You are on page 1of 9

Applied Energy 88 (2011) 1848–1856

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Estimation of wind energy potential using different probability density functions


Tian Pau Chang ⇑
Department of Computer Science and Information Engineering, Nankai University of Technology, Nantou 542, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: In addition to the probability density function (pdf) derived with maximum entropy principle (MEP), sev-
Received 1 July 2010 eral kinds of mixture probability functions have already been applied to estimate wind energy potential
Received in revised form 26 October 2010 in scientific literature, such as the bimodal Weibull function (WW) and truncated Normal Weibull func-
Accepted 6 November 2010
tion (NW). In this paper, two other mixture functions are proposed for the first time to wind energy field,
Available online 4 December 2010
i.e. the mixture Gamma–Weibull function (GW) and mixture truncated normal function (NN). These five
functions will be reviewed and compared together with conventional Weibull function. Wind speed data
Keywords:
measured from 2006 to 2008 at three wind farms experiencing different climatic environments in Taiwan
Wind speed
Wind power density
are selected as sample data to test their performance. Judgment criteria include four kinds of statistical
Probability density function errors, i.e. the max error in Kolmogorov–Smirnov test, root mean square error, Chi-square error and rel-
Weibull function ative error of wind potential energy. The results show that all the mixture functions and the maximum
Mixture function entropy function describe wind characterizations better than the conventional Weibull function if wind
Statistical error regime presents two humps on it, irrespective of wind speed and power density. For wind speed distri-
butions, the proposed GW pdf describes best according to the Kolmogorov–Smirnov test followed by the
NW and WW pdfs, while the NN pdf performs worst. As for wind power density, the MEP and GW pdfs
perform best followed by the WW and NW pdfs. The GW pdf could be a useful alternative to the conven-
tional Weibull function in estimating wind energy potential.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction as the two-component mixture Weibull function (Weibull–


Weibull) and singly truncated normal Weibull mixture function
The utilization of wind resource plays a more and more impor- (Normal–Weibull) [1,21,24–27]. Besides, the probability density
tant role in energy supply currently in the world. Wind speed dis- function derived with maximum entropy principle (MEP) has also
tribution for a particular location determines the wind energy been verified that it has some strength [22,23,28–32].
available and the performance of an energy conversion system. Systematical analysis about the suitability of these functions in
Once the probability distribution of wind speed is obtained, the a single study has rarely been found in the literature. In this paper,
wind energy potential can be determined accordingly. For this rea- two other kinds of mixture probability functions available in statis-
son, a variety of probability density functions (pdf) have been used tics are originally applied to estimate wind energy, i.e. the mixture
in literature to describe wind speed distributions including beta Gamma–Weibull function (Gamma–Weibull) and the mixture
function, Gamma function, lognormal function, logistical function, truncated normal function (Normal–Normal). All these mixture
Rayleigh function, Weibull function, etc. Among these, the Weibull functions and the MEP function will be introduced and compared
function is most widely adopted because of its two flexible param- together with the conventional Weibull function. Wind speed data
eters [1–20]. Weibull shape parameter describes the width of data selected are measured each 10 min from 2006 to 2008 at three
distribution, while scale parameter controls the abscissa scale of a wind farms experiencing different climatic environments in
plot of data distribution. However, it must be noted that the Wei- Taiwan, which are handled by the Ministry of Economic Affairs.
bull function is unable to represent all the wind structures encoun- The 10-min wind speed measurements are transferred to hourly
tered in nature [21–23]. One main limitation of the conventional data and are averaged over the 3 years before doing subsequent
Weibull function is that it does not accurately model calm winds calculations. The selected wind data are considered here as sample
and those with bimodal or even multimodal distributions resulted data just for testing the performance of various probability func-
from special climatic conditions. In this context, some mixture tions. Fig. 1 shows the geographical location of the three wind
functions of simple unimodal distributions can be considered, such farms with their longitude and latitude. The first wind farm,
Hengchun (named station A), locates at the southern peninsula
⇑ Fax: +886 49 2561408. that experiences more stable weather condition throughout the
E-mail address: t118@nkut.edu.tw year; the second one, Chungtun (station B), is at a small island in

0306-2619/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2010.11.010
T.P. Chang / Applied Energy 88 (2011) 1848–1856 1849

Nomenclature

c scale parameter of Weibull function, m/s v3 mean of wind speed cubes, m3/s3
f() Weibull, probability density function vi wind speed in time stage i, m/s
F() cumulative Weibull function w weight of mixture distribution, dimensionless
g() gamma pdf yi actual probability at time stage i
G() cumulative Gamma function yic computed probability at time stage i
h() Gamma–Weibull pdf Z() gradient matrix of Lagrange multiplier
H() cumulative Gamma–Weibull function
I() normalization factor of truncated distribution Greek letters
k shape parameter of Weibull function, dimensionless C Gamma function
KS Kolmogorov–Smirnov test f shape parameter of Gamma distribution, dimensionless
LL logarithm of likelihood function b scale parameter of Gamma distribution, m/s
n number of data point l mean of wind speed, m/s
N number of moment constraint r standard deviation of wind speed, m/s
O() cumulative frequency of observed data w() Normal–Weibull pdf
P wind power density, W/m2 W() cumulative Normal–Weibull function
q() truncated normal pdf h() Weibull–Weibull pdf
Q() cumulative truncated normal function H() cumulative Weibull–Weibull function
r() Normal–Normal pdf q air density, kg/m3
R() cumulative Normal–Normal function an Lagrange multiplier, dimensionless
RMSE root mean square error kn moment of distribution
S entropy, dimensionless /n() power of wind speed, (m/s)n
2
T() cumulative frequency of theoretical function v Chi-square error
v wind speed, m/s

the Taiwan Strait having the highest wind in winter and spring function and actual time-series data. All the computation proce-
months; the third one, Dayuan (station C), locates at the north- dures are implemented in MATLAB software package [33].
western plain of Taiwan having strong wind in winter.
To test how accurate a probability function models the observa- 2. Mathematical models
tion data, four kinds of statistical errors are considered as the judg-
ment criteria, i.e. the max error in the Kolmogorov–Smirnov test, In mathematical statistics, the probability density function of a
root mean square error (RMSE), Chi-square error as well as the rel- random variable is a mathematical model that describes the rela-
ative percent error of wind energy potential between theoretical tive likelihood for this random variable to occur at a given point
in the observation space. The probability for a random variable to
fall within a given interval is equal to the integral of its density
26 over the interval. The cumulative distribution function (cdf) de-
scribes the probability that a random variable will be found less
Station C
than or equal to a specific value, i.e. the cumulative area of the
121.1o E probability density function below the specific value. In this
25.1o N context, if a distribution has no explicit form of cumulative distri-
25 bution function to calculate the probability, the numerical integra-
tion technique is used in this paper. On the other hand, a linear
combination of two or more probability distributions with appro-
Latitude (degree, North)

Station B priate weighting values, named mixture distribution, has already


119.6o E been applied to many fields as mentioned by Carta et al. [1].
24
23.6o N

2.1. Weibull distribution (W pdf)

The conventional (two-parameter) Weibull probability density


23 function has widely been used for describing wind regimes written
as [26]:
  k 
k v k1 v
f ðv ; k; cÞ ¼ exp  for v > 0 and k; c > 0 ð1Þ
c c c
22 where v is the wind speed; k is the shape parameter (dimension-
Station A less) and c is the scale parameter having the same dimension as
120.7 o E 0 50 km wind speed. Note that the Weibull distribution is related to a num-
21.9 o N
ber of other probability distributions; in particular, it becomes the
21
exponential distribution (if shape parameter k = 1), and the Ray-
119 120 121 122 123 leigh distribution (if k = 2). It is found that the Weibull shape
Longitude (degree, East) parameters are always greater than 1 in real wind speed distribu-
tions [1,2,13,17,19]. The Weibull cumulative distribution function
Fig. 1. Geographical location of the three wind farms analyzed in this study. is given by:
1850 T.P. Chang / Applied Energy 88 (2011) 1848–1856

  k 
v of the truncated distribution to one; which is just the cumulative
Fðv ; k; cÞ ¼ 1  exp  ð2Þ
c distribution function evaluated at its domain of definition ex-
pressed by the following equation:
Weibull shape and scale parameters can be calculated using the " #
Z
maximum likelihood method [2,19]: 1 1
ð v  lÞ 2
"P #1 Iðl; rÞ ¼ pffiffiffiffiffiffiffi exp  dv ð11Þ
n Pn r 2p 0 2r 2
i¼1 v i lnðv i Þ i¼1 lnðv i Þ
k
k¼ Pn k  ð3Þ
i¼1 v i
n The cumulative truncated normal distribution function is given
by:
!1=k Z v " #
1X n
1 ð v  lÞ 2
c¼ vk ð4Þ Qðv ; l; rÞ ¼ pffiffiffiffiffiffiffi exp  dv ð12Þ
n i¼1 i 0 Iðl; rÞr 2p 2r 2

where vi is the wind speed in time step i and n is the number of data As a result, the mixture function of two-component truncated
points. normal distribution can be expressed as the following equation,
which is applied for the first time to the wind energy field:
2.2. Mixture Gamma and Weibull distribution (GW pdf)
rðv ; w; l1 ; r1 ; l2 ; r2 Þ ¼ wqðv ; l1 ; r1 Þ þ ð1  wÞqðv ; l2 ; r2 Þ ð13Þ
The probability density function of Gamma distribution can be Corresponding cumulative distribution function is shown as:
expressed with the following function [34]:
  Rðv ; w; l1 ; r1 ; l2 ; r2 Þ ¼ wQ ðv ; l1 ; r1 Þ þ ð1  wÞQ ðv ; l2 ; r2 Þ ð14Þ
v f1 v
gðv ; f; bÞ ¼ f exp  for v > 0 and f; b > 0 ð5Þ As mentioned above, the five parameters of the mixture distribution
b CðfÞ b
can be estimated by maximizing the logarithm of likelihood func-
where f and b are the shape and scale parameters, respectively; C() tion given as:
is the Gamma function. The cumulative Gamma distribution func-
tion is given as: X
n
 
LL ¼ ln wqðv i ; l1 ; r1 Þ þ ð1  wÞqðv i ; l2 ; r2 Þ ð15Þ
Z  
v f1 v i¼1
Gðv ; f; bÞ ¼ exp  dv ð6Þ
bf CðfÞ b
2.4. Mixture normal and Weibull distribution (NW pdf)
A distribution function mixed with the gamma and conven-
tional Weibull distribution can be expressed as below that is orig-
The probability density function mixed with the truncated nor-
inally applied in wind energy assessment:
mal distribution and conventional Weibull distribution can be
hðv ; w; f; b; k; cÞ ¼ wgðv ; f; bÞ þ ð1  wÞf ðv ; k; cÞ ð7Þ written as [27]:
Corresponding cumulative distribution function is shown as: wðv ; w; l; r; k; cÞ ¼ wqðv ; l; rÞ þ ð1  wÞf ðv ; k; cÞ ð16Þ
Hðv ; w; f; b; k; cÞ ¼ wGðv ; f; bÞ þ ð1  wÞFðv ; k; cÞ ð8Þ Its cumulative distribution function is as:
where 0 6 w 6 1 is the weight parameter indicating the mixed Wðv ; w; l; r; k; cÞ ¼ wQ ðv ; l; rÞ þ ð1  wÞFðv ; k; cÞ ð17Þ
proportion of a component distribution. The five parameters of a
Relevant likelihood function is:
mixture distribution (for example here, w, f, b, k, c) can be esti-
mated using the graphical method, maximum likelihood method X
n

and moment method [1,6,25]. In this study the maximum likelihood LL ¼ lnfwqðv i ; l; rÞ þ ð1  wÞf ðv i ; k; cÞg ð18Þ
i¼1
method is adopted that maximizes the logarithm of likelihood func-
tion given as:
2.5. Mixture Weibull distribution (WW pdf)
LL ¼ ln Lðv i ; w; f; b; k; cÞ
n
¼ ln P v
i¼1 fwgð i ; f; bÞ þ ð1  wÞf ðv i ; k; cÞg The probability density function of a mixture Weibull distribu-
X
n tion can be written as [21,27]:
¼ ln fwgðv i ; f; bÞ þ ð1  wÞf ðv i ; k; cÞg ð9Þ
i¼1 hðv ; w; k1 ; c1 ; k2 ; c2 Þ ¼ wf ðv ; k1 ; c1 Þ þ ð1  wÞf ðv ; k2 ; c2 Þ ð19Þ
where vi is the wind speed in time step i and n is the number of Its cumulative distribution function is given by:
data.
Hðv ; w; k1 ; c1 ; k2 ; c2 Þ ¼ wFðv ; k1 ; c1 Þ þ ð1  wÞFðv ; k2 ; c2 Þ ð20Þ
2.3. Mixture normal distribution (NN pdf) Relevant likelihood function is:
X
n
The truncated normal distribution is the probability distribu- LL ¼ lnfwf ðv i ; k1 ; c1 Þ þ ð1  wÞf ðv i ; k2 ; c2 Þg ð21Þ
tion of a normally distributed random variable whose value is i¼1
either bounded below or above (or both). Since wind speed is not
below zero, the one considered in this paper is the singly truncated
2.6. Maximum entropy principle distribution (MEP pdf)
normal distribution, suitable for non negative wind speeds, defined
as [26]:
" # The concept of maximum entropy has commonly been applied
1 ðv  lÞ2 in communication theory. The entropy of a probability density
qðv ; l; rÞ ¼ pffiffiffiffiffiffiffi exp  for v0 ð10Þ
Iðl; rÞr 2p 2r2 function f(x) is defined as [35]:
Z
where l and r are the mean and standard deviation of data, respec- S¼ f ðxÞ ln f ðxÞdx ð22Þ
tively. I(l, r) is the normalization factor that leads the integration
T.P. Chang / Applied Energy 88 (2011) 1848–1856 1851

Z " #
Maximizing the entropy subject to some constraints enables X
N

one to find the most likely probability density function if the infor- Z n ðaÞ ¼ v n exp  am v m dv ¼ kn for n ¼ 0; 1; . . . ; N ð28Þ
m¼0
mation available is provided by moment functions. The (N + 1) con-
straints for a physical system can be written as: In this study, the considered probability density function with the
Z maximum entropy principle is the fourth order which has four mo-
Ef/n ðxÞg ¼ /n ðxÞ f ðxÞdx ¼ kn for n ¼ 0; 1; . . . ; N ð23Þ ment constraints.

where /n(x), n = 0, 1, . . . , N with /0(x) = 1, are the known functions 3. Wind power density
for the system; kn , n = 0, 1, . . . , N with k0 ¼ 1, are the expectation
data. Wind power density is proportional to the cube of wind speed,
The classic solution of the maximum entropy problem can be for a specified theoretical probability distribution f(v), it can be cal-
written by: culated by the following integration:
" # Z 1
X
N 1
P1 ¼ qv 3 f ðv Þdv ð29Þ
f ðxÞ ¼ exp  an /n ðxÞ ð24Þ 2 0
n¼0
As for actual time-series data, wind power density is calculated
where an are the Lagrange multipliers that can be obtained by solv- by:
ing the following (N + 1) nonlinear equations: 1 3
Z " # P2 ¼ qv ð30Þ
X
N 2
Z n ðaÞ ¼ /n ðxÞ exp  an /n ðxÞ dx ¼ kn for n ¼ 0; 1; . . . ; N
n¼0
where q is the air density; v 3 is the mean of wind speed cubes.
ð25Þ
4. Accuracy judgment criteria
For the analysis of wind distribution, we can let /n(x) be the
powers of wind speed (v) such that: To show how a theoretical probability function matches with
the observation data, four kinds of statistical errors are considered
/n ðv Þ ¼ v for n ¼ 0; 1; . . . ; N
n
ð26Þ
as the judgment criteria. Generally the smaller the errors the better
Then kn , n = 0, 1, . . . , N with k0 ¼ 1, are the moments of the distribu- the fit is.
tion representing the mean values of n powers of wind speed, which The first one is the Kolmogorov–Smirnov test (KS), which is de-
can be calculated from wind observation data. Consequently the fined as the max error in cumulative distribution functions [36].
corresponding expressions are given by: KS ¼ max jTðv Þ  Oðv Þj ð31Þ
" #
X
N where T(v) and O(v) are the cumulative distribution functions (cdf)
f ðv Þ ¼ exp  am v m ð27Þ for wind speed not exceeding v in the theoretical and observed data
m¼0 set, respectively.

0.16
observed
Station A GW pdf
0.14 NN pdf
NW pdf
WW pdf
0.12 MEP pdf
W pdf
GW cdf
Wind speed frequency

0.1 NN cdf
NW cdf
WW cdf
MEP cdf
0.08
W cdf
Cumulative distribution function

observed cdf

0.06 1

0.8
0.04
0.6

0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 2. Annual wind speed distributions and related cumulative distribution functions for station A.
1852 T.P. Chang / Applied Energy 88 (2011) 1848–1856

The second one is the root mean square error (RMSE) defined as: where yi is the actual value at time stage i, yic is the value computed
" #1=2 from correlation expression for the same stage, n is the number of
1X n
data.
RMSE ¼ ðyi  yic Þ2 ð32Þ
n i¼1 The third judgment criterion is the Chi-square error given as:

0.16
observed
Station B GW pdf
0.14 NN pdf
NW pdf
WW pdf
0.12 MEP pdf
W pdf
GW cdf
Wind speed frequency

0.1 NN cdf
NW cdf
WW cdf
MEP cdf
0.08
W cdf

Cumulative distribution function


observed cdf

0.06 1

0.8
0.04
0.6

0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 3. Annual wind speed distributions and related cumulative distribution functions for station B.

0.16
observed
Station C GW pdf
0.14 NN pdf
NW pdf
WW pdf
0.12 MEP pdf
W pdf
GW cdf
Wind speed frequency

0.1 NN cdf
NW cdf
WW cdf
MEP cdf
0.08
W cdf
Cumulative distribution function

observed cdf

0.06 1

0.8
0.04
0.6

0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 4. Annual wind speed distributions and related cumulative distribution functions for station C.
T.P. Chang / Applied Energy 88 (2011) 1848–1856 1853

n 
X 
1 presenting relatively smaller statistical errors; the conventional
v2 ¼ ðyi  yic Þ2 ð33Þ
yic Weibull function seems to be unsuitable for describing the distri-
i¼1
butions. For example, at station B (Fig. 3), the appearance probabil-
The fourth judgment is the relative percent error between the wind ity of wind speed 4–7 m/s and 17–22 m/s is underestimated,
potential energy calculated from actual time-series data and that whereas it is overestimated for speed 8–12 m/s when using the
from theoretical probability function, which might be most mean- conventional Weibull pdf. The relative percent errors of potential
ingful in wind energy assessment. energy for the conventional Weibull pdf reach 5.4% and 5.2% above
for the stations B and C, respectively. Concerning the max error in
5. Results and discussion cumulative distribution function, the proposed GW pdf has the
smallest one followed by the NW pdf and WW pdf, whereas the
To illustrate the suitability of the probability density functions NN pdf gets the largest except the Weibull pdf. If RMSE or Chi-
presented, various comparisons have been made based on mea- square error is considered, there is no significant superiority found
sured wind speed data. Figs. 2–4 show the annual wind speed fre- for any particular mixture function and/or maximum entropy func-
quency distributions for three stations experiencing different tion, while the Weibull pdf performs poorly for both stations. This
weather conditions in Taiwan. The curves of cumulative distribu- situation is much similar to those concluded by Zhou et al. [34] as
tion functions are plotted and referred to the right ordinate. Table well as by Akpinar and Akpinar [27], that means different judg-
1 lists the relevant parameter values computed for different prob- ment criterions may lead to inconsistent ranking-orders of fit per-
ability functions for the three stations. Various statistical errors are formance among various candidate pdfs.
summarized in Table 2. Figs. 5–7 show the annual wind power frequency distributions
It is shown that for station A, where only one hump is found on for the three stations. These figures exhibit similar results with
it, all the probability functions match well with the observed histo- those of wind speed, i.e. both the mixture functions and the max-
gram; the cumulative distribution functions are quite consistent imum entropy function outperform the conventional Weibull func-
with that of observed one. As summarized in Table 2, for station tion for stations where wind regimes have bimodal distributions,
A, the discrepancy of statistical errors among different functions especially for high speed range that is quite crucial for wind energy
is basically not very significant. application as concluded by Li and Li [28]. Among the functions,
While for station B and station C, in which wind regimes exhibit the MEP pdf and GW pdf perform best followed by the WW pdf
two significant humps on it, the mixture functions and the maxi- and NW pdf; since MEP pdf used in this paper has four moment
mum entropy function match still very well with the observations constraints of wind speed. As illustrated in Refs. [22,28,31], the

Table 1
Parameter values computed for different probability density functions.

Stations Gamma–Weibull Normal–Normal Normal–Weibull


w f b (m/s) k c (m/s) w l1 (m/s) r1 (m/s) l2 (m/s) r2 (m/s) w l (m/s) r (m/s) k c (m/s)
Station A 0.4125 11.2939 0.9334 2.3746 6.1607 0.4184 10.2836 3.4868 5.5895 2.4092 0.0397 15.6317 1.3950 2.2602 8.1553
Station B 0.3779 20.3196 0.8255 2.2481 6.6478 0.5045 14.8898 4.7167 5.0151 2.0723 0.4929 4.9779 2.0624 3.5716 16.5335
Station C 0.4117 3.8510 1.0085 4.4964 12.0941 0.6699 10.4947 3.0671 3.2109 1.2899 0.6007 11.0237 2.7400 2.3366 4.1546

Stations Weibull–Weibull Maximum entropy principle Weibull


w k1 k2 c1 (m/s) c2 (m/s) a0 a1 a2 a3 a4 k c (m/s)
Station A 0.0468 9.980 2.263 15.949 8.109 5.10085163 1.27133026 0.18858305 0.01110984 0.00025591 2.150 8.539
Station B 0.4426 4.151 2.354 17.334 6.149 5.02276188 1.27182598 0.20579344 0.01212952 0.00024082 1.689 11.231
Station C 0.6375 4.166 2.445 11.824 3.880 4.46372824 1.37253895 0.31984164 0.02864190 0.00086949 1.944 9.115

Table 2
Statistical errors for different probability density functions.

Stations Functions Wind speed Power density Potential energy


Max error RMSE v2 Max error RMSE v2 Percent error (%)
Station A GW 0.01260 0.00515 0.01327 0.02745 0.00756 0.03768 0.0426
NN 0.06139 0.00702 0.01856 0.03484 0.00853 0.04392 0.1480
NW 0.01282 0.00603 0.01642 0.03627 0.00878 0.03911 0.0965
WW 0.02218 0.00592 0.01209 0.02829 0.00764 0.03851 0.0452
MEP 0.02346 0.00584 0.02203 0.02666 0.00701 0.02769 1.6937e005
W 0.05418 0.00628 0.01699 0.04261 0.00976 0.07022 0.1780
Station B GW 0.01594 0.00712 0.10830 0.03554 0.00813 0.08209 0.0061
NN 0.04271 0.01284 0.12560 0.05165 0.01128 0.09552 0.2140
NW 0.01673 0.00827 0.09836 0.04804 0.00990 0.07539 0.1267
WW 0.01681 0.00759 0.04219 0.04209 0.00891 0.08362 0.0721
MEP 0.01998 0.00845 0.03829 0.03237 0.00785 0.07293 1.5773e006
W 0.10716 0.01342 0.14683 0.14007 0.02126 0.32164 5.4088
Station C GW 0.00772 0.00842 0.06213 0.01986 0.00431 0.00584 0.0214
NN 0.04065 0.00912 0.07144 0.04331 0.00873 0.04302 0.3044
NW 0.03004 0.00867 0.06129 0.04289 0.01160 0.01984 0.1157
WW 0.03120 0.00832 0.05861 0.04144 0.01493 0.01743 0.0623
MEP 0.03245 0.00820 0.06095 0.03886 0.01064 0.00473 6.3593e006
W 0.10541 0.01806 0.19421 0.23255 0.02729 0.40675 5.1927
1854 T.P. Chang / Applied Energy 88 (2011) 1848–1856

0.2
observed
Station A GW pdf
0.18
NN pdf
NW pdf
0.16 WW pdf
MEP pdf
W pdf
0.14
GW cdf
Wind power frequency

NN cdf
0.12 NW cdf
WW cdf
MEP cdf
0.1 W cdf
observed cdf

Cumulative distribution function


0.08
1

0.06 0.8

0.6
0.04
0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 5. Annual wind power density distributions and related cumulative distribution functions for station A.

0.2

Station B
0.18
observed
0.16 GW pdf
NN pdf
NW pdf
0.14 WW pdf
Wind power frequency

MEP pdf
W pdf
0.12
GW cdf
NN cdf
0.1 NW cdf
WW cdf
Cumulative distribution function

MEP cdf
0.08
W cdf 1
observed cdf
0.06 0.8

0.6
0.04
0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 6. Annual wind power density distributions and related cumulative distribution functions for station B.

higher the considered order is, the better the MEP pdf fits the data; characterizations better than the Weibull-series pdfs [28,29,32],
generally a quite good pdf can be generated when using four mo- the distribution curve may start with a non-zero value when wind
ment constraints [1]. Besides the MEP pdf describes null wind speed approaches zero. The percent errors of potential energy cal-
T.P. Chang / Applied Energy 88 (2011) 1848–1856 1855

0.2
observed
Station C GW pdf
0.18 NN pdf
NW pdf
0.16 WW pdf
MEP pdf
W pdf
Wind power frequency 0.14 GW cdf
NN cdf
0.12 NW cdf
WW cdf
MEP cdf
0.1 W cdf
observed cdf

Cumulative distribution function


0.08
1

0.06 0.8

0.6
0.04
0.4
0.02
0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Wind speed (m/s)
Fig. 7. Annual wind power density distributions and related cumulative distribution functions for station C.

culated by using the MEP pdf are below 0.01%, independent of sta- whereas the NN pdf performs worst except the Weibull
tions. However, some disadvantages concerning the MEP pdf must pdf. It is found that no particular mixture pdfs or MEP pdf
be noted here; for example, its mathematical expression is more outperform others in terms of the analysis of RMSE or Chi-
complex to establish computer program, the determination of square error. As for wind power density, the MEP and GW
function parameters needs more computation time, and its cumu- pdfs describe best followed by the WW and NW pdfs espe-
lative distribution function could not be expressed in a close form. cially in high speed ranges; the relative percent errors of
On the other hand, as summarized in Table 2, for a given prob- potential energy calculated by using the MEP pdf are even
ability function, the statistical errors of wind power density are below 0.01% because the MEP pdf involves 4 moment con-
commonly larger than those of wind speed, implying that choosing straints of wind speed besides the appearance probability
an appropriate probability function based on the analysis of wind of null winds.
power density would be of importance for wind energy (c) Different judgment criterions may result in inconsistent
assessment. ranking-orders of fit performance. Generally speaking, the
GW pdf proposed by the present study shows good ability
6. Conclusions to estimate both wind speed and power density, and there-
fore it can be considered as a useful alternative to the con-
In this paper, the suitability of four mixture probability density ventional Weibull function for wind energy applications.
functions and a function derived with maximum entropy principle The performance differences between the WW pdf and NW
in describing wind characterizations was analyzed and compared pdf are not significant.
together with the conventional Weibull function. Among these,
both the mixture GW pdf and NN pdf were applied originally to
wind energy field. Judgment criteria regarding the suitability in- Acknowledgments
clude the max error in Kolmogorov–Smirnov test, root mean
square error, Chi-square error and relative error of wind potential The author would deeply thank the Taipower Company for pro-
energy. The conclusions can be summarized as follows: viding observation data and appreciate Dr. Wu C.F. and Dr. Huang
M.W., researchers of the Institute of Earth Sciences, Academia Sini-
(a) If wind speed distribution is unimodal as in station A stud- ca, Taiwan, and international reviewers for their helpful sugges-
ied, the superiority of the four mixture pdfs and the MEP tions. This work was supported partly by the National Science
pdf relative to the conventional Weibull pdf is not Council under the Contract of NSC99-2221-E-252-011.
significant.
(b) While the distribution is bimodal as in stations B and C stud- References
ied, all the mixture pdfs and the MEP pdf describe better
wind characterizations than the Weibull pdf, irrespective [1] Carta JA, Ramirez P, Velazquez S. A review of wind speed probability
of wind speed and power density analyses. For wind speed distributions used in wind energy analysis Case studies in the Canary
Islands. Renew Sustain Energy Rev 2009;13:933–55.
analyses, the proposed GW pdf performs best in the Kol- [2] Chang TP. Performance comparison of six numerical methods in estimating
mogorov–Smirnov test followed by the NW and WW pdfs, Weibull parameters for wind energy application. Appl Energy 2011;88:272–82.
1856 T.P. Chang / Applied Energy 88 (2011) 1848–1856

[3] Bekele G, Palm B. Wind energy potential assessment at four typical locations in [20] Jamil M, Parsa S, Majidi M. Wind power statistics and an evaluation of wind
Ethiopia. Appl Energy 2009;86:388–96. energy density. Renew Energy 1995;6(5):623–8.
[4] Kwon SD. Uncertainty analysis of wind energy potential assessment. Appl [21] Jaramillo OA, Borja MA. Wind speed analysis in La Ventosa, Mexico: a bimodal
Energy 2010;87:856–65. probability distribution case. Renew Energy 2004;29:1613–30.
[5] Fyrippis I, Axaopoulos PJ, Panayiotou G. Wind energy potential assessment in [22] Ramirez P, Carta JA. The use of wind probability distributions derived from the
Naxos Island, Greece. Appl Energy 2010;87:577–86. maximum entropy principle in the analysis of wind energy. A case study.
[6] Akdag SA, Bagiorgas HS, Mihalakakou G. Use of two-component Weibull Energy Convers Manage 2006;47:2564–77.
mixtures in the analysis of wind speed in the Eastern Mediterranean. Appl [23] Li M, Li X. On the probabilistic distribution of wind speeds: theoretical
Energy 2010;87:2566–73. development and comparison with data. Int J Exergy 2004;1:237–55.
[7] Celik AN, Makkawi A, Muneer T. Critical evaluation of wind speed frequency [24] Carta JA, Mentado D. A continuous bivariate model for wind power density and
distribution functions. J Renew Sustain Energy 2010;013102:1–16. wind turbine energy output estimations. Energy Convers Manage
[8] Jowder FAL. Wind power analysis and site matching of wind turbine 2007;48:420–32.
generators in Kingdom of Bahrain. Appl Energy 2009;86:538–45. [25] Carta JA, Ramirez P. Analysis of two-component mixture Weibull statistics for
[9] Raichle BW, Carson WR. Wind resource assessment of the Southern estimation of wind speed distributions. Renew Energy 2007;32:518–31.
Appalachian Ridges in the Southeastern United States. Renew Sustain Energy [26] Carta JA, Ramirez P. Use of finite mixture distribution models in the analysis of
Rev 2009;13:1104–10. wind energy in the Canarian Archipelago. Energy Convers Manage
[10] Ucar A, Balo F. Evaluation of wind energy potential and electricity generation 2007;48:281–91.
at six locations in Turkey. Appl Energy 2009;86:1864–72. [27] Akpinar S, Akpinar EK. Estimation of wind energy potential using finite
[11] Ucar A, Balo F. Investigation of wind characteristics and assessment of wind- mixture distribution models. Energy Convers Manage 2009;50:877–84.
generation potentiality in Uludag-Bursa, Turkey. Appl Energy 2009;86:333–9. [28] Li M, Li X. MEP-type distribution function: a better alternative to Weibull
[12] Kantar YM, Senoglu B. A comparative study for the location and scale function for wind speed distributions. Renew Energy 2005;30:1221–40.
parameters of the Weibull distribution with given shape parameter. Comput [29] Li M, Li X. Investigation of wind characteristics and assessment of wind energy
Geosci 2008;34:1900–9. potential for Waterloo region, Canada. Energy Convers Manage
[13] Zhou W, Yang HX, Fang ZH. Wind power potential and characteristic analysis 2005;46:3014–33.
of the Pearl River Delta region, China. Renew Energy 2006;31:739–53. [30] Akpinar S, Akpinar EK. Wind energy analysis based on maximum entropy
[14] Genc A, Erisoglu M, Pekgor A, Oturanc G, Hepbasli A, Ulgen K. Estimation of principle (MEP)-type distribution function. Energy Convers Manage
wind power potential using Weibull distribution. Energy Sources 2007;48:1140–9.
2005;27:809–22. [31] Shamilov A, Kantar YM, Usta I. Use of MinMaxEnt distributions defined on
[15] Bechrakis DA, Deane JP, McKeogh EJ. Wind resource assessment of an area basis of MaxEnt method in wind power study. Energy Convers Manage
using short term data correlated to a long term data set. Sol Energy 2008;49:660–77.
2004;76:725–32. [32] Kantar YM, Usta I. Analysis of wind speed distributions: wind distribution
[16] Celik AN. On the distributional parameters used in assessment of the function derived from minimum cross entropy principles as better alternative
suitability of wind speed probability density functions. Energy Convers to Weibull function. Energy Convers Manage 2008;49:962–73.
Manage 2004;45:1735–47. [33] Lindfield G, Penny J. Numeric methods using MATLAB. 2nd ed. New
[17] Chang TJ, Wu YT, Hsu HY, Chu CR, Liao CM. Assessment of wind characteristics Jersey: Prentice Hall Inc; 2000.
and wind turbine characteristics in Taiwan. Renew Energy 2003;28:851–71. [34] Zhou J, Erdem E, Li G, Shi J. Comprehensive evaluation of wind speed
[18] Yang HX, Lu L, Burnett J. Weather data and probability analysis of hybrid distribution models: a case study for North Dakota sites. Energy Convers
photovoltaic-wind power generation systems in Hong Kong. Renew Energy Manage 2010;51:1449–58.
2003;28:1813–24. [35] Shannon CE. A mathematical theory of communication. Bell Syst Tech J
[19] Seguro JV, Lambert TW. Modern estimation of the parameters of the Weibull 1948;27:379–423.
wind speed distribution for wind energy analysis. J Wind Eng Ind Aerodynam [36] Sulaiman MY, Akaak AM, Wahab MA, Zakaria A, Sulaiman ZA, Suradi J. Wind
2000;85:75–84. characteristics of Oman. Energy 2002;27:35–46.

You might also like