Content uploaded by Iain Staffell
Author content
All content in this area was uploaded by Iain Staffell on Apr 15, 2014
Content may be subject to copyright.
How does wind farm performance decline with age?
q
Iain Staffell
*
, Richard Green
Imperial College Business School, Imperial College London, London SW7 2AZ, UK
article info
Article history:
Received 17 May 2013
Accepted 25 October 2013
Keywords:
Wind farm
Load factor
Degradation
Ageing
Reanalysis
Levelised cost
abstract
Ageing is a fact of life. Just as with conventional forms of power generation, the energy produced by a
wind farm gradually decreases over its lifetime, perhaps due to falling availability, aerodynamic per-
formance or conversion efficiency. Understanding these factors is however complicated by the highly
variable availability of the wind.
This paper reveals the rate of ageing of a national fleet of wind turbines using free public data for the
actual and theoretical ideal load factors from the UK’s 282 wind farms. Actual load factors are recorded
monthly for the period of 2002e2012, covering 1686 farm-years of operation. Ideal load factors are
derived from a high resolution wind resource assessment made using NASA data to estimate the hourly
wind speed at the location and hub height of each wind farm, accounting for the particular models of
turbine installed.
By accounting for individual site conditions we confirm that load factors do decline with age, at a
similar rate to other rotating machinery. Wind turbines are found to lose 1.6 0.2% of their output per
year, with average load factors declining from 28.5% when new to 21% at age 19. This trend is consistent
for different generations of turbine design and individual wind farms. This level of degradation reduces a
wind farm’s output by 12% over a twenty year lifetime, increasing the levelised cost of electricity by 9%.
Ó2014 The Authors. Published by Elsevier Ltd. All rights reserved.
1. Introduction
Ageing is a fact of life. Its effects are inevitable for all kinds of
machinery, reducing the efficiency, output and availability of steam
and gas turbines, solar PV modules, batteries and automobiles alike.
Previous work on wind turbines has considered the reliability of
individual components and the effect of ageing on availability, but
any impact on the energy production of turbines or farms has not
been widely reported.
If load factors (also known as capacity factors) decrease signif-
icantly with age, wind farms will produce a lower cumulative
lifetime output, increasing the levelised cost of electricity from the
plants. If the rate of degradation were too great, it could become
worthwhile to prematurely replace the turbines with new models,
implying that the economic life of the turbine was shorter than its
technical life, further increasing its cost.
This could have significant policy implications for the desir-
ability of investing in wind power, as argued in a recent report by
Hughes for the Renewable Energy Foundation (REF) [1]. That report
suggested that the load factors of wind farms in the UK have
declined by 5e13% per year, normalising for month-by-month
variations in wind speeds. These findings could represent a sig-
nificant hurdle for the wind industry, but they require replication.
Several factors can confound the relationship between age and
observed output in a fleet of wind farms, given that a turbine’s
output is dependent on wind speeds at its site and the efficiency
with which it captures the energy in that wind. For example, if wind
speeds have fallen slightly over time, farms would have lower load
factors in recent months, when they were at their oldest, giving a
spurious correlation between age and poor performance. If im-
provements in design increase a turbine’s output relative to ca-
pacity (its power coefficient) then newer turbines (of the improved
design) will have higher load factors than old turbines, so that
turbine output appears to decline with age, when really it improves
with newer generations. On the other hand, if the best (windiest)
sites were occupied first, then old farms could have higher load
factors than new ones built on inferior sites, so that turbines would
appear to improve with age.
This paper uses public domain data to infer the hour-by-hour
wind speeds at the site of every wind farm in the UK, and the po-
wer curve for each farm’s model of turbine to estimate the output
that they would ideally produce. This technique corrects for the
q
This is an open-access article distributed under the terms of the Creative
Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original author and source are credited.
*Corresponding author.
E-mail addresses: staffell@gmail.com,i.staffell@imperial.ac.uk (I. Staffell).
Contents lists available at ScienceDirect
Renewable Energy
journal homepage: www.elsevier.com/locate/renene
0960-1481/$ esee front matter Ó2014 The Authors. Published by Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.renene.2013.10.041
Renewable Energy 66 (2014) 775e786
confounding factors (wind patterns, turbine model and site qual-
ity), and validates well for farms that report their half-hourly
output to National Grid. Simulated ideal outputs are compared
with actual monthly load factors from a large portion of the UK’s
fleet over the last decade (282 wind farms, 4.5 GW, 53 TWh),
yielding the normalised performance of each wind farm accounting
for its wind resource availability, and a set of weather-corrected
load factors which reveal the effects of ageing. We measure the
level of age-related degradation at the national level, accounting for
the vintage of turbine and local site conditions at each wind farm.
We test different generations of technology and individual wind
farms to confirm that specific units experienced similar declines in
performance. We find the ageing effect to be present, but much
smaller than predicted by Hughes, in line with experience of other
rotating machinery. The specific causes of this performance loss
and their relative contribution are not considered in this paper,
although an overview of potential reasons is given in the discussion
and conclusions.
Due to the amount of data and processing required for this study
we provide online supplementary material which documents our
sources and their validation in greater depth, along with down-
loadable datasets ofUK wind farms and their energy output histories.
2. Previous studies
All machinery experiences an unrecoverable loss in perfor-
mance over time. Gas turbine efficiency suffers an unrecoverable
decline of 0.3e0.6% per year despite regular washing and compo-
nent replacement, or by 0.75e2.25% without [2]. Similarly, the
output of solar photovoltaic panels declines by 0.5% per year on
average [3]. This loss in performance is not routinely accounted for
in studies of the levelised cost of electricity (LCOE) of wind power.
Recent studies by Mott MacDonald, Parsons Brinckerhoff and Arup
accounted for the efficiency of conventional plants falling by 0.15e
0.55% per year, but omitted any such factor for wind turbines [4e6].
Previous studies of wind turbines have focussed on availability
and reliability [7e9]. There appear to be no long term fleet-level
studies into loss of output from wind farms in open literature.
Regardless of technology, quantifying performance degradation is
difficult because consistent and validated field data is hard to
obtain [2]. The recent study by Hughes [1] is therefore significant, in
that we believe it is the first to attempt to estimate the rate of
decline in wind farm load factors on a national scale.
Hughes analysed over 10 years of operating data from the British
and Danish fleets of turbines, finding rates of performance degra-
dation that are much higher than for other technologies, and which
vary remarkably between the UK and Denmark, and between
onshore and offshore turbines. This was based on econometric
analysis of monthly load factors, using a regression which corrected
for the quality of each wind farm’s location, the monthly variation
in national wind conditions, and the age of each farm. Hughes ar-
gues (and shows mathematically) that accounting for monthly
wind conditions with a set of ‘fixed effects’determined by the
regression is econometrically superior to using a measure of
average wind speeds across the country, since site-specific condi-
tions differ from the national average and the output of wind tur-
bines depends non-linearly on the wind speed at every moment in
time, which is very poorly captured by its average over a month.
We therefore use wind speed datawith high temporal and spatial
resolution, and measure the performance of wind farms by esti-
mating their theoretical potential output over the course of a month
and comparing this with the actual reported load factors. While we
believe we are the first researchers to assess wind farm performance
with this kind of ex-post data, a number of papers present techniques
to estimate output levels from time series of wind data.
Many studies have used hourly wind speed data recorded by
met masts; for example investigations into wind variability by
Pöyry [10] and SKM [11], and estimates of future national output by
Green et al. [12,13] and Sturt and Strbac [14]. Hourly met mast
speeds have been directly compared to metered wind farm load
factors in Northern Spain [15] and Scotland [16], showing that ac-
curate estimates can be made for monthly energy generation, but
not for hourly power outputs.
More recentstudies use reanalysesas a sourceof wind speed data:
atmospheric boundary layer models which process physical obser-
vations from met masts and other sources into a coherent and
spatially complete dataset, and are widely used to produce wind
atlases. Kiss et al. [17] were first to compare the European ERA-40
reanalysis to nacelle measurements of wind speed and power
output at two turbines in Hungary, finding “surprisingly good”
agreement.Hawkins et al. [18]were able toreplicate UK monthlyload
factors using a custom reanalysis model, while Kubik et al. [19]
compared the global NASA reanalysis to half-hourly farm output in
Northern Ireland, findingit to be more accurate than met mastdata.
The first practical application appears tohave been made by Ofgemto
estimate the equivalent firm capacity of the UK’swindfleet during
winter peaks in demand [20]. Both Hawkins and Ofgem noted that
the reanalysis outputs need to be scaled down by a constant factor
(29% and 20% respectively) in order to match actualproduction in the
UK, a finding which we elaborate upon in this paper.
3. Data sources
Predicting a given wind farm’s output is far from being a new
science: on-site monitoring of conditions using wind turbine
SCADA systems is commonplace; and software tools such as WaSP
or consultancies such as GL Garrad Hassan are widely used in the
field. Data is not made publicly available, and these services come at
a price of several thousand Euros.
On the other hand, national average data cannot reveal what is
happening at individual wind farms. We therefore employ farm-
specific data for output and site-specific data for wind speeds,
taken from free and publicly accessible datasets. The primary data
used in our main analysis are described in this section, and addi-
tional data used for validation are described in Section 4. Further
information on our data is given in the Supplementary Material.
3.1. Ofgem/REF output data
All wind farms enrolled in the UK government’s incentive
scheme, the Renewables Obligation, publish their monthly outputs
(in MWh) in the Ofgem Renewables and CHP Register.
1
Hughes
extracted and cleaned this data, cross-linking outputs with details
about each wind farm (its capacity and date of commissioning), and
ensured that each wind farm contained only the same model and
vintage of turbine [1]. This cleaned dataset was published on the
internet by the Renewable Energy Foundation (REF) [21].Weare
very grateful to Prof. Hughes and the REF for making this rich data
source available to the community.
We further validated this dataset, corrected the commissioning
date for 15 wind farms (whichwere incorrectly reported by Ofgem),
integrated further meta-data for each farm (the geographical
location, wind turbine model and hub height), and extended the
time-series by 8 months, adding data from April to December 2012.
Our modified dataset is provided as Supplementary Material to
this paper. It contains 1687 farm-years of load factor data, covering
onshore turbines built from 1991 onwards and spanning 11 years of
1
www.renewablesandchp.ofgem.gov.uk.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786776
operation. The study was restricted to onshore wind farms, as only
a small amount of data was available for the UK’s 20 offshore farms,
and these would need to be considered separately as they face a
very different operating environment and maintenance issues to
onshore farms. Table 1 and Fig. 1 provide a selection of summary
statistics for the data.
Figs. 2 and 3 show what has happened to the load factor of these
wind farms as they get older. Fig. 2 simply plots the distribution of
all observed load factors against age, showing a steady decline
of 0.44 0.04 absolute percentage points per year past age one
(1.69 0.17% loss per year relative to the UK mean load factor).
As explained in the introduction this is not necessarily due to
ageing if newer turbine models are more efficient at extracting
energy from the wind. The oldest turbines in the sample (aged 15e
19 years) were built in the early- to mid-1990s; typically 300e
500 kW two or three bladed machines on 25e50 m towers. These
will clearly be outperformed by the latest generation of 2e3MW
turbines which are no older than 5 years.
To control for technology effects, Fig. 3a charts the individual his-
tories of the 53 farms which have more than tenyears of data, using a
12-month moving average to smooth out seasonal variations in the
wind. Load factors tend to rise during the first yearof operation while
turbines are still being commissioned until the farm achieves full
operation and teething problems are ironed out. Fig. 3b summarises
the annual degradation rates, estimated by running individual linear
regressionsoneachfarm’s unsmoothed load factors, excluding the
first year. Theareas with the darkest shading are derived from farms
which are old enough to give ten years of data; lighter areas add the
more recent farms which have fewer observations.
The degradation rates of individual farms are predominantly
bunched around 0 to 1 percentage points of absolute load factor
per year, but several outliers make the overall distribution fat-
tailed. A Cauchy (or Lorentz) distribution therefore provides a
better fit than a normal distribution, centred on 0.48 with a half
width at half maximum of 0.36. The distribution becomes wider for
farms with fewer observations; and would stretch all the way
from 25 to þ15 points per year if farms with less than five years’
data were included. With an absolute degradation rate
of 0.48 0.36 points per year, a typical wind farm loses
1.81 1.32% of its output per year on average. The range of
degradation rates for individual farms is clearly much greater than
uncertainty on the average rate for the dataset as a whole.
Neither Fig. 2 nor Fig. 3 corrects for weather effects. To extract
the true rate of degradation in the UK’s wind farms, we account for
variations in the weather over the last decade using a detailed wind
resource assessment, and the rate of technology improvement by
modelling the specific turbines installed at every wind farm.
3.2. NASA wind speed data
A database of wind speeds for the British Isles was created using
NAA’s MERRA dataset: a historical reanalysis of global atmospheric
observations assimilated and processed using the Goddard Earth
Observing System (GEOS-5) [22,23]. Wind observations with “fairly
complete global coverage”are taken from weather stations, bal-
loons, aircraft, ships, buoys and satellites, and processed by the
model to give data with hourly resolution on a ½
latitude and
2
/
3
longitude grid (approx. 55 by 44 km), at heights of 2 and 10 m above
the surface displacement height (d, the point at which a logarithmic
wind profile would tend to zero) and at 50 m above ground.
We acquired data for the UK, Ireland and surrounding waters
(15
1
/
3
to 10
E, 46.5
to 65.5
N) from 1993 to 2012, giving a
database ith 1.06 billion observations (175,320 temporal 1521
geographic three speed variables plus displacement height). This
database was then processed using R [24] as follows:
For each hour, the east (u) and north (v) components of wind
speed were extracted at all three heights;
A nonparametric polynomial surface was fitted to each set of
spatially gridded observations using a 2-dimensional LOESS
regression. This allowed wind speeds at any coordinates to be
locally interpolated using the nearest twelve observations, as in
Fig. 4;
The magnitude of the wind speed vector (w) was calculated at
each location from w¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
2
þv
2
p;
Wind speed was extrapolated from the three observation
heights to the hub height of each farm (averaging 60 14 m)
using the log law described in the next section.
4. Data processing and validation
The aim of our method was to use NASA modelled wind speed
data as a predictor for the monthly energy output of wind farms.
Two forms of validation were performed to give confidence that
this technique can estimate load factors for a given wind farm that
are broadly representative and free of seasonal or inter-annual bias.
The first test was whether interpolated values from the NASA
GEOS-5 model accurately represent actual wind speeds measured
at a particular location; and the second was that these speeds,
when extrapolated to hub height and transformed using a repre-
sentative power curve, accurately match the actual metered output
from a wind farm.
The data processing and validation was split into five stages:
The NASA speeds at 10 m were compared to ground-based ob-
servations from the Met Office to validate our use of the GEOS-5
model (reported in our Supplementary Material);
Wind speeds at each farm’s location were extrapolated from
50 m to the hub height of that farm to account for wind shear;
These extrapolated speeds were transformed into estimated
ideal load factors using the power curve of the installed turbine
model;
The ideal hourly load factors were compared to the half-hourly
metered output data for farms where this was available, to check
that the preceding methods were robust;
Ideal load factors were aggregated for each farm monthly over
the period of 2002e12, for comparison against the Ofgem/REF
dataset.
The following sections briefly summarise our methods and
findings. An extended validation section which covers each topic in
greater depth is provided as supplementary material.
4.1. Extrapolating wind speeds to turbine hub height
Extrapolating wind speeds from the height of measurement
stations to the much higher hub height of wind turbines is “prob-
ably one of the most critical uncertainty factors affecting the wind
power assessment at a site”[25]. The change in horizontal wind
speeds with height (known as wind shear or the wind profile) is
Table 1
Summary statistics for the extended UK wind output dataset.
Number of observations 20,243
Temporal resolution Monthly
Number of wind farms 282 onshore
Farm capacity
(median)
(mean)
0.5e322 MW
6.5 MW
13.5 MW
Total capacity 4.4 GW
Construction dates 1991e2012
Data recording period 2002e2012
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786 777
generated by friction from the earth’s surface, and so is highly
dependent on the specific site conditions: the surface roughness of
the terrain, air temperature, season, atmospheric stratification, and
the wind speed itself.
Many simplifications for extrapolating wind speeds exist, most
notably the empirically derived power law and the theoretically
derived log law (Eq. (1))[25,26]
w
x
¼Alogh
x
d
z
0
(1)
The simplifiedlog law assumesthat wind speed, w,is relatedto the
logarithm of its height, h, under the assumption of neutral atmo-
sphericstability. The logarithmic windprofile tends to zeroat a height
of the roughness length, z
0
, plus the surface displacement height, d,
and is scaled bya constant, A, which equals the friction velocity (u*)
divided by 0.4 (von Karman’s constant). The log law provides robust
extrapolations of wind speeds over a range of wind speeds, locations
and altitudes[25,27]. The latter is especially important as turbinehub
heights range from 25 to 100 m, and the Hellman exponent used in
the power law decreases non-uniformly with height.
The NASA data give simultaneous wind speeds at three heights,
along with d, allowing the coefficients for wind shear to be calcu-
lated for the specific site and time of the observation. Eq. (1) was
linearised to Eq. (2), allowing the coefficients Aand z
0
to be esti-
mated by a least-squares regression of the three NASA observations.
The coefficient values were independently estimated for each site
and time period with no smoothing or prior values, but were found
to be temporally and spatially stable.
w
x
¼Alogðh
x
dÞAlogðz
0
Þ(2)
At the average UK hub height of 60 m, wind speeds are 6e9%
larger than at 50 m and 32e41% larger than at 10 m. The NASA data
greatly reduces the extrapolation uncertainty compared with using
10 m met mast data.
4.2. Converting wind speed to power
The power curve for an ‘ideal’wind farm was applied to the
speed data at each site, estimating the potential output from a farm
with perfect availability, perfect calibration, and no site-related
performance loss (e.g. turbulence from surrounding geography or
wake effects from other turbines). We consider this to be the
maximum possible attainment and name it the ideal yield.
The meta-data that we integrated into the Ofgem/REF dataset
gave the model of turbine used at each of the 282 wind farms. We
compiled the power curves for 50 of these turbine models, ac-
counting for 92% of the wind farms in the UK. For the remaining
farms, the best match was found from the known curves, based on
the installed turbine’s capacity and power density (peak power
divided by swept area). Supplementary Table 1 gives details of all
the turbines considered.
0
5
10
15
20
25
30
35
1990 1995 2000 2005 2010
Number of farms
Commissioning Year
(a)
0
500
1000
1500
2000
2500
3000
3500
1990 1995 2000 2005 2010
Number of observations
Observation Year
(b)
0
50
100
150
200
250
0 5 10 15 20
Number of observations
Farm Age (years)
(c)
0
100
200
300
400
500
600
700
800
0% 20% 40% 60% 80% 100%
Number of observations
Monthl
y
Load Factor
Mean: 26.3%
(d)
0%
10%
20%
30%
40%
50%
60%
70%
80%
2002 2004 2006 2008 2010 2012
Load Factor
10th – 90th percentile 30th – 70th percentile Mean
(e)
Fig. 1. Histograms summarising the vintage of UK wind farms in our dataset (a), the distribution of monthly observations from this fleet (bec), and the distribution of load factors
for individual farms over the last eleven years (dee).
0%
10%
20%
30%
40%
50%
60%
0 5 10 15 20
Load factor
Farm a
g
e
(y
ears
)
10–90th percentile
30–70th percentile
Trend: 29.18 – 0.44 · age
Sample means
Fig. 2. The mean and distribution of load factors across all farms as they age.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786778
The power curve for a single turbine is shown in Fig. 5 with the
aggregate power curve for a typical farm of these turbines. The
multi-turbine power curve accounts for the fact that wind speeds at
the location of each individual turbine within a farm will vary ac-
cording to a normal distribution. Following the practice of
[10,20,28], the turbine power curve for each farm was convoluted
by a normal distribution. The standard deviations were determined
by the estimated geographic area that each farm covers [28], based
on observations that UK farms occupy 100 m
2
of land per kW ca-
pacity [29]. The mean for each farm was chosen to normalise the
total energy production to that of the individual turbine curve.
4.3. Comparing predicted and metered output
The Ofgem/REF output data has monthly resolution, which is too
coarse to see whether the structure of hourly wind variations
accurately replicates output from operating turbines. Our second
measure of wind farm energy production is the half-hourly metered
output from all transmission-connected generators published on
Elexon’s TIBCO relay service.
2
Data from 2005 to 2012 covers
25 TWh of output from the 47 wind farms highlighted in Fig. 6a.
Fig. 6b and c compare the hourly metered output from Black Law
wind farm with NASA wind speeds interpolated at its location at
80 m above ground, and the ideal energy yields derived from these
speeds.
Each point in Fig. 6b represents one hour’s operation to give the
empirical power curve for that farm. This empirical curve follows
the features of the farm-aggregated Siemens SWT-2.3e82 power
curve (solid line), albeit shifted to the right (as NASA speeds are the
theoretical maximum, and actual speeds will be lower due to local
site features) and downwards (due to downtime and sub-optimal
turbine calibration), with substantial scatter (as local conditions
like turbulence vary over time).
Fig. 6c shows 1000 h of metered output together with the
simulated ideal output, which was scaled by a constant factor of
0.698 for reasons explained in the next section. The simulated ideal
load factors for other farms compare similarly well to their metered
outputs, with plots given in Supplementary Fig. 9.
5. Results
5.1. Simulated wind speeds
The average monthly wind speed at UKonshore wind farms was
estimated to be 7.5 1.5 m/s at the location and hub height of each
farm, which average 62 m above ground. This average speed has
experienced a slight decline over the last 12 years, although the
trend is not statistically significant (0.23 0.37 m/s per decade).
Fig. 7 shows the correlation between the simulated monthly NASA
wind speeds, averaged overall operating sites, and the national
average load factors reported to Ofgem. The correlation is very high
as the simulation is able to represent the actual sites of generation
and the evolution of this site population over time, reaffirming the
strong linear relationship between speed and output at monthly
resolution. The correlation between monthly average speed and
reported load factor is also high for individual wind farms (aver-
aging 0.84), as shown in Supplementary Fig. 12.
0%
10%
20%
30%
40%
50%
60%
0 5 10 15 20
Monthly load factor
Farm age (years)
Combined trend:
29.86 – 0.48 · age
(1.81% per year)
(a)
0
5
10
15
20
25
30
-3 -2 -1 0 1 2 3
Number of farms
Individual farm de
g
radation rates
(
absolute
)
5 years
7.5 years
10 years
Cauchy distribution
μ = –0.48, γ = 0.36
Farms operating
for at least:
(b)
Fig. 3. Lines tracing the output of individual farms as they age (a), and a histogram summarising the linear trends for individual farms, giving the distribution of annual degradation
rates (b).
Fig. 4. Colour map showing the interpolated wind speeds for a given hour over the
British Isles, highlighting the locations of MERRA grid points. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of
this article.)
2
www.bmreports.com.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786 779
If performance declines with wind farm age, we would expect
the reported load factor to fall relative tothe simulated wind speed
over the sample period. This is in fact the case although it is hard to
see in Fig. 7, because so many new farms have been built that the
average age of the UK fleet has only risen only from 4.8 to 7.2 years
between 2002 and 2012. The degradation rates derived below
imply that this would reduce the fleet’s average load factor by only
1 percentage point.
5.2. Ideal load factor and performance ratio
The ideal load factors derived from these wind speeds ignore a
number of factors that will reduce the actual output attained by a
wind farm at the given location and hub height. Three are well-
understood:
1. Machine availability: analysis of national fleets suggests 4e7%
downtime for farms and the electrical infrastructure they rely
upon [9,30], which translates to an 11% reduction in energy
output as turbines on average fail in windier than average
conditions [31];
2. Operating efficiency: sub-optimal control systems, misaligned
components and electrical losses within the farm are found to
reduce output by 2% in well-performing field installations
relative to the turbine’s supplied power curve [30];
3. Wake effects: wind farms suffer from power loss as interactions
between neighbouring turbines increase turbulence and reduce
wind speeds; for relatively small (up to 20 turbine) onshore
farms estimates are in the region of 5e15% [25,32e35];
and two are less well understood:
4. Turbine ageing: based on the findings from Fig. 2 (and presented
later in this paper), energy outputfrom the UK’sfleet is 7.5% lower
than it would be for a fleet of the same turbines as-new, due to
their average age being 5.9 years across the sample period
3
;
5. Site conditions: imperfections in a turbine’s surroundings are not
considered in our model; for example: turbulence intensity,
terrain slope, blockage effects, blade fouling (by dirt, ice, insects,
etc.), or masking by surrounding terrain. These impacts are
highly site specific and hard to quantify with a single factor, with
the only source we found estimating that they reduce output by
2e5%, plus 1% per 3% increase in turbulence intensity [36].
Combining the first four terms, we could expect the ratio of
observed to ideal load factors to be 0.89$0.98$0.90$0.925 ¼0.725.
We call this metric the Performance Ratio (PR), which is analogous
to availability, except it deals with output rather than uptime. Fig. 8
plots the relationship between actual and ideal load factors,
showing that the performance ratio is unbiased across the range of
simulated wind conditions.
Based on the simulated wind speeds and the model of turbine
installed at each farm, the ideal load factor of UK onshore wind
farms should average 38.4%, whereas the mean observed load
factor for these farms from 2002 to 12 has been 26.3%. The average
performance ratio of the farms is therefore 68 19%, confirming
previous work from Ref. [18] which found that a scale factor of 0.69
gave good correlation between load factors derived from a custom
reanalysis and the Ofgem ROC data.
This result does not imply that UK wind farms produce only
two-thirds of what they ought to, for the ideal yield represents a
hypothetical turbine sited on perfectly flat and smooth terrain,
several kilometres from other turbines, foliage or buildings. The
real-world factors 1e4 listed above suggest that a performance
ratio of 0.725 should be expected, which leaves a reduction of 4
percentage points (6%) attributable to the specific site conditions
for UK turbines.
5.3. Weather-corrected load factors
Combining the observed and ideal load factor data allows us to
calculate a weather-corrected load factor (WCLF
f,t
) for every
observation, giving a time series for each farm that should not be
affected by wind conditions changing from month to month. The
actual load factor (LF
f,t
) for a given farm (f) and month (t) is divided
by the ratio of its ideal load factor (ILF
f,t
) for that month to the
farm’s mean ideal load factor over the entire 2002e12 period
(whether or not the farm existed throughout the whole period), as
in Eq. (3). This can be simplified to the performance ratio (PR) for
each month multiplied by the farm’s average ILF.
WCLF
f;t
¼LF
f;t
O ILF
f;t
ILF
f
!¼PR
f;t
ILF
f
(3)
The WCLF represents what a particular farm would have pro-
duced each month if wind conditions followed their long-term
mean distribution. As with the uncorrected load factors, the abso-
lute value encompasses the available wind resource, the quality of
the local site conditions, and the turbine model installed, but the
variation over time is no longer dominated by seasonal weather
patterns.
This is demonstrated in Fig. 9a: removing the weather noise
reveals a gradual decline in this farm’s conversion efficacy, and
allows periods of low availability to be easily identified. Fig. 9b
shows the WCLF averaged across all farms of a given age (in
months), revealing the aggregate level of degradation without the
need for smoothing. The reduction in scatter reduces the uncer-
tainty on this degradation rate from 0.04 when using nominal load
factors (as in Fig. 2), to 0.01 points per year.
Weather correction is notoriously difficult and has the potential
to skew results as it makes large changes in the month-to-month
0%
20%
40%
60%
80%
100%
0 5 10 15 20 25 30
Load factor
Wind speed (m/s)
Aggregated Farm
Individual Turbine
Fig. 5. The power curve for a single Vestas V80 2 MW wind turbine and for a 50 MW
farm of these turbines.
3
In practice ageing is not a separate issue from availability and efficiency, as
these likely fall over a turbine’s lifetime from the as-new values listed in points 1
and 2, producing the ageing effect that we observe.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786780
load factor values. We find this correction procedure to be unbiased
with wind speed (as in Fig. 8b); and comparing Figs. 2 and 9b shows
that it has almost no effect on the fleet-average degradation rate (a
change of 0.01 points per year). The distribution of WCLF decline
rates at individual farms exhibits little change from that of the
unmodified load factors presented in Fig. 3b, averaging 0.45
instead of 0.48 points per year for farms with more than ten years
of data. As we find in Supplementary Section 5.3, the weather
correction process can clean up the short-term fluctuations without
having a systematic effect on the long-term trends.
5.4. Technology improvement over generations
Having rejected the idea that an underlying change in national
wind speeds has distorted the results, we look at the evolution of
the population of turbines. The oldest farms in our sample are the
earliest to have been built, using (presumably) the worst technol-
ogy, and hence are likely to have the lowest load factors.
Fig. 10 shows the individual degradation rate for each farm
against the year it began operating, for all farms with more than 5
years of data. Bubble size is proportional to capacity, bubble
colour represents the number of observations, and horizontal bars
depict the standard error on each decline rate. Black lines show
the best fit to the data using a capacity-weighted loess regression,
showing the central estimate and the range that covers 95% of
observations.
The central fit to the raw load factors in Fig. 10a has a mean
of 0.50 0.07 points per year of operation, and although it shows
noticeable variation between start years there is no long term
trend. A few modern farms have seen increasing load factors over
the first few years of their lives, but this is offset by the majority
having declining outputs.
Correcting for the weather in Fig. 10b greatly reduces the un-
certainty on individual farms and the scatter between them. It also
reduces the mean decline rate to 0.36 0.05 points overall farms.
This decline rate appears stable until 2002, after which it reduces
for more recently commissioned turbines. Farms built before 2003
have an average decline rate of 0.49 0.05 points per year,
whereas those built afterwards average 0.16 0.08.
The impact of weather correction is strongest on the farms
commissioned most recently as they have the fewest years of data,
and so the noise introduced by year-on-year variations in wind
speed is strongest. This is most evident in the group of farms
commissioned after 2003 which were pulled downwards by the
low wind speeds experienced in 2010, and so see a notable
improvement in Fig. 10b compared to a.
Fig. 6. (a) A map showing the location of the UK’s wind farms, highlighting those which were used for validation, with charts showing (b) the hourly metered output from one farm
against simulated wind speeds, and (c) a short time series comparing the observed and simulated ideal outputs.
y = 0.0586x - 0.1938
R² = 0.972
5%
15%
25%
35%
45%
55%
4681012
Reported load factor
Simulated wind speed (m/s)
4
6
8
10
12
5%
15%
25%
35%
45%
55%
2002 2004 2006 2008 2010 2012
Reported load factor
Simulated wind speed (m/s)
Fig. 7. Correlation between the national monthly average NASA wind speeds and load factors reported to Ofgem.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786 781
5.5. Full regression of national fleet performance
Four systematic factors determine the actual output of a wind
farm: wind speeds at the site, the quality of individual turbine lo-
cations (with regards to turbulence and masking), the model of
turbine installed and age-related deterioration of performance.
Other factors, such as the number of turbines suffering faults in a
given month or undergoing planned maintenance, are less
systematic.
As demonstrated by Hughes in Ref. [1], it is possible to separate
the impact of turbine ageing from these other factors by using an
error components model with fixed effects. The observed load
factor, LF
f,t
, of wind farm fin month tis estimated by least squares
regression against the ideal load factor, ILF
f,t
, with fixed effects for
each site, s
f
, and age of the turbine, A
f,t
, minimising the sum of the
squares of the error component, ε
f,t
, as in Eq. (4):
LF
f;t
¼
a
þ
b
ILF
f;t
þs
f
þA
f;t
þε
f;t
(4)
The fixed effects (s
f
and A
f,t
) are the freeform equivalent of a
linear trend. A numeric constant is determined for each site by the
regression to control for any farm-specific factors which affect the
actual wind speeds experienced at the site relative to the NASA
estimates. Similarly, constant modifiers are determined for each
age of turbine (in years) to assess the impact of ageing. A linear or
quadratic regression against age would not capture complex non-
linear behaviour, and so ageing can be better understood through
fixed effects. For it to be possible to solve this model, the fixed effect
for one site (chosen at random) is held constant at zero, as is the
effect for turbines aged 1, which together act as the reference
point.
4
In the model chosen by Hughes [1], the site fixed effects had to
account for the model of turbine as well as local site conditions, and
period fixed effects were used for each observation month to ac-
count for available wind resource (but with the same impact on
every farm in that month). In our model, the systematic effect of
location-specific wind speeds and turbine model are both incor-
porated in the ideal load factor data, while the site dummy vari-
ables measure the extent to which each farm’s surroundings and
layout systematically give it more or less wind than the simulation
Fig. 8. Comparison of observed and ideal load factors for all farms and periods (a), and the dependence of performance ratio on the ideal load factor (b).
0%
10%
20%
30%
40%
50%
60%
70%
80%
2003 2005 2007 2009 2011
Reported monthly load factors
Weather corrected load factors
Trend: –0.123 points per year
2013
(a)
10%
15%
20%
25%
30%
35%
0 5 10 15 20
Farm a
g
e
(y
ears
)
Monthly Mean WCLF
Trend: 29.64 – 0.45 · age
(b)
(1.68% per year)
Fig. 9. (a) An example of reported and weather corrected load factors for a single wind farm (Burradale 2), highlighting periods of partial downtime; and (b) the decline in weather-
corrected load factor across all farms aggregated by age.
4
In the results presented, the performance of Shooters Bottom (in Somerset)
aged 1 is used as a reference.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786782
predicts. The error term picks up the unsystematic factors affecting
each observation.
The regression produced a constant offset of
a
¼1.634 1.529
percentage points of load factor
5
and an estimated coefficient on
the ideal load factor of
b
¼0.755 0.003. This coefficient is above
our average performance ratio of 68% as it is for one-year-old wind
farms, and the age effects show that performance declines over
time. This model, which uses a highly specific wind resource
assessment, provides a better fit to the observed data than using
period fixed effects, giving an R
2
of 0.802 compared to 0.657
attained in Ref. [1]. The full regression results are given in
Supplementary Section 5.6, along with alternative model formu-
lations which we tested.
Fig. 11 plots the fixed effects produced by this regression, re-
centred to give actual load factors as opposed to deviations from
the reference point. Fig. 11a shows the impact of turbine age on the
load factor, taking account of spatial and temporal differences in the
available wind resource, technology installed and local site
conditions.
From Fig. 11a it can be seen that the uncertainty on the agefixed
effects is small enough to be confident that there is indeed degra-
dation, but is too large to be able to discern whether the trend
beyond age 1 is linear, exponential, or some more complex func-
tion. For simplicity’s sake we fit a linear trend to these effects,
which falls from 28.5% at age 1 to the national average load factor of
26.3% for farms at the national average age (5.9 years), reaching
21.0% for the oldest farms aged 19.
Fig. 11b plots the site fixed effects added to the regression offset
and the individual contributions from the ideal load factor at each
farm (i.e.
a
þ
b
ILF
f
þs
f
). This combination yields what we call the
individual farm effects, which are the model’s estimate of each
farm’s load factor at age 1, accounting for location (estimated wind
resource), technology (turbine model and its hub height), and
surroundings (local site quality). Regressing these farm effects
against the year each farm was built yields no significant trend,
implying that any technical improvement has been masked by
diminishing site quality.
Many variations on this model can be considered; for example,
using an exponential rather than linear fit, substituting the ex-
pected load factor with wind speeds (w
f,t
), or estimating the farms’
performance ratio rather than load factor. These models werefound
to give very similar results to the linear model in Eq. (2), with
annual degradation rates that range from 1.5 to 1.9% per year.
Supplementary Section 5.6 details these results and shows that our
extension of the original REF dataset did not skew the rate of
decline.
6. Discussion
The finding that wind turbines lose around 1.6% of their output
each year poses three questions:
1. What are the reasons for this deterioration?
2. Can we expect it to continue in future?
3. What are its wider impacts?
6.1. Reasons for declining output
The degradation rate we observe is perhaps to be expected, as it
lies in the middle of the range experienced by gas turbine tech-
nologies: 0.75e2.25% per year [2]. As with gas turbines and other
aerodynamic rotating machinery, a portion of the unrecoverable
loss could be attributed to gradual deterioration, such as fouling of
the blades (which will impede the aerodynamic performance) and
a gradual reduction in component efficiencies (gearbox, bearings,
generator). These may not be recoverable by maintenance pro-
cedures, but only by component replacement.
A (potentially larger) contribution could come from availability
declining with age, either because older turbines fail more
frequently or because they take longer to bring back online.
Possible reasons for the latter are the likelihood that older ma-
chines suffer more serious failures, difficulty in obtaining compo-
nents for obsolete models, and operators being less likely to hold
comprehensive maintenance contracts. Availability will depend on
the amount of effort the owner is willing to invest in maintenance,
which may naturally fall over time as the asset is paid down, and
will depend on electricity prices and O&M costs. The manufac-
turer’s availability warranty provided with a new turbine (which
may for example guarantee 97% uptime) is also likely to exceed the
standard provided by third party O&M providers in later life.
Early turbine death is a third contributing reason. If one tur-
bine in a farm of four fails completely at age 17, the farm will
continue operating at a maximum of 75% of its original load factor,
Fig. 10. Bubble plots showing the decline rate of individual farms against the year they began operating, considering (a) unmodified load factors (LF) and (b) weather-corrected load
factors (WCLF). Thick vertical lines denote the central fit to each set of points and 95% interval.
5
This is made up from the offset relative to the reference farm (2.926) plus the
average of all farm site effects (4.560).
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786 783
which would translate to an annualised degradation rate of
around 1.6%.
Fig. 9a highlights these first two reasons as a gradual down-
wards slope in the bulk of weather-corrected load factor observa-
tions, and isolated periods of very low output due to downtime
which are concentrated towards higher ages. Fig. 3a shows an
example of the third: Blyth Harbour stands out at the lower-right of
the chart as its load factor declined from 12% to just 2% between the
ages of 12 and 17. By the end of its life only one of the nine turbines
was generating, giving it the worst degradation rate of the farms we
observed.
6.2. Future trends
What of the future? Although our main calculations are based
on linear rates of decline, we cannot yet know if this assumption is
correct. Figs. 9b and 11a suggest a slightly lower rate of ageing in
the first six or seven years of a farm’s life. This early period is
typically covered by comprehensive warranties which guarantee
near-maximal output, whereas older farms may be less intensively
maintained and thus deteriorate more rapidly. The pattern seen in
these Figures could be produced if all turbines (of whatever cohort)
had an initial period of slow decline, followed by accelerating
degradation as they become older.
Alternatively, it could be that advances in technology mean that
recent turbines experience relatively little decline and will
continue to do so in the future, whereas earlier models have
declined relatively faster throughout their lives. As those early
turbines were new in the 1990s their output is not available in our
dataset, and so the experience of those years does not show up in
Figs. 9b and 11a. If those early turbines had declined rapidly when
new, the kink visible in these Figures would have been lessened or
even removed.
Although the improvement in WCLF decline rates seen in farms
commissioned after 2003 is statistically significant, that is no
guarantee that this result will still hold if we were to repeat this
study in a few years’time. Based on this, and further analysis in
Supplementary Section 5.5, we must conclude that even though the
initial signs are good, more data are required before we could say
for sure that modern turbines are declining less rapidly than earlier
cohorts.
6.3. Wider impacts
The cumulative lifetime output of a 100 MW wind farm with a
28.5% load factor would be 4.99 TWh over 20 years. If this farm
suffers a linear annual deterioration of 0.41 points after the first
year, its lifetime output reduces to 4.37 TWh, a fall of 12.5%. This
will increase the cost of electricity from wind generators, as less
electricity is produced to recover the costs of construction. The
economic value of the lost output is relatively low as it mostly oc-
curs in the far future. With a discount rate of 10%, degradation in-
creases the levelised cost of electricity by 9%, from approximately
£90 [4,5] to £98 per MWh. This impact becomes greater if the
economic lifetime increases or the discount rate decreases.
A second impact is that more capacity will need to be installed
to produce a given level of output. The UK has a target for energy
production from renewable sources (15% of all final energy by
2020), as opposed to a target for peak capacity. If turbine build rates
peak in the coming years, the average age of the UK’s wind farms
will creep upwards, and so the output from a fixed capacity can be
expected to decline. For every year the fleet ages, an additional
435 MW (4 large farms) would need to be brought online to
maintain the original capacity of the UK’s anticipated 30 GW fleet.
7. Conclusions
This paper demonstrates a generic and broadly applicable
method for predicting a wind farm’s monthly load factor, ac-
counting for its location, hub height and the particular model of
turbine installed. We use this to estimate the ideal monthly load
factors for 282 of the UK’s wind farms over the last decade, and
compare these to the actual outputs over this period. This allows us
to correct for the rapid improvement in wind turbine technology
over the last two decades and the huge seasonal variability in wind
speeds, thus revealing the subtle rate of degradation.
We find evidence of important, but not disastrous, performance
degradation over time in a large sample of UK wind farms. When
variations in the weather and improvement in turbine design are
accounted for, we find that the load factors of UK wind farms fall by
1.57% (0.41 percentage points) per year. This degradation rate ap-
pears consistent for different vintages of turbines and for individual
wind farms, ranging from those built in the early 1990s to early
2010s.
We use six methods of increasing complexity to find the
following rates of degradation in absolute percentage points per
year; and relative to the UK mean wind farm:
Simple regression of all load factors against age (0.44 0.04
absolute) (1.69 0.17% relative);
Average trend in load factors for individual farms (0.48 0.36
absolute) (1.81 1.32% relative);
15%
20%
25%
30%
05101520
Fleet average load factor
Farm a
g
e
(y
ears
)
Age Fixed Effects
Standard Error
Trend: 28.90 – 0.41 · age
Average LF: 26.3%
Average age: 5.9 years
(1.57% per year)
(a)
0%
10%
20%
30%
40%
50%
1992 1996 2000 2004 2008 2012
As-new load factor
Start Year
Individual farm effects
Trend: 0.01 points per year
(b)
Fig. 11. Fixed effects from the regression of observed load factor against ideal load factors.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786784
Correct for wind resource and regress weather-corrected load
factor against age (0.45 0.01 absolute) (1.6 8 0.05%
relative);
Trend in weather-corrected load factors for individual farms
against their age (0.45 0.22 absolute) (1.7 0 0.82%
relative);
Capacity-weighted fit to individual farms against their year of
commissioning (0.50 nominal, 0.36 weather-corrected)
(1.90% and 1.24% relative);
Full fixed effects regression, accounting for site-specific wind
speeds, turbine model and site quality (0.41 0.01 absolute)
(1.57 0.06% relative).
The combined average of these measures is 0.43 0.05 per-
centage points per year, giving 1.6 0.2% annual degradation. The
similarity of results from different methods gives us confidence
that the underlying trend is robust: the decline in load factor with
age is neither an artefact of systematic variation in wind speeds nor
of the continual improvement in technology. Questions do however
remain as to the exact form of this degradation, for example
whether it is linear, quadratic or logarithmic with age; or how
degradation rates are changing over time and whether they will be
lower in the future. Access to data from more farms, and a more
detailed wind resource assessment for each site will be funda-
mental to furthering our understanding of these issues.
The level of degradation we find is not insignificant, yet it is not
unusual compared to conventional generation technologies. The
fact that it has been omitted from calculations of the levelised cost
of electricity from wind means that these estimates are around 9%
below the true value (depending on assumed discount rate and
economic lifetime). This is unlikely to be large enough to change
the business case for wind power, but nonetheless it needs to be
accounted for to give an accurate picture of its cost.
Acknowledgements
The authors would like to thank:
Gordon Hughes, John Constable and the Renewable Energy
Foundation, who enabled this research by generously making their
data publicly available;
Christopher Crabtree, Andrew Garrad and Paul Gardner (GL
Garrad Hassan), Matthew Hanson (Pöyry), David MacKay (DECC),
David Newbery, Jim Oswald (Oswald Consulting), Jim Platts and
Peter Tavner for insightful discussions and feedback on our earlier
drafts;
The British Atmospheric Data Centre and NASA GoddardFlight
Centre for the excellent services they provide;
Anabelle Guillory from the Centre for Environmental Data
Archival (CEDA) for help in acquiring the MIDAS marine data;
Alasdair Skea and Simon Vosper from the UK Met Office for
improving our understanding of the MIDAS equipment and data;
The Alan Howard Charitable Trust and the EPSRC Grand
Challenge Project ‘Transforming the Top and Tail of the Energy
Networks’(EP/I031707/1) for funding this work.
Appendix A. Supplementary data
Supplementary data related to this article can be found at http://
dx.doi.org/10.1016/j.renene.2013.10.041.
References
[1] Hughes G. The performance of wind farms in the United Kingdom and
Denmark. London: Renewable Energy Foundation; 2012. URL: http://tinyurl.
com/cn5qnqg.
[2] Brooks FJ. GE gas turbine performance characteristics. GE Power Systems;
2000.
[3] Jordan DC, Kurtz SR. Photovoltaic degradation ratesean analytical review.
Prog Photovoltaics: Res Appl 2013;21(1):12e29.
[4] MacDonald Mott. UK electricity generation costs update; 2010.
[5] Arup. Review of the generation costs and deployment potential of renewable
electricity technologies in the UK. London: Department of Energy and Climate
Change; 2011.
[6] Brinckerhoff Parsons. Electricity generation cost modele2011 update revision
1. London: Department of Energy and Climate Change; 2011.
[7] Ribrant J, Bertling LM. Survey of failures in wind power systems with focus on
Swedish wind power plants during 1997e2005. IEEE Trans Energy Convers
2007;22(1):167e73.
[8] Tavner PJ, Xiang J, Spinato F. Reliability analysis for wind turbines. Wind
Energy 2007;10(1):1e18.
[9] Arwade SR, Lackner MA, Grigoriu MD. Probabilistic models for wind turbine
and wind farm performance. J Solar Energy Eng 2011;133(4):041006e9.
[10] Pöyry Energy Consulting. Implications of intermittency. Oxford. URL: http://
www.uwig.org/ImpactofIntermittency.pdf; 2009.
[11] Sinclair Knight Merz. Growth scenarios for UK renewables generation and
implications for future developments and operation of electricity networks;
2008. URN 08/1021.
[12] Green R, Vasilakos N. Market behaviour with large amounts of intermittent
generation. Energy Policy 2010;38(7):3211e20.
[13] Green R, Staffell I, Vasilakos N. Divide and conquer? k-means clustering of
demand data allows rapid and accurate simulations of the British electricity
system. IEEE Trans Eng Manage 2013 [in press].
[14] Sturt A, Strbac G. A times series model for the aggregate GB wind output circa
2030. In: Presented at renewable power generation (RPG 2011), IET confer-
ence 2011.
[15] García-Bustamante E, González-Rouco JF, Jiménez PA, Navarro J, Montávez JP.
A comparison of methodologies for monthly wind energy estimation. Wind
Energy 2009;12(7):640e59.
[16] Kubik ML, Coker PJ, Barlow JF, Hunt C. A study into the accuracy of using
meteorological wind data to estimate turbine generation output. Renew En-
ergy 2013;51(0):153e8.
[17] Kiss P, Varga L, Janosi IM. Comparison of wind power estimates from the
ECMWF reanalyses with direct turbine measurements. J Renew Sustain En-
ergy 2009;1(3):033105e11.
[18] Hawkins S, Eager D, Harrison GP. Characterising the reliability of production
from future British offshore wind fleets. In: Presented at IET conference on
renewable power generation (RPG 2011) 2011. Edinburgh.
[19] Kubik M, Brayshaw D, Coker P. Reanalysis: an improved data set for simu-
lating wind generation?. In: WREF 2012. Denver, CO. URL: http://tinyurl.com/
c4ge72x; 2012.
[20] Ofgem. Electricity capacity assessment. Ref 126/12: London. URL: http://
tinyurl.com/97cuuxt; 2012.
[21] Renewable Energy Foundation. Analysis of wind farm performance in UK and
Denmark [cited Dec 2012]; URL: http://tinyurl.com/azu3m97; 2012.
[22] Lucchesi R. File specification for MERRA products. GMAO office note No. 1
(Version 2.3). NASA Goddard Space Flight Centre; 2012. URL: http://gmao.gsfc.
nasa.gov/pubs/office_notes/.
[23] Rienecker MM, Suarez MJ, Gelaro R, Todling R, Bacmeister J, Liu E, et al.
MERRA: NASA’s modern-era retrospective analysis for research and applica-
tions. J Clim 2011;24(14):3624e48.
[24] R Development Core Team. R: a language and environment for statistical
computing. Vienna, Austria. URL: http://www.R-project.org; 2013.
[25] Schallenberg-Rodriguez J. A methodological review to estimate techno-
economical wind energy production. Renew Sustain Energy Rev 2013;21(0):
272e87.
[26] Bañuelos-Ruedas F, Angeles-Camacho C, Rios-Marcuello S. Methodologies
used in the extrapolation of wind speed data at different heights and its
impact in the wind energy resource assessment in a region. In:
Suvire GO, editor. Wind farmetechnical regulations, potential estimation
and siting assessment. Rijeka, Croatia: InTech; 2011. http://dx.doi.org/
10.5772/20669.
[27] Fox NI. A tall tower study of Missouri winds. Renew Energy 2011;36(1):
330e7.
[28] Nørgaard P, Holttinen H. A multi-turbine power curve approach. In: Presented
at Nordic wind power conference 2004. Gothenburg, Sweden.
[29] MacKay DJC. Could energy-intensive industries be powered by carbon-free
electricity? Philosophical transactions of the Royal Society a: mathematical.
Phys Eng Sci 2013;371(1986).
[30] Harman K. Availability and operating efficiency ehow well do wind farms
actually perform?. In: Renewable UK conference. Glasgow: GL Garrad Hassan;
2010. URL: http://tinyurl.com/adz94sp.
[31] Conroy N, Deane JP, Ó Gallachóir BP. Wind turbine availability: should it be
time or energy based? ea case study in Ireland. Renew Energy 2011;36(11):
2967e71.
[32] Johnson C, Graves A, Tindal A, Cox S, Schlez W, Neubert A. New developments
in wake models for large wind farms. In: AWEA windpower conference.
Chicago, IL: Garrad Hassan America Inc; 2009. URL: http://tinyurl.com/
dxlkatz.
[33] Phillips JL, Cox SD, Henderson AR, Gill JP. Wake effects within and between
large wind projects: the challenge of scale, density and neighbourseonshore
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786 785
and offshore. In: EWEC conference. Warsaw, Poland: GL Garrad Hassan; 2010.
URL: http://tinyurl.com/bm9nhk9.
[34] Barthelmie R, Larsen G, Pryor S, Jørgensen H, Bergström H, Schlez W, et al.
ENDOW (efficient development of offshore wind farms): modelling wake and
boundary layer interactions. Wind Energy 2004;7(3):225e45.
[35] Ali M, Matevosyan J, Milanovi&cacute JV. Probabilistic assessment of wind
farm annual energy production. Electric Power Syst Res 2012;89(0):70e9.
[36] Johnson C, Tindal A, Harman K, Graves A. Validation of energy predictions by
comparison to actual performance. In: WINDPOWER conference. Garrad
Hassan. URL: http://tinyurl.com/aqkagoy; 2008.
I. Staffell, R. Green / Renewable Energy 66 (2014) 775e786786