Efficiency of Wind Power Production and its Determinants

Simone Pieralli, Matthias Ritter, Martin Odening

Paper prepared for presentation at the 2015 Agricultural & Applied Economics Association and
Western Agricultural Economics Association Annual Meeting, San Francisco, CA, July 26-28

Copyright 2015 by Simone Pieralli, Matthias Ritter, Martin Odening . All rights reserved.
Readers may make verbatim copies of this document for non -commercial purposes by
any means, provided that this copyright notice appears on all such copies.

1
Efficiency of Wind Power Production and its Determinants*

Simone Pieralli a ** , Matthias Ritter a , Martin Odening a

Abstract

This article examines the efficiency of wind energy production. Using non-convex efficiency
analysis, we quantify production losses for 19 wind turbines in four wind parks across Germany.
In a second stage regression, we adapt the linear regression results of Kneip, Simar, and Wilson
(2014) to explain electricity losses by means of a bias-corrected truncated regression analysis.
The results show that electricity losses amount to 27% of the maximal producible electricity.
Most of these losses are from changing wind conditions, while 6% are from turbine errors.

Key words: wind energy, efficiency, free disposal hull, bias correction.
JEL codes: D20, D21, Q42.

1 Introduction

Renewable energy production has experienced rapid growth over the last two decades and this
growth is likely to continue. Wind energy production contributed a significant share to this
expansion and has attracted institutional investors. The profitability of wind energy production is
determined by generation costs, energy prices, and turbine productivity. In the past, investments
in wind parks were able to attain comparably high returns on investments. In many countries,
such as Germany and Spain, producers receive guaranteed prices for wind energy that are above
market prices. Generation costs are also fairly stable since operating costs are relatively low and
installation costs are rather transparent. Thus, productivity is the crucial driver for the
profitability of wind energy production. Productivity, in turn, heavily depends on wind

*
Financial support from the German Research Foundation via CRC 649 “Economic Risk” is gratefully
acknowledged.
a
Humboldt-Universität zu Berlin, Department of Agricultural Economics, Philippstr. 13, 10115 Berlin, Germany.
**
Corresponding author: simone.pieralli@agrar.hu-berlin.de, Tel.: +49 (0) 30209346847, Fax: +49 (0)
30209346841.

2
conditions, i.e., wind speed and its variability, at the production site. In fact, a careful assessment
of wind conditions precedes any investment in wind parks. Given the importance of wind
production, it is not surprising that a lot of effort has been devoted to developing models to
predict how much of the installed capacity will actually be used during the investment period
(e.g., Kusiak et al., 2013).

A second determinant of productivity, however, has received very little attention in the literature,
namely, the efficiency of wind energy production which is the distance between the actual
energy and maximal energy output under a certain level of production factors. In the context of
wind energy, the maximal producible power as a function of wind speed is depicted by a power
curve. Power curves are usually calculated by turbine producers for a specific turbine type under
ideal conditions.1 In reality, wind production does not take place under ideal conditions and
therefore actual energy production regularly deviates from the power curve. For example,
shortfalls can be caused by rainfall, icing, suboptimal adjustment of the pitch angle and nacelle
position under changing wind conditions, technical failures, and scheduled maintenance. Under
marginal wind conditions or a scenario of declining subsidies, these production inefficiencies can
diminish the profitability of wind power plants.

A few empirical papers analyze the productivity and efficiency of wind power generation.
Homola et al. (2009) analyze wind park data in Norway and suggest a correction for power curve
estimation. Ilinca (2011) estimates that power losses due to icing conditions amount to as much
as 50% of total annual production. Hughes (2012) and Staffell and Green (2014) indicate
declining turbine performance due to increasing age of turbines in Denmark and the UK. Some
other papers apply nonparametric methods to estimate the wind energy production frontier.
Kusiak et al. (2012) use data envelopment analysis (DEA) to assess the performance of wind
turbines in the presence of faults. They identify turbine downtime as the major reason for power
curtailment. Iribarren et al. (2014) analyze the entire process of wind energy production and
include further production factors, such as land and investment cost in their DEA model. To the
best of our knowledge, Carvalho et al. (2009) is the only study that applies DEA to estimate a
power curve based on high frequency production data.

The objective of this paper is twofold. First, we estimate the wind energy production frontier
based on production data and quantify production losses that occur relative to this benchmark. In
contrast to most other wind efficiency studies, we base frontier estimations directly on high
frequency production data and do not aggregate the data for a single turbine or wind park. This
sheds light on the emergence of production losses over time and avoids information losses
through data smoothing. Carvalho et al. (2009) pursue a similar approach; however, they use a
DEA model and thus estimate efficiency by assuming a convex production technology. This
approach ignores the non-convex shape of a typical wind power curve and thus overestimates

1
The industry standard for power curve estimation is IEC 61400-12-1 (Wächter et al., 2009; Homola et al., 2009).

3
inefficiency over some range of the production frontier. To avoid this flaw of DEA, we resort to
a free disposal hull (FDH) estimation of the frontier, which does not assume convexity.

The second objective of this paper is to explain the magnitude of the observed production losses
and to trace them back to factors which may or may not be under the control of wind park
operators. To this end, we apply a truncated regression model that accounts for biases in the
regression of estimated efficiency scores in the first step of our analysis (Kneip et al., 2014).
From an applied perspective, our findings help improve the assessment of wind energy
production under real world conditions.

In the following section, we explain in greater detail how we estimate the wind energy
production frontier and derive the corresponding production losses. Moreover, we present the
bias correcting regression model. These methods are then applied to high frequency production
data from four wind parks in Germany. Section 3 describes the data base and Section 4 presents
results. The final section summarizes and draws conclusion for improving the productivity of
wind energy generation.

2 Methodology

The amount of wind’s kinetic energy (𝐸𝑘 ) available to be converted into electricity can be
described by the following function (Hennessey, 1977; Gunturu and Schlosser, 2012):

𝐸𝑘 = 0.5𝜋𝑟 2 𝑑𝑤 3 , (1)

where r is the rotor size, so that the rotor swept area is 𝜋𝑟 2, 𝑑 is the air density, and 𝑤 denotes
wind speed. Air density is directly proportional to air pressure and inversely proportional to air
temperature. Kinetic wind energy increases with wind speed and air density. It is important to
note that according to Eq. (1), kinetic wind energy is a cubic function of wind speed. This
characteristic results in a non-convex technology for wind speeds lower than the rated wind
speed and has implications for the estimation of the production frontier. Air density is directly
proportional to air pressure and inversely proportional to air temperature. Density is higher in the
winter when the temperature is colder and is lower in the summer when the temperature is
warmer. Air pressure causes variability in this general trend: High air pressure increases air
density and low air pressure decreases density. However, only a portion of the wind’s kinetic
energy can be transformed into electricity. The efficiency of this transformation process depends
on various technical and managerial factors and is the subject of this study.

In general terms, the production process is characterized by a production technology, which is

defined as the set of all inputs (in our case: wind speed and air density) that are feasible to
produce electric power:

𝑇 = {𝑤, 𝑑, 𝑒: (𝑤, 𝑑) can produce 𝑒}, (2)

4
where 𝑤 is wind speed, 𝑑 is air density, and 𝑒 is wind electricity.

As mentioned above, wind speed is monotonically related to the amount of electrical power
produced, but the rate of transformation is non-constant and increasing up to the rated wind
speed. However, to preserve the machine equipment from destructive centrifugal forces, the
rotational speed and thus power production are limited for wind speeds greater than the rated
wind speed. These features of the production technology process can be captured by a non-
convex free disposal hull (FDH) for a sample of 𝑛 observation points {𝑤𝑖 , 𝑑𝑖 , 𝑒𝑖 }𝑛𝑖=1 :

𝑇̂FDH = {𝑤, 𝑑, 𝑒 ∶ 𝑤 ≥ 𝑤𝑖 , 𝑑 ≥ 𝑑𝑖 , 𝑒 ≤ 𝑒𝑖 , ∀𝑖 = 1, … , 𝑛}. (3)

The FDH technology set creates an outer envelope of the data points included in technology 𝑇
without assuming convexity. As a measure of the efficiency of the turbines in exploiting wind
and air density conditions, we measure the nonparametric distance between each point and the
frontier envelope. Since inputs cannot be controlled by producers and instead are determined by
nature, it is reasonable to measure distance in the direction of the outputs. We define this
efficiency measure for unit (𝑤0 , 𝑑0 , 𝑒0 ) as follows:

𝜆̂FDH (𝑤0 , 𝑑0 , 𝑒0 ) = sup{𝜆: (𝑤0 , 𝑑0 , 𝜆𝑒0 ) ∈ 𝑇̂FDH }. (4)

An estimate 𝜆̂FDH can be computed by means of a sorting algorithm that identifies all units that
dominate unit (𝑤0 , 𝑑0 , 𝑒0 ), i.e., all units that use less or equal inputs to produce equal or more
output than unit (𝑤0 , 𝑑0 , 𝑒0 ).2 The dominating unit with the highest output is the potential
producible electric power, 𝑒̂0 . The efficiency measure 𝜆̂FDH is then a ratio of the potential
producible electric power 𝑒̂0 and the actual electric power produced 𝑒0 . This measure of
inefficiency is a conservative measure compared to convex hull measures of inefficiency, such as
DEA. Given this efficiency measure, we define the production loss of electric power EL for
every observation as the difference between the production potential and the electric production
observed:

̂ 0 = 𝜆̂FDH 𝑒0 − 𝑒0 = 𝑒̂0 − 𝑒0 .
EL (5)

Previous efficiency analyses on wind energy production consider single turbines or wind parks
as “decision making units” and calculate efficiency scores for these units (e.g., Iribarren et al.,
2014; Iglesias et al., 2010). This kind of analysis requires aggregating inputs and outputs to
annual values. Production factors may include capital and labor. Here, we pursue a different
approach. Efficiency scores are assigned to production intervals consisting of ten minutes. Each
observation in our sample relates the electric power produced to the average wind speed and
average air density in a ten-minute interval. Thus, we derive an efficient production function that

2
We are thankful to Professor Dr. Simar who provided the sorting algorithm to calculate the FDH efficiency scores.
However, we adapted the algorithm received to calculate the frontier in case of missing or infinite values. Efficiency
scores 𝜆 can be alternatively determined by a mixed integer program (Deprins et al., 1984).

5
characterizes the technology of the wind turbine under differing wind and air density conditions.
This production function resembles a power curve. Production factors other than wind and air
density are not considered. Pooling high frequency production data from different wind parks in
an efficiency analysis makes sense if electricity is produced with the same technology, i.e.,
turbine type and rotor size, which is the case in our study. Productivity under this perspective can
be understood as electricity output under given weather conditions. This definition differs from
productivity of a wind turbine or wind farm per unit of time, such as annual production. In our
analysis, we may find that a wind turbine is efficient because it converts wind energy optimally
into electricity, but that the produced power is low due to low inputs, i.e., unfavorable wind
conditions. Thus, our efficiency analysis cannot support decisions on the location of wind parks
or the choice among production technologies. Instead, it provides useful information on the
magnitude of production losses due to suboptimal utilization of wind energy. Such production
losses may be caused by unfavorable weather conditions (apart from wind speed and air density),
such as icing or turbine faults. Though most of these factors are out of immediate managerial
control, it is helpful to understand their contribution to electricity losses.

The second step of our analysis focuses on explaining observed production inefficiency by
̂ on a set of explanatory variables 𝒗. While this two-
regressing the estimated electricity loss EL
step procedure is standard for nonparametric efficiency analyses, it ignores a methodical problem
̂ result from a nonparametric estimation of
(Kneip et al., 2014). The fact that electricity losses EL
the technology frontier entails problems to use them as a dependent variable in a second stage
regression. In fact, the estimated efficiencies are biased measures of the true electricity losses
because in a full population sample, there could be observations lying above the sample frontier.
Thus, the estimated inefficiency measures derived from the sample frontier represent a lower
bound of the true measures. This bias will lead to a biased estimation of the regression
coefficients 𝜽. To account for this, we adapt the procedure proposed by Kneip et al. (2014). A
further aspect that has to be accounted for in the regression is the limited range of the calculated
electricity losses: They are non-negative and cannot exceed the maximum capacity of the turbine
(2.365 MW in our sample). Thus, we apply a truncated regression model. Assuming that the
latent variable EL is normally distributed,

EL𝑖 |𝒗𝑖 ∼ 𝑁(𝒗′𝑖 𝜽, 𝜎 2 ), (6)

̂ can be obtained by maximizing the following likelihood

the estimates of the parameters 𝜽
function:
1 ̂ −𝒗 𝜽
EL
φ( 𝑖 𝑖 )
σ 𝜎
ℒ1 = ∏ni=1 ( UB−𝒗𝑖𝜽 LB−𝒗𝑖𝜽 ) (7)
Φ( )−Φ( )
𝜎 𝜎

where φ(∙) is the standard normal density function and Φ(∙) is the cumulative standard normal
density function. The arguments of the normal density function and cumulative density function

6
are derived from the conditional truncation points of the regression model. UB and LB in the
argument of the cumulative distribution function are the upper bound (2,365) and lower bound
(0) of the dependent variable, respectively, and σ is the variance of the error term.

To correct for the estimation bias, we divide the sample into two parts and recalculate the
̂ 1 and EL
efficiency losses EL ̂ 2, as in Kneip et al. (2014).3 We stack these variables together to
̂
create a column vector of n elements 𝐄𝐋 ̂ 𝑠 = (𝐄𝐋1 ). Accordingly, we stack the respective m
̂2
𝐄𝐋
𝒗𝟏
explanatory variables to obtain an n×m matrix 𝒗𝑠 = (𝒗 ), in which the order of the observations
𝟐
is rearranged to match the dependent variable. Similarly to Eq. (7), we calculate a new estimator
̂ 𝑠 in a regression of 𝐄𝐋
𝜽 ̂ 𝑠 on 𝒗𝒔 by maximizing the following likelihood function:

1 ̂ 𝑠−𝒗𝑠 𝜽𝑠
EL
φ( 𝑖 𝑖 )
𝜎 𝜎
ℒ2 = ∏ni=1 ( 𝑠
UB−𝒗𝑖 𝜽 𝑠 LB−𝒗𝑠𝑖 𝜽𝑠
). (8)
Φ( )−Φ( )
𝜎 𝜎

Under Theorem 5.2 in Kneip et al. (2014), the convergence result for an FDH convergence rate
of 𝑛−𝜉 with 𝜉 ≥ 1/3 can be obtained by:
−1 ℒ
̂ − (2𝜉 − 1)
√𝑛 [𝜽 ̂𝑠 − 𝜽
(𝜽 ̂) − 𝜽] → 𝒩(0, 𝜎 2 𝑸) (9)

̂ is the biased estimate and (2𝜉 − 1)−1 (𝜽

as 𝑛 → ∞. In Eq. (9), 𝜽 ̂𝑠 − 𝜽
̂ ) is the bias correction.
This correction depends on 𝜉, which is the inverse sum of the number of inputs (𝑝) and outputs
(𝑞), i.e., 𝜉 = 1/(𝑝 + 𝑞). That is, the higher the number of inputs and outputs, the higher the bias
correction.

Asymptotic confidence intervals for the vector of parameters can be determined from Eq. (9) by:

̂ − (2𝜉 − 1)−1 (𝜽
𝜽 ̂𝑠 − 𝜽
̂ ) ± 𝑧 𝛼 𝜎̂𝑛 𝑞̂𝑚𝑚 /√𝑛 (10)
1−
2

where 𝑞̂𝑚𝑚 = (𝑍 ′𝑍)−1 /𝑛 are diagonal elements of the matrix 𝑸.

3 Data and model variables

The production data in this study are for four wind parks situated in different regions in western,
central, and eastern Germany.4 There is a total of 19 turbines in the sample, with each wind park
containing up to 7 turbines, which are all of the same type and capacity, namely 2.365 MW. The
3
To maintain a representative sample in this estimation, we assign observations in odd positions (1, 3, 5, etc.) to the
first sub-sample and in even positions (2, 4, 6, etc.) to the second sub-sample.
4
The authors thank 4initia GmbH for providing the data. The names and exact locations of the wind parks are
concealed for confidentiality reasons.

7
average amount of power produced in kilowatts is reported for 10-minute intervals from July 1st,
2013 to June 30th, 2014. There are 989,175 observations for the 19 turbines in the sample. The
dataset also includes observations of average mast wind for each 10-minute interval which
constitutes the first (non-controllable) input in our efficiency analysis. It should be noted that the
averaging of electricity output and wind speed over a ten-minute interval may lead to
measurement errors in case of short-term fluctuations in wind speed. Since wind electricity is a
non-linear function of wind speed, the mean value of wind electricity differs from the wind
electricity rated at mean wind speed due to Jensen’s inequality. This flaw of standard methods of
power curve estimation is well-known and has led to modifications, such as dynamical power
curve estimation (cf. Gottschall and Peinke, 2008; Homola et al., 2009). We do not adjust the
measurement in our efficiency analysis. Instead, we include wind variability as an explanatory
variable for inefficiency in the regression analysis.

The second input for wind electricity production is air density. We obtain this variable from a
reanalysis dataset often used in wind power analysis, namely the Modern-Era Retrospective
Analysis for Research and Applications (MERRA) data provided by NASA (Rienecker et al.,
2011). MERRA reanalysis data reconstruct the atmospheric state by integrating data from
different sources, such as conventional and satellite data. They offer a complete worldwide grid
of weather data at a spatial resolution of 1/2° latitude and 2/3° longitude (about 45 km × 54 km
in Germany). We interpolate the surface air density data of the four nearest grid points weighted
by their distance to each wind park.5 The data are available daily at one-hour intervals, beginning
at 12:30 a.m., which we linearly interpolate to obtain observations for each 10-minute interval.

The summary statistics of input and output variables used for the estimation of the FDH and
production losses are presented in Table 1. Summary statistics of the electricity produced in all
single turbines is provided in Table A.1 in the appendix.

Table 1: Summary statistics of input and output variables

Mean Standard Minimum Maximum
deviation
Inputs
Wind speed (m/s) 5.90 3.04 0.00 28.20
Air density (kg/m³) 1.20 0.03 1.10 1.34
Output
Electricity produced (kW) 507.46 622.76 0.00 2,365.00
Number of observations 989,175

5
The air density variable is called RHOA in the “MERRA IAU 2d surface turbulent flux diagnostics
(AT1NXFLX)”. More details on MERRA products can be found in Lucchesi (2012).

8
The upper part of Figure 1 depicts the average electricity production in the 19 turbines plotted
against wind speed during the 10-minute intervals.6 This figure provides a first impression of the
range of productivity in observed wind electricity given a certain level of wind speed.
Apparently, the distance between the lowest and highest output varies with the wind speed. The
highest variation in productivity can be observed at moderate wind speeds between 5 and 12 m/s,
whereas the productivity is less variable for calm wind conditions as well as for observations
above the rated wind speed. This is plausible since the production potential under calm wind
conditions is low for technical reasons. On the other hand, the frequency of observations at
moderate wind speed is high (cf. Fig. 1(b)), so that heterogeneity in outcomes is expected to be
high in the presence of other production factors or measurement errors. The wind distribution in
Fig. 1(b) reveals that the behavior at moderate wind speeds is important for the overall efficiency
of wind turbines.

Fig. 1. (a) Power produced against wind speed; (b) Frequency of observed wind speed

6
Note that production is measured in kW and not in kWh. This allows for a direct comparison with the installed
capacity, but needs careful interpretation in the context of production losses. Production losses reported in kW for a
10-minute interval can be converted to kWh by dividing by six.

9
Our data set provides further information that can be used to derive explanatory variables in the
second stage regression. We hypothesize that variability of wind conditions, i.e., wind speed and
wind direction, affect measured productivity for two reasons. First, changing wind conditions
require adjustments of the turbine’s operation, e.g., the rotor pitch angle or the nacelle position.
Because these adjustments are not frictionless and are realized with some delay, they decrease
power generation compared to a situation of stable wind conditions. Here, we use the range of
the wind speed in a 10-minute interval as an indicator of wind speed variability. Moreover, as
mentioned above, changing wind speed will lead to errors in measurement of the average power
production during 10-minute intervals. We further control for the speed of adjustment of the
machine to different wind speeds by considering the difference in the average rotor speed
between two consecutive 10-minute intervals. Changes in the wind direction are approximated
by the absolute change (either to the left or right) in nacelle positions between two subsequent
observations. The impact of this variable on productivity, however, is not clear a priori. On the
one hand, higher direction variability is supposed to decrease the capability of a stable
production of electric energy. On the other hand, the capability to adapt to changing wind
directions can be regarded as an efficiency improving feature of the turbine. To ensure
convergence of the maximum likelihood estimation, we use the cubic root of the absolute
change.

Moreover, a detailed report for the turbine status is available at each point in time. This includes
the occurrence of various error types with their starting times. In the data processing, we
combine this information with the 10-minute interval data, i.e., we assign an error to all 10-
minute intervals between the beginning of an error and the restart of the turbine. If there are
concurrent errors, we consider only the error that occurred first. In the regression model, we
consider three error categories that result in the disruptions of electricity production: the presence
of an ice alert on the turbine, the presence of turbine maintenance, and a residual error category
that includes all other errors. To put this in perspective, errors occur in about 3% of the
observations in our sample, 30,363 observations. About half of these errors, 15,077, are ice
alerts, with the remaining errors resulting from machine errors (9,775 cases or 32%) and
maintenance (5,511 cases or 18%). All dummy variables for errors are interacted with wind
speed to weigh the occurrence of an error by the wind energy.

Finally, to account for location specific heterogeneity, we include a full set of 19 turbine
dummies (𝑫turbine) which capture, for example, exposition to particular wind conditions due to
the position of the turbine within a wind park.

We specify the regression model of Eq. (6) for the production loss EL by:

EL = 𝑫turbine 𝜶 + 𝒁𝜷 + 𝑤𝑯𝜹 + 𝜖 (11)

where 𝒁 = [Wind range, Nacelle position change, Rotor speed difference]

and 𝑯 = [Icing error, Maintenance error, Machine error]

10
Summary statistics of the dependent and explanatory variables used in the second stage are in
Table 2.

Table 2: Summary statistics of second stage regression variables

Mean Standard Minimum Maximum
deviation
Second stage variables
Electricity loss (kW) 190.78 150.55 0.00 2,365.00
Wind range (m/s) 3.51 2.09 0.00 30.60
Wind speed (m/s) 5.90 3.04 0.00 28.20
3
Nacelle position change (˚, √|𝑥 |) 1.18 0.91 0.00 6.70
Rotor speed difference (rpm) 0.0004 1.02 -17.36 16.65
Number of observations 989,175

4 Results

The resulting estimate of the FDH technology is depicted in Fig. 2. As required by the free
disposability assumption on the technology, increasing amounts of both inputs are related to
higher potential electricity production. The impact of the two factors on electricity production,
however, is different: Low wind speed renders any air density amount unimportant for power
production, whereas low air density diminishes electricity production only marginally. For this
reason, we focus on wind speed as the most important production factor in the subsequent
figures, i.e., we focus on the power curve.

Fig. 3 depicts two power curves. The broken line represents a power curve that we calculated
following the industry standard IEC 61400-12-1 (Homola et al., 2009). It reflects the average
produced power for wind speed bins of 0.1 m/s width, where wind speeds are adjusted to an air
density of 1.225 kg/m³ and erroneous observations are excluded. The solid line is a cross-section
of the FDH at the same air density of 1.225 kg/m³. It differs from the standard power curve in
two ways. First, it represents an envelope of the production data instead of an average. Second,
the estimation is based on all (non-filtered) observations. Apparently, both curves differ across a
wide range of wind speeds. The difference between the curves amounts to 183 kW per
observation, which corresponds to 8% of the rated capacity. Fig. 3 shows that the distance
between the best performing observations and the average observations for a given wind speed is
higher for wind speeds between 8 and 12 m/s, compared to lower wind speeds. For wind speeds
larger than 15 m/s, both functions converge. Note that there is a deviation between the FDH and
the standard power curve for extreme wind speeds. Storm control in modern wind turbines
results in a decline in power for very high wind speeds. Due to the free disposability assumption,
this particular feature of wind electricity production cannot be mimicked by the FDH. In these

11
cases, our method overestimates electricity losses. However, in our dataset only eight
observations have wind speeds greater than 25 m/s (cf. Fig. 1(a)).

Fig. 2. Estimated free-disposal hull technology

Fig. 3. Power curve vs. production frontier

12
The difference for every observation between the potential power production represented by the
FDH estimated frontier in Fig. 3 and the actual power production defines an electricity loss (EL)
that is plotted against wind speed and air density in Fig. 4. Fig. 4(a) shows that the bulk of
electricity losses occurs when approaching the rated wind speed. One can further realize a
bifurcation of electricity losses. While most losses decline for wind speeds greater than 10 m/s,
there are a few observations that increase linearly with wind speed and are capped at the rated
power capacity. The latter losses represent faults that occur when the turbine is partly or
completely out of operation. In Fig. 4, observations with and without an error in the status code
are represented by grey and black dots, respectively. Fig. 4(b) shows that there is variability in
electricity losses for different air densities. Note that for air densities beyond 1.3 kg/m³ only
observations occur that are affected by errors. These errors result from ice alerts caused by low
temperatures, which correspond to high air density.

Fig. 4. (a) Power losses and wind speed; (b) Power losses and air density;
observations with (gray) and without (black) turbine errors
However, as mentioned for Fig. 4(a), it is also the case for Fig. 4(b) that the frequency of
production losses is not visible because of overlapping data points. To illustrate this more
clearly, Fig. 5 depicts the sum of the losses calculated in correspondence of wind speeds at 0.1
m/s intervals. Thus, Fig. 5 accounts for the severity and frequency of losses occurring at different
wind speeds. It can be seen that the highest losses occur for wind speeds between 4 and 9 m/s,
whereas for lower and higher wind speeds, cumulative losses are much smaller. This reflects the
distribution of wind speeds (Fig. 1(b)).

13
Fig. 5. (a) Cumulative power losses against wind speed

Fig. 5. (b) Cumulative power losses against wind speed for different turbine error groups

14
The quantitative analysis of electricity losses in Table 3 shows that an average loss in a 10-
minute interval is 191 kW, which is associated with an average production of approximately
507 kW. This means that average losses amount to 27% of the amount of electricity that could be
produced. If one aggregates this finding over a year, these losses amount to 9,930 MW per
turbine, which translates to 1,655 MWh. Multiplied by an average spot price of 35.27 €/MWh
attainable on the German electricity spot market, this value results in a yearly loss of 58,372 €
per turbine.

Table 3: Total loss and number of observations for different error groups
Observations % FDH Loss (kW) % FDH Loss/
observation
Icing error 15,077 1.52 7,670,313 4.06 508.74
Maintenance error 5,511 0.56 1,379,068 0.73 250.24
Machine error 9,775 0.99 3,141,902 1.66 321.42
All Errors 30,363 3.07 12,191,283 6.46 401.52
No Error 958,812 96.93 176,525,730 93.54 184.11
Total 989,175 100.00 188,717,013 100.00 190.78
Total FDH potential
690,683,462
production

Table 3 separates losses for the three different error groups. Not surprisingly, in case of an error,
the average loss is much higher than without an error (401.52 kW compared to 184.11 kW).
However, only 6.46% of the total losses are caused by errors because they occur for only 3.07%
of the observations. Among the three error groups, ice has the largest share, with 4.06% of the
total loss. The average loss in case of an icing error equals 508.74 kW, which is very similar to
the overall mean of electricity produced (507.46 kW, Table 1). The average loss of 250.24 kW in
case of a maintenance error is much lower and is not far from the average loss without an error
(184.11 kW). This indicates that maintenance, if possible, should be conducted in periods where
few losses are expected, i.e., during light wind conditions (Fig. 5(b)). The average loss of
321.42 kW in case of a machine error is higher than losses resulting from a maintenance error,
but is lower than losses from icing.

The results of the second stage regressions are presented in Table 4. We ran three models of
different complexity. Model 1 is a parsimonious model that includes the variability of wind
speed and direction, error dummies, and turbine dummies as regressors. All coefficients are
highly significant which is not surprising given the large number of observations. The coefficient
of the wind speed range is negative. This can be explained by the concavity of some parts of the
production frontier: If the wind speed is constantly at or below the cut-in speed, no electricity
will be generated. On the other hand, a mean preserving spread of the wind speed will result in
some positive output during the observed 10-minute interval and thus reduce observed losses.
The coefficient of the wind range should be interpreted in conjunction with the effect of a change

15
of the rotor speed, since the latter is a response to the former and both variables are not
independent. We find that a higher change of the rotor speed over time is positively related to
electricity losses, which reflects the physical energy required to overcome the inertia of the rotor
blades. Surprisingly, adjustments of the nacelle position have an opposite effect, i.e., they reduce
production losses. As expected, all error dummies interacted with wind speed are positively
related to power losses and the magnitude of the coefficients confirms the earlier analysis, i.e.,
icing errors have the highest impact among the three considered error types. Turbine dummy
variables have a positive sign, which indicates a positive average loss depending on turbine
specific wind conditions and the position of the turbine in the park.

The fit of the predicted losses to the observed losses in Model 1, calculated via Pearson’s 𝜌2, is
modest (𝜌2 = 0.249). Though the fit or predictive power is not the key issue in second stage
regression analyses of technical efficiency, we modify the base regression model to attain a
better fit to the data. There is a controversial discussion on whether input factors that are used in
the nonparametric estimation of the production frontier should also enter the second stage
regression (Simar and Wilson, 2007; Kneip et al., 2014). This amounts to the violation of a
separability condition on the production technology. Disregarding these potential theoretical
flaws, we include wind speed in the set of regressors in Model 2. This can be justified by the fact
that potential power output varies with wind speed and thus the impact on wind electricity losses
is different depending on the speed. In fact, the sign of the wind speed is positive while the signs
of all other variables remain the same as in the base model. The inclusion of this variable
increases 𝜌2 considerably, from 0.249 to 0.41.

Inspection of Fig. 4(a) suggests that the impact of factors on realized production losses depends
on whether the wind speed is above or below the rated wind speed. To distinguish these two
regimes, in Model 3 we interact all variables with a dummy variable (DR) indicating a wind
speed at or above the rated wind speed of 11.5 m/s.7 Separating these two wind regimes increases
𝜌2 further to 0.63. Again, the sign of the coefficients remain the same as before, apart from the
rotor speed variable which is now negative. The coefficients associated with the dummy variable
for the high wind speed regime capture the difference between the regression coefficients for
wind speed above the rated speed relative to the coefficients below the rated wind speed. This
difference is positive for wind variability. Varying wind speeds around a high mean value
typically represents gusty and turbulent wind conditions, which may disturb power production,
e.g., via storm control, while maximal capacity cannot be further increased. For the error
variables, this means that their impact on power losses is even stronger than in the low wind
regime. The effect of the wind speed becomes negative under the high wind regime. This is
plausible since losses are lower for increasing wind speed in the high wind regime. Comparing
the turbine dummies reveals considerable differences in productivity at different locations. The

7
This wind speed is the empirically recovered rated wind speed in our sample, that is the first occurrence of an
average production of 2,365 kW in a 10 minutes interval is at 11.5 m/s. A higher wind speed than rated wind speed
occurs in approximately 5% of the observations.

16
best performing turbine in our sample (A1) is located in a wind park in a position that is free
from obstacles in most recurring wind directions, such as other wind turbines or hills. None of
the other turbines in this study have the same advantage.

Table 4: Second stage regression results

Dependent variable: Truncated regression model

Electricity loss (kW) Model 1 Model 2 Model 3
Wind range -1.101 *** -11.792 *** -0.332 ***
3
Nacelle position change(√|𝑥 |) -12.361 *** -8.549 *** -6.817 ***
Icing error*wind speed 76.346 *** 90.358 *** 82.119 ***
Maintenance error*wind speed 25.436 *** 45.661 *** 38.673 ***
Machine error*wind speed 58.217 *** 72.398 *** 58.849 ***
Rotor speed difference 5.124 *** 2.454 *** -0.589 ***
Wind speed 16.568 *** 34.923 ***
Wind range *DR 12.506 ***
Nacelle position change *DR 75.149 ***
Icing error*Wind speed*DR 154.512 ***
Maint. error*Wind speed*DR 64.504 ***
Machine error*Wind speed*DR 75.706 ***
Rot. speed diff. *DR 14.978 **
Wind speed*DR -49.938 ***
A1 434.145 *** 287.684 *** 124.344 ***
B1 437.149 *** 308.448 *** 140.722 ***
B2 431.775 *** 305.745 *** 139.863 ***
B3 430.911 *** 309.061 *** 147.127 ***
B4 434.334 *** 299.450 *** 129.991 ***
B5 451.763 *** 308.846 *** 124.861 ***
B6 438.381 *** 301.511 *** 131.402 ***
B7 429.294 *** 305.203 *** 144.920 ***
C1 424.015 *** 296.618 *** 139.792 ***
C2 420.881 *** 293.061 *** 138.054 ***
C3 423.832 *** 301.221 *** 145.960 ***
C4 428.083 *** 301.607 *** 142.611 ***
C5 417.072 *** 286.462 *** 136.712 ***
C6 420.409 *** 289.573 *** 134.045 ***
D1 422.769 *** 289.764 *** 133.259 ***
D2 419.365 *** 291.877 *** 137.401 ***
D3 431.860 *** 290.234 *** 124.822 ***
D4 419.091 *** 294.521 *** 136.140 ***
D5 411.033 *** 283.066 *** 138.753 ***
Pearson’s 𝜌 2
0.249 0.409 0.631
** and *** denote statistical significance at the 5 and 1 percent level, respectively.

17
5 Conclusions

This article analyzes the productivity and efficiency of wind electricity generation under real
world conditions. Based on a sample of 19 wind turbines located in four wind parks in Germany,
we calculate an efficient production frontier that represents the maximum producible electricity
at a certain level of wind speed and air density. This method results in a production frontier that
is similar to a power curve. With this frontier, we can quantify electricity losses that have been
realized under various wind conditions compared with the efficient frontier benchmark. By
construction, the production frontier dominates the standard power curve since the latter
represents average production under ideal conditions. We find that the average difference
between the frontier and the standard power curve is 183 kW. Thus, the standard power curve
can be regarded as a conservative estimate of electricity output, conditional on exogenous wind
input.

Our results show that production inefficiencies amount to a loss of 27% of the potential amount
of producible electricity. In a subsequent step, we decompose these production losses. It is
noteworthy that turbine errors are responsible for only 6% of the production losses, though they
often cause a complete stop in production. The reason is that only 3% of the observations in our
sample were affected by errors. Among the losses caused by turbine errors, icing had the highest
impact. We employ a regression model to explain the occurrence of production losses in greater
detail. Other than turbine errors, we find that changing wind conditions, i.e., variations in wind
speed and direction, affect the efficiency of electricity production. Moreover, turbine specific
effects exist which can likely be traced to the position of a turbine within a wind park.

From a methodological perspective, to the best of our knowledge, this paper represents the first
empirical estimation of efficiency with high frequency wind electricity production data using
non-convex analysis methods. Moreover, it is the first estimation of bias-corrected explanations
of efficiency in production by adapting the linear regression convergence results in Kneip et al.
(2014) to a truncated regression.

It is not straightforward to draw immediate managerial conclusions from the empirical findings
since most factors (air density, wind speed and variability, temperature) are not controllable.
Weather conditions are entirely exogenous and stochastic. Nevertheless, it is important to
understand how vulnerable wind electricity production is under real world conditions. This
analysis emphasizes that weather conditions should be carefully inspected when making the
decision of where to locate a wind turbine. Moreover, our analysis is helpful to assess the trade-
offs between the benefits from larger distances between turbines within a wind park and higher
costs for land acquisition to build larger wind parks. Finally, our results highlight the gains
arising from flexible adjustments of wind turbines to changing weather conditions. Technical
progress, such as anti-icing and de-icing systems, aims to increase this flexibility. This study
shows that even if such systems are costly, they could prevent important losses in the medium to

18
long-run. Since our analysis considers only one particular turbine type we cannot draw
conclusions on the impact of different technologies on the efficiency of wind energy production.
A comparison of different wind power technologies is recommended as a subject of further
research.

6 References

Carvalho, A., González, M.C., Costa, P., Martins, A. (2009): Issues on Performance of
Wind Systems derived from Exploitation Data. Industrial Electronics, 2009. IECON '09.
35th Annual Conference of IEEE: 3599-3604.

Deprins, D., Simar, L., Tulkens, H. (1984): Measuring Labor Efficiency in Post Offices. In
M. Marchand, P. Pestieau and H. Tulkens (eds.), The Performance of Public Enterprises:
Concepts and Measurements. Amsterdam: North Holland.

Gottschall J., Peinke J. (2008): How to improve the estimation of power curves for wind
turbines. Environmental Research Letters 3(1): 015005 (7pp).

Gunturu U. B., Schlosser, C. A. (2012): Characterization of wind power resource in the

United States. Atmospheric Chemistry and Physics, 12: 9687-9702.

Hennessey Jr., J. (1977): Some Aspects of Wind Power Statistics. Journal of Applied
Meteorology 16: 119-128.

Homola, M. C., Byström, J., Nicklasson, P. J., Sundsbø, P. A. (2009): An improved method
for wind power estimation. Working Paper, Narvik University College, Norway, 1-13.

Hughes, G. (2012): The Performance of Wind Farms in the United Kingdom and Denmark.
Renewable Energy Foundation, London.

Iglesias, G., Castellanos, P., Seijas, A., (2010). Measurement of productive efficiency with
frontier methods: A case study for wind farms. Energy Economics 32 (5): 1199–1208.

Iribarren, D., Vázquez-Rowe, I., Rugani, B., Benetto, E. (2014): On the feasibility of using
emergy analysis as a source of benchmarking criteria through data envelopment
analysis: A case study for wind energy. Energy 67: 527-537.

Ilinca, A., (2011). Analysis and Mitigation of Icing Effects on Wind Turbines, Wind
Turbines, Dr. Ibrahim AlBahadly (Ed.), ISBN: 978-953-307-221-0, InTech, Available
from: http://www.intechopen.com/books/windturbines/analysis-and-mitigation-of-icing-
effects-on-wind-turbines.

19
Kneip, A., Simar, L., Wilson, P. W. (2014): When Bias Kills the Variance: Central Limit
Theorems for DEA and FDH Efficiency Scores. Econometric Theory, FirstView
http://dx.doi.org/10.1017/S0266466614000413.

Kusiak, A., Verma, A., Wei, X. (2012): Wind turbine capacity frontier from SCADA. Wind
Systems Magazine. September 2012: 36-39.

Kusiak, A., Zhang, Z., Verma, A. (2013): Prediction, operations, and condition monitoring
in wind energy. Energy 60: 1-12.

Lucchesi, R. (2012). File specification for MERRA products (version 2.3). GMAO Office
Note No. 1. http://gmao.gsfc.nasa.gov/pubs/docs/Lucchesi528.pdf.

Rienecker, M. M., Suarez, M. J., Gelaro, R., Todling, R., Bacmeister, J., Liu, E., Bosilovich,
M. G., Schubert, S. D., Takacs, L., Kim, G.-K., Bloom, S., Chen, J., Collins, D., Conaty,
A., da Silva, A., Gu, W., Joiner, J., Koster, R. D., Lucchesi, R., Molod, A., Owens, T.,
Pawson, S., Pegion, P., Redder, C. R., Reichle, R., Robertson, F. R., Ruddick, A. G.,
Sienkiewicz, M., Woollen, J. (2011). MERRA: NASA’s Modern-Era Retrospective
Analysis for Research and Applications. Journal of Climate 24(14): 3624-3648.

Simar, L., Wilson. P. W. (2007): Estimation and inference in two-sage semi-parametric

models of production processes. Journal of Econometrics 136: 31-64.

Staffell, I., Green, R. (2014): How does wind farm performance decline with age?,
Renewable Energy, 66: 775-786.

Wächter, M., Gottschall, J., Rettenmeier, A., Peinke, J. (2009): Power curve estimation
using LIDAR measurements. In: Proceedings of the EWEA European Wind Energy
Conference, Marseille, France, 16–19 March 2009.Vol.1: 4473-4477.

20
7 Appendix

Table A.1: Summary statistics of electric power production of individual turbines

Mean Standard Minimum Maximum

deviation
Produced power (kW) 507.46 622.76 0.00 2,365.00
By turbine:
A1 659.69 699.27 0.00 2,365.00
B1 406.14 496.44 0.00 2,365.00
B2 465.19 554.64 0.00 2,365.00
B3 438.52 541.08 0.00 2,365.00
B4 445.77 534.40 0.00 2,365.00
B5 455.51 540.41 0.00 2,365.00
B6 451.54 542.09 0.00 2,365.00
B7 460.88 553.64 0.00 2,365.00
C1 540.56 645.49 0.00 2,365.00
C2 542.04 659.38 0.00 2,365.00
C3 516.36 641.55 0.00 2,365.00
C4 532.76 632.25 0.00 2,365.00
C5 604.09 709.68 0.00 2,365.00
C6 553.62 666.26 0.00 2,365.00
D1 534.79 647.18 0.00 2,365.00
D2 504.27 643.73 0.00 2,364.00
D3 462.71 621.11 0.00 2,364.00
D4 485.51 644.01 0.00 2,364.00
D5 582.42 732.41 0.00 2,365.00
Number of observations 989,175

21