Received: 30 October 2021 Revised: 20 May 2022 Accepted: 23 May 2022

DOI: 10.1002/we.2759

SHORT COMMUNICATION

Limitations of reanalysis data for wind power applications

Michael R. Davidson1,2 | Dev Millstein3

1
Department of Mechanical and Aerospace
Engineering, University of California San Abstract
Diego, San Diego, California, USA Wind energy resource estimates commonly depend on simulated wind speed profiles
2
School of Global Policy and Strategy,
generated by reanalysis or weather models due to the lack of long time series mea-
University of California San Diego, San Diego,
California, USA surements with sufficient coverage at relevant heights (roughly 90 m above ground).
3
Lawrence Berkeley National Laboratory, However, modeled data, including reanalyses, can be noisy and display a wide range
Berkeley, California, USA
of biases and errors, variously attributed to terrain effects, poor coverage of assimi-
Correspondence lated inputs, and model resolution. Wind generation records, if available at high tem-
Michael R. Davidson, Department of
Mechanical and Aerospace Engineering, poral and geographical resolution, can provide a proxy for wind measurements and
University of California San Diego, 9500 allow for evaluation of reanalyses and weather model wind time series. We use a
Gilman Dr, San Diego, CA 92093-0411, USA.
Email: [email protected] 7-year-long data set of hourly, plant-level generation records from over 100 wind
Dev Millstein, Lawrence Berkeley National plants across Texas to evaluate two commonly used reanalysis data sets (MERRA2
Laboratory, 1 Cyclotron Rd, Berkeley, CA and ERA5). Additionally, we use 1-year of records (2019) to evaluate an operational,
94720, USA.
Email: [email protected] high-resolution regional weather modeling product (HRRR v3). We find that across
the region, and across all modeling products, the modeled representation of wind
generation (i.e., wind speeds at hub heights passed through a power curve) has rela-
tively small mean errors when aggregated daily, but that accuracy and hourly correla-
tion have a strong diurnal sensitivity. Accuracy and correlation systematically decline
through the evening and markedly improve after sunrise. These diurnal patterns per-
sist even in the highest resolution model tested (HRRR v3). We hypothesize the
nighttime decline in accuracy is mostly due to poorly represented boundary layer
conditions, perhaps related to model representation of stability, while other uncer-
tainties (such as wake effects) play a secondary role.

KEYWORDS
boundary layer, reanalysis, wind power, wind resource assessment

1 | I N T RO DU CT I O N

Net-zero emissions energy systems will rely upon large increases in wind and solar energy, whose intermittency characteristics create challenges
for power systems planning and operations. Measurements of relevant renewable energy availability metrics (e.g., wind speeds at wind turbine
hub heights, 90 m above ground) are generally sparse geographically and limited in time coverage, necessitating the use of simulated renewable
energy profiles generated by weather models. Planning exercises to manage continent-scale renewable energy and accompanying infrastructure
expansion depend upon not only aggregate availability but also the temporal effects of spatial deployment.1 For the case of wind, however, there
is little consistency in the generation profiles and models adopted by various studies, and frequently little to no validation is done.2 Examination

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,
provided the original work is properly cited.
© 2022 The Authors. Wind Energy published by John Wiley & Sons Ltd.

1646 wileyonlinelibrary.com/journal/we Wind Energy. 2022;25:1646–1653.

DAVIDSON AND MILLSTEIN 1647

of simulated and actual wind power output at high spatial and temporal granularity shows significant biases and errors across a wide range of
weather models and geographies.3
Validation and error correction of these inputs, if performed at all, are often limited to aggregate correction factors such as country-wide
adjustments.2 Sophisticated correction methodologies are sometimes used,4–6 but these approaches would benefit from enhanced validation of
the relevant meteorological models. In general, error-correction approaches can capture persistent, long-term biases in the input weather data
and, to some extent, shorter-term phenomena, though necessarily moderated by the effects of aggregation.
Renewable energy analyses have largely relied upon two types of weather models: assimilation-based reanalysis and numerical weather pre-
diction (NWP). Reanalysis models construct a physically consistent model of the atmosphere driven by an underlying physical model and con-
strained by a wide range of historical observations from both the surface and satellites.7 Their primary advantages lie in the global coverage and
long historical record (exceeding four decades in some cases). NWPs simulate and forecast the atmosphere in real time from a set of initial condi-
tions, though they can also incorporate assimilation processes of available observations.8 NWPs generally have higher geographic resolution and
are more computationally intensive, thus limiting them to specific regions and a shorter historical record.
Key to the challenge of estimating spatially and temporally granular wind power profiles is the representation of the boundary layer, the slice
of the atmosphere from the ground up to where synoptic flows become dominant. In the boundary layer, friction with the ground and varying
thermal stratification (or stability) complicate the description of vertical wind profiles.9 Two commonly used formulae to extrapolate wind speeds
to different heights (e.g., turbine hub height) both rely upon boundary layer parameterizations in terms of wind shear: the “log law” derived from
Monin–Obukhov similarity theory and neutral stratification assumptions (wind shear parameter = 1) and the “power law” heuristic with a wind
shear exponent commonly assumed to be 0.14 for neutral conditions.10,11 Generalizations of the “log law” to nonneutral conditions incorporate a
wind shear parameter that varies with turbulence or vertical heat flux and, in some cases, empirically determined stability functions to account for
further discrepancies.12 Model performance of strongly stratified nocturnal regimes is highly dependent on land–surface interactions.13 Low-lying
jets, for example, are challenging to predict and can result in coherent wind conditions over large areas14 and impact wind speeds at wind energy
relevant heights.15,16
Validations of reanalysis surface wind fields are frequently limited by data availability to low heights (e.g., 10 m) where boundary layer parame-
terization errors would be less pronounced.17 Validation of higher fields has shown stability is a predictor of mean errors.18 Recent work has also
found that interannual variability in recorded wind generation is only weakly correlated with observed surface wind speeds, suggesting that wind
turbines are subject to meteorological phenomena above the surface boundary layer.19 Furthermore, ground observations of surface heat fluxes
(an important diurnally varying signature of stratification) do not provide sufficient geographic coverage to usefully constrain reanalysis models.20
Together, these points indicate that while reanalysis outputs, such as wind speed, are most often validated at the surface (10 m above ground), that
surface validation may not provide much insight into accuracy at wind turbine heights as different phenomena affect meteorology at hub heights.
The lack of validation of reanalysis wind speeds at wind plant hub heights represents a research gap for the wind energy community. To
address this gap, we use a unique set of plant-level, hourly generation records across more than a hundred wind plants in Texas to assess the abil-
ity of two commonly used reanalysis models to represent wind speeds at relevant heights for wind energy. Though not direct measurements of
wind speed, recorded generation here provides useful observational data at many locations and, importantly, reflects winds well above ground
that impact wind generation but cannot necessarily be observed from ground-based anemometers. To compare reanalysis data to recorded gener-
ation, we convert reanalysis wind speeds to energy generation using manufacturer power curves. Though this conversion adds some uncertainty,
differences between recorded and estimated generation are dominated by variation in the underlying wind speed estimates, not from the power
curve conversion.
We examine two reanalysis products: MERRA-2 and ERA5. MERRA2 is produced by the National Aeronautics and Space Administration
(NASA) Global Modeling and Assimilation Office (GMAO), and ERA5 is produced by the European Centre for Medium-Range Weather Forecasts
(ECMWF).7,21 In order to explore issues related to model resolution, we also assess recent data (from 2019) from the High-Resolution Rapid
Refresh (HRRR) v3 NWP of the National Oceanic and Atmospheric Administration (NOAA) Global Systems Laboratory (GSL) and Earth System
Research Laboratories (ESRL).22

2 | MATERIALS AND METHODS

2.1 | Wind generation data

Plant-level, hourly wind generation data were provided directly by ERCOT. For each plant, ERCOT provided both records of hourly generation
sold to markets along with estimates of generation had curtailment not been applied. It is the latter set of records that we compare to reanalysis-
based generation estimates. These generation records have been used in prior publications and provide some of the only hourly, plant-level gener-
ation records that are publicly available in the United States.5,23 Minimal data cleaning was conducted to remove sites with incomplete records
and to truncate capacity factors at nominal power capacity (i.e., CF ≤ 1).
1648 DAVIDSON AND MILLSTEIN

Characteristics of wind plants in the ERCOT region were defined based on the US Wind Turbine Database. Characteristics of interest include
project centroid, average turbine hub height, overall plant capacity, and turbine manufacturer and type.24

2.2 | Wind reanalysis construction

We use hub-height wind speeds from three meteorological/reanalysis models: MERRA2, ERA5, and HRRR v3. We combine these wind speeds
with a manufacturer's power curve, specific to each plant, to estimate hourly generation implied by the meteorological model. Power curves were
selected to match the dominant technology at each wind plant (i.e., either all turbines were the same, or we selected a turbine that matched the
majority of turbines within a plant). A database of power curves was purchased from “thewindpower.net,” and in the rare case where a direct
match of manufacture and type was not found for a power curve, we selected an alternate power curve based on the physical characteristics of
the turbines, based on matching rotor diameter and turbine capacity.
MERRA2 provides hourly wind speeds at 2, 10, and 50 m above ground. We first matched each plant centroid to the closest MERRA2 grid
cell. To extrapolate winds from 50 m to hub heights of 80–90 m (depending on the plant), we used a log law to estimate the scaling factor for each
location and hour, following the approach used by Staffell and Green.25 We then applied the power curve to these hub-height wind speeds to
develop plant-level generation estimates.
ERA5 provides hourly, three-dimensional files which include wind speed at multiple heights. We matched each plant centroid to the clos-
est ERA5 grid cell. We then interpolated between vertical layers, using a cubic spline function, to find the hub-height wind speed at each hour.
The ERA5 grid cell layers, at the heights of interest, ranged from 20 to 40 vertical meters. We note that the interpolation introduces less
uncertainty compared with the extrapolation that was necessary with the MERRA2 data, as the ability to incorporate wind speed estimates
above and below the hub heights of wind plants ensured that the results were internally consistent with the wind shear represented by the
ERA5 model.
For HRRR, we also matched each plant centroid to the closest HRRR grid cell. Given the higher resolution of HRRR, wind plants often span a
larger area than the HRRR grid cells themselves (this is not the case for ERA5 and MERRA2). Thus, matching to a single grid cell includes a simpli-
fying assumption that the grid cell corresponding to the centroid of each plant is generally representative of the overall plant conditions. HRRR
provides hourly wind speeds at 80 meters above ground. We used this output directly, and, to maintain computational tractability, we did not
adjust for the relatively small differences that would be caused by the adjustment from 80 meters to actual wind plant hub heights (usually
between 80 and 90 meters of height).

2.3 | Error and correlation analysis

We evaluate weather data against actual generation along three main dimensions: mean error, correlation coefficient, and root mean squared
error (RMSE). To maintain focus on errors inherent in modeled estimation of wind speed, we have chosen to neglect explicit modeling of a genera-
tion loss factor, which would substantially expand the scope of analysis. Typically, a loss factor would be used to account for mechanical losses,
wind plant wake effects, and maintenance and other downtime.3,26 We expect neglecting losses to result in a small positive bias and higher RMSE,
but a minor or negligible impact on correlation.
We also introduce a new variable for solar time (SolTime), which translates timestamps according to local sunrise (SolTime = 0). This allows us
to highlight the impact of diurnal weather cycles and indeed show much stronger effects compared to typical approaches by hour of day. We
group times into hourly SolTime bins for the purposes of generating statistics.

2.4 | Capacity factor definition

Capacity factor (CF) represents a plant's realized generation divided by the total possible generation assuming the plant was operating at full
capacity in all hours, expressed as a percentage. Throughout, when evaluating CF errors, we use % CF to refer to capacity factor percentages, not
percentage deviations from an absolute wind generation level or relative to a given capacity factor. For example, the difference between a model-
estimated CF of 32% and an actual CF of 35% is 3%. We do this for two reasons: (a) to compare wind plants of different sizes and (b) to account
for power system operators' primary concerns resting with biases and errors in total energy delivered (hence, percentage point differences) as
opposed to relative biases, which would tend to exaggerate the impact of small differences at low wind output.
DAVIDSON AND MILLSTEIN 1649

3 | RESULTS AND DISCUSSION

We find that both ERA5 and MERRA2 represent average levels of the wind resource accurately, with small mean errors (i.e., bias) of 1.1% CF and
6.4% CF, respectively. We expect that inclusion of loss factors would push both models toward small negative average mean errors but that the
rough characterization of “small” average bias would remain acceptable.
Averaging over time dampens differences between modeled and recorded capacity factors. Daily capacity factors, for example, show a similar
positive bias as hourly, but with a tighter distribution. Hourly (daily) standard deviations of MERRA2 and ERA5 errors are 21.5% CF (6.9% CF) and
19.0% CF (7.0% CF), respectively, as shown in Figure 1. Hourly errors for both ERA5 and MERRA2 show a strong peak at 0% CF, corresponding
to a zero mode in both actual and estimated generation when wind speeds are below the cut-in speed of the turbines (see Figure 1). The larger
spread away from 0% error for hourly versus daily comparisons shows that a substantial amount of the mean error derives from diurnal fluctua-
tions that are averaged out when aggregating to the daily level. For example, the probability that the estimated CF is within ±20% CF of actual
hourly (daily) for MERRA2 and ERA5 is 70.3% (97.0%) and 78.0% (99.1%), respectively.
Evidence of the pivotal role of the diurnal cycle to reanalysis model representation of wind energy resources is given by examining three eval-
uation metrics in terms of solar time (see Figure 2). In particular, RMSE and correlation coefficients show a rapid improvement following sunrise
(compared to the nighttime hours). For example, ERA5 RMSE falls from 21.7% CF 4 h prior to sunrise to 15.7% CF 5 h after. Similarly, ERA5

F I G U R E 1 Hourly and daily mean CF errors by location (gray) and entire sample (black). CF error is calculated as model (%CF) minus recorded
generation (%CF)

FIGURE 2 Diurnal mean error, RMSE, and correlation coefficient in local solar time (sunrise = 0)
1650 DAVIDSON AND MILLSTEIN

correlation coefficients rise from 0.78 just prior to sunrise up to 0.87 at 5 h after sunrise. MERRA2 shows even larger improvements following
sunrise. A potential explanation for this diurnal pattern is that reanalysis models systematically underestimate nighttime stability in the boundary
layer, as discussed in Section 4.
The diurnal cycle of mean error is a bit harder to interpret within this context, but we do see that ERA5 has a limited range of mean error over
the day, while MERRA2 shows a strong diurnal trend. Again, when incorporating loss factors, we would expect mean error to shift downward;
thus, the fact that the mean error for both MERRA2 and ERA5 is close to zero in the hours before sunrise (Figure 2) suggests that both models
underestimate wind speeds in those hours. We note the improvement to RMSE and correlation coefficients that occurs after sunrise, suggesting
that issues related to the nighttime representation of the surface boundary layer impact the models' ability to produce accurate wind energy
resource estimates.
Further separating the data into season, in Figure 3, we find that MERRA2 and ERA5 perform worse (as measured by the correlation coeffi-
cient) at nighttime compared to daytime across all seasons. Both models tend to perform worse during summer nighttime compared to winter
nighttime, while seasonal variations in daytime remain somewhat smaller.
One hypothesis for the improved model performance in winter over summer is the possible correlation of model performance with wind
speed. If the models' performance improved at higher wind speeds compared to lower wind speeds, then the increased frequency of high wind
speed hours during winter, compared with summer, could explain the improved model performance during the winter. However, we find that
while model performance improves with the highest wind speeds during winter daytime and nighttime hours (as evidenced by declining RMSE
with high CF; Figure 4), the opposite relationship is found during summer nighttime hours (high CF summer night hours have some of the highest
levels of RMSE during summer; Figure 4). Here, daytime is defined as times between 1 and 8 h after sunrise, and night is defined as 8 to 1 h
before sunrise. This seasonal pattern is intriguing, as seasonal meteorological differences might provide further clues as to what is driving model
performance, but a focused study with a larger set of atmospheric variables would be required to fully explore the causes of this seasonal pattern.

FIGURE 3 Diurnal correlations by season in local solar time (sunrise = 0)

F I G U R E 4 RMSE as a function of recorded CF bins for winter (DJF) and summer (JJA) and nighttime (8–1 h before sunrise) and daytime
(1–8 h after sunrise)
DAVIDSON AND MILLSTEIN 1651

FIGURE 5 Correlation coefficients by location

FIGURE 6 Diurnal correlations by season in solar time (sunrise=0) for MERRA2, ERA5, and HRRR (2019)

Geographic variation among plants in terms of correlation coefficients was somewhat muted. ERA5 correlation coefficients range from 0.63
to 0.88, and MERRA2 correlation coefficients range from 0.50 to 0.83 (see Figure 5). Besides slightly higher correlation coefficients at the coastal
wind plants, there is no clear geographic explanation for this variation. Elevation and rough terrain increase further north into the Texas panhan-
dle, but this is not the main driver for variation in correlation coefficients or the other metrics.
One potential concern of using MERRA2 and ERA5 wind speeds to estimate wind generation is the coarse model resolution (50 and 30
km, respectively). This resolution is improved compared to past global reanalysis products, but is still coarse compared to regional meteorological
modeling products. To address the potential influence of model resolution, we analyzed for year 2019 the output of the HRRR v3 model at a reso-
lution of 3 km. Hourly correlations between recorded and modeled generation are indeed improved in HRRR compared to the lower-resolution
global analysis models. Nevertheless, a similar decline in HRRR hourly correlations occurs during nighttime hours (Figure 6).

4 | C O N CL U S I O N

We use a unique data set of hourly, plant-level, wind generation records across the ERCOT region in Texas to validate wind speeds at relevant
hub heights from three weather models commonly used in power system studies. We find that, in this region, RMSE and hourly correlation is
dominated by diurnal patterns while errors are reduced substantially with daily averaging. Furthermore, after converting to solar time (setting the
origin at local sunrise), we identify a strong nighttime/daytime sensitivity to all evaluation metrics, with models performing worse in the hours
before sunrise than after. Notably, all models regardless of whether we extrapolate or interpolate to hub height, or use a fixed model layer near
hub height, show the same behavior, indicating systemic errors that affect using derived wind speeds for hourly energy modeling. Enhancing grid
resolution (e.g., moving from 50 km in MERRA2 to 3 km in HRRR v3) does improve evaluation metrics. That said, HRRR nighttime, though
improved over MERRA2 and ERA5 nighttime, still has worse hourly correlation coefficients than the relatively low-resolution MERRA2 does dur-
ing daytime, implying that the diurnal pattern in errors is still an important issue despite the improved resolution of HRRR.
1652 DAVIDSON AND MILLSTEIN

We hypothesize, but do not prove, that the nighttime decline in accuracy of NWP and reanalysis models is due to the incomplete characteri-
zation of boundary layer phenomena, an issue previously identified as a challenge for meteorological modeling. Though nighttime model perfor-
mance is worse than daytime across all seasons, models performed worse during summer nighttime hours. This seasonal difference may be
influenced by model performance at high wind speeds. For example, we found that though model performance (measured by RMSE) improves
with higher wind speeds during winter daytime and nighttime hours, this is not the case during summer nighttime, where RMSE was particularly
high during high wind speed hours. Worse performance at nighttime could be due to an underestimation of stable conditions and poor represen-
tations of phenomena such as low-lying jets in state-of-the-art meteorological models. Small-scale variations in the nocturnal boundary layer
driven by topography could decrease wind speed correlations.27 Furthermore, atmospheric stability and turbulence can directly impact power out-
put even while holding hub-height wind speeds constant.28,29
Our simple approach for modeling generation from reanalysis, or NWP, wind speeds has limitations. Our modeled generation estimates do
not account for issues such as wake losses, shutdowns due to extreme cold, air density changes, and precipitation impacts. We found that correla-
tions were higher during winter than summer, suggesting that extreme cold, at least, was not a dominant driver of RMSE. We hypothesize that
the simplicity of our modeling approach does not drive the diurnal or seasonal trends in RMSE and correlation and that more complicated model-
ing efforts would find similar results.
As the wind energy community seeks to answer increasingly difficult questions about power systems operations under high penetration sce-
narios, more attention should be paid to the renewable resource profile inputs. Annual bias correction and plant-specific loss factors can improve
the utility of weather model-derived wind speeds, but this work points to more challenging structural deficiencies in the weather models that are
not easily removed through simple correction factors or improved model resolution. The multi-decadal trend toward taller towers and longer
blade sizes23 means that wind turbines tips are reaching ever higher heights, in some case above 200 m,30 and thus more frequently interact with
winds above the surface boundary layer. In order to best serve the wind energy community, providers of global and regional weather models
should, in the short term, work toward outputting hourly variables at multiple heights to eliminate the potential for extrapolation-based errors.
For the longer term, directing research toward improving the representation of nighttime surface boundary layer and wind flows at wind energy
relevant heights (i.e., order of 100 m) would be helpful in improving overall representation of wind energy resources.

ACKNOWLEDGEMEN TS
The authors thank Seongeun Jeong and Amos Ancell of Lawrence Berkeley National Laboratory for assistance with preparation of the meteoro-
logical data. The authors thank ERCOT for sharing plant-level generation data. Michael Davidson would like to thank Graduate Climate Confer-
ence participants and American Geophysical Union Annual Meeting attendees for providing feedback on earlier versions of this work and the MIT
Joint Program on the Science and Policy of Global Change and Tianyu Qi for early inspiration. Support for the work of author Dev Millstein was
provided by the US Department of Energy's Office of Energy Efficiency and Renewable Energy under Lawrence Berkeley National Laboratory
contract no. DE-AC02-05CH11231. The US Government retains, and the publisher, by accepting the article for publication, acknowledges, that
the US Government retains a non-exclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript,
or allow others to do so, for US Government purposes.

P EE R R EV I E W
The peer review history for this article is available at https://publons.com/publon/10.1002/we.2759

DATA AVAI LAB ILITY S TATEMENT

MERRA2, ERA5, and HRRR meteorological data are freely available to the public. Recent, plant-level, hourly generation records in ERCOT are
publicly available, but the long-term generation records used in this analysis were provided directly to Lawrence Berkeley National Laboratory
from ERCOT. Additionally, manufacturer power curves were purchased via license from thewindpower.net. Therefore, not all of the data used can
be posted publicly. However, to the extent possible, datasets will be made available upon request. Please contact author Dev Millstein
([email protected]) for information related to generation and meteorological data and author Michael Davidson ([email protected]) for infor-
mation related to validation scripts.

ORCID
Michael R. Davidson https://orcid.org/0000-0002-5035-8209
Dev Millstein https://orcid.org/0000-0002-8091-0535

RE FE R ENC E S
1. Grams CM, Beerli R, Pfenninger S, Staffell I, Wernli H. Balancing Europe's wind-power output through spatial deployment informed by weather
regimes. Nat Clim Change. 2017;7(8):557-562. doi:10.1038/nclimate3338
DAVIDSON AND MILLSTEIN 1653

2. Staffell I, Pfenninger S. Using bias-corrected reanalysis to simulate current and future wind power output. Energy. 2016;114:1224-1239. doi:10.1016/
j.energy.2016.08.068
3. Jourdier B. Evaluation of ERA5, MERRA-2, COSMO-REA6, NEWA and AROME to simulate wind power production over France. Adv Sci Res. 2020;17:
63-77. doi:10.5194/asr-17-63-2020
4. Olauson J, Bergkvist M. Modelling the Swedish wind power production using MERRA reanalysis data. Renew Energy. 2015;76:717-725.
5. Millstein D, Wiser R, Mills AD, Bolinger M, Seel J, Jeong S. Solar and wind grid system value in the United States: the effect of transmission conges-
tion, generation profiles, and curtailment. Joule. 2021;5(7):1749-1775. doi:10.1016/j.joule.2021.05.009
6. Wiser R, Millstein D, Bolinger M, Jeong S, Mills A. The hidden value of large-rotor, tall-tower wind turbines in the United States. Wind Eng. 2021;
45(4):857-871.
7. Gelaro R, McCarty W, Suárez MJ, et al. The modern-era retrospective analysis for research and applications, version 2 (MERRA-2). J Climate. 2017;
30(14):5419-5454. doi:10.1175/JCLI-D-16-0758.1
8. Alexander C, Dowell DC, Hu M, et al. Rapid refresh (RAP) and high resolution rapid refresh (HRRR) model development. In: 100th American Meteoro-
logical Society Annual Meeting. AMS; 2020.
9. Emeis S. Wind Energy Meteorology. Springer; 2013. Accessed October 30, 2014. http://link.springer.com/10.1007/978-3-642-30523-8
10. Jacobson MZ. Fundamentals of Atmospheric Modeling. Cambridge University Press; 2005.
11. Ll L, Torralba V, Soret A, Ramon J, Doblas-Reyes FJ. Seasonal forecasts of wind power generation. Renew Energy. 2019;143:91-100. doi:10.1016/j.
renene.2019.04.135
12. Optis M, Monahan A. Vertical wind profiles above the surface layer under stable conditions-an analysis of existing model performance. In: European
Wind Energy Association; 2012. Accessed November 1, 2014. http://proceedings.ewea.org/annual2012/allfiles2/1464_EWEA2012presentation.pdf
13. Baas P, van de Wiel BJH, van der Linden SJA, Bosveld FC. From near-neutral to strongly stratified: adequately modelling the clear-sky nocturnal
boundary layer at Cabauw. Boundary-Layer Meteorol. 2018;166(2):217-238. doi:10.1007/s10546-017-0304-8
14. Heppelmann T, Steiner A, Vogt S. Application of numerical weather prediction in wind power forecasting: assessment of the diurnal cycle. Meteorol Z.
2017;26(3):319-331. doi:10.1127/metz/2017/0820
15. Banta RM, Newsom RK, Lundquist JK, Pichugina YL, Coulter RL, Mahrt L. Nocturnal low-level jet characteristics over Kansas during xases-99. Bound-
Lay Meteorol. 2002;105(2):221-252. doi:10.1023/A:1019992330866
16. Vanderwende BJ, Lundquist JK, Rhodes ME, Takle ES, Irvin SL. Observing and simulating the summertime low-level jet in central Iowa. Mon Weather
Rev. 2015;143(6):2319-2336. doi:10.1175/MWR-D-14-00325.1
17. Carvalho D. An assessment of NASA's GMAO MERRA-2 reanalysis surface winds. J Climate. 2019;32(23):8261-8281. doi:10.1175/JCLI-D-19-0199.1
18. Rose S, Apt J. Quantifying sources of uncertainty in reanalysis derived wind speed. Renew Energy. 2016;94:157-165. doi:10.1016/j.renene.2016.
03.028
19. Millstein D, Bolinger M, Wiser R. What can surface wind observations tell us about interannual variation in wind energy output? Wind Energy. 2022;
25(6):1142-1150. doi:10.1002/we.2717
20. Draper CS, Reichle RH, Koster RD. Assessment of MERRA-2 land surface energy flux estimates. J Climate. 2017;31(2):671-691. doi:10.1175/JCLI-D-
17-0121.1
21. Hersbach H, Bell B, Berrisford P, et al. The ERA5 global reanalysis. Q J Roy Meteorol Soc. 2020;146(730):1999-2049. doi:10.1002/qj.3803
22. Benjamin SG, Weygandt SS, Brown JM, et al. A North American hourly assimilation and model forecast cycle: the rapid refresh. Mon Weather Rev.
2016;144(4):1669-1694. doi:10.1175/MWR-D-15-0242.1
23. Wiser R, Bolinger M, Hoen B, et al. Land-based wind market report: 2021 edition. 2021.
24. Rand JT, Kramer LA, Garrity CP, et al. A continuously updated, geospatially rectified database of utility-scale wind turbines in the United States. Sci
Data. 2020;7(1):15. doi:10.1038/s41597-020-0353-6
25. Staffell I, Green R. How does wind farm performance decline with age? Renew Energy. 2014;66:775-786. doi:10.1016/j.renene.2013.10.041
26. Clifton A, Smith A, Fields M. Wind plant preconstruction energy estimates. Current practice and opportunities. National Renewable Energy Lab; 2016.
27. Mahrt L, Pfister L, Thomas CK. Small-scale variability in the nocturnal boundary layer. Boundary-Layer Meteorol. 2020;174(1):81-98. doi:10.1007/
s10546-019-00476-x
28. Wharton S, Lundquist JK. Atmospheric stability affects wind turbine power collection. Environ Res Lett. 2012;7(1):014005. doi:10.1088/1748-9326/
7/1/014005
29. St. Martin CM, Lundquist JK, Clifton A, Poulos GS, Schreck SJ. Wind turbine power production and annual energy production depend on atmospheric
stability and turbulence. Wind Energ Sci. 2016;1(2):221-236. doi:10.5194/wes-1-221-2016
30. Veers P, Dykes K, Lantz E, et al. Grand challenges in the science of wind energy. Science. 2019;366(6464):eaau2027. doi:10.1126/science.aau2027

How to cite this article: Davidson MR, Millstein D. Limitations of reanalysis data for wind power applications. Wind Energy. 2022;25(9):
1646‐1653. doi:10.1002/we.2759