Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Matt Rosenzweig
Abstract
In this paper, we discuss some mathematical methods used to estimate the wind power resources of a
given location. We introduce the two-parameter Weibull distribution as a model for hourly average wind
speeds, prove an existence an existence and uniqueness theorem for the maximum likelihood estimates of
this distribution, and provide an algorithm for their computation. We derive a probability distribution
for the power output of a wind turbine with given cut-in, cut-o, and rated wind speeds and compute the
moments of this distribution, in addition to deriving an expression for the capacity factor as a functon
of these inputs. Lastly, we consider the problem of optimizing the choice of wind turbine for a given
location with known wind speed distribution.
Contents
1 Introduction 2
2 The Weibull Distribution 2
2.1 Denition and Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Likelihood Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.3 Computation of MLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Goodness of Fit 9
4 Single-Site Power Distribution 9
4.1 Ideal Wind Turbine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2 Power Distribution of Ideal Wind Turbine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.3 Turbine-Site Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5 Conclusion 13
A Newton-Raphson Method 14
B Example Graphs 15
1
1 Introduction
Before one decides to a build a wind farm at a given location, he or she needs to know what the distribution
of wind speeds is for the location in order to be able to estimate the amount of electricity the wind farm will
produce. In particular, if the distribution of the wind speed is such that only a small portion of the right
tail is above the cut-in speed for the model of wind turbine, then it would be physically and economically
nonsensical to build a windfarm at the location. Analogously, if the distribution of the wind speed is such that
only a small portion of the left tail is below the cut-o speed, then it would also be nonsensical, in addition
to potentially dangerous, to build a windfarm at the location. Given that dierent geographic locations can
have completely dierent wind speed proles, the distribution of wind speed is an inherently local question.
This paper is intended as an introduction to the use of the two-parameter Weibull distribution to model
single-site hourly average wind speeds (i.e. wind speeds at a given wind farm). Our goal is to write down
an expression for the probability distribution of the power produced by a wind turbine at a xed location,
so that the modeling problem reduces to collecting data to estimate the two parameters of the Weibull
distribution and running some algorithms and statistical tests with standard software (e.g. Matlab). This
reasoning a priori assumes that the Weibuill distribution is an appropriate choice, an assumption which has
support in the literature ([7], [5], [8]) and will not be challenged in this paper. Our focus will be abstract
and mathematical rather than concrete and empirical.
In section 2, we dene the Weibull probability distribution, compute its moments, and prove some other
basic properties. We then discuss the maximum likelihood estimate (MLE) technique for estimating the
unknown parameters of the Weibull distribution. We prove an existence and uniqueness theorem for the
MLE under mild hypotheses and give an algorithm based on the Newton-Raphson root-nding method for
computing the MLE. The mathematical details of the Newton-Raphson method can be found in appendix A.
For analysis of the performance of the MLE technique in comparison to the least-squares method and method
of moments in terms of minimizing mean-squared error (MSE), we refer the reader to [5]. We then use actual
wind speed data from Bahrain to compute the parameters and a give graphical sense of the goodness of t
of the Weibull distribution with estimated parameters for the given data. In section 3, we briey discuss
the general problem of testing the goodness of t of a null hypothesis distribution and provide references
for the reader interested in applying such tests when the parameters of the null hypothesis distribution are
estimated. In section 4, we derive a probability distribution for the eletric power output of a wind turbine
with given cut-in, cut-o, and rated wind speeds. We also compute the moments of this distribution and
derive an expression for the capacity factor of the wind turbine. Lastly, we consider the problem of matching
a wind turbines rated wind speed to the wind speed distribution of a given site.
There are two, not necessarily disjoint, intended audiences for this paper: quantitatively literate per-
sons with a general interest in applied mathematics and a desire to make informed decisions concerning
renewable energy policy and practitioners who need mathematical modeling to estimate the wind resources
as a particular site. The reader is assumed to have a knowledge of real analysis at the level of [15], basic
measure-theoretic probability, elementary statistics to understand all proofs. The general reasoning of this
paper can be followed without attention to all the mathematical details, but to ignore them is contrary to
the intent this paper: to give a rigorous brief exposition of some mathematical methods in wind speed and
power modeling. The reader is only assumed to have knowledge of basic physics.
For questions about notation or any arguments in the proofs presented, the author can be reached by
e-mail at matthew.rosenzweig@college.harvard.edu.
2 The Weibull Distribution
There appears to be a consensus in the literature that the Weibull distribution is a good probabilistic model of
wind speed at one location. In this section, we will introduce the two-parameter Weibull distribution, denoted
Weib(, k), and prove some of its basic properties. We will then survey a few techniques for estimating the
parameters , k. We will not focus on algorithms for the numerical computation of these estimates, as these
can be easily done with software such as R or Matlab. We will defer questions of goodness of t to a later
section of this paper.
2
2.1 Denition and Basic Properties
Let (, F, P) be a probability space. A real-valued random variable X : R is said to have a (two-
parameter) Weibull distribution if it has a probability density function (pdf)
f(x; , k) =
_
k
(
x
)
k1
e
(
x
)
k
x 0
0 x < 0
where k > 0 is a dimensionless shape parameter and > 0 is the scale parameter. As an exercise, the
reader can check by means of integration that this is indeed a pdf. For our modeling purposes below, x will
correspond to the wind velocity v. We will use the notation X Weib(, k) to denote that X has Weibull
distribution with parameters , k.
Proposition 2.1. The Weibull distribution with paramters , k > 0 has cummulative distribution function
(cdf ) given by
F(x; , k) =
_
0 x < 0
1 e
(
x
)
k
x 0
Proof. By the fundamental theorem of calculus,
P(X x) =
_
x
f(y; , x)dy =
_
x
0
k
(
y
)
k1
e
(
y
)
k
=
_
x
0
d
dy
_
e
(
y
)
k
_
dy =
_
e
(
x
)
k
1
_
= 1 e
(
x
)
k
Figure 1: Plots of the Weibull pdf and cdf, respectively, for various shape paramters k
Proposition 2.2. Let X Weib(, k). Then X has m
th
moment given by
E[X
m
] =
m
(1 +
m
k
)
In particular, X has mean, variance, and maximum respectively given by
1. E[X] = (1 +
1
k
),
3
2. Var(X) =
2
_
(1 +
2
k
)
2
(1 +
1
k
)
_
,
where () denotes the gamma function.
Proof. Recall Eulers integral for the gamma function, (z) =
_
0
e
t
t
z1
dt for Re z > 0. Hence, for
m Z
1
,
E[X
m
] =
_
R
x
m
f(x; , k)dx =
_
0
x
m
k
(
x
)
k1
e
(
x
)
k
dx
We make the change of variable y = (
x
)
k
to obtain
=
_
0
m
y
m
k
e
y
dy = (1 +
m
k
)
Setting m = 1 gives the expression for E[X]. For Var(X), by denition
Var(X) = E[X
2
] E
2
[X] =
2
_
(1 +
2
k
)
2
(1 +
1
k
)
_
The following proposition tells us that the wind speed follows a Weibull distribution regardless of the
choice of units.
Proposition 2.3. Let > 0 and X Weib(, k). Then X =: Y Weib(, k).
Proof. Using Proposition 2.1, we see that
F
Y
(x) = P(X x) = P(X
x
) = 1 e
(
x
)
k
,
which is the cdf of a Weib(, k) random variable. Since the distribution function uniquely characterizes the
law of the random variable, the conclusion follows immediately.
The following result on the minimum order statistic of Weibull random variates will be useful in giving a
rough answer to questions, such as given nitely many wind farms, what is the probability that all of them
are producing more than P watts. In reality, wind speeds as dispersed locations are correlated, so that is
what makes the following a rough answer.
Proposition 2.4. Let X
1
, , X
n
be independent random variables with X
i
Weib(
i
, k
i
), for 1 i n.
Then
P(min(X
1
, , X
n
) > x) = exp
_
i=1
(
x
i
)
ki
_
If k
1
= = k
n
= k, then min(X
1
, , X
n
) Weib(
min
, k), where
min
:=
_
n
i=1
k
i
_
1
k
Proof. Clearly, min(X
1
, , X
n
) > x X
i
> x i = 1, , n. Since the X
i
are independent,
P(X
1
> x, , X
n
> x) =
n
i=1
P(X
i
> x) =
n
i=1
_
1
_
1 e
(
x
i
)
k
i
__
=
n
i=1
e
(
x
i
)
k
i
= exp
_
i=1
(
x
i
)
ki
_
4
This completes the proof of the rst assertion. Now suppose that k
i
= k, 1 i n. Then we may write the
above as
= exp
_
i=1
x
k
k
i
_
= exp
_
x
k
_
n
i=1
k
i
_
1
_
= exp(
x
k
k
min
),
min
:=
_
n
i=1
k
i
_
1
k
We therefore have that
P(min(X
1
, , X
n
) x) = 1 P(min(X
1
, , X
n
) > x) = 1 exp
_
(
x
min
)
k
_
,
which shows that min(X
1
, , X
n
) Weib(
min
, k) by the uniqueness of distribution functions and Propo-
sition 2.1.
2.2 Parameter Estimation
Having established some basic properties of the Weibull distribution and found a probability distribution for
the power output at a single site, we now turn to the problem of estimating the parameters , k.
2.2.1 Likelihood Function
Let (X
1
, , X
n
) : R
n
be a random variable with probability density function f(x; ) for a k-tuple of
parameters R
k
. Recall that for a sample X() = x R, the likelihood function of , denoted by
x
() = (), is dened by
: R, () = f
(x)
Let denote the topological closure of in R. If
satises (
) = sup
(), then
we say that
is a maximum likelihood estimator (MLE) of .
We dene the log-likelihood function to be log (). We rst note that this denition makes sense since
we may asume that () > 0 for all . The following lemma often simplies computations of MLEs since the
logarithm converts products to sums.
Lemma 2.5.
is an MLE if and only if log (
) = sup
(log ).
Proof. This is immediate from the fact that log x is a strictly increasing function on (0, ).
2.2.2 Maximum Likelihood
It is not true that every probability density function has a nice closed-form expression for the MLE
. Unfor-
tunately for us, this is also the case for the Weibull distribution. However, with the use of numerical software
such as Matlab, we can use an iterative scheme (or trial and error, if we just want a rough approximation)
to approximate
.
Let X
1
, , X
n
be independent identically distibuted (i.i.d.) Weib(, k) samples. Since the joint density
function of independent random variables factors, we have that
(, k) =
n
i=1
f(x
i
; , k) =
n
i=1
(
k
)(
x
i
)
k1
e
(
x
i
)
k
= (
k
)
n
exp
_
k
n
i=1
x
k
i
_
n
i=1
x
k1
i
n(k1)
Taking the natural logarithm of both sides, we obtain that the log-likelihood function is
log (, k) = nlog(k) nlog()
1
k
_
n
i=1
x
k
i
_
+
n
i=1
(k 1) log(x
i
) n(k 1) log()
= nlog(k) nk log()
1
k
_
n
i=1
x
k
i
_
+
n
i=1
(k 1) log(x
i
)
5
Taking the partial derivatives with respect to , k, we obtain
[log (, k)] =
nk
nlog() +
k
k+1
_
n
i=1
x
k
i
_
and
k
[log (, k)] =
n
k
+
log
k
_
n
i=1
x
k
i
_
k
_
n
i=1
log(x
i
)x
k
i
_
+
n
i=1
log(x
i
)
Proposition 2.6. The MLE (
+
k
k+1
_
n
i=1
x
k
i
_
=
_
n
i=1
x
k
i
n
_1
k
It is clear that, for k xed,
is the unique root of g
k
() =
nk
+
k
k+1
(
n
i=1
x
k
i
). I claim that g
k
() 0 for
all
. Indeed, since
k+1
= o(
k
) as 0, g
k
() < 0 for all suciently small. If g
k
() > 0 for some
<
implies by the intermediate value theorem that there exists
, 0 <
< <
such that g
k
(
) = 0,
which is a contradiction. Noting that = o(
k+1
) as and therefore g
k
() < 0 for all suciently
large, the same argument shows that g
k
() 0 for
. We conclude that
is the unique global maximum
of g
k
.
Similarly, we have
0 =
n
k
nlog(
) +
log
k
_
n
i=1
x
k
i
_
k
_
n
i=1
log(x
i
)x
k
i
_
+
n
i=1
log(x
i
)
=
n
k
nlog(
) +
nlog(
k
_
n
i=1
log(x
i
)x
k
i
_
+
n
i=1
log(x
i
)
=
n
k
n
n
i=1
x
k
i
log(x
i
)
n
i=1
x
k
i
+
n
i=1
log(x
i
)
n
i=1
x
k
i
log(x
i
)
n
i=1
x
k
i
k
=
n
i=1
log(x
i
)
n
I claim that this equation has a unique solution. For existence, observe that
h(k) = log (
, k) = nlog(k)
n
k
k log
_
n
i=1
x
k
i
n
_
n
n
i=1
x
k
i
n
i=1
x
k
i
+
n
i=1
(k 1) log(x
i
)
= nlog(k) nlog
_
n
i=1
x
k
i
n
_
n +
n
i=1
(k 1) log(x
i
)
It is evident that h(k) as k 0. We now consider h(k) for large values of k, in particular k 2. Since
the function x x
k
is strictly convex, for k 2, by Jensens inequality,
n
i=1
x
k
i
n
>
_
n
i=1
x
i
n
_
k
nlog
_
n
i=1
x
k
i
n
_
< nlog
_
_
_
n
i=1
x
i
n
_
k
_
_
= nk log
_
n
i=1
x
i
n
_
Since the function x log(x) is strictly convex and by our hypothesis that min(x
1
, , x
n
) < max(x
1
, , x
n
),
we have by another application of Jensens inequality that
nlog
_
n
i=1
x
k
i
n
_
< nk log
_
n
i=1
x
i
n
_
< nk
n
i=1
log(x
i
)
n
= k
n
i=1
log(x
i
),
6
which implies that
k
_
n
i=1
log(x
i
)
_
nlog
_
n
i=1
x
k
i
n
_
< k
_
n
i=1
log(x
i
) nlog
_
n
i=1
x
i
n
__
. .
C
< 0
Hence, C < 0. Since log(k) = o(k) as k , it follows that h(k) as k . Hence, there exists
k
1
, k
2
with 0 < k
1
< k
2
< , such that sup
k[k1,k2]
h(k) = sup
k(0,)
h(k). By Weierstrass extreme value
theorem, there exists
k [k
1
, k
2
] such that
h(
k) = sup
k[k1,k2]
h(k) = sup
k(0,)
h(k),
which implies that
k is a global maximum of h(k). By adjusting k
1
and k
2
, if necessary, we may assume that
h(k) < h(
k), k (0, ) \ (k
1
, k
2
)
Since (0, ) is an open subset of R, we have by Fermats lemma for critical points that
0 = h
k) =
(
, k)
k
(
k) =
n
k
n
n
i=1
x
k
i
log(x
i
)
n
i=1
x
k
i
+
n
i=1
log(x
i
)
For uniqueness, we note that since
k is a local maximum, there exists an > 0 such that h
k ,
k) and h
k,
k + ). Suppose that
k
>
k,
then h
(k) = 0
for some
k (
k,
k
(k):
h
(k) =
n
k
2
n
n
i=1
x
k
i
(log x
i
)
2
n
i=1
x
k
i
+ n
_
n
i=1
x
k
i
log(x
i
)
_
_
n
j=1
x
k
i
log(x
i
)
_
_
n
i=1
x
k
i
_
2
=
n
k
2
n
i<j
(x
i
x
j
)
k
_
log(x
i
) log(
xi
xj
) + log(x
j
) log(
xj
xi
)
_
_
n
i=1
x
k
i
_
2
=
n
k
2
n
i<j
(x
i
x
j
)
k
_
(log x
i
)
2
2(log x
i
)(log x
j
) + (log x
j
)
2
_
n
i=1
x
k
i
_
2
=
n
k
2
n
i<j
(x
i
x
j
)
k
(log(x
i
) log(x
j
))
2
_
n
i=1
x
k
i
_
2
< 0
But if
k is a local minimum, by the second derivative test, h
<
k is
completely analogous.
I now claim that (
)
(0, )
2
is another global maximum. We showed above that log (
, k) < log (
, k) for
=
, so we have
=
. Since (
, k) = (
, k) is a global maximum, h(
) = log (
, k) = log (
k) = h(
=
k, since we showed that h(k) has a unique global maximum.
2.2.3 Computation of MLE
The preceding existence and uniqueness result does not help us much in practice, since we are still left with
the problem of computing
k, or obtaining a reasonable approximation. We now give a Newton-Raphson
algorithm for nding
k. Dene
g : (0, ) R, g(k) :=
n
k
n
n
i=1
x
k
i
log(x
i
)
n
i=1
x
k
i
+
n
i=1
log(x
i
)
7
Then
g
(k) =
n
k
2
n
n
i=1
x
k
i
(log x
i
)
2
n
i=1
x
k
i
+ n
_
n
i=1
x
k
i
log(x
i
)
_
_
n
j=1
x
k
i
log(x
i
)
_
_
n
i=1
x
k
i
_
2
=
n
k
2
n
i<j
(x
i
x
j
)
k
(log(x
i
) log(x
j
))
2
_
n
i=1
x
k
i
_
2
< 0,
as shown above; and
g
(k) = 2
n
k
3
n
_
n
i=1
x
k
i
_
2
_
i<j
log(x
i
x
j
)(x
i
x
j
)
k
(log(x
i
) log(x
j
))
2
_
_
n
i=1
x
k
i
_
4
+ n
2
_
i<j
(x
i
x
j
)
k
(log(x
i
) log(x
j
))
2
_
_
n
k=1
log(x
i
)x
k
i
_
_
n
i=1
x
k
i
_
4
,
which is evidently bounded on compact subsets of (0, ), being continuous. Fix 0 < a <
k < b < such
that g(a) > 0, g(b) < 0, 0 < |g
(k)| and |g
k, b)
denote the initial choice. Then m
th
iteration of the Newton-Raphson method, for m 1, is given by
k
m
=
k
m1
n
km1
n
n
i=1
x
k
m1
i
log(xi)
n
i=1
x
k
m1
i
+
n
i=1
log(x
i
)
k
2
m1
n
i<j
(xixj)
k
m1
(log(xi)log(xj))
2
n
i=1
x
k
m1
i
2
We leave it as an exercise for the reader to implement the above algorithm in Matlab or some other compu-
tational mathematics software.
2.3 An Example
To give concreteness to our discussion of the Weibull distribution above, we now consider actual wind speed
data. The following data is reproduced from [8], in which the author attempted to estimate the wind power
potential in the kingdom of Bahrain. Hourly average wind speeds were collected from January 2003 through
December 2005 at the Bahrain International Airport at a height of 10m. The wind speeds were rounded,
resulting in integer classes of wind speeds. Table 1 gives the number of hours each wind speed class was
observed for each month over the period of data collection. Table 2 lists the maximum likelihood estimates
for the Weibull paramters and k for the distribution of hourly average wind speed class for each month.
Strictly speaking, we cannot the MLE technique with observations taking the value 0, since the likelihood
function would therefore 0. This is an unfortunate consequence of using rounded data. As a way around this
problem, we replace the 0 with a small value, say .1, in the computation of the MLEs.
To give the reader a graphical sense of how the Weibull probability density function with estimated
parameters approximates the empirical probability density function for each monthh, we have included graphs
of the empirical pdfs and Weibull pdfs for each month in Appendix B.
8
Wind Speed (
m
s
) Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
0 21 31 36 29 20 24 43 42 51 78 41 74
1 138 93 96 126 109 93 170 212 204 245 129 201
2 277 171 216 245 305 183 244 381 341 389 349 313
3 339 305 318 347 388 306 407 440 417 530 407 292
4 331 294 337 381 432 309 376 375 368 383 320 274
5 315 279 328 316 291 282 322 250 285 230 236 255
6 256 243 260 249 230 283 262 203 192 132 200 240
7 203 197 229 174 193 250 190 159 151 91 169 201
8 166 157 172 122 108 218 152 101 98 72 119 163
9 95 104 122 69 77 124 55 49 41 51 94 132
10 58 97 64 37 39 43 11 14 8 25 48 59
11 22 41 28 30 22 34 0 3 4 5 33 24
12 8 15 14 16 14 8 0 3 0 1 10 3
13 3 10 7 14 4 3 0 0 0 0 4 1
14 0 3 3 4 0 0 0 0 0 0 1 0
15 0 0 1 0 0 0 0 0 0 0 0 0
16 0 0 1 1 0 0 0 0 0 0 0 0
Table 1: Frequency (number of hours) of wind speed classes according to month from 2003-2005
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
5.4017 5.9035 5.7283 5.3186 5.1571 5.8659 4.8827 4.4430 4.4203 4.0345 5.0466 5.1341
k 2.0042 1.9946 2.0209 1.9239 2.0171 2.1825 2.0318 1.8459 1.8616 1.6810 1.7876 1.6707
Table 2: Maximum Likelihood Estimates for the Weibull parameters
,
k
_
2x
A
_k
3
_
and pdf
f
P
(x) =
kx
k
3
1
3
k
_
2
A
_k
3
exp
_
_
2x
A
_k
3
_
Proof. By Proposition 2.1, we have that
F
P
(x) = P(P x) = P
_
V
_
2x
A
_1
3
_
= F
_
_
2x
A
_1
3
_
= 1 exp
_
_
2x
A
_k
3
_
Dierentiating both sides with repect to x, we obtain that P has probability density function f
p
given by
f
P
(x) =
k
3
k
2
A
_
2x
A
_k
3
1
exp
_
_
2x
A
_k
3
_
=
kx
k
3
1
3
k
_
2
A
_k
3
exp
_
_
2x
A
_k
3
_
10
A wind turbine, though, does not operate at all wind speeds V . The implied wind speed of wind turbine
with power coecient C
P
and eciency coecient producing power P is given by
V
turbine
=
_
_
0 V
turbine
< v
cut-in
V v
cut-in
V v
rated
v
rated
v
rated
< V < v
cut-o
0 V v
cut-o
where v
cut-in
< v
rated
< v
cut-o
are specied by the manufacturer. For example, the GE 1.5 MW SLE
wind turbine has cut-in wind speed v
cut-in
= 3.5
m
s
, rated wind speed v
rated
= 14
m
s
, and cut-out wind speed
v
cut-o
= 25
m
s
(see [1] for the technical details of the 1.5 SLE model). Analogously,
P
turbine
=
_
_
0 V
turbine
< v
cut-in
1
2
AC
P
V
3
v
cut-in
V v
rated
1
2
AC
P
v
3
rated
v
rated
< V < v
cut-o
0 V v
cut-o
It is evident V
turbine
and P
turbine
are discontinuous random variable, but we can still compute their distribution
functions and moments. They will fail to have a probability density function, though, their laws are not
absolutely continuous with respect to the Lebesgue measure: the law of V
turbine
assigns nonzero probability
to the singleton {0}, which is a set of Lebesgue measure zero.
Proposition 4.2. P
turbine
has cdf
F
P
turbine
(x) =
_
_
0 < x < 0
1 + e
(
v
cut-off
)
k
e
(
v
cut-in
)
k
0 x <
1
2
C
P
Av
3
cut-in
1 e
k
(
2x
C
P
A
)
k
3
+ e
(
v
cut-off
)
k
1
2
C
P
Av
3
cut-in
x <
1
2
C
P
Av
3
rated
1
1
2
C
P
Av
3
rated
x <
Proof. It is evident that P(P
turbine
0) is
P({V v
cut-in
} {V v
cut-o
}) = F(v
cut-in
) + (1 F(v
cut-o
)) = (1 e
(
v
cut-in
)
k
) + (1 (1 e
(
v
cut-off
)
k
))
= 1 + e
(
v
cut-off
)
k
e
(
v
cut-in
)
k
So,
F
P
turbine
(x) = 1 + e
(
v
cut-off
)
k
e
(
v
cut-in
)
k
, 0 x <
1
2
C
P
Av
3
cut-in
For 0 < x =
1
2
C
P
Av
3
1
2
C
P
Av
3
r
, note that x
1
2
C
P
Av
3
r
v v
r
, which implies that
F
P
turbine
(x) = P(0 < P
turbine
x) +P(P
turbine
= 0)
= P(
1
2
C
P
Av
3
cut-in
P
turbine
x) + 1 + e
(
v
cut-off
)
k
e
(
v
cut-in
)
k
=
_
1 e
k
(
2x
C
P
A
)
k
3
_
_
1 e
k
(
2
1
2
C
P
Av
3
cut-in
C
P
A
)
k
3
_
+ 1 + e
(
v
cut-off
)
k
e
(
v
cut-in
)
k
= 1 e
k
(
2x
C
P
A
)
k
3
+ e
(
v
cut-off
)
k
For
1
2
C
P
Av
3
rated
x < , it is evident that F
P
turbine
(x) = 1.
We will use the notation (s, x) =
_
x
0
t
s1
e
t
dt to denote the lower incomplete Gamma function.
Proposition 4.3.
E[P
m
turbine
] =
_
1
2
C
P
A
_
m
_
3m
_
(1 +
3m
k
, (
v
rated
)
k
) (1 +
3m
k
, (
v
cut-in
)
k
)
_
+ v
3m
rated
_
e
(
v
rated
)
k
e
(
v
cut-off
)
k
_
_
11
Proof. By Theorem 1 pg. 192 and Theorem 7 pg. 196 in [17],
E[P
m
turbine
] =
_
P
m
turbine
d(F
P
turbine
(x))
=
_
v
rated
vcut-in
_
1
2
C
P
Av
3
_
m
(
k
)(
v
)
k1
e
(
v
)
k
dv +
_
v
cut-off
v
rated
_
1
2
C
P
Av
3
rated
_
m
(
k
)(
v
)
k1
e
(
v
)
k
dv
We make the change of variable y = (
v
)
k
to obtain
E[P
m
turbine
] =
_
1
2
C
P
A
_
m
_
3m
_
(
v
rated
)
k
(
v
cut-in
)
k
y
3m
k
e
y
dy +
_
(
v
cut-off
)
k
(
v
rated
)
k
v
3m
rated
e
y
dy
_
=
_
1
2
C
P
A
_
m
_
3m
_
(1 +
3m
k
, (
v
rated
)
k
) (1 +
3m
k
, (
v
cut-in
)
k
)
_
+ v
3m
rated
_
e
(
v
rated
)
k
e
(
v
cut-off
)
k
_
_
In particular, P
turbine
has mean
E[P
turbine
] =
1
2
C
P
A
_
3
_
(1 +
3
k
, (
v
rated
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+ v
3
rated
_
e
(
v
rated
)
k
e
(
v
cut-off
)
k
_
_
We can dene the capacity factor of a wind turbine to be the ratio
CF =
E[P
turbine
]
1
2
C
p
Av
3
rated
Using our preceding work, we can obtain a closed-form expression for CF:
CF =
1
v
3
rated
_
3
_
(1 +
3
k
, (
v
rated
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+ v
3
rated
_
e
(
v
rated
)
k
e
(
v
cut-off
)
k
_
_
=
_
v
rated
_
3
_
(1 +
3
k
, (
v
rated
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+ e
(
v
rated
)
k
e
(
v
cut-off
)
k
4.3 Turbine-Site Matching
Suppose we have chosen a location where the wind speed V for a given month has a Weib(, k) distribution.
Furthermore, suppose a manufacturer can manufacture wind turbines with xedcut-in and cut-o speeds
v
cut-in
and v
cut-o
, respectively, but can adjust the rated wind speed v
r
[v
cut-in
, v
cut-o
). Suppose also that
the power coecient C
P
, the eciency coecient and the rotor area A are independent of v
r
. What should
the rated wind speed be if we want to maximize the capacity factor CF? What should the rated wind speed
be if we want to maximize average power of the wind turbine E[P
turbine
]? Are these two wind speeds equal?
These questions can be answered by use of the calculus. Dene a function CF : [v
cut-in
, v
cut-o
) R by
CF(v) :=
_
v
_
3
_
(1 +
3
k
, (
v
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+ e
(
v
)
k
e
(
v
cut-off
)
k
Then
CF
(v) =
3
3
v
4
_
(1 +
3
k
, (
v
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+
_
v
_
3
_
(
v
)
k
3
k
e
(
v
)
k k
(
v
)
k1
_
(
v
)
k1
e
(
v
)
k
=
3
3
v
4
_
(1 +
3
k
, (
v
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
12
We see that the last expression has a root precisely at v = v
cut-in
and since CF
3
_
(1 +
3
k
, (
v
)
k
) (1 +
3
k
, (
v
cut-in
)
k
)
_
+ v
3
_
e
(
v
)
k
e
(
v
cut-off
)
k
_
_
Dierentiating with respect to v,
P
avg
(v) =
1
2
C
P
A
_
3
(
v
)
k
3
k
e
(
v
)
k k
(
v
)
k1
+ 3v
2
_
e
(
v
)
k
e
(
v
cut-off
)
k
_
v
3
k
(
v
)
k1
e
(
v
)
k
_
= 3v
2
_
e
(
v
)
k
e
(
v
cut-off
)
k
_
Setting the RHS equal to 0, we see that P
avg
has no critical points in the domain [v
cut-in
, v
cut-o
). Rather,
since P
avg
> 0 on [v
cut-in
, v
cut-o
), we only conclude that sup
v[vcutin,v
cut-off
)
P
avg
(v) =
1
2
C
P
Av
3
cut-o
.
These results indicate that by having lowly rated wind turbines, we can maximize the capacity factor at
the cost of minimizing expected power output. Conversely, we can increase expected power output at the
cost of decreasing the capacity factor. It is not a priori clear which approach is better, since we have not, nor
will we, discussed the economics of wind power. One obvious downside to using lowly rated wind turbines is
increased land usage, which would also raise costs if the wind farm operator is leasing the land. For example,
in a study of wind power potential in Bahrain, the author of [8] noted that using a large number of low-rated
turbines would not necessarily be ideal given that the kindgom of Bahrain is only 665km
2
.
5 Conclusion
In the preceding sections, we have given a blueprint for an analyst to estimate the wind resources of a xed
location. With a collection of, say, hourly average wind speeds for a year, he or she can use the maximum
likelihood method to determine Weibull parameters for the monthly wind speed distributions. He or she can
then use the results of section 4 to compute the expected power output of a given model of wind turbine,
with specifed cut-in, cut-o, and rated speeds, and can also compute the capacity factor. We emphasize that
the end result of this analysis is an estimate, and a rough one at that. To make computations tractable, we
have introduced a number of simplications. In reality, the wind is not always normal to the rotor plane,
nor is the power for wind speeds in between the cut-in and cut-o speeds a simple cubic function of the wind
speed; it depends on the specic wind turbines power curve (see [1] for the power curves of GE 1.5 MW
turbine models). Furthermore, our model ignores variations in wind speeds over large areas, over which a
suciently large wind farm would be spread. Perhaps most importantly, the Weibull distribution is not a
perfect description of wind speed distributions, as one can tell by the graphs in Appendix B. However, even if
the absolute estimates given by the model are to be interpreted with caution, the model nevertheless remains
very useful for comparing the wind resources of one site to another.
Having discussed the limitations of the model mathematically described in this paper, it is natural to
ask in what ways it can be improved and generalized. In terms of improvement, the obvious answer is to
more precisely account for the aerodynamics of wind turbines. But the remaining limitation in this direction
is that we still ignore the fact that we are still only really considering one wind turbine at one location.
Even within the same wind farm, the power output of any individual wind turbine is a random variable,
not perfectly correlated with the power of other turbines: the wind does not strike each turbine at the same
angle, nor blow at the same speed for turbines far apart from one another, and each turbine has a failure rate
that is not independent of the failure rate of other wind turbines. We encounter similar modeling diculties
when we consider multiple dispersed wind farms. Modeling dispersed wind farms is particularly important
because a question with which we should be concerned is how valid a criticism is it that the wind doesnt
blow all the time and when it does blow, it doesnt blow at a uniform, predictable speed? Does this criticism
lose its merits under the model of a wind farms dispersed over multiple high wind speed locations, where the
wind blows at peak speed at dierent times, providing electricity over an integrated system. Of course, such
a grid system does not exist in the status quo; but it is not inconceivable that a federal initiative, analogous
to that of the interstate highway system, could eect such a grid in the not too distant future.
How is one to tackle such a dicult modeling problem? One might naturally start by considering a
multivariate extension of the Weibull distribution, but there is not a unique multivariate distribution with
13
prescribed Weibull marginal distributions (see Section 4, Chapter 47 in [9] for several classes of multivariate
Weibull distributions). Moreover, when one does start to look at the various multivariate extensions of the
Weibull distribution, he or she sees that they can be quite ugly, as the reader can see in [11]. In [2], Carlin and
Haslett used the multivariate normal distribution as an approximation to a joint distribution of correlated
Weibull random variables, but adopting their approach seems like a step in the wrong direction. Computing
power has become much more ecient and cheaper since their paper was written, and thus using numerical
and Monte Carlo methods to handle seemingly intractable computations seems the better approach. It
suces to say that multi-site wind power modeling and estimation remains a dicult open problem in need
of further research.
A Newton-Raphson Method
Theorem A.1. Let [a, b] R such that the function f : [a, b] R is twice dierentiable, has a unique zero
[a, b], and satises |f
(x)| > 0, |f
(x
n
)
, n 1
Then lim
n
x
n
= and moreover,
0 x
n+1
1
A
(A(x
1
))
2
n
, A :=
M
2
Proof. Since [a, b] is connected, either f
(x
n
)
> 0 x
n+1
< x
n
n 1
Since f(x
n
) > 0 n 1, we must have that x
n
for all n. By the monotone limit theorem, x
n
. I
claim that
= . Indeed, since f, f
= lim
n
x
n+1
= lim
n
_
x
n
f(x
n
)
f
(x
n
)
_
= lim
n
x
n
lim
n
f(x
n
)
lim
n
f
(x
n
)
=
f(
)
f
)
,
which implies that
f(
)
f
)
= 0 f(
.
For the second assertion, we have by Taylors theorem with remainder (See Theorem 5.15 in [15]) that for
each n 0,
0 = f() = f(x
n
) + f
(x
n
)( x
n
) +
f
(
n
)
2
( x
n
)
2
,
for some
n
(, x
n
). Since f
(x
n
) = 0 and x
n+1
= x
n
f(xn)
f
(xn)
, we may write
0 =
f(x
n
)
f
(x
n
)
+ ( x
n
) +
f
(
n
)
2f
(x
n
)
( x
n
)
2
= (x
n
x
n+1
) + ( x
n
) +
f
(
n
)
2f
(x
n
)
( x
n
)
2
= ( x
n+1
) +
f
(
n
)
2f
(x
n
)
( x
n
)
2
(x
n+1
) =
f
(
n
)
2f
(x
n
)
( x
n
)
2
|x
n+1
|
M
2
| x
n
|
2
By induction, we obtain that
|x
n+1
|
. .
n+1
_
M
2
_
2
n
1
| x
1
|
2
n
=
1
A
(A| x
1
|)
2
n
If we choose our initial guess x
0
so that A|x
1
| < 1, then the above gives us an upper bound for
n+1
which converges to 0 as n .
14
B Example Graphs
15
16
17
18
19
20
References
[1] http://geosci.uchicago.edu/
~
moyer/GEOS24705/Readings/GEA14954C15-MW-Broch.pdf
[2] Carlin, John and John Haslett. The Probability Distribution of Wind Power from a Dispersed Array of
Wind Turbine Generators, Journal of Applied Meteorology 21 (1982), no. 3, 303-313.
[3] Chandra, M., N.D. Singpurwalla, and M.A. Stephens. Kolmogorov Statistics for Tests of Fit for the
Extreme Value and Weibull Distributions, Journal of Amer. Stat. Assoc., 76 (1981), no. 375, 729-731.
[4] Gasch, Robert and Jochen Twele. Wind Power Plants: Fundamentals, Design, Construction, and Oper-
ation (Second Edition). Springer, 2012.
[5] Genc, Asir, Murat Erisoglu, Ahmet Pekgor, Galip Oturanc, Arif Hepbasli. Estimation of Wind Power
Potential Using Weibull Distribution, Energy Sources 27 (2005), no. 9, 809-822.
[6] Hansen, Martin O.L. Aerodynamics of Wind Turbines (Second Edition). Earthscan, 2008.
[7] Hennessey, Joseph P. Jr. Some Aspects of Wind Power Statistics, Journal of Applied Meteorology 16
(1977), no. 2, 119-128.
[8] Jowder, Fawzi A.L. Wind Power Analysis and Site Matching of Wind Turbine Generators in Kingdom
of Bahrain, Applied Energy 86 (2009), 538-545.
[9] Kotz, Samuel, N. Balakrishnan, and Norman L. Johnson. Continuous Multivariate Distributions, Vol. 1:
Models and Applications (Second Edition). John Wiley & Sons, 2000.
[10] Kotz, Samuel, N. Balakrishnan, and Norman L. Johnson. Continuous Univariate Distributions, Vol. 1
(Second Edition). John Wiley & Sons, 1994.
[11] Lee, Cheng K. and Miin-Jye Wen. A Multivariate Weibull Distribution. http://arxiv.org/ftp/math/
papers/0609/0609585.pdf.
[12] Manwell, J.F., J.G. McGowan, and A.L. Rogers. Wind Energy Explained: Theory, Design, and Appli-
cation (Second Edition). John Wiley & Sons, 2009.
[13] Patel, Mukund R. Wind and Solar Power Systems: Design, Analysis, and Operation (Second Edition).
Taylor & Francis, 1972.
[14] Jangamshetti, Suresh H. and V. Guruprasada Rau. Site Matching of Wind Turbine Generators: A Case
Study, IEEE Trans. on Energy Conversion 14 (1999), No. 4, 1537-1543.
[15] Rudin, Walter. Principles of Mathematical Analysis (Third Edition). McGraw-Hill, 1964.
[16] Shao, Jun. Mathematical Statistics (Second Edition). Springer, 2003.
[17] Shiryaev, A.N. Probability (Second Edition). Springer, 1995.
21