as well as on the measurement height hobs. At the onshore MTs, the amplitudes of the mean diurnal variations range from 0.5 to 1.9 m s−1 near the surface. The minimal diurnal change (i.e., the most constant wind conditions throughout the day) is not found at the top but is found at medium heights [hυmin)] of about 60–100 m on average. Above hυmin), the diurnal variation increases again so that the uppermost platforms reveal nearly similar amplitudes as observed near the ground. This is depicted by the example of the Cabauw MT in Fig. 4 for the different seasons. Figure 4 also shows that not only the amplitude but also the phase of the diurnal variation depends on height. Below hυmin), the wind speed is minimal during the night and the maximum roughly coincides with the surface maximum temperature in the (early) afternoon. The strongest speedup is between 0600 and 0900 UTC (UTC = LST − x h, with −2 ≤ x ≤ 0 for the available measurement sites) in spring, summer, and autumn. In the cold season, the speedup is much weaker and is about 3 h later. Above hυmin), the diurnal cycle features a minimum during the morning hours (0900 UTC) and a nocturnal maximum. The latter presumably can be attributed to the lowest part of the nocturnal low-level jet (e.g., Baas et al. 2009). In both the annual and seasonal averages, offshore sites reveal only very small changes of the wind velocity throughout the day because of the generally small thermal variations. At all heights, the amplitude is less than 0.5 m s−1, with higher wind speeds during the night.

Fig. 4.

Mean diurnal deviation from seasonal average (at 0.2 m s−1 intervals) at Cabauw. Gray areas indicate that the wind exceeds mean wind speed at a given time: (top left) December–February, (top right) March–May, (bottom left) June–August, and (bottom right) September–November. The averages are calculated separately for each level.

Fig. 4.

Mean diurnal deviation from seasonal average (at 0.2 m s−1 intervals) at Cabauw. Gray areas indicate that the wind exceeds mean wind speed at a given time: (top left) December–February, (top right) March–May, (bottom left) June–August, and (bottom right) September–November. The averages are calculated separately for each level.

The seasonal changes in wind speed were examined for the available 3-yr (2007–09) datasets. The results shown in Fig. 5a and 5b for the examples at Cabauw and Karlsruhe MTs, respectively, unveil the expected relation to synoptic events connected with the position of the jet stream. In general and quasi independent from hobs, the wind exceeds the total mean during the cold season (approximately from November until March) by up to 40%. During the warmer seasons (April–October), wind velocities are mostly 10%– 20% lower than the total mean.

Fig. 5.

Seasonal variation of wind speed at the (a) Cabauw and (b) Karlsruhe MTs. The monthly averages at the different heights are normalized by the total mean at the individual height for the years 2006–08.

Fig. 5.

Seasonal variation of wind speed at the (a) Cabauw and (b) Karlsruhe MTs. The monthly averages at the different heights are normalized by the total mean at the individual height for the years 2006–08.

4. Verification of model wind

On the basis of several verification parameters, the skill of the global model in representing WEL wind as described in the previous section is presented. First, the performance of diverse model wind products listed above is investigated to reveal which corresponds best to the observations. This “best” product is subsequently used to highlight the model performance in absolute and relative values, depending on site characteristics and observation height.

a. Performance of model wind products

Model performance of deterministic short-term forecasts is typically determined by the bias Bi, the MAE, and the RMSE, respectively. For practical use, Bi is irrelevant since it can be easily corrected if known. MAE and RMSE are qualitatively similar, but the latter emphasizes (undesired) large errors. Therefore, it is chosen over MAE. Because the error also depends on the magnitude of the wind velocity, it has to be normalized by the mean wind speed of the individual measurement dataset. Thus, the skill of the model wind products listed above will be studied in the following by the root-mean square of the bias-corrected error, normalized by the mean wind speed, which is denoted hereinafter by Erel. With the definition ,

 
formula

where υm is the model value.

In this study, the focus is on the model skill for wind energy applications. Therefore, only the results for 50 ≤ hobs ≤ 110 m are considered, which is typical hub height.

Averaging over all MTs (over all hobs ≥ 50 m), the overall best results in terms of minimum error are achieved by the linear interpolation method from neighboring model levels (lin), as well as by the near-surface model level L90, with Erel being slightly smaller than 24%. Almost similar values of less than 26% are obtained when using a model wind product that includes the 100-m wind information (100 m, ln100, or 10 ln100) and for the model levels L89 and L88. The largest errors, with Erel ≥ 32% m, occur for logarithmic extrapolation of the 10-m wind to hobs (ln10) and for the two highest model levels (L85 and L86), which are more than 100 m above typical hobs.

Analyzing the TB dataset separately yields very similar results. The best products are lin, L89, and L90, with Erel ≈ 26%, which confirms the findings above. With quasi-identical values as with MT data, the poorest results are again obtained by ln10, L85, and L86.

For a more detailed analysis of the performance of diverse model wind products, MT observations are grouped for hobs in steps of 10 m and using the four different terrain classes defined previously: 1) offshore, 2) flat, 3) hilly, and 4) moderately complex terrain (see section 2d and Table 2). The group-averaged results are summarized in Fig. 6.

Fig. 6.

Relative, bias-corrected RMSE (Erel) between model and MT observations, using various model wind products (see section 2c for details) to reveal their performance. The Erel is normalized by mean observed wind speed for four different terrain classes (offshore, flat, hilly, and moderately complex terrain; section 2d) and is averaged for various groups of hobs ± 5 m. Numbers in the legend connected with heights indicate the available number of different MT platforms at the respective height. Note that the range of the y axis varies among the panels.

Fig. 6.

Relative, bias-corrected RMSE (Erel) between model and MT observations, using various model wind products (see section 2c for details) to reveal their performance. The Erel is normalized by mean observed wind speed for four different terrain classes (offshore, flat, hilly, and moderately complex terrain; section 2d) and is averaged for various groups of hobs ± 5 m. Numbers in the legend connected with heights indicate the available number of different MT platforms at the respective height. Note that the range of the y axis varies among the panels.

For noncomplex terrain (Figs. 6a–c), best results with respect to minimum values of Erel are provided by both a neighboring model level (mostly the level beneath hobs) and by interpolation methods that use wind information from above the surface (100 m, ln100, 10 ln100, and lin). In mountainous regions, however, the smallest deviations between model and observations are achieved using wind from near-surface model levels and the 10-m wind itself. For any given model product, the relative difference between model and the observation is smallest offshore and increases with increasing complexity of terrain and lower measurement heights. This is confirmed by a correlation coefficient of 0.87 between the bias-corrected RMSE, normalized by the mean wind speed, Erel, and the topography given by the standard deviation of the topography σ(t30). For the different hobs of the offshore sites, the error Erel of the best product ranges around 15%–16% of the individual mean wind speed. In flat terrain, Erel reaches 20%–23%, in hilly environments it is 22%–30%, and with complex terrain Erel exceeds 32% for all hobs. Except for mountainous conditions or low hobs ≤ 60 m, the new, operationally available model product “100m” achieves satisfying results in the sense of yielding results that are almost similar to the particular best method. In particular for offshore sites and for hobs > 60 m above flat terrain, the ln10 wind ranks among the worst information for these sites.

Again, the analysis of the TB dataset (Fig. 7) leads to very similar results for the available terrain classes. This shows that, despite the poorer data quality, measurements from wind turbines can provide useful qualitative information.

Fig. 7.

As in Fig. 6, but for the TB dataset.

Fig. 7.

As in Fig. 6, but for the TB dataset.

b. Model skill using linearly interpolated model wind

The analysis of the bias and the (relative) RMSE provides additional details on the model skill as shown in Fig. 8. We use the overall best model product independent from hobs and terrain, and the wind linearly interpolated from neighboring model levels, lin, is used. Model wind speeds are higher than observed speeds over land and lower over the ocean (Fig. 8a). The bias-corrected RMSE increases with height (Fig. 8b) since wind speeds also increase with height. When it is normalized by mean observed wind speeds (Fig. 8c), the RMSE remains approximately constant with height or even decreases.

Fig. 8.

(a) Bias (), (b) bias-corrected RMSE (E), and (c) bias-corrected RMSE normalized by mean observed wind speed (Erel), for 2009 wind data. Profiles of MTs are indicated by circles connected by lines (Table 1), and TBs are shown by plus signs. Values exceeding the limits of the abscissa are marked by the appropriate sign with the actual value written to the side.

Fig. 8.

(a) Bias (), (b) bias-corrected RMSE (E), and (c) bias-corrected RMSE normalized by mean observed wind speed (Erel), for 2009 wind data. Profiles of MTs are indicated by circles connected by lines (Table 1), and TBs are shown by plus signs. Values exceeding the limits of the abscissa are marked by the appropriate sign with the actual value written to the side.

The bias Bi as given by ranges between −0.9 and 2.7 m s−1 (Fig. 8a). For both offshore sites, as well as at the elevated hobs of the flat-terrain-located Cabauw, the actual wind is underestimated (Bi < 0), but by less than 0.4 m s−1. For the MTs at Karlsruhe, Garching, Cardington, and Lindenberg farther inland, the urban Hamburg, as well as the near-ground measurements at Cabauw, the model overestimates the observed wind speed (Bi > 0). Except for one outlier TB, all values exceeding 0.8 m s−1 stem from the Karlsruhe and Hamburg MTs at hobs ≤ 100 m, that is, where the “extreme” surface-roughness conditions of the forest and the city, respectively, are not represented adequately by the model. The bias for most TBs is of a similar order to that for MTs and ranges between ±0.9 m s−1. Although measured wind speed behind a working turbine is smaller than the actual incoming wind, the model still underestimates the observations at 11 of 17 TBs. This result presumably can be attributed to the typical placement of TBs at locations with winds that are higher than the regional average (e.g., on ridges).

Relating Bi to (Birel, not shown) reveals that for the Horns Rev, Fino, Cabauw, and Lindenberg MTs, and all except two TBs, the model deviates from the observed mean wind speed by not more than ±10% on average. At hobs < 70 m at Garching, Hamburg, and Karlsruhe, the error exceeds 15% and even reaches 290% for the lowermost observation height of the Karlsruhe MT.

The root-mean square of the bias-corrected error E depends on the magnitude of the velocity and thus also on height, as well as on surface roughness. As the profiles of the MTs show (Fig. 8b), E increases with height (i.e., with mean wind speed). For the Karlsruhe site with its low wind but large roughness, it is much higher than, for example, the Cabauw and Lindenberg sites with their strong wind but low roughness. In absolute terms, E ranges between 0.9 and 2.2 m s−1 and concentrates around 1.4 ± 0.2 m s−1. For each height, the lowest deviations are found for the Cabauw, Lindenberg, and Hamburg MTs. The profile of the Karlsruhe MT has the largest errors, with approximately 1.5 m s−1 at hobs = 50 m and 1.9 m s−1 at hobs = 200 m. Large errors (>1.6 m s−1) are also found for six of the TBs, all of which are located in terrain that ranks among the six most complex examples according to σ(t30) as listed in Table 2.

Analyzing the normalized error Erel, similarly defined as in the previous section, yields a relative error from 14% up to 160% of the observed wind velocity (Fig. 8c). Using Erel better highlights the influence of surface conditions. The lowest errors are found at the offshore sites, closely followed by the flat-terrain, quasi-coastal Cabauw MT. The Erelincreases for Lindenberg, Cardington, and Hamburg with increasing hilliness (see section 2) and/or because of the influence of the strong surface roughness emerging from the nearby city. Karlsruhe has the largest errors, because its site is strongly affected by both the complex terrain and the high vegetation. The decrease of these influences with height is shown by the decrease of Erel with hobs. On the other hand, the lower the surface roughness is, the more constant are the profiles with height. For 11 of 17 TBs, Erel is of a similar order to those for Lindenberg, Cardington, and Hamburg (22%–25%). Only for the six TBs already mentioned above does the relative error reach 30%–36%, which can be now clearly attributed to the terrain.

To scrutinize how well the model depicts the diurnal variations for each season, modeled and observed mean wind speeds at the respective time of the day, corrected for the (seasonal) bias, are compared. The example of the Cabauw MT is shown in Fig. 9, but other locations will also be discussed in the text. A comparison of the model wind speed with the actual diurnal cycle (see section 3 and Fig. 4) throughout the year and for all MTs reveals that the diurnal variations are underestimated by the model.

Fig. 9.

Bias-corrected differences between modeled and observed diurnal variations of the wind speed at the Cabauw MT (at 0.2 m s−1 intervals) for the different seasons, similar to Fig. 4. Gray areas indicate model winds exceeding measured wind speeds.

Fig. 9.

Bias-corrected differences between modeled and observed diurnal variations of the wind speed at the Cabauw MT (at 0.2 m s−1 intervals) for the different seasons, similar to Fig. 4. Gray areas indicate model winds exceeding measured wind speeds.

At the lowest measurement heights at all onshore sites except for Karlsruhe, the model’s wind speed is smaller than is observed during the day (about 0900–1500 UTC) and is higher during the night by up to ±0.4 m s−1 (bias corrected). For Karlsruhe, the model overestimates the wind velocities during the day and partly in the early evening whereas (the low) nocturnal wind speeds quasi agree.

Above hυmin) (heights at which minimal diurnal variations are observed; see section 3), the actual nocturnal wind maximum is underestimated by the model by up to 0.6 m s−1. During the day the situation is reversed and the model wind velocities are too high by up to 0.8 m s−1. These features are especially well marked during summer but are also found in spring and autumn; they are weak during winter months. Although the amplitude of the diurnal cycle is much weaker in the model than is measured, the timing is relatively well depicted by the model at all heights (except for Karlsruhe). Offshore, no height dependence is found. Even the relatively small diurnal variations are slightly underestimated by the model for Fino and Horns Rev.

Of course, this diurnally varying model skill is seen in Bi and E (bias-corrected RMSE), when calculated either for 0000 (Fig. 10) or 1200 UTC (Fig. 11). Onshore and near the surface, the bias is positive and larger at 0000 UTC than at 1200 UTC. During the night and above about 40 m, the difference between modeled and observed wind speeds decreases with height and can even become negative. During the day, on the other hand, that difference remains almost constant with height (with the exception of Karlsruhe). Offshore, the mean actual wind speed is always underestimated without a clear height dependence, but more so during the night than during the day. The E increases strongly with height during the night as absolute wind speeds also increase. During the day, the RMSE increase with height is much weaker. Near the surface, the error is smaller during the night than during the day for all MTs, and vice versa for the upper part of the profiles. Because it is normalized by the mean wind speed, Erel shows only small diurnal changes for the individual MTs. Karlsruhe, Hamburg, and Lindenberg reveal a slightly stronger decrease in Erel with height at 0000 UTC than at 1200 UTC. The other MTs do not show any significant difference.

Fig. 10.

As in Fig. 8, but for 0000 UTC values only.

Fig. 10.

As in Fig. 8, but for 0000 UTC values only.

Fig. 11.

As in Fig. 8, but for 1200 UTC values only.

Fig. 11.

As in Fig. 8, but for 1200 UTC values only.

5. Summary and conclusions

Observations of the wind speed in the lowest 20–250 m of the troposphere are analyzed and are compared with deterministic short-term forecasts of the global NWP model from ECMWF for 2009. This part of the troposphere is of special interest for the use of wind energy questions and is called wind energy layer in the context of this study. The observations stem from eight European meteorological towers, with two of them offshore. This high-quality dataset is enlarged by measurements from wind turbines at 17 different locations. Owing to the lower quality of these data, they are used only for qualitative analyses. The observation sites are offshore, in coastal, flat, hilly, and mountainous environments, with low and high vegetation, as well as with urban influences. The strongly differing site characteristics modulate the relative contributions of the synoptic-scale and smaller-scale forcings to local wind conditions. This influence finds expression in the mean wind speed and the diurnal and seasonal variations, as well as in the performance of the NWP model.

At Karlsruhe MT, where measurements are above forest vegetation and the location is surrounded by relatively complex terrain, the lowest mean WEL wind velocities are found. They range from 2.2 to 6.1 m s−1 at 20 ≤ hobs ≤ 200 m and reveal strong diurnal variations. For offshore sites, on the other hand, the sea dampens much of the diurnal thermal cycle, and synoptic driving forces dominate the wind climatological characteristics. The low roughness length translates into mean wind speeds exceeding 9 m s−1 above 70 m. Because the (onshore) diurnal variations are related to thermal mixing and decoupling, their amplitude and timing depend on height. On average, the layer below approximately 60–100 m is decoupled from elevated levels during the night, resulting in a nocturnal minimum of the wind speed. Higher levels experience a maximum during the night. This presumably can be related to the lower part of a nocturnal low-level jet (Baas et al. 2009), which has its maximum around midnight. The minimum at these higher levels is reached during the morning hours, as a result of thermal mixing of horizontal momentum downward to the surface, where the wind maximum coincides with the temperature maximum. The largest diurnal amplitudes are found for the lowest as well as for the topmost hobs. Relative to the mean wind speed, the decoupling and mixing effects are especially strong at Karlsruhe with largest near-surface amplitudes, since Karlsruhe is located in a valley with the accordingly strong diurnal variability of stability. The relative amplitude of the low-level jet is found to be similar for all onshore MTs, indicating the nocturnal decoupling of higher levels.

Deterministic ECMWF short-term forecasts with lead times of 0–9 h and with time steps of 3 h are analyzed with respect to their performance in reproducing these observations. Diverse model wind products such as the frequently used 10-m wind logarithmically extrapolated to higher elevations, the “new” 100-m model wind, and linearly interpolated wind from neighboring model levels, respectively, are compared to identify which best represents the WEL winds.

Averaged over all sites, the smallest relative bias-corrected RMSE is reached by model-level wind L90 (approximately at 34 m) as well as by the wind linearly interpolated from neighboring model levels lin with less than 24% of the mean observed wind speed, whereas it reaches 33% with ln10, the logarithmically extrapolated 10-m wind. In (very) stable conditions, which are commonly characterized by low surface winds, ln10 underestimates elevated wind speed by far (Motta et al. 2005).

For typical wind turbine heights (>70 m) in noncomplex terrain, the interpolation methods that use elevated wind information (100, ln100, 10 ln100, and lin) yield results that are similar to the best model level. The main drawback of using model levels is the required larger dataset and testing necessary to identify the best choice, which is just mostly the level beneath hobs. If a “low-cost method” is preferred that uses only one model product, the wind logarithmically interpolated from the 100-m wind to hub height is best. The bias-corrected RMSE clearly reveals that the 10-m wind products (10 m, ln10) are unsuitable for estimating WEL winds. When high-accuracy wind predictions are required (e.g., in the wind energy sector), operational model output is commonly statistically postprocessed to match the individual site characteristics. This becomes more important the more complex the terrain is. We recommend using the logarithmically extrapolated 100-m wind ln100 for all terrain. It strikes the best compromise among low cost for data, availability even when the vertical resolution of the model changes, and high-accuracy input data for postprocessing procedures. The use of near-surface model levels in complex terrain, which models are closer to reality for the wrong reason (the terrain is too smooth in the model), is not recommended given that the optimal level has to be redetermined with every change in the vertical model resolution.

The comparison of individual MT and TB sites using the linearly interpolated wind from neighboring model levels (lin) unveils a relatively strong dependence of model skill on the actual topography. Although the model performs very well for offshore and coastal, flat-terrain conditions, the deviations increase with increasingly hilly terrain. Also insufficiently resolved influences like urban or wooded environments (such as at Hamburg or Karlsruhe) diminish model skill for these sites. An increase in model resolution does not guarantee increased model skill (Rife et al. 2004). In general, the statistical postprocessing just mentioned above is more effective and thus is usually applied to match the site characteristics. Furthermore, the postprocessing can also eliminate other systematic errors such as the diurnal bias. This diurnal bias results from an underestimation of the actual diurnal cycle of the wind speed by the model.

Measurements on wind turbines provide valuable WEL wind information in addition to that coming from the sparse meteorological towers. The double-peak structure found in most TB wind distributions suggests that these data must be treated carefully by accounting for the effects of wind attenuation downstream of the rotors when cut-in wind speed is reached, because of the extraction of kinetic energy by the turbine. The comparison with the MT observations and model winds proved their general suitability for studies on WEL wind.

To maximize the energy yield, actual wind turbines are in general built at locations within a model grid box at which the highest wind speeds occur. The model wind speed, on the other hand, is an average speed over the whole grid box. The TB siting therefore in general compensates for the observational wind deficit caused by anemometer positions that are leeward of the rotor. Liu et al. (2010) suggest the use of TB observations in, for example, data assimilation. They showed that these data can (significantly) improve (0–6 h) 12-h wind power prediction.

Acknowledgments

This study was supported by the Austrian Science Fund FWF under Grant L615-N10. We are deeply indebted to all providers of observation data from meteorological towers: the Royal Netherlands Meteorological Institute (KNMI) for the Cabauw tower data obtained by Cabauw Tower Experimental Site for Atmospheric Research (CESAR); the Met Office for Cardington data; the Meteorological Institute of the University of Munich, which supplied Garching wind data; the German Bundesministerium für Umwelt, Naturschutz und Reaktorsicherheit (BMU) and Projektträger Jülich (PTJ) for FINO data; the Meteorological Institute of the University of Hamburg for providing data from the Hamburg Weather Mast; Vattenfall for supplying data from Horns Rev; the Karlsruhe Institute of Technology (KIT); and the Lindenberg Meteorological Observatory of the Deutscher Wetterdienst. Furthermore, thanks are given to all of the providers of wind speed measurements from wind turbines. Because some of them preferred to stay anonymous, they are not listed by name here. Very special thanks are given to Pierre Pinson, Gregor Giebel, Thomas Petroliagis, Carsten Maass, and Anton Beljaars, who kindly answered questions on model wind and how it is commonly used in commercial wind power prediction.

REFERENCES

REFERENCES
Baas
,
P.
,
F.
Bosveld
,
H.
Klein Baltink
, and
A.
Holtslag
,
2009
:
A climatology of nocturnal low-level jets at Cabauw
.
J. Appl. Meteor. Climatol
,
48
,
1627
1642
.
Champeaux
,
J.
,
V.
Masson
, and
F.
Chauvin
,
2005
:
ECOCLIMAP: A global database of land surface parameters at 1 km resolution
.
Meteor. Appl.
,
12
,
29
32
.
Dankert
,
H.
, and
J.
Horstmann
,
2007
:
A marine radar wind sensor
.
J. Atmos. Oceanic Technol.
,
24
,
1629
1642
.
ECMWF
,
2008
:
IFS documentation—Cy33r1. Part VI: Technical and computational procedures. ECMWF Tech. Rep., 237 pp. [Available online at http://www.ecmwf.int/research/ifsdocs/CY33r1/TECHNICAL/IFSPart6.pdf.]
Elliot
,
D. L.
,
C. G.
Holladay
,
W. R.
Barchet
,
H. P.
Foote
, and
W. F.
Sandusky
,
1986
:
Wind energy resource atlas of the United States. U.S. Dept. of Energy Rep. DOE/CH 10093-4, 210 pp
.
Frandsen
,
S.
, and
Coauthors
,
2007
:
Summary report: The shadow effect of large wind farms: Measurements, data analysis and modelling. Risø National Laboratory Tech. Rep. Risø-R-1615(EN), 36 pp. [Available online at http://www.risoe.dk/rispubl/reports/ris-r-1615.pdf.]
Giebel
,
G.
,
R.
Brownsword
,
G.
Kariniotakis
,
M.
Denhard
, and
C.
Draxl
,
2011
:
The state-of-the-art in short-term prediction of wind power: A literature overview. ANEMOS Project Deliv. Rep. D1.1, 36 pp. [Available online at http://www.risoe.dk/rispubl/vea/veapdf/ANEMOS_giebel.pdf.]
Global Wind Energy Council
,
2010
:
Görsdorf
,
U.
,
F.
Beyrich
,
H.
Dier
, and
U.
Leiterer
,
2002
:
Composite wind and temperature profiles obtained from a complex of in-situ and remote sensing measurement systems for the forcing of a boundary layer model
.
Theor. Appl. Climatol.
,
73
,
97
105
.
Gryning
,
S.
,
E.
Batchvarova
,
B.
Brümmer
,
H.
Jørgensen
, and
S.
Larsen
,
2007
:
On the extension of the wind profile over homogeneous terrain beyond the surface boundary layer
.
Bound.-Layer Meteor.
,
124
,
251
268
.
Jiménez
,
P. A.
,
J. F.
González-Rouco
,
J.
Navarro
,
J. P.
Montávez
, and
E.
García-Bustamante
,
2010
:
Quality assurance of surface wind observations from automated weather stations
.
J. Atmos. Oceanic Technol.
,
27
,
1101
1122
.
Kelly
,
M.
, and
S.-E.
Gryning
,
2010
:
Long-term mean wind profiles based on similarity theory
.
Bound.-Layer Meteor.
,
136
,
377
390
.
Lapworth
,
A.
,
2003
:
Factors determining the decrease in surface wind speed following the evening transition
.
Quart. J. Roy. Meteor. Soc.
,
129
,
1945
1968
.
Liu
,
Y.
, and
Coauthors
,
2010
:
Improving short-term wind energy prediction with wind farm data using the NCAR WRF-RTFDDA models. Abstracts, 10th European Conf. on Applications of Meteorology, Zürich, Switzerland, European Meteorological Society, EMS2010-390. [Available online at http://www.emetsoc.org/fileadmin/ems/dokumente/annual_meetings/2010/AW14_Liu.pdf.]
Méchali
,
M.
,
R.
Barthelmie
,
S.
Frandsen
,
L.
Jensen
, and
P.-E.
Réthoré
,
2006
:
Wake effects at Horns Rev and their influence on energy production. European Wind Energy Conf. (EWEC 2006), Athens, Greece, European Wind Energy Association, 10 pp. [Available online at http://johnstonanalytics.com/yahoo_site_admin/assets/docs/WakeEffectsWindTurbines.21172331.pdf.]
Motta
,
M.
,
R. J.
Barthelmie
, and
P.
Vlund
,
2005
:
The influence of non-logarithmic wind speed profiles on potential power output at Danish offshore sites
.
Wind Energy
,
8
,
219
236
.
Petersen
,
E. L.
,
N. G.
Mortensen
,
L.
Lars
,
J.
Højstrup
, and
H. P.
Frank
,
1997
:
Wind power meteorology. Risø National Laboratory Tech. Rep. Risø-I-1206(EN), 46 pp. [Available online at http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.123.5357.]
Peterson
,
E. W.
, and
J. P.
Hennessey
,
1978
:
On the use of power laws for estimates of wind power potential
.
J. Appl. Meteor.
,
17
,
390
394
.
Petroliagis
,
T.
,
M.
Denhard
,
J.
Tambke
,
R.
Hagedorn
, and
D.
Heinemann
, Eds.,
2010
:
How well we can forecast winds at different heights? An assessment of ECMWF IFS & EPS skill of forecasting wind fields at different model levels. Proc. EWEC 2010, Warsaw, Poland, European Wind Energy Association
.
Pinson
,
P.
, and
R.
Hagedorn
,
2012
:
Verification of the ECMWF ensemble forecasts of wind speed against observations
.
Meteor. Appl.
,
doi:10.1002/met.283, in press
.
Rife
,
D.
,
C.
Davis
,
Y.
Liu
, and
T.
Warner
,
2004
:
Predictability of low-level winds by mesoscale meteorological models
.
Mon. Wea. Rev.
,
132
,
2553
2569
.
Stull
,
R. B.
,
2008
:
An Introduction to Boundary Layer Meteorology. Atmospheric and Oceanographic Sciences Library, Vol. 13, Springer, 670 pp
.
Troen
,
I.
, and
E.
Petersen
,
1989
:
European Wind Atlas. Risø National Laboratory, 656 pp
.
Wenzel
,
A.
,
N.
Kalthoff
, and
V.
Horlacher
,
1997
:
On the profiles of wind velocity in the roughness sublayer above a coniferous forest
.
Bound.-Layer Meteor.
,
84
,
219
230
.