AbstractWind power output is always uncertain but, in a

sufficiently long time interval, the output exhibits statistical behavior

that is meaningful enough to be characterized by probability
distribution. The aim of this paper is to develop a model for
probabilistic wind power generation. In particular, we successfully
derive the analytical expression and statistics up to the fourth order of
the wind power density function. The work also extends the
modeling of wind power output up to a regional scale by Gram-
Charlier series. Model results are checked by empirical power data
and Monte Carlo simulation. This paper discusses some applications
of the wind power statistics such as probabilistic production costing
and reliability evaluation in power system literature.

Keywords Gram-Charlier series, production cost, reliability,
wind power cumulant/moment
I. INTRODUCTION
LECTRICAL energy is increasingly produced from
renewable sources worldwide for a number of reasons such
as environmental concern, elevating fossil fuel prices, political
motive, etc. Wind power is one of the most popular forms of
renewable energy. Yet wind power is intermittent and virtually
uncontrollable. Its large-scale deployment would influence
power system in unprecedented ways. High penetration wind
power poses a need of refinement to a number of existing
methodologies for power system analysis.
Power system analysis has a lot to do with probabilistic
evaluation, notably, production costing and reliability. In order
to incorporate wind power generation into existing analytical
framework, probabilistic wind power model is highly desirable.
Such model shall represent wind power generator as a multi-
state (capacity) unit. Early attempt did not consider failure and
repair characteristics of wind turbine [1]. It was improved to
incorporate wind turbine availability by discrete convolution
[2][3]. In [4], wind speed is modeled by Markov chain, then
power curve and availability of individual wind turbines are all
taken into account to form the probability distribution of wind
farm output capacity. Building the wind power generation
model using the Markov method theoretically requires a
stochastic process for the wind speed. From a different

This work is supported by Research Grant Council, Hong Kong SAR,
under grant HKU7186/08E. The first author is also supported by the CLP
Fellowship in Electrical Engineering, awarded by the CLP Holdings Limited.
Henry Cheng, Yunhe Hou and Felix Wu are with the Department of
Electrical and Electronics Engineering, The University of Hong Kong, Hong
Kong, China (email: mkcheng@eee.hku.hk; yhhou@eee.hku.hk;
ffwu@eee.hku.hk)

perspective, wind energy production in a sufficiently long time
interval shall be able to be statistically characterized if the
wind speed distribution and power curve representation are
reasonably accurate. Say, in a year, wind power generation
should follow certain probability density function (PDF).
Simulated wind power PDF was first noted in [5].
Mathematically, analytical expression of the wind power PDF
based on wind speed distribution and linearized power curve
can be derived; reference [6] and [7] works on Weibull and
Rayleigh wind speed respectively, and both are able to provide
first order statistic (mean) of the wind power PDF. More
complicated works based on Weibull wind speed and
nonlinear power curve simultaneously are also available [8][9].
Consideration of wind directions in the wind power
distribution is reported in [10].
The objective of this work is also to build a probabilistic
wind power model. Polished from our earlier paper [11], this
paper aligns with previous references, but works in a
simplified domain that permits relatively straightforward
analytical derivation of higher order statistics of the wind
power PDF. The wind power statistics have important
applications in conjunction with production costing and
reliability evaluation. Production costing of conventional
generators can be performed by cumulant method [12][13].
Those higher order statistics can supplement the cumulant
method to determine expected cost of the conventional
generation net of wind power. The wind power statistics can
also be utilized in reliability evaluation, which is similar to the
handling of generators forced outage rate by cumulant method
[14][15].
Previous works tend to be confined by modeling the output
of single wind turbine or wind farm, but are not readily
generalized to wind power output in regional scale, which is
needed for production costing and reliability evaluation in a
national perspective. The wind power statistics make the
modeling of regional wind power possible by furnishing
analytical approximation of the total wind power PDF using
Gram-Charlier series. If wind power output is treated as a
random variable, then any total regional wind power is
obtained by recursively adding all individual distributions.
Correlations due to intra-farm and inter-farm proximity are all
efficiently rendered by the correlated cumulant method [12].
The resultant regional wind power distribution is anticipated to
be bell-shaped suggested by the Central Limit Theorem.
This paper is organized as follow. Section II offers the
proposed wind power model, with most of its derivation kept
in the Appendix. Section III highlights the data sets of wind
Probabilistic wind power generation model:
Derivation and applications
Henry Cheng, Yunhe Hou, and Felix Wu
E
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
17
speed and power. Pre-processing of data is carefully explained
in Section IV. Section V of this paper is a collection of
simulated and empirical results.
II. IDEAL WIND POWER MODEL
Most probabilistic wind generation models start with
discussion on the wind speed distribution. Several distributions
other than Gaussian had been early suggested [16]. Among
them, the Weibull and its special case Rayleigh, have generally
been recognized and are employed in engineering literature
[17][18]. On the other hand, the wind turbine power curve is
often made linear for simplicity. A generic wind generation
model based on those simplifying assumptions is always nice,
yet ideal. It may not be able to address a number of issues such
as empirical distribution mismatch, power curve actual
performance, wake effect when scaling up power from single
turbine to wind farm, and some minor problems like wind
direction and turbine outage.
Aerodynamic principle states that the power in airflow
follows a cubic relationship with the free wind speed w:
3
1
2
air
P Aw = (1)
where is the air density and A is the area swept by the rotor.
Not all of this power is extractable by a wind turbine as
mechanical power, apart from the power loss during
conversion to electrical power. In fact if the turbine output
power P
wt
is measured and divided by P
air
, the ratio is referred
as power coefficient C
P
:
wt
P
air
P
C
P
=
(2)
It is observed that no matter how the wind turbine is designed,
there is a maximum value for C
P
, known as the Betz Limit and
C
PMax
equals 0.593. In other words one can never extract more
than 60% of the power in airflow. Furthermore, power
coefficient is a nonlinear function of tip speed ratio [19], in
which is defined as
r
w

=
(3)
where r is the radius of the blade and is the angular speed of
the rotor. To achieve the maximum power coefficient, the tip
speed ratio has to remain constant at a particular value. It
implies that the desirable value occurs at a particular wind
speed for fixed speed wind generator. For variable speed wind
generator its rotational speed can be adjusted to track the wind
speed in order to operate the wind turbine at optimal tip speed
ratio, hence preserving maximum power coefficient [20].
Knowing that the power in airflow follows a cubic
relationship with free wind speed, the ascending segment of
the power curve should better be modelled as a curve.
However, it does not necessarily mean the output power is
proportional to w
3
because the actual performance is
complicated by the power coefficient. The power curve should
theoretically be represented by general nonlinear function. So
in order for analytical formulation to be carried on, this work
proposes to break down the ascending segment into many
smaller linear pieces, as represented by the following
expression:
1 1 1
~
2 2 1 2
1
0 0
in out
in
n n n r
r out m
w or w
w w
a w b w
w w
a w b w
w w
g
a w b w
w w
w g
w w

+ <

+ <

+ < <

<

(4)
where
1 0 1 0 1
1 1
1 1
2 2 2 1 1 1
2 2
2 1 2 1
1 1 1
1 1
,
,
,
in
in in
r m m n n n
n n
r n r n
w g g g g w
a b
w w w w
g g g w g w
a b
w w w w
g g g w g w
a b
w w w w



= =


= =


= =

,
and g
0
=0, g
1
, g
2
, , g
n-1
, g
m
correspond to the wind turbine
output at w
in
, w
1
, w
2
, , w
n-1
, w
r
respectively.
~
g is the random
wind power, defined by constants a
i
and b
i
{ [1, ]} i n , g
m
is
the maximum net output, and w
in
, w
r
and w
out
are the cut-in,
rated and cut-out wind speed respectively. Without loss of
generality, consider breaking the ascending segment into only
two segments, with (w
1
, g
1
) a knee point freely chosen on the
curve part.
On the other hand, the probability density function (PDF)
and cumulative distribution function (CDF) of Rayleigh
distribution of wind speed
~
w are respectively
2
~
( )
2
2
( ) 0
w
w
w
w w f
e

= (5)
and
2 2
~
( ) ( )
2
0
2
{ } 1
w
v w
v
P w w dv
e e

= =

(6)
where is the scale parameter. By integration with respect to
each wind speed partition, CDF of the wind power is
determined. Upon differentiating the CDF, the wind power
PDF
~
( )
g
f x is obtained as follow.
2 2
2 1
1
~
2 2
2
2 2
( ) ( )
( )
1
2 2
1 1
( ) 1
2
2 2
2
( ) ( )
(1 ) ( ) 0
2( )
0
( )
2( )
( ) ( )
in out
r out
w w
x b
a
x b
g
m
a
w w
m m
e e x x
x b
e
x g a
f x
g x g
x b
e
a
e e x g x g

+ =

<

=

< <

(7)
where (x) is a delta function to maintain the derivative-
integral relation between CDF and PDF.
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
18
Based on the analytical wind power PDF, it is tempting to
ask for its statistics. Derivation of formulae of higher order
statistics is tedious and lengthy; therefore results of the four
statistics, i.e. mean, variance, skewness and kurtosis are left in
the Appendix. Effectively, the four statistics are cumulants of
wind power PDF and they complete the statistical
characterization of the wind power generation model.

III. WIND SPEED AND POWER DATA
Overall this work is based on two sources of hourly wind
data. The first one is a multi-decade, speed-only database of
tens of Dutch locations [21]. The second one [22] is less
extensive, but with corresponding output from a real wind
turbine, which is most valuable. Brief descriptions of them are
as follow.

A. Royal Netherlands Meteorological Institute

There are more than 50 measuring stations, onshore or
offshore, each has a name and ID. Every station measures its
hourly wind speed, with figures of measuring height and
roughness length. Data entry may be void (missing) or
negative (faulty). Out of all 31 stations are chosen that are
most complete in the past 19 years (1991-2009). Void and
negative data entries are assigned a very small positive value.
All 29
th
Feb in leap year are ignored for easy programming.
Effectively, each location has 8,760 x 19 equals 166,440 hour
wind speed.

B. Vermont Small-scale Wind Energy Demonstration
Program

In this dataset, wind speed and corresponding output of
medium scale wind turbine (nominal 10kW) are available from
a few sites. Data are recorded at hourly intervals. Within each
hour, average wind speed and total energy production are
known. Maximum and minimum wind speed and output power
per hour are also given. It provides better insight because the
average wind speed of an hour is weak to show any intra-hour
highs and lows. Most importantly, those speed-power pairs can
be used to determine empirically the wind turbine power
curve. Separately, specifications of the actual wind turbine are
found in [23].
IV. DATA PRE-PROCESSING
A. Wind speed measuring height

Normally, wind speed is measured at a convenient height
above ground. It has to be interpolated to wind turbine hub
level for power conversion. In general wind profile follows a
logarithmic relation with its height above ground [25]:
0
ln( )
h
h
w
z
(8)
where h is the elevation, w
h
is the wind speed at the elevation
concerned and z
0
is so-called the roughness length. Roughness
length indicates surface friction and is defined as the height
drops to where the mean wind speed becomes zero. It is
necessary to up-scale measured wind speed to the hub level.
All Dutch wind speeds in section III. A are scaled to 100m
high.

B. Wind speed partitions and the parameter lambda

Cut-in, rated and cut-out speeds of a wind turbine are
designed to suit the prospective wind regime where it is
located. The rule of thumb of design is based on the annual
average wind speed
_
w onsite: w
in
equals 0.6
_
w; w
r
equals 1.5-
1.75
_
w and w
out
equals 3
_
w [20]. Next, value of the scale
parameter is needed for generating Rayleigh random variates.
Such value can be estimated from a set of historical wind
speed data. Alternatively, if the long-term average wind speed
is known, the lambda could be determined passively by the
following equation
_
2
w

= , (9)
which relates and the Rayleigh mean.

C. Power curve

Power curve measures the conversion ability of a wind
turbine. A simplified power curve consists of linear segments
is described in (4). In general the ramp segment should rise
nonlinearly because aerodynamic principle states that the
power in airflow follows a cubic relationship with free wind
speed. Meanwhile, its actual performance depends on
generator type (fixed or variable speed) and control method
(pitch or stall controlled). In practice, power curve data are
empirical and given by wind turbine manufacturer. It consists
of power values (and losses) across a range of operating wind
speed, e.g. in pp. 68 of [18]. Hence an empirical power curve
may not have a proper mathematical form.
Even wind speeds at different hours are the same, the
energies produced could be different. Arguably, the speed is
only average of an hour and intra-hour fluctuations could be
significantly different. Wind turbine performance varies with
wind gust, turbulence and wind directions to different extents.
It is reported in commercial software WAsP [26] that wind
speed within a particular directional sector does not normally
show Weibull distribution, fortunately, improvement made to
the estimation of power output by considering directional
effect is little in most cases.
The approach here is to run regression on numerous real
speed-power data to determine an empirical power curve [27].
The empirical power curve has theoretically captured any long
run site characteristic, e.g. terrain, wind direction etc. Prior to
the regression process, it is necessary to have a hypothesis of
what the underlying function of the power curve is. And since
a power curve can be reasonably partitioned into an ascending
segment and a plateau, it is logical to have two underlying
functions: cubic function and constant function of speed
respectively. Then we use linear regression subroutine built in
Matlab to obtain the fitted curves.
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
19
We make reference to the data source mentioned in section
III. B. The site picked for demonstration is called RicLin. The
span of data is typically a year of hourly speed-power data
points, with omission of faulty measurements. Fig. 1 shows the
power curve under trial regression.

Fig. 1 Empirical power curve determined by regressing real data

D. Scaling and wake effect

When many wind turbines are assembled in an array,
downstream wind speed is reduced by upstream wind turbines.
This is called wake effect. Hence the total power of wind farm
is less than the proportional scaling of single wind turbine
power. To account for the speed deficit, a wake effect model
consolidated in [28] is applied. Let w
x
denote the reduced wind
speed x metre behind the upstream turbine, w
0
denote the
undisturbed wind speed, then
2
0
[1 (1 1 )( ) ]
x T
r
w w C
r kx
=
+
, (10)
where r is the upstream turbine rotor radius, k is the so-called
wake decay constant and C
T
is the thrust coefficient. k and C
T

depends on wind farm terrain and wind turbine type
respectively. Downwind distance x is normally 12r to 24r.
Though waked wind speed is different from the undisturbed
one in magnitude, it still demonstrates a reasonable shaped
Rayleigh distribution. Fig. 2 shows a sample waked wind
speed profile for wind turbine with typical C
T
value [26].
We believe existing methods of how to incorporate wake
effect into wind farm expected output could be improved. Our
approach is conceptually very simple. If wind power is treated
as a random variable, then any total wind power is obtained by
recursively adding all individual output distributions. There is
no need to distinguish wind turbine out of a wind farm in the
summation process for the reason that, given any downwind
turbine speed reduced by wake effect, the waked speed profile
could be modeled by a new, separate distribution. Correlations
due to intra-farm and inter-farm proximity are rendered by the
correlated cumulant method.

Fig. 2 Waked wind speed density function (b) compared with its original
Rayleigh source (a)

V. MODEL RESULTS
This paper does not preclude the assumptions behind the
generic wind generation model, rather, it investigates how
much different the model prediction is compared with
empirical result. Empirical result is based on real data of wind
speed and power. Availability of comprehensive real data has
been a difficult task in course of the research. In particular,
wind speed and corresponding power data of large wind farm
are very difficult to find in open source. This work makes use
of the data as mentioned in Section III. We tested locations
for their analytical wind power distributions, which fit the
same set of hourly empirical wind data satisfactorily. Results
followed are divided into three parts.

A. Verification of the statistical formulae by Monte Carlo
simulation

The formulae of four wind power statistics are checked by
Monte Carlo simulation. Result variations due to power curve
(three or four segments) and simulation runs are demonstrated
in TABLE I. It can be seen that deviations between
corresponding analytic and simulated values for the case of 3-
segment power curve is larger than those of 4-segment power
curve, indicating the latter is a more accurate model.
Computation is implemented in Matlab 7.0 on a common
desktop computer with 2.66GHz Intel Core 2 processor
and 2GB RAM.



INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
20
TABLE I
Checking wind power statistics by Monte Carlo simulation
Rated power = 3000kW
g

skewness kurtosis
excess
3-segment
power
curve
Monte Carlo
10000 runs
1190.7 1131.9 0.4209 -1.3377
Analytic 1181.7 1127.7 0.4358 -1.316
Deviations
(%)
1.86 1.07 5.73 3.26
4-segment
power
curve
Monte Carlo
2500 runs
1069 1117.8 0.6963 -1.0321
Monte Carlo
10000 runs
1041 1106.2 0.7297 -0.964
Analytic 1053.1 1103.4 0.7094 -0.9861
Deviations
(%)
1.15 0.26 2.86 2.24

Wind power PDFs are also plotted for visual inspection.
Fig. 3 shows the PDF of the case of 3-segment (simple) power
curve, which is smooth in the middle. Fig. 4 shows the case of
4-segment (improved) power curve in which the PDF shows a
jump. Theoretically, the model accuracy increases as the
power curve is partitioned more.


Fig. 3 Wind power PDF built from simple power curve


Fig. 4 Wind power PDF built from improved power curve

B. Comparison between analytical and empirical wind
power density functions

The fact that analytical and simulated results in Section V.
A are matched only means the formulae are correctly derived,
it does not guarantee the same for analytical and any empirical
data. Analytical model uses only parameter of wind speed
distribution and simplified power curve to calculate wind
power. Empirical data are what actually be measured from the
wind turbine under the same speed distribution. Accuracy of
the analytical model lies on the fitness of distribution envelope
and precision of power curve. The key effort of this part is to
substantiate the proposed wind power model by benchmarking
with some real speed-power data. Annual data of the Harvest
Hill Farm, another site from the Vermont dataset, are used. Its
analytical and simulated wind power PDFs are presented in
Fig. 5(a) as control, the critical point is whether the analytical
PDF models the empirical power data well in Fig. 5(b).
Compared with Fig. 4, the empirical power distribution does
not show any frequency at the right end of the horizontal axis,
although the empirical power curve in Fig. 1 clearly shows
occurrence of high power output. It is because the empirical
power curve is made up of maximum values of speed and
power within an hour whereas the wind power PDF is derived
from hourly average values. It suggests that only persistently
large wind speed can generate high power output.

Fig. 5 Modeling empirical wind power by analytical PDF

The wind power distributions in terms of CDF for the case
of RicLin and Harvest Hill Farm are also plotted in Fig. 6 and
Fig. 7 respectively. It can be seen that the empirical and
analytical distributions for concurrent year data are reasonably
matched. More exhaustive empirical work would reinforce the
robustness of the result. In terms of chronological simulation,
[24] reported good tracking ability of its testing wind turbine
for varying wind speed condition and anticipated power output.
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
21

Fig. 6 Wind power distribution CDFs for RicLin (Oct 07 Sept 08)

Fig. 7 Wind power distribution CDFs for Harvest Hill Farm (Mar 07 Feb
08)

It is beyond the scope of this paper to tell any dependence
of inter-year wind speed, hence wind power distributions for
each site. The research would involve enormous amount of
empirical data, though availability of multi-year real wind
power may be difficult to obtain. Apparently, inter-year wind
speed distributions have distribution parameters in closed
range, suggested by some trial work on the Dutch wind speed
database, but assertion cannot be made until significant future
work is done. Perhaps the annual average is the attribute that is
the easiest to check for any long-term trend. We plot the
annual average wind power, converted from hourly wind speed
data of about two decades, for a Dutch location as in Fig. 8.


Fig. 8 Trend of multi-year annual average wind power for Valkenburg

C. Regional wind power distribution

An extended numerical example of this work is to
synthesize the probability distribution of regional wind power
output for production costing and reliability evaluation in large
scale. As worked out, individual wind turbine output is a
random variable characterized by probability distribution.
Mathematically, any total wind power is the convolution sum
of wind turbine random outputs. But obviously it is too
overwhelming if thousands of wind turbines, each with
numerous capacity states, are convoluted to yield the resultant
distribution. We seek to employ the Gram-Charlier series to
approximate the resultant wind power density function, but
soon the substantial correlation among wind speeds is realized
and the independence assumption of Gram-Charlier approach
is potentially violated. Fortunately, the correlated cumulant
method is readily applicable, which is the blueprint of this
section. The cumulant method requires only one-off
calculation of cumulants, and the value of analytical statistics
derived earlier comes in place. No matter how individual wind
turbines are different in terms of availabilities, ratings and
wind profiles, etc., each wind turbine possesses a set of
characteristic cumulants for the summation process.
From Fig. 3 to Fig. 5, it is shown that the power density
function of one wind turbine has concentration at zero output.
When many of such PDFs of different capacities are added
together, one would anticipate by Central Limit Theorem the
resultant density function to be bell-shaped. We verify this
belief by enumeration of real data. The resultant wind power
density function can be obtained by two ways. On one hand, it
is done by successive convolution of individual outputs, which
is rendered by Matlab subroutine. On the other hand, the
Gram-Charlier series gives an analytical approximation to the
resultant wind power PDF by processing the cumulants of all
individual outputs. The two approaches should match in
outcome when sufficient quantities of wind turbine are
grouped together. However, extensive real wind power data
could be difficult to obtain as an open source. For research
trial and illustration, we build a hypothetical scenario with
reference to real data as much as possible.
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
22
Denmark is used as the backdrop of our hypothetical
scenario. Denmark has over 3,000 MW wind power installed
capacity in 2008 [29], shared by slightly more than 5,000 wind
turbines in 2009, with breakdown by ratings as grouped in
TABLE II [30]. The ideal exposition of this numerical
example would be to synthesize a power density function of all
wind turbines using empirical wind power data. As a
compromise, wind power data are converted by wind speed,
and hourly wind speed records of tens of Dutch locations for
any single year are available. But obviously there are not
enough wind speed profiles compared with the number of
wind turbines, and by no means these wind speed profiles are
matched with the actual wind turbine locations. Therefore we
only conduct test example.
TABLE II
Wind turbine breakdown by capacities in Denmark 2009
Ratings (kW) 0-225 226-
499
500-
999
1,000+ Total
Numbers 1427 306 2596 749 5078

Specific formulations of the test example follow. Thirty one
wind speed profiles for a particular year from the Dutch
database are extracted. Each profile of hourly wind speed is
tagged with a wind turbine of randomised rating to enumerate
the respective annual wind power distribution. All these power
distributions exhibit randomness to various degrees, and are
mildly correlated due to wind speed correlation. It is
interesting and meaningful to check how closed to bell shape
the resultant density function would be when individual wind
power PDFs are convoluted successively. To visualize the
improvements of large number, Fig. 9 plots the wind power
PDFs generated by an occasion of six successive convolutions
stage by stage. It is demonstrated that the resultant density
function approaches bell shape when the number of sites
convoluted increases. The corresponding cumulative wind
power capacity increases along the horizontal axis.
Fig. 9 Successive convolution of individual wind turbine outputs

Apart from numerical convolution, the resultant wind power,
in terms of standardized density function, can be approximated
by the correlated cumulant method. Fig. 10 shows the
analytical approximation of the resultant wind power PDF to
the same setting as by convolution. The envelope is laid on a
histogram of simulation of a standard normal random variable
for visual checking. It could be seen that the analytical PDF
and the Monte Carlo simulation result are similar, but not yet
matched for the fact that only seven sets of cumulants are
available to synthesize the curve. The result is improved
considerably when the number of wind turbines increases, say
up to 31, as shown in Fig. 11. Then the analytical envelope is
very much alike a standard normal PDF.

Fig. 10 Standardized PDF of Gram-Charlier series of 7 variables


Fig. 11 Standardized PDF of Gram-Charlier series of 31 variables

As an attempt we have compared resultant wind power PDF
by numerical convolution and by analytical approximation.
Some possible reasons of mismatch are aware of. Apparently,
faulty wind speed data explain partially. The fact that only
cumulants up to the fourth order are used also leads to less
precision.
VI. CONCLUSION
This paper has provided a probabilistic wind power model
based on Rayleigh wind speed distribution and linearized
power curve. Formulae of wind power statistics up to the
fourth order are completely derived. The model is not only
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
23
checked by simulation, but also benchmarked with empirical
data, and shows satisfactory performance. The model can help
evaluate production costing and reliability with wind power.
VII. APPENDIX
Suppose f(x) is the wind power density function, its r
th

moment (about zero) and r
th
central moment are given
respectively as
( )
r
r
m x f x dx

and (VII.1)
1
( ) ( )
r
r
c x m f x dx

(VII.2)
In particular, m
1
is the mean and c
1
=0. For completeness we
may also define m
0
=c
0
=1. By binomial theorem, any r
th

moment can be expressed in terms of the r
th
and lower order
central moments [31]. Therefore,
1
0
r
j
r r j
j
r
m c m
j

=
| |
=
|
\

(VII.3)
or vice versa,
1
0
( )
r
j
r r j
j
r
c m m
j

=
| |
=
|
\

(VII.4)
They can be deduced interchangeably and it is more
convenient to start with the mean m
1
and then all higher order
moments can be generated. Consider (VI.1),
2 2
2 2 1 2
1
1 2
0 1
2 2
2 1
1
1
0 1
0 ( ) ( )
0
( ) ( )
1 2
2 2 2 2
1 2
( ) ( )
( )
1
2 2
1
( ) 0 (1 ) ( )
2( ) 2( )
( ) ( )
2( )
in out
m
r out
m
m
m
w w
r r
x b x b
g g
a a r r
g g
w w
g
r
m m
g
x b
g g
a r r
g g
x f x dx e e x dx
x b x b
x e dx x e dx
a a
g e e x g dx
x b
x e dx x
a

= +

+ +
+

= +

2 2
2
2 2
( )
2
2 2
2
( ) ( )
2( )
( )
r out
x b
a
w w
r
m
x b
e dx
a
g e e

+
(VII.5)
The first two terms are the same except with different
parameters for different power curve segments, implying that
the formula derived for one segment should work for others.
For simplicity, we abuse the notation of a and b by omitting
their subscripts. Let
x b
y
a

= , then
2
2
1
1
2 2
1
( )
2 2
2( ) 2 ( )
2 ( )
m
m
g b x b
g
r r y
a a
g b
g
a
k
r y
k
x b
x e dx y ya b e dy
a
y ya b e dy

= +
= +

(VII.6)
where the lower and upper limits of the integration domain are
denoted by k
1
and k
2
respectively for convenience (for the
segment g
0
to g
1
, let the limits be l
1
and l
2
respectively). For an
integral of the form
2
n y
y e dy

, denote it by
n
I . Let
2
1
, 2
2
n
y
y dh
g ye
dx

= = and consider integration by parts,

then recursively
2
1
2
1
2
2 2
n
y
n n
y n
I e I for n

= + (VII.7)
For n=1, it is observed that
2
1
1
2
y
I e

= (VII.8)
For n=0,
2 2 2
2 ( )
0
( )
u v r
R P
I e dudv e rdrd
+
= =

(VII.9)
After considering the integration domain,
2 2
2
2 1 2
1
1
1 1
[ ] ( e )
2 2
k y
k
k k
I e e

= = (VII.10)
2 2 2 2 2
1 1
0
0
0
k k
y y y
k k
I e dy e dy e dy

= = +

(VII.11)
Hence
0
I can be expressed in terms of error function erf(.):
0 2 1
{ ( ) ( )}
2
I erf k erf k

= (VII.12)
Although an error function does not have closed form solution,
one can expand the integrand
2
y
e

by Taylor series and

integrate term by term to obtain an approximation. In modern
computer programs, it is readily generated by numerical
integration.
By now we have obtained a recursive formula of any higher
order of the integral
n
I once
1
I and
0
I are generated. Let us
show the derivation up to
3
I for the calculation of the mean
and variance.

A. Mean

The mean of the random wind power is obtained by setting
r=1 in (VI.1). From (VI.6),

2 2
1
2 2 2 2
1 1
2 2
2
2 1 2
1
2 2
1 2
2
2 1
0
2 1 1
2 ( )
2 2
2 2
([ ] ) ( e )
{ ( ) ( )}
2
k
y
k
k k
y y
k k
k y
k
m
y ya b e dy
a y e dy b ye dy
a I bI
k k
a ye I b e
a
k k
erf k erf k g e g e

+
= +
= +
= + +
= +

(VII.13)
Adding the results of two segments together, the mean or first
order moment m
1
is
2 2
1 2
2 2
2 2
1 2
_
1
1
2 1 0 1
( ) ( )
2
2 1 1
( )
{ ( ) ( )}
2
{ ( ) ( )} ( )
2
r out w w
m m
x m xf x dx
a
l l
erf l erf l g e g e
a
k k
erf k erf k g e g e g e e

= =
= +
+ + +

(VII.14)

B. Variance
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
24

The variance of
~
x , or the second order central moment, is
given by (VI.2). Conveniently, from (VI.4) we can use
2 2
2 2 1
m
c m
= = (VII.15)
Therefore we need m
2
first. Setting r=2 in (VII.6),
2 2 2 2
1 1
2 2
2 2
1 1
2
1
2
2
2 2 2 3 2 2
2 2 2
3 2 1
2 2 2 2
1 0 1
2 2 2 2 2 2
1 1
2 2 2 2 2 2
2 0 2
2
1
2 (2 4 2 ) ( )
2 4 2
([ ] 2 ) 2 ([ ] ) 2
( 2 )
( 2 ) 2
{
k k
y y
k k
k k y y
k k
y e dy a y ab y b y e dy ya b
a I ab I b I
a y e I ab ye I b I
k
a a ab k b e
k
k
a a ab k b e ab I
k
g

= + + +
= + +
= + + + +
= + + +
+ + + +
= +

2 2
1 2
2 2 2
2 1
( ) } { ( ) } { ( ) ( )}
m
k k
a e a e ab erf k erf k g

+ +
(VII.16)
Adding the results of two segments together, then m
2
:
2 2
1 2
2 2
1 2
2 2
2
2 2 2 2
2 2 2 2 2 1 1
2 2 2 2
1 1 1 1 2 1 0 1
( ) ( )
2
{ ( ) } { ( ) } { ( ) ( )}
{ ( ) } { ( ) } { ( ) ( )}
( )
r out
m
w w
m
m
k k
a e a e a b erf k erf k g g
l l
a e a e a b erf l erf l g g
g e e






=
+ + +
+ + + +
+
(VII.17)
The formula of variance follows when m
1
and m
2
are known.

C. Skewness and Kurtosis

It becomes clear that any higher order moments or central
moments can be built for better characterization of the PDF.
Also eventually any orders of cumulant can be calculated by
(VI.5) and those cumulants could readily be used in the Gram-
Charlier series. Without showing the steps, the third and fourth
cumulants are stated as follow.
Third cumulant:
3 3
3
3 2 1 1
3 2
c
m m m m
=
= +
(VII.18)
where
2
1
2
2
2
1
2
2
2
3 2 2
3 2 1 2 1
2 2
3 2 2 2 2
2 2 2 2 2 1
3 2 2
1 0 1 0
2 2
3 2 2 2 1
1 1 1 1 1 2 1 1
( ) (
3
3
{ ( )}
2
3
{ ( )} 3 ( ) { ( ) ( )}
2 2 2
3
{ ( )}
2
3
{ ( )} 3 ( ) { ( ) ( )}
2 2 2
(
r out
m m
w w
m
k
m a g b e g
a
k
a g b e a b erf k erf k g
l
a g b e g
a
l
a g b e a b erf l erf l g
g e e

= + +
+ + + +
+ + +
+ + + +
+
2
)
)
(VII.19)
Fourth cumulant:
2
4 4 2
2 4 2 2
4 3 1 2 1 1 2 1
3
( 4 6 3 ) 3( )
c c
m m m m m m m m
=
= +
(VII.20)
where
2
1
2
2
2
1
2
2
4 2 2 2 2 2
4 1 2 1 2 1 2 2
4 2 2 2 2 2
2 2 2 2
3 3 3
2 2 2 2 2 1
4 2 2 2 2 2
0 1 0 1 0 1 1
4 2 2 2 2 2
1 1 1 1 1 1 1
1
{ 2( ) ( )}
{ 2( ) ( )}
(3 2 ) { ( ) ( )}
{ 2( ) ( )}
{ 2( ) ( )}
(3
m m m
k
m g a g b g a b e
k
g a g b g a b e
a b a b erf k erf k
l
g a g b g a b e
l
g a g b g a b e
a

= + + + +
+ + + +
+ +
+ + + + +
+ + + +
+
2 2
3 3 3
1 1 1 2 1
( ) ( )
4
2 ) { ( ) ( )}
( )
r out w w
m
b a b erf l erf l
g e e



+
+
(VII.21)
By definition, we can also find skewness and kurtosis excess
respectively as:
3
1
3
c

=
(VII.22)

4
2
4
3
c

=
(VII.23)
VIII. REFERENCES
[1] R. G. Deshmukh and R. Ramakumar, Reliability analysis of combined
wind-electric and conventional generation systems, Solar Energy, Vol.
28, No. 4, pp. 345-352, 1982.
[2] P. Giorsetto and K. F. Utsurogi, Development of a new procedure for
reliability modeling of wind turbine generators, IEEE Trans. PAS., Vol.
PAS-102, No. 1, pp. 134-143, Jan 1983.
[3] X. Wang, H. Dai, and R. J. Thomas, Reliability modeling of large wind
farms and electric utility interface system, IEEE Trans. PAS., Vol.
PAS-103, No. 3, pp. 569-575, Mar 1984.
[4] F. Castro Sayas and R. N. Allan, Generation availability assessment of
wind farms, IEE Proc.-Gen. Trans. Distrib., Vol. 134, No. 5, pp. 507-
518, Sept. 1996.
[5] G. Papaefthymiou, P. H. Schavemaker, L. van der Sluis, W. L. Kling, D.
Kurowicka and R. M. Cooke, Integration of stochastic generation in
power systems, International Journal of Electrical Power & Energy
Systems., Vol. 28, No. 9, pp. 655-667, 2006.
[6] Ciprian Nemes and Florin Munteanu, The wind energy system
performance overview: capacity factor vs. technical efficiency,
International Journal of Mathematical Models and Methods in Applied
Sciences, Vol. 5, No. 1, pp. 159-166, 2011.
[7] Hui Zhou, Yunhe Hou, Yaowu Wu, Haiqiong Yi, Chengxiong Mao, and
Gonggui Cheng, Analytical assessment of wind power generation asset
in restructured electricity industry, in Proc. the Universities Power
Engineering Conference, UPEC, 2007, pp. 1086-1092.
[8] M. H. Albadia and E. F. El-Saadanyb, New method for estimating CF
of pitched-regulated wind turbines, Electric Power Systems Research,
80, pp. 1182-1188, 2010.
[9] D. Villanueva and A. Feijoo, Wind power distributions: A review of
their applications, Renewable and Sustainable Energy Reviews, 14, pp.
1490-1495, 2010.
[10] Mario Ar Talaia, Analysis of wind power potential of a region of
Aveiro, Portugal, WSEAS Transactions on Environment and
Development, Vol. 6, No. 7, pp. 519-528, Jul 2010.
[11] Henry M. K. Cheng, Yunhe Hou, and Felix Wu, Probability
distribution of the output power of wind turbine, in Proc. The 16
th

International Conference on Electrical Engineering, Busan, Korea, 11-
14 Jul, 2010.
[12] J. P. Stremel, R. T. Jenkins, R. A. Babb, and W. D. Bayless,
Production costing using the cumulant method of representing the
equivalent load curve, IEEE Trans. on PAS., vol. PAS-99, no. 5, pp.
1947-1956, Sept/Oct 1980.
[13] Jeremy A. Bloom, Probabilistic production costing with dependent
generating sources, IEEE Trans. Power App. Syst., Vol. PAS-104, No.
8, pp. 2064-2071, Aug. 1985.
INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
25
[14] C. Singh and A. Lago-Conzalez, Reliability modeling of generation
systems including unconventional energy sources, IEEE Trans. PAS.,
Vol. PAS-104, No. 5, pp. 1049-1056, May 1985.
[15] C. Singh and Y. Kim, An efficient technique for reliability analysis of
power systems including time dependent sources, IEEE Trans. Power
Syst., Vol. 3, No. 3, pp. 1090-1096, Aug 1988.
[16] Barbara G. Brown, Richard W. Katz, and Allan H. Murphy, Time
series models to simulate and forecast wind speed and wind power,
Journal of Climate and Applied Meteorology, Vol. 23, pp. 1184-1195,
May 1984.
[17] L. L. Freris, Wind Energy Conversion Systems, NJ: Prentice-Hall, 1990.
[18] David A. Spera, Wind Turbine Technology, Fundamental Concepts of
Wind Turbine Engineering, NY: ASME Press, 1994.
[19] Lata Gidwani and Harpal Tiwari, Dynamic analysis of grid connected
wind turbine with a permanent magnet synchronous generator during
fault conditions, WSEAS Transactions on Power Systems, Vol. 5, No.
4, pp. 287-296, Oct 2010.
[20] Brendan Fox, Damian Flynn, Leslie Bryans, Nick Jenkins, David
Miborrow, Mark OMalley, Richard Watson, and Olimpo Anaya-Lara,
Wind Power Integration, Connection and System Operation Aspects,
London: IET Power and Energy Series, 2007.
[21] Royal Netherlands Meteorological Institute. [Online]. Available:
http://www.knmi.nl/klimatologie/onderzoeksgegevens/potentiele_wind/
[22] Vermont Small-scale Wind Energy Demonstration Program. [Online].
Available: http://www.vtwindprogram.org/
[23] Specifications of a Bergey xls wind turbine. . [Online]. Available:
http://www.windenergydirect.org/Windgenerators-bergy-xls10kw.php
[24] Jeff Huang, Zulati Litifu, and Ken Nagasaka, Installation of wind
power into a weak power system considering system fluctuation and
limitation, WSEAS Transactions on Environment and Development,
Vol. 4, No. 1, pp. 33-528, Jan 2008.
[25] Lars Landberg, A mathematical look at a physical power prediction
model, Wind Energy, 1, pp. 23-28, 1998.
[26] Wind Atlas Analysis and Application Program. [Online]. Available:
http://www.wasp.dk/index.htm
[27] Cameron W. Potter, Hugo A. Gil, and Jim McCaa, Wind power data
for grid integration studies, in Proc. 2007 IEEE PES General Meeting,
Tampa, 24-28 June 2007.
[28] Julija Matevosyan and Lennart Soder, Minimization of imbalance cost
trading wind power on the short-term power market, IEEE Trans. on
Power Systems, vol. 21, no.3, pp. 1396-1404, Aug 2006.
[29] World Wind Energy Report 2008. World Wind Energy Association,
Bonn, Germany. [Online]. Available:
http://www.wwindea.org/home/images/stories/worldwindenergyreport20
08_s.pdf
[30] Wind Turbines in Denmark. Danish Energy Agency, Amaliegade,
Copenhagen, Nov 2009.
[31] A. Stuart and J. K. Ord, Kendalls Advanced Theory of Statistics: Vol.
1. Distribution Theory, sixth edition, London: Edward Arnold, 1994.



Henry Cheng (M 05) received his B. Eng. degree (2003) from the
Department of Electrical and Electronic Engineering, and master degree in
economics (2006), both at the University of Hong Kong. He is currently a
Ph.D. candidate in the Center of Electrical Energy Systems at the same
university.

Yunhe Hou (M 07) received his B.E. (1999), M.E. (2002) and Ph.D.
(2005) degrees from the Huazhong University of Science and Technology,
China. He worked as postdoctoral research fellow at the Tsinghua University
from 2005 to 2007. He is now a research assistant professor with the
Department of Electrical and Electronic Engineering at the University of
Hong Kong.

Felix Wu (M74-F89) is the Philip Wong Wilson Wong Professor in
Electrical Engineering at the University of Hong Kong, Hong Kong, where he
served as Pro Vice Chancellor (Vice President) from 1997 to 2001. He is also
a Professor Emeritus at the University of California, Berkeley, where he has
been on the Faculty since 1974.


INTERNATIONAL JOURNAL OF ENERGY, Issue 2, Vol. 5, 2011
26