SEARCH

SEARCH BY CITATION

Keywords:

  • boundary layer parameterizations;
  • wind energy;
  • wind shear;
  • wind power forecasting;
  • WRF

ABSTRACT

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

The existence of vertical wind shear in the atmosphere close to the ground requires that wind resource assessment and prediction with numerical weather prediction (NWP) models use wind forecasts at levels within the full rotor span of modern large wind turbines. The performance of NWP models regarding wind energy at these levels partly depends on the formulation and implementation of planetary boundary layer (PBL) parameterizations in these models. This study evaluates wind speeds and vertical wind shears simulated by the Weather Research and Forecasting model using seven sets of simulations with different PBL parameterizations at one coastal site over western Denmark. The evaluation focuses on determining which PBL parameterization performs best for wind energy forecasting, and presenting a validation methodology that takes into account wind speed at different heights.

Winds speeds at heights ranging from 10 to 160 m, wind shears, temperatures and surface turbulent fluxes from seven sets of hindcasts are evaluated against observations at Høvsøre, Denmark. The ability of these hindcast sets to simulate mean wind speeds, wind shear, and their time variability strongly depends on atmospheric static stability. Wind speed hindcasts using the Yonsei University PBL scheme compared best with observations during unstable atmospheric conditions, whereas the Asymmetric Convective Model version 2 PBL scheme did so during near-stable and neutral conditions, and the Mellor–Yamada–Janjic PBL scheme prevailed during stable and very stable conditions. The evaluation of the simulated wind speed errors and how these vary with height clearly indicates that for wind power forecasting and wind resource assessment, validation against 10 m wind speeds alone is not sufficient. Copyright © 2012 John Wiley & Sons, Ltd.

INTRODUCTION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

Electricity generated from wind has a share of more than 20% in Denmark's electricity market. Consequently, accurate wind and wind power forecasts are important for both wind farm and transmission system operators: They bid their expected power generation on the energy market for the day ahead and have the possibility to correct their forecast up to half an hour before delivery. Deviations from the bid power result in penalty costs. Improving short-term wind forecasts is therefore of great economical interest to bidders of wind power.

Wind forecasts, however, should not only focus on the wind speed at hub height, neither only on that at 10 m, as sometimes performed. Forecasting the change in wind speed with height, in addition to wind speeds at a particular height, is required for wind energy applications for several reasons. Firstly, wind shear is responsible for the development of turbulent motions associated with shear-induced instability. [1] These turbulent motions can lead to vibration or damaging loading events and thus to shorter life times of wind turbines. Secondly, wind shear impacts the extraction of energy from the wind over the area of the rotor disk. [2] Lastly, the wake generated by a wind farm is affected by atmospheric stability conditions. [3] Information about the wind shear, which gives an indication for atmospheric stability, is then important when assessing the overall power losses by a wind farm. Furthermore, when the output from mesoscale models is used for wind resource assessment, information about the wind profile across the rotor area will lead to more accurate resource estimates.

Because the grid resolution in mesoscale numerical weather prediction (NWP) models is too large to explicitly resolve the processes responsible for small-scale fluxes in the planetary boundary layer (PBL), sub-grid-scale turbulent fluxes, which are mostly unknown, and vertical mixing are calculated by PBL parameterizations. [4] PBL parameterizations use the distribution of wind, temperature, and water vapor mixing ratio with height and the surface fluxes calculated from a land surface model (LSM) and/or surface layer scheme to determine, amongst many others, the time tendencies of wind, temperature and water vapor mixing ratio. The choice of PBL scheme thus plays a significant role in the evolution of the low-level wind structure and therefore can heavily impact the quality of the forecast winds.

Many aspects of the mesoscale model determine the quality of the forecast PBL structure. In the Weather Research and Forecasting (WRF) model, surface momentum, heat and moisture fluxes are calculated by a surface layer scheme that is coupled to a LSM, which in turn provides input for the PBL scheme. Because each PBL scheme is expected to run with a particular surface layer scheme, an evaluation of the performance of the different PBL schemes alone is difficult. [5] In this paper, we will show differences between simulations using the different PBL schemes while evaluating properties of the surface layer and LSMs as well. The accuracy of the forecast also depends on the accuracy of the initial and boundary conditions, domain sizes, the number and spacing of the vertical levels, and the horizontal grid spacing. The height of the first model level is apparently also important [6, 7] because it determines the minimum height of the surface layer and thus influences the surface fluxes. These factors are beyond the scope of this manuscript. Moreover, the purpose of the evaluation is to evaluate the performance of a typical real-time system, where factors like grid spacing, vertical levels or topographical features cannot be adapted a priori to the site used for validation.

The results of the PBL parameterization comparison can further depend on the physical characteristics of the site used for comparison and the grid point chosen for evaluation. Hu et al., [8] for example, concluded in their study over Texas that two of the first-order PBL schemes were outperforming a parameterization based on a turbulent kinetic energy (TKE) closure. Peña and Hahmann [9] showed that schemes performed quite differently for sites over the sea than over the land. The site used for validation in this study is situated near the coast and, although we chose the nearest model grid point over land to be compared with the measurements, it is likely that there are influences of internal boundary layers (IBLs), low-level jets and sub-layer structures not captured by the model resolution. In fact, the validation of these phenomena is an important research area in itself.

Mesoscale model simulations using five of the PBL parameterizations were compared previously by Zhang and Zheng [10] for surface wind and temperature in the Mesoscale Model Version 5 (MM5), and by Shin and Hong [5] in WRF. Previous studies tackled the comparison with fewer schemes in WRF with a different focus: Li and Pu [11] and Nolan et al. [12] studied sensitivity during hurricane events, Jankov et al. [13] discussed different WRF configurations in rainy conditions, Hu et al. [8] focused on three PBL schemes; their wind profile analysis did not meet the required resolution for wind energy purposes, however. A vast comparison of different PBL parameterizations, using single-column models, was carried out during the GEWEX Atmospheric Boundary Layer Study. [14, 15] Studies related to wind energy comprise a comparison of two boundary layer schemes with an emphasis on forecasting low-level jets, [16] as well as an analysis of the wind shear over the United States Great Plains with four PBL parameterizations. [17]

Because of the large variation in forecast wind shear in time and height among the different PBL parameterizations, 10 m winds do not represent wind conditions within the full rotor span of modern large wind turbines. It is therefore evident that validation against 10 m winds alone, as sometimes performed, is not an appropriate metric for wind energy applications. With this paper, we present a more extensive comparison as carried out before of seven PBL parameterizations available in WRFV3.1 and focus on the novel approach of analyzing not only wind speeds at hub height but also their vertical variation with height (i.e., wind shear) within that part of the boundary layer that covers the entire rotor of most modern and large wind turbines.

With this paper, our purpose is twofold. (i) By analyzing seven PBL schemes of the WRF model, we provide recommendations on which PBL parameterization performs best for wind energy purposes under similar climatological and geographical conditions as the ones from our study. (ii) We present a validation methodology for mesoscale models used for wind energy applications that takes winds at several heights into account. The paper is structured as follows: we describe the WRF model setup, the observations and the validation methods in the next section. Results are presented in Section 4. We discuss the influence of the model setup on the results and the interaction of analyzed parameters in Section 5. Finally, we conclude in Section 6.

DATA

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

WRF model setup

Our WRF model setup (Advanced Research WRF, Version 3.1) [18] consists of a main grid with horizontal grid spacing of 18 km and two nested domains (with 6 and 2 km spacing), the innermost domain covering most of Denmark (Figure 1). The model was initialized and forced at the boundaries by 1°  ×  1° US National Center for Environmental Prediction (NCEP) Global Forecast System analyses at 6 h intervals. Hence, the model simulations are hindcasts rather than forecasts. The sea surface temperature fields are also obtained from NCEP analyses at a horizontal resolution of 0.5°  ×  0.5°. Land use categories come from the United States Geological Service. We used two-way nesting between domains and 37 vertical levels, with eight levels within the lowest 500 m. The lowest levels important for wind energy applications were at approximately 14, 53, 105 and 164 m above ground level during the studied period. The model physics options included the following: Thompson microphysics scheme, Kain–Fritsch cumulus parameterization, sixth-order numerical diffusion, and positive definite advection of moisture and scalars. The rapid radiative transfer model and Dudhia schemes are used for longwave and shortwave radiation calculations, respectively. No data assimilation or grid nudging was used in the forecasts. These choices were based on experience from previous modeling systems [19] and short sensitivity experiments. A similar model setup has been used in real-time WRF forecasts at the Technical University of Denmark, Department of Wind Energy, since May 2009.

image

Figure 1. Domain configuration and terrain elevation (m) of the WRF model setup. The black squares indicate the boundaries of two nested domains.

Download figure to PowerPoint

Seven sets of 30 h hindcasts were carried out for the period 1–30 October 2009. Each hindcast was initialized at 12:00 Coordinated Universal Time (UTC  =  local standard time  −  1 h). For all the evaluations presented here, we only used hindcasts with lead times between 7 and 30 h (i.e., 24 time samples) to avoid using forecasts within the model spin-up period. The seven sets of hindcasts differ only in the PBL schemes, their associated LSMs and surface layer physics as recommended by Wang et al., [18] and are described in Table 1. All other model aspects, such as convective and radiation parameterizations, remained the same.

Table 1. Description of the seven sets of hindcasts: PBL parameterizations, their closure type (TKE), associated LSMs and surface layer physics schemes, as recommended by Wang et al. [18]
Hindcast setPBL parameterizationClosure typeLand surface modelSurface layer scheme
ACM2Asymmetric Convective Model version 2 [20]First-order closurePleim-XuPleim-Xu
MRFMedium Range Forecast Model [21]Non-local-K mixingUnified Noah LSMMonin-Obukhov
MYJMellor–Yamada–Janjic [22]TKE 1.5-orderUnified Noah LSMEta similarity
MYNN2Mellor–Yamada–Nakanishi–Niino Level 2.5 [23]TKE 1.5-orderUnified Noah LSMMYNN
MYNN3Mellor–Yamada–Nakanishi–Niino Level 3 [24]TKE 2nd-orderUnified Noah LSMMYNN
QNSEQuasi-Normal Scale Elimination [25]TKE 1.5-orderUnified Noah LSMQNSE
YSUYonsei University Scheme [26]Non-local-K mixingUnified Noah LSMMonin-Obukhov

We chose the month of October 2009 because of its variable and representative weather conditions: The synoptic situation in Denmark was characterized by a low-pressure system over Scandinavia during the first few days, interrupted by a ridge of high pressure on 5 October. From 9 October and for about 10 days onwards, anticyclonic conditions prevailed, leading to stable conditions at night and unstable conditions during the day. After that, an upper-level low-pressure system over western Europe with neutral and slightly unstable conditions determined the weather conditions over Scandinavia again, followed by a high-pressure system and characterized by stable conditions during nighttimes by the end of the month. Stable atmospheric conditions in the lower boundary layer occur often in Denmark. [27] In response to these large-scale conditions, the low-level flow at the validation site is mainly northwesterly (from the sea) during the first half of October and easterly (from land) during the second half.

Observations

The WRF model simulations were compared against measurements from a meteorological mast and a light tower at Høvsøre, situated on the northwest coast of Denmark, 1.7 km inland (Figure 2(a)). DTU Wind Energy manages the national test station for large wind turbines there, where up to five wind turbines are usually in testing. The terrain around the site is flat and homogeneous, and the prevailing wind directions are west and northwest. In our study, we use measurements of 10 min averaged wind speeds at heights of 10, 40, 60, 80, 100 and 116 m on the meteorological mast and at 160 m on the light tower; further, the temperature measurements at 2 and 100 m as well as kinematic heat fluxes were measured by a sonic anemometer at 10 m. The friction velocity u *  and the Obukhov length L were computed from the sonic fluxes. When winds are from the north (330°–30°), the mast is located in the wake of a row wind turbines (Figure 2(a)). Data within this range of directions were not used for the results presented here. The percentage of wake-free data in each atmospheric stability class (defined within intervals of L as proposed by Gryning et al. [28]) relative to all the data per stability class varies from 72–93% (Table 3) during the time of the simulations.

image

Figure 2. (Left) Map of Høvsøre, Denmark. The large dots show the location of the wind turbines, squares show the location of the meteorological mast and light tower. The lines indicate the wake zone of the wind turbines (330°–30°). The inset shows Denmark; the arrow points at Høvsøre. (Right) Location of Høvsøre (red plus sign) relative to the WRF model grid (2 km grid spacing). The green cross shows the location of the model grid point used for validation. Ocean grid squares are blue; yellow grid boxes are dryland cropland and pasture; the brown box is cropland/grassland mosaic.

Download figure to PowerPoint

METHODS OF VALIDATION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

In this paper, we propose a validation method for mesoscale models for wind energy applications. Because the hub height of most large wind turbines is between 60 and 100 m, a mesoscale model must be verified at these heights at least. Furthermore, information about the wind profile is crucial for wind energy forecasting because rotor diameters of large wind turbines are between 30 and 130 m. Within this interval, the turbines might experience a high wind speed difference across the rotor area. This is often observed in regions with shallow stable boundary layers, such as in the North Sea, where the boundary layer top can be around hub height. [29] The lower part of the rotor will then be below and the upper part above the boundary layer top.

In our comparison, we concentrate thus on verifying the model hindcasts at different model levels (Section 4.1), the wind shear (Section 4.2) and its dependence on atmospheric stability conditions (Section 4.2 and Table 3). Error metrics such as the root-mean-square-error and a newly defined ‘wind profile error’ (WPE) (Section 4.4) lead to our conclusions.

Because the WRF mass grid point closest to Høvsøre is classified as water (Figure 2(b)), the next grid point eastward is chosen for comparison against observations. This grid point displays similar distance to the coast as in reality and is classified as ‘dryland cropland and pasture’, with a surface roughness length of 0.15 m during 1–14 October and 0.05 m after. The observational estimates of surface roughness length at the site are in the range of 0.01–0.02 m. [28, 27]

We use the power exponent parameter α defined from the power law:

  • display math(1)

where u1 and u2 are the wind speeds at levels z1 and z2, respectively, to evaluate the wind shear in the observations and model simulations.

The value of α in the surface layer can be estimated from Monin–Obukhov similarity theory (MOST, e.g., Stull [30]). For a given set of heights, the value of α depends on atmospheric stability and roughness length. In our calculations, z1 and z2 represent the measurement heights of 10 and 60 m. For the model data, we use the wind speed at the first and second model levels and their respective heights at approximately 14 and 53 m. Replacing u1 and u2 in equation (1) with the logarithmic wind law and for a surface roughness of z0 = 0.02 m, which is close to the observed value at Høvsøre, [27] α ≃ 0.14. On the basis of the roughness length description of the WRF setup and under neutral conditions, the values of alpha are slightly larger: α ≃ 0.16 and α ≃ 0.19 for z0 = 0.05 m (winter value) and z0 = 0.15 m (summer value), respectively. Therefore, the value of α is quite insensitive to the local surface roughness.

Gryning et al. [28] showed, using data from Høvsøre, that on a semi-log graph, the wind profile in statically neutral situations ( | L | > 500 m) appears as a straight line, whereas wind profiles in stable (L < 500 m) and unstable (L < − 500 m) boundary layers curve downward/negatively and upward/positively with height. We thus compare the model-derived wind profiles under distinct classes of L to evaluate the ability of the various PBL schemes to represent such behavior. We calculate the WPE by averaging the root-mean-square errors for the observed wind speeds at 60, 100 and 160 m, and the second, third and fourth model levels (at 53, 105 and 164 m).

VALIDATION OF LOW-LEVEL WINDS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

The simulated wind speed was verified against observations at model levels one to four; observed temperatures against those at 2 m and on the second model level. No vertical interpolation of the model values to match the measurement heights has been carried out because the height of the model levels is often quite close to that of the observations (only at the second level, the average difference in height is 7 m). In addition, we used the observed values at Høvsøre of the shear exponent, α; the friction velocity, u * ; the kinematic heat flux at 10 m, inline image; and the Obukhov length, L. From the measurements, inline image. T0 is a fixed temperature of 20 °C, and κ and g are the von Kármán constant (κ = 0.4) and the acceleration due to gravity, respectively. For the model, the same relationship is used but with the model-derived values that are instantaneous and output hourly.

The overall error statistics between the simulated values from the seven sets of hindcasts and the observations are presented in Table 2. Biases in wind speed are mostly positive (i.e., the simulated wind speeds are higher than those observed), except for the QNSE-based (at 10 and 60 m), the MYJ-based (at 60 m) and YSU-based (at 160 m) hindcasts that have small negative biases. The absolute values of the biases are larger at the first model level than at 160 m. Smaller errors at 10 m are sometimes, but not always, correlated to smaller errors at 60 and 160 m. Temperature biases are within ± 1 °C, except for the YSU-based hindcasts at 2 m. Thus, the hindcasts tend to be slightly warmer than the observations at 2 m, an effect that could also be explained by incorrect soil moisture initialization. The biases in α are mostly positive; i.e., the values of α from the hindcasts are larger than those diagnosed from the observations. This is expected under neutral conditions, because of the differences in height (10 and 60 m in the observations versus 14 and 53 m in the simulations) and the higher surface roughness length in the model.

Table 2. Error statistics between WRF hindcast sets and the observations at Høvsøre for the period 1–30 October 2009. Only forecasts with lead times between 7 and 30 h are considered. The levels used in the wind statistics for the observations and model are 10 and 14 m, 60 and 56 m, and 160 and 164 m. The lowest BIAS or RMSE error per hindcast set is in bold.
BIAS (WRF observations) Hindcast setACM2MRFYSUMYJMYNN2MYNN3QNSE
Wind speed 10 m (ms  − 1)0.680.321.190.130.340.59 − 0.24
Wind speed 60 m (ms  − 1)0.160.340.37 − 0.200.250.37 − 0.33
Wind speed 160 m (ms  − 1)0.120.07 − 0.270.040.000.080.23
Temperature 2 m ( °C)0.680.931.210.220.400.440.29
Temperature 100 m ( °C) − 0.030.320.20 − 0.56 − 0.28 − 0.31 − 0.65
α 10/60 m ( × 10 − 1) − 0.090.33 − 0.590.380.420.220.89
RMSE (WRF observations) Hindcast setACM2MRFYSUMYJMYNN2MYNN3QNSE
Wind speed 10 m (ms  − 1)2.022.152.431.952.062.062.06
Wind speed 60 m (ms  − 1)2.522.692.632.492.662.572.56
Wind speed 160 m (ms  − 1)2.492.672.532.462.562.472.65
Temperature 2 m ( °C)1.501.551.781.261.241.281.30
Temperature 100 m ( °C)1.321.421.241.431.381.431.50
α 10/60 m ( × 10 − 1)1.301.851.831.651.791.772.17

Very small differences in RMSE of wind speeds are seen among the various sets of hindcasts, and the spread among them becomes even smaller with increasing height. The RMSE in temperature is also similar among the sets of hindcasts at 100 m, but quite different at 2 m. The RMSE in α varies between 0.130 for the ACM2-based hindcasts and 0.217 for the QNSE-based hindcasts.

Diurnal cycle

Figure 3 compares the diurnal variations in the observations at Høvsøre with those predicted by the seven sets of WRF hindcasts (Table 1). Most hindcast sets overestimated the wind speed at 10 m, especially after sunset (16:00–22:00 UTC). The overall lowest bias (0.13 ms  − 1) is that of the MYJ simulation; the highest that of the YSU simulation (1.19 ms  − 1). At 60 m, the wind speeds simulated by the various PBL-based hindcasts are within 1 ms  − 1 of the observations and most tend to underestimate its value during 1:00–7:00 UTC and overestimate it during 11:00–16:00 UTC. At 160 m, the spread in the wind speed among the various PBL-based hindcasts is much reduced from that at 10 and 60 m, but all the hindcast sets fail to reproduce the diurnal range in wind speed in the observations ( ∼ 2 ms  − 1).

image

Figure 3. Comparison of (a) wind speed (ms  − 1) at 10 m and first model level ( ∼ 14 m), (b) wind speed (ms  − 1) at 60 m and second model level ( ∼ 54 m), (c) wind speed (ms  − 1) at 160 m and fourth model level ( ∼ 164 m), (d) heat flux (Wm  − 2) at 10 m, (e) temperature (K) at 2 m; (f) temperature (K) at 100 m and third model level ( ∼ 105 m); (g) friction velocity (u * ; ms  − 1) at 10 m and (h) α as a function of time of the day (UTC hours). Model values are extracted from the land grid point closest to Høvsøre (Figure 2(a)) for each of the simulations for the period 1–30 October 2009. Only 7–30 h of hindcasts are averaged.

Download figure to PowerPoint

All the hindcast sets overestimated temperatures at 2 m (by up to 2K) compared with observations, the YSU-based hindcasts being the warmest and the MYJ-based hindcasts the coldest. Biases are generally smaller during daytime than during nighttime. At the third model level ( ∼ 105 m), the simulated temperatures are much closer to those observed at 100 m than at 2 m. The hindcasts using TKE-based PBL schemes underestimate the temperatures at this level, whereas the non-TKE-based hindcasts overestimate them. The ACM2-based hindcasts are nearly perfect with an overall bias of only 0.03 K. The phase of the diurnal cycle simulated at the third model level agrees well with the observations, whereas the phase of that simulated at 2 m seems to be 1 h ahead of that observed. Consistently, the simulated surface heat flux shows the same lag as the 2 m temperature and is under-predicted at night and over-predicts the afternoon maximum by ∼ 0.3Wm  − 2 for the QNSE-based and YSU-based hindcasts. The spread among the hindcast sets is small in terms of heat flux.

The value and diurnal evolution of the friction velocity is poorly represented in all the hindcast sets. The spread among the hindcast sets is smaller than their averaged bias to the observations. The morning and evening transitions in u *  in the observations, at about 9:00 and 15:00 UTC, exist at approximately the right time in the simulations, but the magnitudes of change of u *  in the course of the day are larger in all the hindcast sets than in the observations. The values of u *  for the YSU-based hindcasts have the highest biases of all the sets, but show very little variation during the day in better accordance with the observations.

The relationship between the values of α computed from observations and those in the hindcast sets has an interesting pattern through the day. During daytime (9:00–16:00 UTC), most sets follow quite closely the time evolution of the observed values. In the morning and at night, the spread among the hindcast sets is much larger than during the day. The two outliers are the QNSE- and YSU-based hindcasts: The first over-forecast the general evolution of α during the day, whereas the second ones show a nearly constant value throughout the day. A value close to that is expected from MOST under neutral stability conditions.

The relationship between the biases in 2 m temperature and in the temperature at the second model level is consistent with the errors in wind shear. Although most simulated 2 m temperatures are warmer than observed, the temperatures at the second model level are under-forecast. Therefore, the thermal stratification in the hindcast sets is lower than that observed (so it is closer to MOST neutral conditions).

Wind shear exponent

We now examine the value of α as a way of diagnosing the wind shear, which is an important parameter in wind energy. We focus on α between 10 and 60 m to try to avoid possible contamination with the IBL resulting from the sea–land transition at Høvsøre. The performance above this level will be examined in the next section. As shown by Floors et al., [31] from different IBL models and observations, the kink in the wind profile at Høvsøre occurs at 60–70 m. Also, this is the layer where the wind shear varies the most because of the presence of the Earth's surface.

Figure 4 shows the evolution of α during the period 1–30 October 2009 for all hindcast sets and the observations. For every hindcast set, α > 0.15 dominate during the night and morning hours, whereas α < 0.15 prevail during daytime. The distribution of α is fairly well captured by all the hindcast sets, except for the YSU-based ones, which simulate neutral conditions for most of the month at all times of the day. The block of α > 0.20 nighttime conditions during 9–19 October was captured best by the hindcasts using the TKE-based parameterizations MYNN2 and MYNN3. The QNSE-based and MYJ-based sets simulated too large values of α during that period. However, the MYNN2-based and MYNN3-based sets showed larger values for the period from about 19–26 October than those derived from the observations. During this period, the QNSE-based and ACM2-based hindcasts show better agreement with the observations. On the other hand, the MYNN2-based and MYNN3-based sets did capture the large values of α that occur during the last day of the month better than any other hindcast sets. The very low observed values of α during daytime are not well captured by any of the sets. However, some of the negative values observed during the period 10–18 October might be influenced by the wake of the wind turbines (last panel in Figure 4), which for these heights will tend to produce a more mixed boundary layer and thus lower values of α.

image

Figure 4. Comparison of the wind shear parameter α for Høvsøre as a function of time of the day (UTC, x-axis) and day of the month (y-axis) for 1–30 October 2009 computed from the wind speed values at 14 and 54 m in the WRF hindcast sets and 10 and 60 m in the observations. The title above each subplot indicates the set name (Table 1). The bottom middle panel represents the observations. The bottom right panel displays by black squares the hours when the observed wind directions correspond to those in the wake of the wind turbines (330–30°). Only 7–30 h of hindcasts are used.

Download figure to PowerPoint

Curvature of the wind speed profile

The representation of the modeled α-parameter depends on how well the models prognose the wind speed profile, but because of the selected α levels, the aforementioned wind shear analysis is valid for the layer close to the surface only. Hence, to examine the wind speed profile further up, we calculated averaged wind profiles for 1–30 October 2009 (Figure 5). The averages were separated into stability classes according to the observed Obukhov length, as indicated in Table 3. We based this analysis on the work of Gryning et al. [28] and Peña et al. [32] at Høvsøre. They showed that when plotting the wind speed normalized by the surface friction velocity as a function of the logarithm of height, the ‘stability’ of each profile is easily distinguishable: a more or less straight, curving downward and slightly curving upward line for neutral, stable and unstable conditions, respectively. This behavior is expected from similarity theory and is easily seen in the observed wind profiles (black solid lines) in Figure 5. Because the mean surface friction velocity in the WRF simulations is always larger than that observed, there is a systematic offset between the observed and modeled normalized wind profiles. This is not relevant to the analysis presented in this section because we are mainly interested in comparing the shape of the profiles. The profiles in this figure include the wind speed averages at 10 m in the simulations, which is a diagnostic quantity of each PBL scheme. Note that a larger velocity range is used in the last panel to represent larger disparity among the PBL schemes under very stable conditions.

image

Figure 5. Comparison of the observed normalized and averaged wind speed profiles (solid black thick line) and those simulated by the seven sets of hindcasts for Høvsøre. The profiles are grouped into five stability classes as described in Table 3 and indicated in the title of the plots. The thin black solid line serves as a reference for a neutral profile with z0 = 0.05 m. N indicates the number of observed cases in each plot. Times with wind directions from 330–30°are excluded from the calculations.

Download figure to PowerPoint

Table 3. Boundaries of stability classes based on the observed Obukhov length. The classes were used in the grouping of wind profiles and computation of error metrics.
Stability classObukhov length range (m)Number of samples (% of wake-free data)
Unstable − 500 ≤ L < − 5055 (72)
Neutral | L | > 500200 (93)
Near stable200 < L ≤ 500118 (93)
Stable50 < L ≤ 200136 (83)
Very stable10 < L ≤ 5072 (74)

Figure 5 shows that the difference between observed and modeled normalized wind speeds increases with height in all stability classes, but especially for stability classes stable and very stable, because the winds simulated in most of the hindcast sets are more typical of a neutral atmosphere. The hindcast set that best captured the observed curvature of the average wind profiles is the one based on QNSE. The hindcast set that had the most different average wind profile from the observed one is the one based on YSU; it shows the greatest deviation from the measured profile during stable conditions and produced ‘neutral’ wind profiles in every stability class. The other sets lie in between these two with MYJ-based, MYNN2-based, MYNN3-based and ACM2-based ones performing similarly. The largest spread among the sets is seen in the very stable class, but might not be as significant as in the other classes because of the reduced number of samples. The MYJ-based and QNSE-based hindcasts display higher winds at 10 m than at 14 m (Figure 5), which is a consequence of the diagnostic scheme used in these parameterizations.

Error metrics

When averaging wind speeds as in Figure 5, information about the daily performance of the simulations is included only if errors are systematic. However, in wind power forecasting, one single poorly forecast event can result in huge costs for transmission system operators. Therefore, a metric is needed that also includes information about outliers and penalizes them. We therefore have calculated the RMSE for wind speeds at heights 60 and 100 m, for the α-value, and the WPE for every stability class in Table 4 to objectively find the best performing hindcast set for wind profile forecasting.

Table 4. RMSE for wind speed at 60 and 100 m, for the wind profile, and the α-value for the seven sets of PBL-based hindcasts in the five stability classes (Table 3) during observed wind directions between 30 and 330°. The lowest RMSE error per stability class and PBL-based hindcast set is in bold; the highest is underlined.
 ACM2MRFYSUMYJMYNN2MYNN3QNSE
Unstable
Wind 60 m (ms  − 1)2.2122.3722.0142.2122.2292.2642.792
Wind 100 m (ms  − 1)2.4092.5562.1692.3942.3772.4103.011
WPE (ms  − 1)2.4152.5612.1942.4212.3902.4273.008
α0.0780.0840.1070.0740.0790.0770.069
Neutral
Wind 60 m (ms  − 1)2.3552.3842.4492.5792.5872.3802.750
Wind 100 m (ms  − 1)2.3772.3892.4622.4522.5012.3062.640
WPE (ms  − 1)2.3702.3792.4632.4812.5102.3212.676
α0.0480.1070.0540.0520.1030.0850.082
Near stable
Wind 60 m (ms  − 1)1.9962.0442.0782.1182.3212.0402.154
Wind 100 m (ms  − 1)2.0482.2602.1942.0622.2882.0752.074
WPE (ms  − 1)2.0602.2212.1982.1212.3292.1112.171
α0.0680.1240.0740.1010.1450.1310.143
Stable
Wind 60 m (ms  − 1)2.1022.2381.9441.8202.1272.0351.922
Wind 100 m (ms  − 1)2.2372.4812.0191.9852.1602.1022.117
WPE (ms  − 1)2.2572.4622.0932.0112.2152.1422.1647
α0.1110.1180.1430.1410.1420.1310.249
Very stable
Wind 60 m (ms  − 1)1.6191.7531.5991.4461.5381.5511.518
Wind 100 m (ms  − 1)1.8332.1102.0021.4531.6431.6191.590
WPE (ms  − 1)1.8112.0601.9511.5311.6851.7051.671
α0.1720.2200.3100.1810.2100.2070.233

The error metrics show that the performance of each hindcast set depends on the analyzed parameter, and in particular the observed atmospheric stability. For unstable conditions, the YSU-based hindcasts have the lowest RMSE for wind speed and the lowest WPE of all the simulations. For neutral and near-stable conditions, the ACM2-based hindcasts are lowest in most wind speed and α RMSE. For stable and very stable conditions, the MYJ-based hindcasts are lowest in wind speed RMSE. Despite showing very good agreement with observations in simulating the mean wind speed profile, the QNSE-based hindcasts, especially for classes unstable and neutral, have the highest wind speed RMSE and the largest WPE of all the sets.

Effect of land surface and surface layer schemes

Figures 3 and 4 showed that the results from the hindcast set using the YSU parameterization gives an almost constant value of α throughout the day and a wind profile typical of neutral conditions most of the time, whereas the other sets show variations of the α-parameter more similar to those observed. The obvious question is as follows: Are the YSU-based hindcasts, and to a lesser degree all the other hindcasts sets, deficient in the response of the wind profile to the stability conditions or are these conditions already misdiagnosed in the simulations? To answer this question, Figures 6 and 7 show the relationship in surface heat flux, surface friction velocity, Obukhov length and α between the observations and the hindcast sets. Figure 6 is for the hindcasts using the TKE-based parameterizations, whereas Figure 7 is for the first-order and non-local-K-based PBL parameterizations. All the hindcast sets use the Noah LSM, except for ACM2 that uses the Pleim-Xu LSM and a similar surface layer based on similarity theory, but with slightly different parameters and stability functions.

image

Figure 6. Comparison of observed (x-axis) versus WRF-simulated (y-axis) fields for the sets of hindcasts using TKE-based PBL parameterizations: (a)–(d) MYJ, (e)–(h) QNSE, (i)–(l) MYNN2, and (m)–(p) MYNN3. The four columns from left to right show: heat flux (Wm  − 2), surface friction velocity (ms  − 1), Obukhov Length (1 ∕ L) and α-parameter. The label bar shows the frequency of occurrence (%) of each 2D bin. Only 7–30 h of hindcasts are used; times with observed wind directions from 330–30°are excluded from the calculations.

Download figure to PowerPoint

image

Figure 7. Same as Figure 6, but for the hindcast sets using the first-order and non-local-K-based PBL parameterizations: (a)–(d) MRF, (e)–(h) YSU, and (i)–(l) ACM2.

Download figure to PowerPoint

The simulated values of heat flux in the seven hindcast sets are overall lower than those observed. But the range of heat flux in the simulations is larger than that observed. The relationship between observed and simulated friction velocity is reversed: the hindcasts overestimate u * . In the MYNN2-based, MYNN3-based, MRF-based and YSU-based hindcasts, the value of u *  is constrained to a minimum value of 0.1 by the PBL parameterization implementation. Because L is calculated as the ratio of these two previous quantities, their errors tend to compensate each other resulting in a good relationship between their observed and simulated values, especially for the simulations using the TKE-based PBL schemes (Figure 6). Reinforcing the results from Figure 3, the relationship between observed and simulated heat flux and surface friction velocity panels in Figures 6 and 7 is very similar for all the hindcast sets. In terms of 1 ∕ L, the sets using the TKE-based PBL schemes show higher correlations between observed and simulated values than those using the other PBL schemes. In terms of α, the relationship between observations and simulations varies greatly from simulation to simulation showing considerable spread. The highest correlations between observed and simulated α-values are in the ACM2 (r = 0.51) and MYJ (r = 0.42) hindcasts; the lowest are in the YSU-based ones (r = − 0.16).

DISCUSSION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

Boundary layer winds simulated by the WRF model using seven different PBL schemes have been evaluated against observations at a coastal site in Denmark with a focus on the representation of the wind shear under different atmospheric conditions. The results show that although the time-averaged wind speeds predicted by a NWP model agree well with those observed, the details in the simulated variations in time and in height, which are important in wind energy applications, are much less in agreement with the observations.

Taking into account all the metrics in Tables 2 and 4, the best performing hindcast set for this site and this period is the MYJ-based set. It shows the lowest WPE under stable and very stable conditions. Under other stability conditions, its WPE is usually very close to the lowest value. Although the QNSE-based hindcasts produced average wind profiles with a curvature similar to that of the observations, its RMSE values (Table 4) show that the individual profiles deviated considerably from those observed. Similarly, the probability distribution of α (Figure 6h) is biased towards higher values.

The results seen in the validation of the simulations reported here are in line with previous studies. Storm and Basu [17] found that simulations using the YSU parameterization produce ‘too neutral’ wind profiles. They explain its failure by noting that the parameterization has excessive mixing in the stable PBL, which destroys the near-surface shear. Shin and Hong [6] also mentioned excessive mixing in the ACM2 and YSU schemes. They further explained that in all their simulations for a single day, the value of u *  is overestimated and the schemes are incapable of reproducing the decoupling of the air at the surface and aloft at night, and thus the occurrence of low-level jets. Our study shows the same problem: too high winds near the surface at night (Figure 3(a)) except for the QNSE scheme, which is the only one trying to ‘bend back’ the wind profile above 100 m as seen in the observations (Figure 5; very stable). Hu et al. [8] concluded that the differences between their evaluated PBL schemes ACM2, YSU and MYJ were due to differences in vertical mixing strength and entrainment of air from above the PBL. The latter was not evaluated in this study because we focused on low-level boundary layer winds available from the tower measurements at Høvsøre. Validation of winds from WRF simulations against lidar measurements up to 2 km is currently underway at Høvsøre.

The observed temporal variations in wind shear clearly indicate that using 10 m winds alone is not sufficient when verifying mesoscale model simulations for wind energy applications. Firstly, an accurate simulation of winds at 10 m does not guarantee an accurate simulation of wind conditions at hub height, nor across the whole rotor area. It has also been shown in Figure 5 that the 10 m winds diagnosed by the QNSE and MYJ parameterizations are higher than those that could be expected from similarity theory (Jimmy Dudhia, personal communication). Secondly, the wind field near the ground (e.g., 10 m) is greatly affected by the local topography, roughness and presence of obstacles. Therefore, the representativeness error of wind measurements is larger than that from other atmospheric variables. Because of the inherent design and limited resolution of most NWP mesoscale models, such effects are not adequately represented in their simulations. It is thus questionable to directly compare the raw wind output of a mesoscale model with site observations at heights lower than 40 m, unless the site and model surroundings are very flat and homogeneous or it is located offshore. This fact is partly demonstrated by the larger bias in the 10 m wind than those at 60 or 160 m. A post-processing technique that uses the output from a micro-scale simulation of the site conditions using the Wind Atlas Analysis and Application Program (e.g., Troen et al. [33]) can be applied to the mesoscale model-derived winds to compensate for the features not taken into account by the NWP model as carried out by Larsén et al. [34, 35] We did not apply this technique to the results presented here because the terrain is very flat and homogeneous, except for the coastline.

Because the choice of PBL scheme in WRF is tied to a certain land surface and surface layer scheme, the results of this study are not only related to the choice of PBL scheme. Shin and Hong [5] found that the sensible heat flux depends on the surface layer parameterization rather than on the temperature gradient near the surface (i.e., on the PBL scheme). They stated that the value of the forecast 2 m temperature depends on the LSM but the wind and temperature profiles are determined by the PBL mixing algorithms. Their results are confirmed by the analysis of the simulations presented here. In addition, the difference in temperature bias between 2 and 100 m can be explained by inadequate grid spacing to resolve the transition zone between the coast and the nearest grid point to the measurement mast on land. If the flow is from western directions, which are prevailing in this climate, air parcels moving from the ocean to land, have no chance to cool as they are advected over the adjacent grid point used for validation. Land surface variability impacts turbulent surface fluxes. The variability of surface fluxes is correlated to roughness length, atmospheric stability and advection. [36] A more accurate roughness length representation could therefore result in better forecasts. Indeed, the roughness length in the WRF model changes on 15 October from 0.15 to 0.05 m, which is in the middle of the evaluation period. Figure (8) shows the time-averaged wind speed as a function of height for the observations and the simulations for the periods. The figure also shows the wind roses for these two periods derived from the observations at 60 m. The first half of October 2009 is dominated by strong offshore winds, whereas during the second half, easterly winds dominate. Therefore, mean wind speeds are stronger during the first half, when surface roughness lengths are larger, than during the second half, when surface roughness lengths are smaller. The effect of the change in roughness on 15 October is thus minimized because of these opposing changes. Very little sensitivity is found from additional simulations modifying the WRF surface roughness of the grid cells at and around the site to that observed. [37] This is partly due to the strong horizontal diffusion in WRF in this particular experiment, which makes the grid point chosen for validation more like ocean than land, especially when the wind comes from the sea, which is common at the site. Moreover, the sensitivity of the WRF model at Høvsøre during easterly and westerly winds separately has been studied additionally. [37] This study confirms that although an internal PBL forms during westerly winds, the results of our study are robust.

image

Figure 8. Time-averaged wind speed (ms  − 1) as a function of height for observations and hindcasts at Høvsøre during the periods: (a) 1–14 October 2009 and (b) 15–30 October 2009. Only values from hindcast lead times of 7–30 h are used. The insets on each plot show the wind rose based on observed winds at 60 m for these same periods. The rings are 10, 20 and 30% frequencies.

Download figure to PowerPoint

The value of z0 used in the simulations affects the value of the friction velocity. Because the value of the roughness length specified in the WRF LSM is larger than that observed at Høvsøre, the model physics compensate the value of other parameters to produce wind simulations closer to reality. This is done by artificially enhancing the value of u *  (Figure 3(g); second column of Figures 6 and 7). Changing z0 in the model to more realistic values could therefore improve the simulations of u * , but might result in unrealistic values of the simulated surface sensible and latent heat fluxes.

The Obukhov length is fairly well predicted by all the seven PBL schemes (last column of Figures 6 and 7). Furthermore, these Figures show that the heat flux, u *  and L are fairly well correlated in most of the schemes, but this is not the case for α. Separating the wind profiles in classes according to observed Obukhov length reveals that the PBL schemes tend to diffuse the surface momentum fluxes upward in a way that is not always consistent with the MOST. This is partly because the surface momentum and heat fluxes are not realistically predicted, although their combination produces relatively reasonable L values. Similar results are found in a paper by Peña and Hahmann. [38]

It can be argued that some of the problems in the representation of the wind shear in the WRF simulations could be due to the relatively low vertical resolution of the model grid. We have conducted additional simulations with enhanced vertical resolution especially in the PBL (63 vertical levels with approximate heights important for wind energy at 13, 36, 53, 68, 84, 99, 114, 130, 145, 160 and 176 m; not shown). The obtained results in terms of the mean shape of the wind profile are almost identical to those presented here. However, this conclusion might not hold over more complex terrain.

SUMMARY AND CONCLUSIONS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

This paper analyzes the results of numerical simulations with the WRF model using seven different PBL parameterizations for their performance in wind energy forecasting. We further present a validation methodology for the use of mesoscale models for wind energy that takes the vertical structure of the wind into account. Validation metrics include averaged time series, the α-parameter (a measure of the vertical wind shear), averaged wind profiles and RMSE, which were grouped into atmospheric stability classes. The results clearly indicate that for wind energy applications, i.e., wind power forecasting and wind resource assessment, validation against 10 m wind speeds alone is not sufficient. Validating an NWP model at different heights will lead to more truthful results. This becomes even more important in the future, as wind turbines get taller.

The ability of the seven hindcast sets to simulate the mean wind speed and its time variability depends strongly on atmospheric static stability. None of the PBL schemes is able to outperform the others under the range of stability conditions experienced at Høvsøre. The MYJ parameterization performs best during stable and very stable atmospheric conditions, and the ACM2 during neutral and near stable cases. The YSU scheme outperforms the others during unstable conditions. Average wind profile shapes of the QNSE scheme compare best of all schemes with the average observed wind profiles, but individual profiles actually differ the most from those observed. Especially, the YSU PBL parameterization does not exhibit correct diurnal variations and curvature of the average wind profiles and produces wind shears typical of the neutral atmosphere most of the time because of strong vertical mixing near the surface during nighttime. [5, 8]

The choice of the best model setup for a forecasting system for a particular region will thus depend on the typical distribution of atmospheric stability conditions at the site. For regions where stable conditions prevail, we recommend the MYJ PBL scheme. For regions where neutral and near stable cases dominate, we recommend the ACM2 PBL scheme. These conclusions are of course not universally applicable, because the validation presented in this study focuses on a very flat and homogenous land site. Similar conclusions were drawn for nearly the same WRF simulations over the North and Baltic Seas in a paper by Hahmann and Peña. [39] The performance of the various schemes over complex terrain is not possible to determine from the results presented here. In addition, for an evaluation of results from NWP models in more complex terrain, the representativeness of measurements and model output would have to be taken into account.

The analysis shows that the ability of the seven PBL schemes to capture the shape of the wind profile also depends on atmospheric stability. If the curvature of the wind profile simulated by all the schemes diverted more from the observation, the more stable the atmosphere is. This deviation increases at 100 and 160 m, because of the schemes’ tendency to produce profiles that are more mixed and thus with less vertical shear than those observed. An overestimation of u *  and 2 m temperature, and an underestimation of the second model level temperature were found to be partly the reason for the schemes to produce wind profiles with vertical shear expected from a neutral atmosphere. All the schemes used in the simulations underestimated the wind around turbine hub height during the night and overestimated it during the day. This diurnal compensation is coincidentally good for wind energy resource assessment on the basis of NWP model output at 60–100 m height. It is also true that the wind speed simulated by the NWP model will, somewhere in the atmosphere, match the observations. This will not happen close to the surface because of the over-prediction of u *  but somewhere higher from the ground, within the height of common wind turbines. This is important for accurate wind power forecasting, as it will pay off to find the best performing level at each site.

In this study, we focused on simulating wind and its vertical structure using seven WRF PBL parameterizations. Analyses concerning model behavior in predicting eddy diffusivity, TKE, Prandtl number, mixing length, virtual potential temperature profiles and boundary layer heights can lead to more explanations on the different behavior of the schemes. This would be a suggestion for further studies and was partly conducted for the YSU, ACM2, MYJ and QNSE PBL parameterizations by Shin and Hong. [5] Furthermore, because a PBL parameterization is not applied in isolation from the other settings of the model, ideas for future work would be to explore other aspects of the model within the validation framework that is proposed in this paper. Evaluating the hindcasts at different locations could lead to interesting results as well. A big shortcoming is the lack of tall measurement masts and the low availability of flux measurements at most sites. A longer period of study would have provided more solid conclusions; however, additional simulations at Høvsøre and other sites have shown that the conclusions are robust.

ACKNOWLEDGEMENTS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES

This work is funded by the EU project SafeWind under contract number 213740 and the Danish PSO project Radar@Sea, which is within the Danish ForskEL program under contract number 2009-1-0226. Funding from the Danish Council for Strategic Research Project number 2104-08-0025 ‘Tall Wind’ and from the EU project contract TREN-FP-7EN-219048 ‘NORSEWind’ is also acknowledged. Many thanks to Gregory Roux for programming assistance as well as Joseph Olson, Joakim Refslund, Mark C. Kelly, Rogier Floors and Sven-Erik Gryning for fruitful discussions. We also would like to thank three anonymous reviewers who provided many invaluable comments and suggestions that helped improve the quality of the manuscript.

REFERENCES

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. DATA
  5. METHODS OF VALIDATION
  6. VALIDATION OF LOW-LEVEL WINDS
  7. DISCUSSION
  8. SUMMARY AND CONCLUSIONS
  9. ACKNOWLEDGEMENTS
  10. REFERENCES
  • 1
    Kelley N, Shirazi M, Jager D, Wilde S, Adams J, Buhl M, Sullivan P, Patton E. Lamar low-level jet project interim report, Technical Report NREL/TP-500-34593, National Renewable Energy Laboratory, Golden, CO, USA, 2004.
  • 2
    Wagner R, Antoniou I, Pedersen SM, Courtney MS, Jørgensen HE. The influence of the wind speed profile on wind turbine performance measurements. Wind Energy 2009; 12(4): 348362. DOI: 10.1002/we.297.
  • 3
    Christiansen MB, Hasager CB. Wake effects of large offshore wind farms identified from satellite SAR. Remote Sensing of Environment 2005; 98: 251268.
  • 4
    Pielke RAS. Mesoscale Meteorological Modeling, 2nd edn. International Geophysics Series, Vol. 78, Academic Press: San Diego, 2002.
  • 5
    Shin HH, Hong SY. Intercomparison of planetary boundary-layer parametrizations in the WRF model for a single day from CASES-99. Boundary Layer Meteorology 2011; 139: 261281.
  • 6
    Shin HH, Hong SY, Dudhia J. Impacts of the lowest model level height on the performance of planetary boundary layer parameterizations. Monthly Weather Review 2011; 140: 664682. DOI: 10.1175/MWR-D-11-00027.1.
  • 7
    Zängl G, Gohm A, Obleitner F. The impact of the PBL scheme and the vertical distribution of model layers on simulations of Alpine Foehn. Meteorology and Atmospheric Physics 2008; 99: 105128.
  • 8
    Hu XM, Nielsen-Gammon JW, Zhang F. Evaluation of three planetary boundary layer schemes in the WRF model. Journal of Applied Meteorology and Climatology 2010; 49(9): 18311844.
  • 9
    Peña A, Hahmann AN, Hasager CB, Bingöl F, Karagali I, Badger J, Badger M, Clausen NE. South Baltic Wind Atlas. South Baltic Offshore Wind Energy Regions Project. ISBN: 978-87-550-3899-8 Risø-R-1775(EN), Risø DTU, May 2011.
  • 10
    Zhang DL, Zheng WZ. Diurnal cycles of surface winds and temperatures as simulated by five boundary layer parametrisations. Journal of Applied Meteorology 2004; 43: 157169.
  • 11
    Li X, Pu Z. Sensitivity of numerical simulation of early rapid intensification of hurricane Emily (2005) to cloud microphysical and planetary boundary layer parameterizations. Monthly Weather Review 2008; 136(12): 48194838.
  • 12
    Nolan DS, Zhang JA, Stern D. Evaluation of planetary boundary layer parameterizations in tropical cyclones by comparison of in situ observations and high-resolution simulations of Hurricane Isabel (2003). Part I: Initialization, maximum winds, and the outer-core boundary layer. Monthly Weather Review 2009; 137: 36513674. DOI: 10.1175/2009MWR2785.1.
  • 13
    Jankov I, Gallus WA, Segal M, Shaw B, Koch SE. The impact of different WRF model physical parameterizations and their interactions on warm season MCS rainfall. Weather and Forecasting 2005; 20(6): 10481060.
  • 14
    Bosveld F, Bruijn d EIF, Holtslag A. Intercomparison of single-column models for GABLS3: preliminary results, 18th AMS workshop on Boundary Layer and Turbulence, 9/6/2008-13/6/2008, Stockholm, 2008.
  • 15
    Cuxart J, Holtslag A, Beare R, Bazile E, Beljaars A, Cheng A, Conangla L, Ek M, Freedman F, Hamdi R, et al. Single-column model intercomparison for a stably stratified atmospheric boundary layer. Boundary Layer Meteorology 2006; 118: 273303.
  • 16
    Storm B, Dudhia J, Basu S, Swift A, Giammanco I. Evaluation of the Weather Research and Forecasting model on forecasting low-level jets: implications for wind energy. Wind Energy 2008; 12: 8190.
  • 17
    Storm B, Basu S. The WRF model forecast-derived low-level wind shear climatology over the United States Great Plains. Energies 2010; 3: 258276.
  • 18
    Wang W, Bruyère C, Duda M, Dudhia J, Gill D, Lin HC, Michaelakes J, Rizvi S, Zhang X. WRF-ARW Version 3 Modeling System User's Guide, Mesoscale & Microscale Meteorology Division, National Center for Atmospheric Research: Boulder, USA, 2010.
  • 19
    Liu Y, Warner TT, Bowers JF, Carson LP, Chen F, Clough CA, Davis CA, Egeland CH, Halvorson SF, Huck TW, et al. The operational mesogamma-scale analysis and forecast system of the U.S. army test and evaluation command. Part I: Overview of the modeling system, the forecast products, and how the products are used. Journal of Applied Meteorology and Climatology 2008; 47: 10771092. DOI: 10.1175/2007JAMC1653.1.
  • 20
    Pleim J. A combined local and nonlocal closure model for the atmospheric boundary layer. Part i: Model description and testing. Journal of Applied Meteorology and Climatology 2007; 46: 13831395.
  • 21
    Troen I, Mahrt L. A simple model of the atmospheric boundary layer; sensitivity to surface evaporation. Boundary Layer Meteorology 1986; 37: 129148.
  • 22
    Mellor GL, Yamada T. Development of a turbulence closure model for geophysical fluid problems. Reviews of Geophysics and Space Physics 1982; 20: 851875.
  • 23
    Janjic ZI. Nonsingular implementation of the Mellor–Yamada level 2.5 scheme in the NCEP Meso model. Technical Report, National Centers for Environmental Prediction, Camp Springs, MD, USA, 2001.
  • 24
    Nakanishi M, Niino H. An improved Mellor-Yamada Level-3 model: its numerical stability and application to a regional prediction of advection fog. Boundary Layer Meteorology 2006; 119: 397407.
  • 25
    Sukoriansky S, Galperin B, Perov V. A quasi-normal scale elimination model of turbulence and its application to stably stratified flows. Nonlinear Processes in Geophysics 2006; 13: 922.
  • 26
    Hong SY, Noh Y. Dudhia. A new vertical diffusion package with an explicit treatment of entrainment processes. Monthly Weather Review 2006; 134: 23182341.
  • 27
    Peña AD. Sensing the wind profile. ISBN: 978-87-550-3709-0 Risø-PhD-45(EN), Risø DTU, 2009.
  • 28
    Gryning SE, Batchvarova E, Brümmer B, Jørgensen H Larsen. On the extension of the wind profile over homogeneous terrain beyond the surface layer. Boundary Layer Meteorology 2007; 124: 251268.
  • 29
    Giebel G, Gryning SE. Shear and stability in high met masts, and how WAsP treats it, EWEA Special Topic Conference “The Science of Making Torque from Wind”, Delft (NL), 2004; 356363.
  • 30
    Stull RB. An Introduction to Boundary Layer Meteorology, Atmospheric and Oceanographic Sciences Library, Vol. 13, chap. 9, Springer, 1988: 670. ISSN 978-90-277-2769-5.
  • 31
    Floors R, Gryning SE, Peña A, Batchvarova E. Analysis of diabatic flow modification in the internal boundary layer. Meteorologische Zeitschrift 2011; 20: 649659. DOI: 10.1127/0941-2948/2011/0290.
  • 32
    Peña A, Gryning SE, Hasager CB. Comparing mixing-length models of the diabatic wind profile over homogeneous terrain. Theoretical and Applied Climatology 2010; 100: 325335.
  • 33
    Troen I, Petersen EL. European Wind Atlas, Risø National Laboratory: Roskilde, Denmark, 1989. ISBN: 87-550-1482-8.
  • 34
    Larsén XG, Mann J, Berg J, Göttel H, Jacob D. Wind climate from the regional climate model REMO. Wind Energy 2010; 13: 279296.
  • 35
    Larsén XG, Ott S, Badger J, Hahmann AN, Mann J. Recipes for correcting the impact of effective mesoscale resolution on the estimation of extreme winds. Journal of Applied Meteorology and Climatology 2012; 51: 521533.
  • 36
    Talbot C, Bou-Zeid E, Smith JA. Impact of land-surface variability on turbulent surface fluxes in multi-scale atmospheric simulations, 19th Symposium on Boundary Layers and Turbulence, Keystone, CO , USA, 2010.
  • 37
    Floors R, Vincent CL, Gryning SE, Batchvarova E, Peña A. Wind profile in the coastal boundary layer: wind lidar measurements and WRF modelling. in review Boundary Layer Meteorology 2012.
  • 38
    Peña A, Hahmann AN. Atmospheric stability and turbulence fluxes at Horns Rev—an intercomparison of sonic, bulk and WRF model data. Wind Energy 2012; 15: 717731. DOI: 10.1002/we.500.
  • 39
    Hahmann AN, Peña A. Validation of boundary-layer winds from WRF mesoscale forecasts over Denmark, European Wind Energy Conference and Exhibition, Warsaw, Poland, 2010.