Accepted Manuscript

ERA5: The new champion of wind power modelling?

Jon Olauson

PII: S0960-1481(18)30367-7
DOI: 10.1016/j.renene.2018.03.056
Reference: RENE 9925

To appear in: Renewable Energy

Received Date: 27 September 2017

Revised Date: 28 February 2018
Accepted Date: 22 March 2018

Please cite this article as: Olauson J, ERA5: The new champion of wind power modelling?, Renewable
Energy (2018), doi: 10.1016/j.renene.2018.03.056.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

ERA5: The new champion of wind power modelling?

Jon Olausona,b,1
a Division of Electricity, Department of Engineering Sciences, Uppsala University, Sweden
b Integration of Renewable Energy Sources Group, KTH Royal Institute of Technology, Stockholm, Sweden

PT
Abstract

RI
Output from meteorological reanalyses are used extensively in both academia and industry for modelling
wind power. Recently, the first batch of the new ERA5 reanalysis was released. The main purpose of this
paper is to compare the performance of ERA5 and MERRA-2 (a commonly used reanalysis today) in terms

SC
of modelling i) the aggregated wind generation in five different countries and ii) the generation for 1051
individual wind turbines in Sweden. The modelled wind power generation was compared to measurements.
In conclusion, ERA5 performs better than MERRA-2 in all analysed aspects; correlations are higher, mean
absolute and root mean square errors are in average around 20% lower and distributions of both hourly

U
data and changes in hourly data are more similar to those for measurements. It is also shown that the
uncertainty related to long-term correction (using one year of measurements and reanalysis data to predict
AN
the energy production during the remaining 1–5 years) is 20% lower for ERA5. In fact, using one year sample
data and ERA5 gives slightly more accurate estimates than using two years of sample data and MERRA-2.
Additionally, a new metric for quantifying the system size and dispersion of wind farms is proposed.
Keywords: ERA5, MERRA-2, Wind power modelling, Long-term correction, Reanalyses
M
D

1. Introduction

Quoting Ref. [1], “reanalysis is the process whereby an unchanging data assimilation system is used to
TE

provide a consistent reprocessing of meteorological observations, typically spanning an extended segment

of the historical data record”. The products from a reanalysis include many variables such as wind speeds,
5 temperature and atmospheric pressure and are used extensively for modelling wind power both in academia
and industry. Examples of applications include generation of long wind power time series for grid integration
EP

studies and long-term correction (LTC) of wind speed measurements conducted prior to the construction of
wind farms. An important advantage of reanalyses is that these are often freely and globally available.
Over the decades, new reanalyses have gradually become more high-resolved in time and space [2].
10 During the last years, MERRA [3] and MERRA-2 [1] (both produced by NASA) have been very popular for
C

modelling wind power generation [4–14] due to, e.g., the hourly temporal resolution and adequate height for
wind speeds (50 m). A general conclusion from the studies cited above is that MERRA gives good results
AC

in terms of country-wise wind power generation (relatively low errors when comparing to measurements).
During 2017–2018, a new reanalysis named ERA5 [15] will be released. Contrary to its predecessor
15 ERA-Interim [16], ERA5 has hourly resolution and wind speed data is freely available at 100 m height, i.e.
relevant for modern wind turbines (WTs). The horizontal resolution is around 31 km, which is better than
ERA-Interim and MERRA. The main objective of this paper is to compare the performance of MERRA-2
and ERA5 in terms of modelling wind power, both for countries and for individual WTs. For the former, the

1 Kungliga Tekniska Högskolan, SE-100 44, Stockholm, Sweden

Tel.: +46 720 411636

E-mail: olauson@gmail.com

Preprint submitted to Renewable Energy March 23, 2018

 ACCEPTED MANUSCRIPT

generation in Germany, Denmark, France, Sweden and Bonneville Power Administration (BPA) in north-
20 west USA were analysed. For simplicity, BPA is denoted a “country” in the remaining of this paper. For
the analysis of individual WTs, hourly time series for more than 1000 Swedish units were obtained.
The most common approach for modelling country-wise wind power from reanalyses is to i) generate wind
speed time series for each site, ii) transform wind speed to power via a power curve (PC) and iii) aggregate.
Regarding point i), the reanalysis wind speeds can either be scaled so that a desired mean generation is
25 met [9, 10] or calculated from reanalysis data at different model heights [4, 12] (possibly using statistical [6]

PT
or dynamical [17] downscaling). When transforming wind speed to power, some authors [4, 9] use one
generic PC while others [6, 12] use farm specific ditto. It is common to alter the original PCs, e.g. using
smoother “multi-turbine” power curves [4, 12]. Especially for larger countries, it may be important to handle
the maximum country-wise output, which can be considerably lower than the installed capacity [4, 10, 12].

RI
30 Post-processing of the outputs at any of steps i-iii) is possible for improving the results. Examples include
correction of seasonal and diurnal biases [10] and noise addition for a better representation of observed wind
power variability.

SC
The remainder of the paper is structured in the following manner. First, the two reanalyses and all
measurement datasets are described. In Section 3, the methods used are presented, including a suggestion
35 for a new metric describing the geographical dimensions of a wind power system. Subsequently the results
are given and the paper is concluded with a short discussion.

2. Data
U
AN
Several different datasets were retrieved for the purpose of this study. The two reanalyses are described
in Section 2.1 while measurements and metadata are described in Section 2.2.
M

40 2.1. Reanalyses
In this work, the performance of MERRA-2 and ERA5 are compared. Other freely available reanalyses
also exist (e.g. ERA-Interim and JRA-55), but have lower spatiotemporal resolutions and heights for wind
D

speed data.
The first version of MERRA, the “Modern-Era Retrospective Analysis for Research and Applications”,
45 was released in 2009. In early 2016 the production stopped and was replaced by MERRA-2 which is now
TE

updated in “near real-time”, i.e. data is available with a couple of months lag. For the purpose of modelling
wind power, these two reanalyses perform similarly [12].
The first batch of the ERA5 dataset, produced by the European Centre for Medium-Range Weather
Forecasts (ECMWF), was released in July 2017 and covers the period 2010 - 2016. Subsequently, ERA5
EP

50 will span 1950 to near real-time. Some important properties of MERRA-2 and ERA5 are shown in Table 1.
For comparison, data for ERA-Interim, the predecessor of ERA5, are also given. ECMWF reanalyses are
produced on reduced Gaussian grids, i.e. using different number of grid points along different latitudes and
thus keeping the grid point separation in metres approximately constant. Since the data is downloaded on a
C

user-defined regular lat/lon grid, the resolutions given in Table 1 are only approximate. Downloading with
55 a higher resolution is possible, but this merely implies an interpolation of the data.
AC

The spatial resolutions of ERA5 and MERRA-2 are illustrated qualitatively in Fig. 1. For comparison,
mean wind speeds from the meso-scale model MIUU [18] (spatial resolution 0.5 km) are also shown. Note
that data from the native model heights are shown (50 m for MERRA-2 and 100 m for ERA-I) and that
the colors thus represent different ranges of mean wind speeds. Although ERA5 gives more correct patterns
60 than MERRA-2 e.g. along the coast, it is obvious that none of the reanalyses are capable of resolving the
local variations, especially in more complex terrain. Using reanalyses directly for determining the mean
wind speed at a site is therefore not recommended.

2.2. Wind power measurements and metadata

In order to model country-wise generation and to evaluate the results, the following measurements and
65 metadata were used:

2
 ACCEPTED MANUSCRIPT

Table 1: Information on the two reanalyses considered in this study (MERRA-2 and ERA5). For comparison, data for
ERA-Interim, the predecessor of ERA5, is also shown. The spatial resolutions of the ERA reanalyses are approximate (see

PT
Section 2.1). In general, wind speed data are available at several heights or pressure levels. Only the most proper for modelling
wind power is given here.
MERRA-2 ERA-Interim ERA5

RI
Temporal resolution 1 hour 6 hours 1 hour
Spatial resolution 0.5◦ × 0.625◦ 79 km (≈ 0.75◦ ) 31 km (≈ 0.28◦ )
2010–2016,
Coverage 1980– 1979–
soon 1950–

SC
Wind speed height 50 m 10 m 100 m

U
AN
M
D
TE
C EP
AC

Figure 1: Mean wind speeds for ERA5 (at 100 m above ground) and MERRA-2 (50 m). For comparison, mean wind speeds
from the meso-scale model MIUU [18] with a spatial resolution of 0.5 km are also shown. No numeric color bars are given since
the comparison should only be seen as qualitative (one particular color does not represent the same mean wind speed in the
three panels).

3
 ACCEPTED MANUSCRIPT

1. Coordinates, capacity and, if available, commissioning and decommissioning dates of each farm.
2. Hourly measurements of nation-wide generation.
3. Time series of the installed capacity corresponding to the measured generation. Since the wind farm
database might be incomplete and that commissioning/decommissioning dates might not be available,
70 this time series can differ from the one calculated from the wind farm data (item one in this list).
The measured and modelled generation time series were therefore normalised to their corresponding
capacity time series in order to get comparable per unit generation; pp.u. [t] = pM W [t]/pinst [t].

PT
For Germany, Denmark and France, data were retrieved from the “Open Power System Data” (OPSD)
database (data packages “Renewable power plants” version 2017-07-03 and “Time series” version 2017-07-
09). Primary data are from various sources, for a complete list see Ref. [19]. Since coordinates for German

RI
75

offshore farms were lacking or seemed suspicious, these were instead obtained from wikipedia.org. The
latter source was also used for country-wise installed capacities (linear interpolation of end-of-year values
was employed to calculate pinst [t]). BPA generation time series and farm metadata except coordinates were

SC
retrieved from Ref. [20]. Coordinates for BPA farms were obtained from online sources; thewindpower.net,
80 wikipedia.org and farm developers’ web pages. The BPA generation time series is given with five-minute
resolution, so the data were averaged into hourly blocks.
Generation measurements for individual Swedish WTs were obtained from the Cesar database. This

U
dataset can be retrieved by request from the Swedish Energy Agency and a more thorough description can
be found in Ref. [21]. Only wind turbines commissioned year 2000 or later with a capacity of ≥ 1 MW and
AN
85 at least one year records were considered. Wind turbines for which the rated capacity changed during the
measurement period and measurements taken for farms in aggregation were also excluded. In total, 1–6
year long time series for 1051 WTs remained to be analysed.
For Sweden, generation time series for individual WTs but no coordinates were available. The latter
were therefore estimated from the reanalyses grid points with the highest correlation between modelled and
M

90 measured generation. Since choosing the best fitting data could potentially give overly optimistic results,
a test was performed. Relatively exact coordinates for ten randomly chosen WTs were found from the
developers/owners web pages and reanalyses time series were computed for these sites. The corresponding
D

correlations to measurements were in principle identical to those obtained from the best fitting time series;
the largest difference for a single WT was 0.001 and the average correlations were the same within three
TE

95 digits precision. It was therefore concluded that estimating coordinates from the best fitting reanalyses grid
points does not skew the results (neither positively nor negatively). It can be mentioned that the mean
distances between the correct and the estimated locations were 36 km for MERRA-2 and 15 km for ERA5.
Summary statistics of the country-wide measurements are shown in Table 2. “Capacity 1” and “Capacity
2” give the first and last samples of the capacity time series that were used to normalise measured and
EP

100 modelled generation, respectively. These values can sometimes, as described in the numbered list opening
this section, differ. Note that for France, no commissioning and decommissioning dates were available
(Capacity 2 is constant), so all farms were included in the simulations for the entire period. The “equivalent
system radius” is a new metric for quantifying the system size and wind farm dispersion, which can be
C

important for wind power variability as well as the difficulty of modelling the generation, see Section 3.4.
105 The capacity factor (CF) is the average p.u.-generation during the whole period. Note that aggregated
AC

Swedish generation was only studied during 2014–2015 while time series for individual WTs were studied
for the whole concurrent periods of measurements and ERA5 data (between one and six years during 2010–
2015).

3. Methods

110 This section contains four parts. Firstly, the methodology for modelling hourly generation is described;
secondly, the method for identifying downtime of Swedish WTs is given; thirdly, long-term correction is
described and lastly a new metric for quantifying the size of a wind power system is proposed.

4
 ACCEPTED MANUSCRIPT

Table 2: Information on country-wise wind power measurements. “Capacity 1” and “Capacity 2” give the first and last samples
of the capacity time series that were used to normalise measured and modelled generation (see main text). The “equivalent
system radius” is a metric describing the system size and wind farm dispersion, see Section 3.4.
Germany Denmark France Sweden BPA
Country code DE DK FR SE
Time period 2016 2015–2016 2015–2016 2014–2015 2014–2015

PT
Temporal resolution Hourly Hourly Hourly Hourly 5 min
Capacity 1 (GW) 44–50 4.9–5.2 9.3–12.1 2.0–2.3 4.5–4.8
Capacity 2 (GW) 43–48 4.7–5.2 10.2 2.0–2.3 4.5–4.8
Equivalent system

RI
329 166 376 471 91
radius, req (km)
Capacity factor 0.186 0.303 0.215 0.329 0.266

SC
3.1. Modelling hourly generation
The hourly energy production for each unit (WT or farm) was modelled from reanalyses wind speeds
115 and subsequently aggregated to country-wise time series. The model was deliberately kept relatively simple

U
and the same parameters (or rules for determining parameters) were used for all countries. Because of this,
no separate training, validation and test sets were used. No formal optimisation procedure was employed
AN
for trimming parameter values, only manual trial end error. All model parameters are described below and
the chosen values are given in Table 3.
120 In horizontal, bi-linear interpolation of the four neighbouring grid points was employed, thus neglecting
the impact from local topography. Based on earlier studies [2, 6, 17, 21], statistical or dynamical downscaling
M

does, however, not necessarily implies better results in terms of e.g. correlations with measurements. As
described below, the mean wind speed was also not taken directly from reanalysis data.
Since no information on rotor diameters were available, a power curve corresponding to a specific rating
125 (rating over swept rotor area) of 360 W/m2 was used for all units. The power curve was first smoothed,
D

i.e., transformed to a “multi-turbine” power curve, using a Gaussian distribution with standard deviation
1.5 m/s [10].
TE

There are several types of losses in a wind farm. Wake losses, mild icing, degraded blades etc. shift
the power curve (a higher wind speed is required in order to give a certain output) but does not affect the
130 maximum generation level. Downtime and some electrical losses can however lead to a lower maximum
generation, at least for a country. Furthermore, reanalyses generally underestimate the diversity of wind
EP

speeds; in reality, all farms in a country will never have wind speeds above rated simultaneously and thus
the generation in countries like Germany, France and Sweden will seldom exceed 0.8 p.u. In order to mimic
these two different phenomena, denoted “internal” and “external” losses in Ref. [10], the incoming energy in
135 the wind was reduced by a fixed amount and the resulting energy time series were multiplied by a factor.
C

As has been showed in Section 2.1, reanalyses are generally not suitable for determining the mean wind
speed at a particular site. In the present study, hub heights were also not available, i.e. it was not possible
to interpolate/extrapolate the wind speeds from different reanalysis model heights. The mean wind speeds
AC

were therefore determined from the estimated CF of each farm (CFi ), which was calculated as

CFi = CFcountry + β(ūi − ūcountry ), (1)

140 where CFcountry is the measured CF of the country in question, ūi and ūcountry are the mean reanalyses wind
speeds for the particular site and the capacity-weighted mean for all sites in the country, respectively, and
β is a parameter determining how much the mean wind speeds influence the CFs. Note that ūi and ūcountry
were computed for the reanalyses model heights, i.e. 50 m for MERRA-2 and 100 m for ERA. For Denmark
and Germany, separate calculations were performed for onshore and offshore farms. The wind speed time
145 series were subsequently linearly scaled to the desired means (the mean wind speeds that corresponds to

5
 ACCEPTED MANUSCRIPT

Table 3: Parameter values used for modelling hourly generation for countries and individual wind turbines (WTs). See
Section 3.1 for parameter descriptions. No formal optimisation procedure was employed for finding parameter values, only
manual trial end error.
Countries Individual WTs
Specific power 360 W/m2 360 W/m2
Multi-turbine smoothing 1.5 m/s 0.5 m/s

PT
Internal losses 10% 15%
External loss factor 1.01 · measured maximum 1
β 1/40 1/40

RI
the desired CFs), i.e. assuming constant wind shears. Although this is obviously not entirely correct, using
time-varying wind shears give little or no improvement of the model performance [10].

SC
For the modelling of generation for individual Swedish WTs, internal losses were increased to 15% and
the external loss factor was set to unity. The reason for not using any external losses was, of course, that the
150 generation for a single WT can sometimes be 1 p.u. The multi-turbine smoothing parameter was lowered
to 0.5 m/s since this gives a somewhat better representation of the output from individual WTs.

U
As a final note, it should be recognised that the method outlined above make use of the measured average
CF and maximum generation, i.e. the modelled time series will have the correct averages and maxima.
AN
3.2. Identifying downtime
155 When evaluating the performance of reanalyses in terms of modelling generation for individual WTs,
downtime (stops due to e.g. maintenance and technical faults) will make the comparison less fair. A method
M

for automatic detection of downtime was therefore used and these data samples were removed. The basic
idea was that if the modelled generation is relatively high while the measured generation is zero for an
extended period of time, this is due to downtime.
160 For each WT, a generation time series was first calculated with the method given in the previous section.
D

For periods with zero actual generation for at least five hours, the corresponding mean reanalyses generation
was computed. The period was considered downtime if at least one of the following conditions was fulfilled:
TE

1. The period is longer than one week

2. The period is at least twelve hours and the corresponding reanalysis mean is at least 0.05 p.u.
165 3. The corresponding reanalysis mean is at least 0.15 p.u.
EP

In general, there is a trade-off between false positive and false negative downtime classifications. If, for
example, the threshold of five hours was to be lowered to three hours, additional downtime events would
be correctly captured at the expense of increasing the number of events falsely classified as downtime. The
chosen parameters give balanced results; when validating with actual downtime records, there were 9% false
C

170 negative and 8% false positive errors (in relation to the true number of downtime hours) [21].
After the downtime samples were removed, a new calculation of reanalysis generation was performed
AC

based on the new (higher) measured CF. Note that downtime data was only removed for the analysis of
individual WTs, not for the analysis of aggregated Swedish data.

3.3. Long-term correction

175 Long-term correction (LTC) is routinely used for finding the best estimate of the long-term wind speed
distribution at a site given a shorter (one or a few years) measurement. A LTC can also be used to determine
the long-term CF given a shorter generation time series, e.g. to see whether a new farm performs according
to plan or to analyse the value of a farm that is to be sold. All types of LTC methods require some sort
of long, preferably at least ten years, reference time series. A relationship is established between the site
180 measurement and the reference during the period with concurrent data. This relationship is subsequently

6
 ACCEPTED MANUSCRIPT

“fed” with the whole reference time series in order to determine the long-term distribution or mean of the
variable of interest.
In recent years, reanalyses wind speeds have become the de facto standard as reference time series due
to their consistency and long records. Wind speed measurements from meteorological masts, generally 10
185 meters high, can also be used but often have consistency problems due to e.g. gradually growing vegetation
and changes in anemometry.
In this work, the energy errors from using short, long-term corrected generation time series of single WTs

PT
are evaluated. One randomly chosen year of measurements (sample data/training period) and reanalyses
wind speeds were used for predicting the energy production for the remaining 1–5 years (out-of-sample data).
190 Since at least two years of data is thus required, “only” 893 WTs were analysed. A simple “wind index”
method was employed:

RI
1. A one-year period was randomly chosen as sample data and the corresponding CF was calculated.
2. Generation time series for the entire measurement period were modelled from ERA5 and MERRA-2

SC
data. The mean wind speeds were linearly scaled so that the CF during the training period matches
195 measurements.
3. The predicted CF was taken as the average modelled generation for out-of-sample data.
Note that independent LTCs were performed for each WT, i.e., the training periods were not the same.

U
Since the results to some degree depend on the random choice of training periods, ten different runs were
performed.
AN
200 Small errors in such an exercise is not only of academic interest; when, for instance, a farm is to be
sold, the most important figure is often not the predicted generation but rather the generation that will be
achieved with a given certainty, e.g. 75% or 90%. A low uncertainty is thus worth a lot in monetary terms.
M

3.4. Equivalent system radius

There are many possible candidates for quantifying the system size and/or dispersion of wind farms. Such
205 metrics can be used for a general system description or as a proxy for wind power variability and include,
D

e.g., the system dimensions in North-South and West-East [22], the system area [8] and the (weighted) mean
distance between farms [23]. Although these metrics might be adequate, there are some issues:
TE

• System dimensions and areas are the same no matter if the farms are evenly distributed of heavily
concentrated in one part of the system.
210 • The area can be the same for various system shapes, although wind power variability differs.
EP

• The mean distance is influenced by the level of details of the dataset, e.g. whether coordinates are
given for farms or WTs. The mean distance also, in my opinion, gives a less intuitive feeling of the
system dimensions.
C

Because of these drawbacks, a new metric was developed: the “equivalent system radius” req . The idea is
215 that a set of wind farms can be described by a disk of uniformly distributed wind power that has the same
AC

combined variance as the actual wind farm fleet, assuming that the correlation between generation of two
farms declines in an exponential fashion with separation distance. Before proceeding with the derivation of
the equation for req , the concept is illustrated in Figure 2.
Let ptot [t] denote the per unit time series of the combined generation for the actual farm distribution.
220 The variance of ptot [t] is
N X
X N
Var(ptot [t]) = Pi Pj σi σj ρi,j , (2)
i=1 j=1

7
 ACCEPTED MANUSCRIPT

Germany
400 Wind farms are shown in dark grey (the
areas are proportional to the capacities).
300
The disks with radius r eq are shown in light
200
Distance from centre (km)

grey.
100

PT
-100 BPA
200
-200

-300 100

RI
-400
0
-500

-600 -100

SC
-400 -200 0 200 400 -400 -300 -200 -100 0 100 200
Distance from centre (km) Distance from centre (km)

Figure 2: Illustration of the equivalent system radius (req ) for Germany and BPA. Note that BPA farms are concentrated in
a relatively small area which gives a small req as compared to the total system dimensions. In particular, the small farm at
around (-300,100) does not impact req much. For Germany, wind turbines with the same coordinates are plotted like one farm

U
although this is not always the case.
AN
P
where N is the number of farms, Pi is farm i’s share of the total capacity (i.e. i Pi = 1), σi is the standard
deviation of pi [t] (the per unit time series for farm i) and ρi,j is the correlation between the outputs for
farms i and j. Assuming the same σi for all farms and using the common [24] correlation model
M

ρi,j = e−di,j /D , (3)

where di,j is the distance between farms i and j and D is a constant, Eq. 2 can be rewritten as
D

N X
X N
Var(ptot [t]) = σ 2 Pi Pj e−di,j /D . (4)
TE

i=1 j=1

225 We now want to find req so that a uniform “wind power disk” with this radius will have the same
variance as that obtained with Eq. 4. Let us denote the total output ptot,eq [t]. The variance of ptot,eq [t] can,
in analogue with Eq. 4 for the discrete case, be expressed as
EP

Z Z
Var(ptot,eq [t]) = σ 2 e−di,j /D dPi dPj , (5)
Aeq Aeq
R
where Aeq is the area of our disk. Since dP = 1, we can write
C

dA
dP = . (6)
AC

2
πreq
In polar coordinates,
dA = rdrdθ and (7)
q
di,j = ri2 + rj2 − 2ri rj cos(θi − θj ) (8)
and Eq. 5 can thus be rewritten as

Zreq Zreq Z2π Z2π √2 2

σ2
Var(ptot,eq [t]) = 2 4 ri rj e− ri +rj −2ri rj cos(θi −θj )/D dri drj dθi dθj (9)
π req
ri =0 rj =0 θi =0 θj =0

8
 ACCEPTED MANUSCRIPT

1200

1000 r eq = - 5793.3*x 5 + 20815*x 4 - 30283*x 3 +

23081*x 2 - 10269*x + 2446.5

800

r eq (km)
Solutions to integral
600 Polynomial fit

400

PT
200

0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

RI
x

Figure 3: Polynomial approximation of the equivalent system radius (req ) as a function of x, the left-hand side of Eq. 10. The
variable x can be interpreted as the per unit variance of a country-wise time series as compared to that for individual wind
turbines (WTs). The approximation is valid for D = 500 km, where D is the parameter determining how fast correlations

SC
decline with WT separation distance.

230 Combining Eq. 4 and Eq. 9 gives us

U
N
N X Zreq Zreq Z2π Z2π √2 2
X 1
Pi Pj e −di,j /D
= 2 4 ri rj e− ri +rj −2ri rj cos(θi −θj )/D dri drj dθi dθj (10)
AN
i=1 j=1
π req
ri =0 rj =0 θi =0 θj =0

Equation 10 was solved numerically for req using D = 500 km, a common value in the literature [24]. For
practical purposes, an approximate solution of Eq. 10 suffice. The left-hand side of Eq. 10 can be computed
M

easily, let us denote this sum x. A good polynomial approximation of req for 0.22 ≤ x ≤ 1 is given in Fig. 3,
using D = 500 km as before. The mean absolute residual of the fit is 0.5 km.
235 Rather than quantifying variability by looking at the variance of generation, one might want to study
the variance of changes in generation. As the correlation of changes is much lower than that for generation
D

itself, a lower value of D must be used. Since D appears on both sides of Eq. 10, the value of req does
however not change very much. As an example, using D = 50 km instead of D = 500 km changes req
TE

with -4% and -8% for Germany and Denmark, respectively. It should be mentioned that a lower value of D
240 requires a higher quality of the farm coordinates since the estimated correlations become more sensitive to
farm separation distances.
EP

4. Results

This section is divided into two subsections. First an evaluation of the modelling of country-wise gener-
ation is presented and second results for individual Swedish WTs are given.
C

245 4.1. Country-wise generation

AC

The challenges of modelling wind power generation in different countries vary with e.g. system size,
installed capacity, terrain complexity and wind regimes. A large system with many WTs is generally easier
to model since the errors for individual WTs are partially cancelled out; the correlations between errors
reduce with separation distance. Because of this, the most important results of this paper are not the
250 absolute errors but rather the difference in performance between MERRA-2 and ERA5.
In order to give a reference frame for the quantitative results that will soon be presented, some graphical
examples (Sweden modelled with ERA5 data) are shown in Fig. 4. As an overall judgment, the model
performance for Sweden is medium good in comparison to all five evaluated countries. Panel a shows a
representative 4000 sample excerpt of the time series of measured and modelled generation. In panel b,
255 all data have been sorted from highest to lowest, forming so called duration curves. Panel c demonstrates
the distribution of changes of hourly generation. The magnitudes of these “step changes” are somewhat

9
 ACCEPTED MANUSCRIPT

a) Time series
1
Generation (p.u.)

0.5

PT
0
500 1000 1500 2000 2500 3000 3500 4000
Hour in year 2015

RI
c) Step changes d) Monthly CFs
b) Duration curves 0.6
1 103
0.5

Measurements
Generation (p.u.)

0.8 Measurements
Model (ERA5) 2

SC
10 0.4
Count

0.6
0.3
0.4
101
0.2
0.2

U
0.1
0 100 0.1 0.2 0.3 0.4 0.5
0 0.5 1 -0.1 -0.05 0 0.05 0.1
ERA5
AN
Exceedance probability One hour change (p.u.)

Figure 4: Illustration of the model performance (ERA5 data used for modelling the Swedish wind power generation). Panel
a shows a representative 4000 sample excerpt of the time series of measured and modelled generation. Panel b gives duration
curves (generation sorted from highest to lowest) of all data. Panel c shows the distribution of one hour step changes, i.e. the
M

changes in average power from one hour to the next. In panel d, a scatter plot of modelled and measured monthly capacity
factors (CFs) is given. The straight line shows the desired one-to-one relationship.
D

underestimated by the model; the standard deviation of ERA5 step changes are around 8% lower than that
for measurements (in relative terms), see Table 4. Panel d, finally, gives a scatter plot of actual versus
predicted monthly CFs. For all months but one, the modelled and measured CFs agree well. For February
TE

260 2014, ERA5 predicts 0.40 but the actual outcome was 0.32. A likely explanation for at least part of the
difference is icing losses (which are not modelled).
Quantitative results for all countries and for both MERRA-2 and ERA5 are presented in Table 4.
The metrics considered are correlations, root-mean-square errors (RMSE), mean absolute errors (MAE),
EP

relative errors in the standard deviation of one hour step changes (σ∆P ) and RMSE of the duration curves
265 (RMSEdur ). Note that RMSE, MAE and RMSEdur are given in absolute terms, so e.g. 2.82% corresponds
to an error of 0.0282 p.u. For σ∆P , the errors are given in relative terms; +12.8% implies that the standard
deviation is 12.8% higher for the model than for measurements (e.g. 0.0338 p.u. versus 0.03 p.u.). All
C

metrics in Table 4 are calculated for hourly data.

For a clearer comparison of the performance of the reanalyses, Fig. 5 shows the RMSE and MAE of
AC

270 ERA5 normalised to those for MERRA-2 (a bar of e.g. height 0.8 thus implies that the errors are 20% lower
for ERA5 than for MERRA-2). In average, the RMSE is 22% lower and the MAE is 24% lower. Large
differences however exist between the five countries. For Denmark, the MERRA-2 and ERA5 results are
very similar. Likely, this can be explained by the flat terrain and lack of forest; the benefit of a higher model
resolution is limited for such conditions. For BPA, on the other hand, the errors are around 50% lower for
275 ERA5. In comparison to Denmark, the equivalent system radius of BPA is much lower and the surrounding
terrain is more complex. As a consequence, the generation is more volatile in the BPA area; the standard
deviation of one hour step changes is for instance 0.048 p.u. as compared to 0.028 p.u. for Denmark. Taken
together, these factors increase the benefits of using a reanalysis with higher spatial resolution.
The performance of MERRA for modelling wind generation in several different European countries has
280 been evaluated earlier [6, 12]. The results in this work (correlation and RMSE) for MERRA-2 are comparable

10
 ACCEPTED MANUSCRIPT

Table 4: Quantitative results for the modelling of hourly, county-wise wind power. The metrics considered are correlations,
root-mean-square errors (RMSE), mean absolute errors (MAE), relative errors in the standard deviation of one hour step

PT
changes (σ∆P ) and RMSE of the duration curves (RMSEdur ). Note that RMSE, MAE and RMSEdur are given in absolute
terms, so e.g. 2.82% corresponds to an error of 0.0282 p.u. For σ∆P , the errors are given in relative terms; +12.8% implies
that the standard deviation is 12.8% higher for the model than for measurements (e.g. 0.0338 p.u. versus 0.03 p.u.).
Country Reanalysis Correlation RMSE MAE Error σ∆P RMSEdur

RI
MERRA-2 0.982 2.82% 2.06% +12.8% 0.43%
Germany
ERA5 0.987 2.35% 1.66% +7.5% 0.44%
MERRA-2 0.973 5.40% 3.75% -11.7% 1.75%

SC
Denmark
ERA5 0.973 5.45% 3.71% -6.0% 1.52%
MERRA-2 0.975 3.49% 2.63% +17.2% 0.79%
France
ERA5 0.982 2.97% 2.25% +14.3% 1.00%
MERRA-2 0.950 6.10% 4.58% -8.7% 1.27%

U
Sweden
ERA5 0.975 4.40% 3.21% -7.9% 0.52%
MERRA-2 0.756 18.4% 13.3% -10.4% 6.09%
AN
BPA
ERA5 0.945 9.10% 6.06% 1.5% 2.34%
M
D
TE

1.2
RMSE
Errors for ERA5 relative MERRA-2

1 MAE
EP

0.8

0.6
C

0.4
AC

0.2

0
DE DK FR SE BPA

Figure 5: Root-mean-square errors (RMSE) and mean absolute errors (MAE) for ERA5 in relation to those for MERRA-2.
A bar of e.g. height 0.8 thus implies that the considered metric is 20% lower for ERA5 than for MERRA-2. In average, the
RMSE is reduced with 22% and the MAE is reduced with 24% by using ERA5 data.

11
 ACCEPTED MANUSCRIPT

Table 5: Results for the modelling of generation for individual wind turbines. Means and medians for the 1051 analysed WTs
are given. For an explanation of the metrics, see Table 4.
Correlation RMSE MAE RMSEdur Monthly correlation
MERRA-2 (mean) 0.802 19.1% 13.2% 2.94% 0.922
ERA5 (mean) 0.868 15.8% 10.6% 1.92% 0.964
MERRA-2 (median) 0.812 18.3% 12.6% 2.71% 0.935

PT
ERA5 (median) 0.874 15.3% 10.3% 1.66% 0.973

RI
or better, which indicates that the methodology and measurement data quality are adequate. One should be
careful in drawing too far-reaching conclusions on the pros and cons of the different modelling approaches
and reanalyses since the studied time periods were not the same and that different WT databases were
used. With that said, it is interesting to note that the performance of ERA5 seem to be better than that

SC
285 of “EMHIRES”, a dataset obtained by statistically downscaling MERRA data [6], for the countries studied
here.
When comparing the Swedish results here to those in Ref. [10], the errors are somewhat higher and the
correlation is lower. The main explanation is that only modern WTs were included in this study. These

U
are generally located in more complex terrain (hilly forests) and are more exposed to icing than the early
290 WTs which were erected mainly in open landscape in southern Sweden. The CF for the WTs studied here is
AN
around 35% higher than for those studied in Ref. [10], which tend to increase the per unit errors. The model
in Ref. [10] was also more detailed, e.g. accounting for seasonal/diurnal biases and direction dependent
losses. Finally, MERRA-2 seem to perform slightly worse than MERRA in Sweden [12]. A final note on
the modelling of Swedish, aggregated generation is that only the average national CF was used as input to
M

295 the model, i.e. not information on the measured CFs of individual WTs. Using this extra information does
however, surprisingly enough, not improve the results at all.

4.2. Individual WTs

D

In this section, results are given for modelling of individual Swedish WTs. As described in Section 3.2,
periods classified as downtime were first removed. Table 5 gives mean and median values of correlations,
TE

300 RMSE, MAE and RMSE of duration curves for the 1051 analysed WTs. In the last column, correlations of
monthly generation are given (only months with at least 300 valid data samples were considered). Boxplots
of the analysed hourly metrics are given in Fig. 6. On each box, the central mark is the median, the edges
of the box represent the 25th and 75th percentiles and whiskers extend to the most extreme data. It is clear
EP

from Table 5 and Fig. 6 that ERA5 performs better than MERRA-2 in all analysed aspects. One can note
305 that the median performance is always better than mean. This is due to very large errors for a few WTs.
Neither MERRA-2 nor ERA5 is capable of reproducing the variability of individual WTs; the measured
fluctuations are much larger in magnitude than those obtained from reanalyses. This is not surprising since
C

coarse models are not able to reproduce the fluctuations at a local site. In average, the standard deviation
of one hour step changes are underestimated by 49% and 43% by the models based on MERRA-2 and ERA5
AC

310 data, respectively. A Welch power spectral density estimate [25] can be used to illustrate the magnitude
of fluctuations of different frequencies. In Fig. 7, a representative example is given, using 50 segments to
reduce the noise and make the graph easier to interpret. For frequencies higher than around (1 day)−1 ,
the spectra for modelled generation start to deviate from that for measurements. At the highest resolvable
frequency, the difference is large, in particular for MERRA-2.
315 Although the variability of an individual WT is underestimated, that for aggregated country-wise gen-
eration might not be. This is because the correlations between different time series are overestimated by
the models and consequently the modelled generation is smoothed more by aggregation. From Table 4,
the variability as quantified by the standard deviation of one hour step changes is overestimated for some
countries and underestimated for some. The absolute mismatch is always lower for ERA5.

12
 ACCEPTED MANUSCRIPT

a) Correlation b) RMSE (p.u.)

0.9 0.3

PT
0.8 0.25
0.2
0.7
0.15
0.6
0.1

RI
MERRA-2 ERA5 MERRA-2 ERA5

c) MAE (p.u.) d) RMSE dur (p.u.)

0.2 0.06

SC
0.15 0.04

0.1 0.02

0.05
MERRA-2 ERA5 MERRA-2 ERA5

U
AN
Figure 6: Boxplots of the results for 1051 Swedish wind turbines. On each box, the central mark is the median, the edges of the
box represent the 25th and 75th percentiles and whiskers extend to the most extreme data. For an explanation of the metrics,
see Table 4.
M
D
TE

Measurements
EP

100
PSD estimate

10-2
C

ERA5
AC

10-4
MERRA-2

(1 m) -1 (1 w) -1 (1 d) -1 (2 h) -1
Frequency

Figure 7: Representative example of Welch power spectral density (PSD) estimates for one wind turbine (WT). It is clear that
the variability of the generation for a single WT is underestimated by both reanalyses (however, somewhat less by ERA5).

13
 ACCEPTED MANUSCRIPT

Table 6: Energy errors in long-term correction for 893 individual Swedish wind turbines (WTs). Both root-mean-square errors
(RMSE) and mean absolute errors (MAE) are given. For each WT, one randomly chosen year of data was used to train a
model which was subsequently used to predict the energy production during the remaining 1–5 years of out-of-sample data.
The Naïve estimator assumes the same capacity factor as for the training period. The results are means from ten realisations
and normalised to the measured generation.
MAE RMSE

PT
Naïve estimate 7.42% 9.26%
MERRA-2 3.07% 4.00%
ERA5 2.43% 3.23%

RI
320 Next, energy errors from long-term correction (LTC) are evaluated. As described in Section 3.3, one
randomly chosen year was used as sample data in order to predict the energy production during the remaining

SC
1–5 years (out-of-sample data). Note that all results given in this section are means from ten different
realisations. Results for the 893 studied WTs are shown in Table 6. For comparison, results are also given
for the “Naïve estimate” which is simply assuming the same capacity factor as for the sample data. All three
325 estimates are in principle unbiased which implies that the standard deviations of the errors are similar to
the RMSEs.

U
First, one should note that the errors are considerably lower with a proper LTC than with the Naïve
estimator. This is of course not surprising, but nevertheless important. In conclusion, the energy errors are
AN
around 20% lower for ERA5 than for MERRA-2. If two years’ worth of data are used for training the models,
330 the errors for both reanalyses are reduced with around 18%. This means that one year of measurements long-
term corrected with ERA5 actually gives slightly more accurate estimates than two years of measurements
long-term corrected with MERRA-2.
M

In order to validate the robustness of the results, two-tailed t-tests were performed, testing the null
hypothesis that the differences in absolute errors, i.e. |EERA5 ,i | − |EMERRA2 ,i |, come from a (normal)
335 distribution with mean equal to zero. Using one year sample data for both reanalyses, the mean difference
D

is -0.64 percentage points (pp) and the 95% confidence interval is (-0.79, -0.49) pp. The null hypothesis
can thus be rejected on the 0.05 significance level (or any other meaningful level) and we can conclude that
ERA5 performs better than MERRA-2. With one year training data for ERA5 and two years training data
TE

for MERRA-2, the corresponding figures are -0.11 (-0.31, 0.09) pp, i.e. the difference is not statistically
340 significant. For verification, 10,000 samples bootstrap tests were also performed; the results were identical
within the precision given above.
In Ref. [21], correlations between measurements and generation modelled from MERRA, ERA-Interim
EP

and ConWx (a dataset dynamically downscaled from ERA-Interim) were analysed. It was concluded that
MERRA and ConWx performed similarly, which was somewhat surprising given the higher spatial resolution
345 of ConWx. Similar results have however been seen also in earlier work [2]; a higher resolution does not
guarantee better performance. As has already been mentioned, the WTs chosen for the current work are
C

more often located in complex terrain and suffer more severe icing losses than Swedish WTs in general. The
average correlation obtain for MERRA-2 here is thus lower than that for MERRA in Ref. [21] (0.80 versus
AC

0.84). Despite the tougher modelling conditions, the average correlation for ERA5 is however higher (0.87).

350 5. Concluding discussion

In this work, MERRA-2 and the new reanalysis ERA5 were compared in terms of modelling wind power,
both for countries and for individual wind turbines. The main conclusion is that ERA5 performs better
in all analysed aspects. As an average for the five studied countries, the mean absolute errors in modelled
hourly generation (when compared to measurements) were 24% lower for ERA5. For the small BPA system,
355 the errors were roughly 50% lower. Correlations were also higher and the mismatches in standard deviations
of one hour step changes, a common variability metric, were lower.

14
 ACCEPTED MANUSCRIPT

ERA5 also performed better in terms of modelling generation for the 1051 individual Swedish WTs that
were studied. The average correlation in hourly data was 6.6 percentage points higher and the average MAE
was 20% lower. A similar reduction was seen for the long-term correction energy errors (using one year of
360 data to predict the energy production during the remaining period). As a matter of fact, using one year of
measurements long-term corrected with ERA5 gives slightly more accurate estimates than using two years of
measurements long-term corrected with MERRA-2. The latter result is however not statistically significant.
This is good news for the wind industry since lower uncertainties imply reduced risks and increased project

PT
values. Apart from the better performance, ERA5 will also soon be available from 1950 to near real-time,
365 i.e. for a longer period than most other reanalyses.
The proposed metric “equivalent system radius” has its advantages as compared to earlier metrics for
quantifying the size of a wind power system. Most importantly, it is based on the smoothing of wind power

RI
fluctuations from geographical dispersion, which is often an important characteristic when the system size
is to be quantified. Unfortunately, the equation for req turned out to be quite cumbersome and had to be
370 solved numerically. Although a polynomial approximation works fine for practical purposes, an analytic

SC
solution to the integral would be a welcome contribution.
In this work, a relatively simple and generic wind power model was chosen and data on e.g. hub heights
and farm specific power curves were not available. An important conclusion is that good results (generally
equally or better than in earlier work) were still obtained. By including bias-correction and trimming the

U
375 model harder for a specific country, somewhat better results can however be achieved [10]. Likely, the
improvement would be larger for MERRA-2 since e.g. the biases in seasonal and diurnal wind patterns are
AN
larger. In practice, the models used in both research and industry are however also relatively simple, i.e.
the results presented here are representative for the benefits of using the more high-resolved ERA5.
A common application of reanalysis based wind power models is to study hypothetical, highly renewable
380 power systems [8, 11]. In a future with more offshore wind power, higher hub heights and lower specific rating,
M

the capacity factors will be higher and the WTs will more often generate at or near rated capacity. Even
for large countries, the maximum generation will thus be closer to 1 p.u. than today. Here, the maximum
generation in a country was determined from measurements (around 0.8 p.u. for larger countries). A better
modelling approach for future scenarios might be to add multivariate noise to the wind speed time series and
D

385 simulate downtime stochastically. If a proper model is set up, the spectra, correlations, maximum generation
etc. will be correct both for individual WTs and country-wise generation (at the expense of increasing the
TE

errors to some extent).

Acknowledgements
EP

This work was conducted within the StandUP for Energy strategic research framework and was funded
390 by Vindforsk through contract 40290-1. Energimyndigheten, ECMWF, NASA, BPA and OPSD provided
data. The Matlab function for projecting latitudes and longitudes onto an azimuthal equidistant image
was written by Kurt von Laven. The method for solving quadruple integrals in Matlab was inspired by
C

Mike Hosea. Dorothée Vallot and Jonatan Samuelsson are acknowledged for their valuable comments on the
equivalent system radius method. Lasse Johansson is acknowledged for the idea of approximating unknown
AC

395 WT coordinates from coordinates of the best fitting reanalyses time series.

References
[1] R. Gelaro, W. McCarty, M. J. Suárez, R. Todling, A. Molod, L. Takacs, C. A. Randles, A. Darmenov, M. G. Bosilovich,
R. Reichle, K. Wargan, L. Coy, R. Cullather, C. Draper, S. Akella, V. Buchard, A. Conaty, A. M. da Silva, W. Gu, G.-K.
Kim, R. Koster, R. Lucchesi, D. Merkova, J. E. Nielsen, G. Partyka, S. Pawson, W. Putman, M. Rienecker, S. D. Schubert,
M. Sienkiewicz, B. Zhao, The Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2),
Journal of Climate 30 (14) (2017) 5419–5454. doi:10.1175/JCLI-D-16-0758.1.
[2] S. Liléo, E. Berge, O. Undheim, R. Klinkert, R. E. Bredesen, Long-term correction of wind measurements - state-of-the-art,
guidelines and future work, Tech. rep., Elforsk report 13:18 (Jan. 2013).

15
 ACCEPTED MANUSCRIPT

[3] M. M. Rienecker, M. J. Suarez, R. Gelaro, R. Todling, J. Bacmeister, E. Liu, M. G. Bosilovich, S. D. Schubert, L. Takacs,
G.-K. Kim, S. Bloom, J. Chen, D. Collins, A. Conaty, A. da Silva, W. Gu, J. Joiner, R. D. Koster, R. Lucchesi, A. Molod,
T. Owens, S. Pawson, P. Pegion, C. R. Redder, R. Reichle, F. R. Robertson, A. G. Ruddick, M. Sienkiewicz, J. Woollen,
MERRA: NASA’s modern-era retrospective analysis for research and applications, Journal of Climate 24 (2011) 3624–3648.
doi:10.1175/JCLI-D-11-00015.1.
[4] D. J. Cannon, D. J. Brayshaw, J. Methven, P. J. Coker, D. Lenaghan, Using reanalysis data to quantify extreme wind
power generation statistics: A 33 year case study in Great Britain, Renewable Energy 75 (2015) 767–778. doi:10.1016/
j.renene.2014.10.024.

PT
[5] A. Cosseron, U. B. Gunturu, C. A. Schlosser, Characterization of the wind power resource in europe and its intermittency,
Energy Procedia 40 (2013) 58 – 66. doi:http://dx.doi.org/10.1016/j.egypro.2013.08.008.
[6] I. González-Aparicio, F. Monforti, P. Volker, A. Zucker, F. Careri, T. Huld, J. Badger, Simulating European wind power
generation applying statistical downscaling to reanalysis data, Applied Energy 199 (2017) 155 – 168. doi:http://dx.doi.
org/10.1016/j.apenergy.2017.04.066.

RI
[7] W. L. Henson, J. G. McGowan, J. F. Manwell, Utilizing reanalysis and synthesis datasets in wind resource characterization
for large-scale wind integration, Wind Engineering 36 (1) (2012) 97–109. doi:10.1260/0309-524X.36.1.97.
[8] M. Huber, D. Dimkova, T. Hamacher, Integration of wind and solar power in Europe: Assessment of flexibility require-
ments, Energy 69 (2014) 236–246. doi:10.1016/j.energy.2014.02.109.

SC
[9] M. L. Kubik, D. J. Brayshaw, P. J. Coker, J. F. Barlow, Exploring the role of reanalysis data in simulating regional wind
generation variability over Northern Ireland, Renewable Energy 57 (2013) 558–561. doi:10.1016/j.renene.2013.02.012.
[10] J. Olauson, M. Bergkvist, Modelling the Swedish wind power production using MERRA reanalysis data, Renewable Energy
76 (2015) 717 – 725. doi:http://dx.doi.org/10.1016/j.renene.2014.11.085.
[11] J. Olauson, M. N. Ayob, M. Bergkvist, N. Carpman, V. Castellucci, A. Goude, D. Lingfors, R. Waters, J. Widén,
Net load variability in Nordic countries with a highly or fully renewable power system, Nature Energy 1 (2016) 16175.

U
doi:10.1038/nenergy.2016.175.
[12] I. Staffell, S. Pfenninger, Using bias-corrected reanalysis to simulate current and future wind power output, Energy 114
AN
(2016) 1224–1239. doi:10.1016/j.energy.2016.08.068.
[13] M. McPherson, T. Sotiropoulos-Michalakakos, L. D. Harvey, B. Karney, An open-access web-based tool to access global,
hourly wind and solar PV generation time-series derived from the MERRA reanalysis dataset, Energies 10 (7) (2017)
1007. doi:10.3390/en10071007.
[14] S. Liao, W. Yao, X. Han, J. Wen, S. Cheng, Chronological operation simulation framework for regional power system
under high penetration of renewable energy using meteorological data, Applied Energy 203 (2017) 816 – 828. doi:
M

10.1016/j.apenergy.2017.06.086.
[15] ERA5 data documentation, Available online: https://software.ecmwf.int/wiki/display/CKB/ERA5+data+
documentation (Accessed: 2017-07-18).
[16] D. P. Dee, S. M. Uppala, A. J. Simmons, P. Berrisford, P. Poli, S. Kobayashi, U. Andrae, M. A. Balmaseda, G. Balsamo,
D

P. Bauer, P. Bechtold, A. C. M. Beljaars, L. van de Berg, J. Bidlot, N. Bormann, C. Delsol, R. Dragani, M. Fuentes, A. J.
Geer, L. Haimberger, S. B. Healy, H. Hersbach, E. V. Hólm, L. Isaksen, P. Kållberg, M. Köhler, M. Matricardi, A. P.
McNally, B. M. Monge-Sanz, J.-J. Morcrette, B.-K. Park, C. Peubey, P. de Rosnay, C. Tavolato, J.-N. Thépaut, F. Vitart,
TE

The ERA-Interim reanalysis: configuration and performance of the data assimilation system, Quarterly Journal of the
Royal Meteorological Society 137 (656). doi:10.1002/qj.828.
[17] E. Nuño, P. Maule, A. Hahmann, N. Cutululis, P. Sørensen, I. Karagali, Simulation of transcontinental wind and solar pv
generation time series, Renewable Energy 118 (2018) 425 – 436. doi:https://doi.org/10.1016/j.renene.2017.11.039.
[18] H. Bergström, Boundary-layer modelling for wind climate estimates, Wind engineering 25 (5) (2001) 289–299. doi:
EP

10.1260/030952401760177864.
[19] Open Power System Data, Available online: https://data.open-power-system-data.org (Accessed: 2017-08-02).
[20] Wind generation and total load in the BPA balancing authority, Available online: https://transmission.bpa.gov/
Business/Operations/Wind/ (Accessed: 2017-08-09).
[21] J. Olauson, P. Edström, J. Rydén, Wind turbine performance decline in Sweden, Wind Energy (2017) 1–5doi:10.1002/
C

we.2132.
[22] H. Holttinen, Hourly wind power variations in the Nordic countries, Wind Energy 8 (2) (2005) 173–195. doi:10.1002/
we.144.
AC

[23] S. M. Fisher, J. T. Schoof, C. L. Lant, M. D. Therrell, The effects of geographical distribution on the reliability of wind
energy, Applied Geography 40 (2013) 83 – 89. doi:http://dx.doi.org/10.1016/j.apgeog.2013.01.010.
[24] C. M. St. Martin, J. K. Lundquist, M. A. Handschy, Variability of interconnected wind plants: correlation length and its
dependence on variability time scale, Environmental Research Letters 10 (4) (2015) 044004. doi:10.1088/1748-9326/10/
4/044004.
[25] P. D. Welch, The use of fast fourier transform for the estimation of power spectra: a method based on time averaging over
short, modified periodograms, IEEE Trans. Audio and Electroacoust. AU-15 (1967) 70–73.

16
 ACCEPTED MANUSCRIPT
 The performances of ERA5 and MERRA-2 for modelling wind power are compared
 Errors in country-wise output are around 20% lower for ERA5
 ERA5 also performs considerably better for 1051 individual wind turbines
 The lower uncertainty translates directly to higher project values
 A new metric for quantifying system size is proposed

PT
RI
U SC
AN
M
D
TE
C EP
AC