Bonfring International Journal of Data Mining, Vol. 2, No.

4, December 2012 11
Comparison of Estimators of Gumbel Distribution
for Modelling Wind Speed Data
N. Vivekanandan and Dr. S.K. Roy


Abstract--- Estimation of extreme wind speed potential at
a region is of importance while designing tall structures such
as cooling towers, stacks, transmission line towers, etc.
Assessment of wind speed in a region can expediently be
carried out by probabilistic modelling of historic wind speed
data using an appropriate extreme value distribution. This
paper illustrates the use of five parameter estimation methods
of Gumbel distribution for modelling Hourly Maximum Wind
Speed (HMWS) data recorded at Delhi and Visakhapatnam
regions. Goodness-of-Fit (GoF) tests involving Anderson-
Darling and Kolmogorov-Smirnov are used for checking the
adequacy of fitting of the method to the recorded data. Root
Mean Square Error (RMSE) is used for selection of a suitable
method for determination of parameters of Gumbel
distribution for modelling HMWS data. The results of GoF
tests and RMSE show that order statistics approach is better
suited for estimation of design wind speed for the regions
under study.
Keywords--- Anderson-Darling, Gumbel, Kolmogorov-
Smirnov, Mean Square Error, Order Statistics, Wind Speed

I. INTRODUCTION
Wind forces, and their static and dynamic effects, need to
be taken into account while designing buildings, structures and
their components thereof. If a structure is tall and slender, the
effect of wind on the structure can be critical. The distribution
of wind speed is also important in determining the
serviceability of buildings. The estimated wind speed at 1000-
year (yr) return period is often used to arrive at the design load
that a structure must withstand during its lifetime [1-2]. For
arriving at such design values, a standard procedure is to
analyse historic wind data over a period of time and arrive at
statistical estimates.
A theoretical analysis of extreme hydrologic phenomena
has led researchers to identify Gumbel distribution as a
standard distribution for frequency analysis of recorded
extreme meteorological data such as rainfall, temperature,
wind speed, evaporation, etc; and hence used in the present
study. Standard analytical procedures such as Method of

N. Vivekanandan, Assistant Research Officer, Central
Water and Power Research Station, Pune 411024,
Maharashtra. E-mail:[email protected].
Dr.S.K. Roy, Chief Research Officer, Central Water and
Power Research Station, Pune 411024, Maharashtra.

DOI: 10.9756/BIJDM.10079
Moments (MOM), Maximum Likelihood Method (MLM),
Method of Least Squares (MLS), Order Statistics Approach
(OSA) and Probability Weighted Moments (PWM) are
commonly available for determination of parameters of
Gumbel distribution [3]. Number of studies has been carried
out by different researchers on analysing the characteristics of
the parameter estimation methods of Gumbel distribution.
Research reports indicated that MOM is a natural and
relatively easy parameter estimation method. MLM is
considered the most efficient method, since it provides the
smallest sampling variance of the estimated parameters and
hence of the estimated quantiles compared to other methods.
But, the method has the disadvantage of frequently giving
biased estimates and often failed to give the desired accuracy
in estimating extremes from hydrological data. It may not
produce good estimators in small samples, especially when the
random variable is restricted to an interval that depends on the
parameters [4-6]. PWM and MLS are much less complicated,
and the computations are simpler. Parameter estimates from
small samples using PWM and MLS are sometimes more
accurate than the MLM estimates for Gumbel distribution. On
the other hand, OSA estimators are unbiased and having
minimum variance [7-10]. Since there is no general agreement
in applying particular method for a region because of the
characteristics of the estimators, an attempt is made to apply
five parameter estimation methods of Gumbel distribution for
modelling HMWS data recorded at Delhi and Visakhapatnam
regions. GoF tests involving Anderson-Darling (AD) and
Kolmogorov-Smirnov (KS) are employed for checking the
adequacy of fitting of the method to the recorded data.
Diagnostic analysis involving RMSE is used for selection of a
suitable method of Gumbel distribution for modelling HMWS
for the regions under study. The methodology adopted in
estimation of design wind speed using all five methods of
Gumbel distribution, GoF tests and diagnostic analysis are
briefly described in the ensuing sections.
II. METHODOLOGY
A) Probability Distribution
The Probability Density Function [PDF; f(W)] and
Cumulative Distribution Function [CDF; F(W)] of Gumbel
distribution is given by:
( )
( )

=


/
i
W
e /
i
W
e e
) W ( f
, W
i
, >0 (1)

( )

=
/
i
W
e
e ) W ( F (2)

ISSN 2277 - 5048 | 2012 Bonfring
Bonfring International Journal of Data Mining, Vol. 2, No.4, December 2012 12
where and are location and scale parameters of the
distribution [11]. The parameters are computed by MOM,
MLM, MLS, OSA and PWM; and further used to estimate
extreme wind speed (W
T
) for different return periods from
+ =
T T
Y W (3)
The Lower and Upper Confidence Limits (LCL and UCL)
of the estimated wind speeds at 95% level are computed from
LCL= W
T
-1.96(SE) and UCL= W
T
+1.96(SE). Here,
and SE is the Standard Error of the
estimated wind speed.
( ) ( ( T / 1 1 Ln Ln Y
T
= ))
A1) Methods for Estimation of Parameters
Method of Moments
= 5772157 . 0 W
and
( )
W
S 6 =
(4)
where Wand S
W
are the mean and standard deviation of
the recorded HMWS.
Maximum Likelihood Method
( ) ( )

=
=

=
N
1 i
i
N
1 i
i i
W exp W exp W W
and
( )

=
=
N W exp log
N
1 i
i
(5)
Method of Least Squares
( ) ( ) ( ) ( ( ) )

=
=

=
N
1 i
i
N
1 i
i
N
1 i
i i
N
1 i
2
i
2
N
1 i
i
P ln ln W P ln ln W N W N W

(6)



( ) ( )

+ =
=
N P ln ln W
N
1 i
i

(7)
where P
i
=(i-0.44)/(N+0.12) and ln(-ln(P
i
)) defines the
cumulative probability of non-exceedance for each W
i .

Order Statistics Approach
OSA is based on the assumption that the set of extreme
values constitutes a statistically independent series of
observations. The OSA estimators of Gumbel distribution are
given by:
'
M
' *
M
*
r r + = and (8)
'
M
' *
M
*
r r + =

where
*
r and
'
r are proportionality factors, which can be
obtained from the selected values of k, n and n

using the
relations as follows:
N / kn r
*
= and (9)

N / ' n r
'
=
Here N is the sample size containing the basic data that are
divided into k sub groups of n elements each leaving n
remainders. and are the distribution parameters of the
groups, and and are the parameters of the remainders,
if any. These can be computed from the following equations:
*
M

'
M

*
M

'
M

( ) ; S k 1
i
n
1 i
ni
*
M

=
=
i
;

' n
1 i
i ' n
'
M
W
=
=
( )
i
n
1 i
ni
*
M
S k 1
=
=
and (
w e

k
i

' n
1 i
i ' n
'
M
W
=
=
10)
her j=1,2,3,..,n.
Prob hted Mo
1 i
, ij i
W S
=
=
ability Weig ments
= 57721 . 0
100
57 M
and ( ) 2 ln M 2 M
101 100

(11)
=
where
W M
100
=
and
( ) ( ) ( )
N
1 i=
ch sam nding
r.
i
1 N N i N W
. Here i
the rank assigned to ea ple arranged in asce
orde

alues of SE of the estimated wind speed using MOM,
LM, MLS and PWM methods may be computed from the
101
M =
is
Computation of Standard Error
The v
M
following equation:
( )
5 . 0
2
T T
Y 1 . 1 Y 1919 . 0 1589 . 1
N
SE + +

=

(12)
The SE of the estimated wind speed by OSA can be
obtained from
( )
2 / 1
' n
'
n
*
W r W r SE + =
(13)
where
( )( )
2 *
r = ( )
2
' '
N n r = .
N kn k 1
and

or
W
n
and W
n
are
defined by the general f m as
( )
2
n G n
2
G n n
C Y B Y A + + =
.
W
The weights
ni

and
ni

used i OSA estima rs,

and values of A
n
, B
n
, and C
n
used r
OSA, are given in AERB safety gu
B) Goodness-of-Fit Tests
The AD and KS statistics are defined by:
n
in computing the SE fo
ide [12].
determining to
( ) ( ) ( )
( )
( ) ( ) ( ) { } + +
) i (
Z 1 Ln i 2 1 N 2 Z
(14)


where Z
(i)
=F(W
i
), for i=1 3,
Also, F (W)=(i-0.44)/(N+0.12) is the empirical CDF of W
i

and
=
N
Ln 1 i 2 N 1 N AD
=1 i
i
( ) ( ) ( )
i D i e
N
1 i
W F W F Max KS =
=
(15)

,2, ,N and W
1
<W
2
< .W
N
.
e i
( )
i D
W F is the co e computed
method
mputed CDF of W
i
. If th
values of GoF tests statistics given by the method are less than
that of theoretical values at the desired significance level, then
the is suitable for determination of parameters of
Gumbel distribution for modelling HMWS [13].
C) Diagnostic Test
Theoretical description of RMSE is given by:
RMSE=
( )
N
N 1

( )
2 / 1
2
*

(16)
where and are the recorded and estimated wind
spe observation given by G mbel
method provides minimum RMSE is considered as the most
sui tho st
B hodol describe ove, a computer
prog HWMS for
program computes the
parameters of Gumbel distribution by five methods, GoF tests
1 i
i i
W W


=

i
W
*
i
W
ed of i
th
u distribution. The
table me d for e imation of design wind speed [14].
III. RESULTS AND DISCUSSIONS
A) Estimation of Design Wind Speed
y adopting the met ogy d ab
ram was developed and used for modelling
Delhi and Visakhapatnam regions. The
ISSN 2277 - 5048 | 2012 Bonfring
Bonfring International Journal of Data Mining, Vol. 2, No.4, December 2012 13
ISSN 2277 - 5048 | 2012 Bonfring
and
e

eturn Periods together with SE gi n by Five Methods of
Gumbel Distribution for Delhi
RMSE. HMWS data recorded at Delhi for the period
1991-2003 and Visakhapatnam for the period 1987-1997 was
used to estimate the extreme wind speed for different return
periods adopting five methods of Gumbel distribution [15].
Tables 1 and 2 give the extreme wind speed estimates for
different return periods together with SE given by fiv
Table 1: Extreme Wind Speed Estimates for Different R
methods of Gumbel distribution for Delhi and Visakhapatnam.

ve
Return
period
Estimated extreme wind speed with SE (km/ hr) using
(yr)
MOM MLM OSA PWM MLS
W
T
SE W
T
SE W
T
SE SE W
T
SE W
T
2 55 3.1 55 3.6 55.5 3.3 .7 .6 3.1 55.8 3.4 55.5
5 6 6 7 7 7 6 .6 5.3 6 .6 4.8 6 .8 5.8 6 .8 5.7 6 .1 5.4
10 73.8 7.1 73.8 6.2 75.7 7.8 76.0 7.6 74.8 7.1
20 80.7 9.0 80.8 7.6 83.3 9.9 83.8 9.4 82.2 8.8
50 89.7 1 1 1 1 1.5 89.8 9.4 93.1 2.7 93.9 2.0 91.8 1.1
1 1 1 1 1 1 1 1 00 96.4 3.4 96.5 0.8 00.5 4.8 01.5 3.9 98.9 2.8
2 1 1 1 1 1 1 00 03.1 15.3 03.3 2.2 07.9 16.9 09.1 15.8 06.0 14.5
500 1 1 1 12.0 17.9 12.1 14.1 117.6 19.6 119.1 18.3 15.4 16.8
1000 118.7 19.8 118.8 15.5 124.9 21.7 126.6 20.3 122.6 18.6

T : Ex ind for D nt R o the SE gi y Fi thods
bel Di butio sakha nam
able 2 treme W Speed Estimates
Gum
iffere
stri
eturn Peri
or Vi
ds toge
pat
r with

ven b ve Me of
n f
Return
period
Estimated extreme wind speed with SE (km/ hr) using
(yr)
MOM OSA PWM MLM MLS
W
T
SE W
T
SE W
T
SE SE W
T
SE W
T
2 55 4.5 55 5.3 55 5.0 .9 .8 4.8 56. 5.0 55. 0 8 .6
5 0 .6 1 .3 2 .5 2 .4 1 .1 7 .3 7 7 .1 7 7 .1 8 7 .7 8 7 .5 8
10 79.9 10.3 81.2 9.4 82.8 11.5 83.8 11.1 82.1 10.6
2 1 1 1 1 1 0 89.1 3.0 90.9 1.5 93.0 4.5 94.6 3.8 92.3 3.1
50 1 1 1 1 1 1 01.0 16.6 03.4 4.3 06.2 18.5 08.5 17.4 05.4 16.5
1 1 1 1 1 1 00 10.0 19.4 12.8 16.4 16.2 21.5 18.9 20.1 15.2 19.1
200 118.9 22.1 122.2 18.5 126.0 24.6 129.3 22.8 125.0 21.7
500 130.6 25.8 134.6 21.4 139.1 28.7 142.9 26.5 137.9 25.1
1000 139.4 28.6 143.9 23.6 148.9 31.8 153.3 29.2 147.7 27.7

From Tables 1 and 2, it may be noted that the extreme
wind spee imate b stentl gher
wh
puted from Eqs. (14-15) using
o and given in Table 3.
d est s given y OSA are consi y hi
en compared with the corresponding values of other four
methods for both the regions.

B) Analysis Based on GoF Tests
oF tests statistics were com G
the estimators of Gumbel distributi n

Table 3: Computed Values of GoF Tests using Five Methods
of Gumbel Distribution for Delhi and Visakhapatnam
Method Computed values of GoF tests for
Delhi Visakhapatnam
AD KS AD KS
M 0.29 OM 5 0.173 0.449 0.205
MLM 0.293 0 0.173 .372 0.195
MLS 0.271 0.160 0.322 0.182
OSA 0.285 0.169 0.302 0.186
PWM 0.274 0.168 0.331 0.190

F ble 3, be that t put es of
GoF tests statisti en by all fiv ods of Gumbel
dis
able f term of e ors bel d ution
re unde dy.
Five Methods of
rom Ta it may
cs giv
noted he com
e meth
ed valu
tribution are less than the theoretical values (AD
0.05
=0.757;
KS
0.05,11
=0.410; KS
0.05,13
=0.377)

at five percent significance
level, and at this level, all five methods are found to be

C) Analysis Based on Diagnostic Test
The RMSE values were computed from Eq. (16) using the
estimators of Gumbel distribution and given in Table 4.

suit
for the
or de
gions
ination
r stu
stimat of Gum istrib
Table 4: RMSE values given by
Gumbel Distribution for Delhi and Visakhapatnam
Method RMSE (km/ hr)
Delhi Visakhapatnam
MOM 3.28 4.45
MLM 3.26 4.03
MLS 2.76 3.61
OSA 2.71 3.41
PWM 2.96 3.77

F ble 4, it may ted that the di ce between
the R values given LS, OSA an M are very
meag wever, by co ring the amou riation in
RM A is identifi the best sui ethod for
det
rom Ta be no fferen
MSE by M d PW
er. Ho
E, OS
nside
ed as
nt of va
table m S
ermination of parameters of Gumbel distribution for
estimation of design wind speed for both the regions. Figures
1 and 2 show the PDF plots of recorded and estimated wind
speeds given by Gumbel (using OSA) for Delhi and
Visakhapatnam.
Bonfring International Journal of Data Mining, Vol. 2, No.4, December 2012 14


Figure 1: Plot of recorded and Estimated Wind Speeds given
by Gumbel distribution (using OSA) for Delhi



Figure 2: Plot of Recorded and Estimated Wind Speeds gi en
by Gumbel Distribution (using SA) for Visakhapatnam
wi el
given b umbel distribution (using OSA) for Delhi and
Vis
v
O

Figures 3 and 4 show the plots of recorded and estimated
nd speeds together with confidence limits at 95% lev
y G
akhapatnam regions.



Figure 3: Plot of Recorded and Estimated Wind Speeds for
Different Return Periods given Gumbel Distribution (using
OSA) together with 95% confidence limits for Delhi
by

Figure 4: Plot of Recorded and Estimated Wind Speeds for
Different Return Periods given by Gumbel Distribution (usi
OSA) together with 95% confidence limits for Visakhapatn
est ated wind speed given by Gumbel distribution (using
OS
er aided procedure for
modelling HM im ti treme wind speeds
for Delhi and adopting Gumbel
dis
Power Research Station, Pune, for provid the research
facilities to carry out are thankful to
M/
y Agency (IAEA), Meteorological events in
site evaluation for nuclear power plants IAEA Safety Guide No. Ns-G-
3.4, International Atomic Energy Agency, Vienna, 2003.

ng
am

From Figures 3 and 4, it can be seen that the recorded
annual HMWS data falls within the confidence limits of the
im
A) for both the regions. The Correlation Coefficient (CC)
between the recorded and estimated wind speeds by all five
methods of Gumbel distribution is computed as 0.977 for
Delhi and 0.979 for Visakhapatnam.
IV. CONCLUSIONS
The paper presented a comput
WS data for est a on of ex
Visakhapatnam regions
tribution using MOM, MLM, MLS, OSA and PWM. GoF
tests results confirmed the use of five methods for
determination of parameters of Gumbel distribution for
modelling HMWS for both the regions. The diagnostic
analysis indicated that OSA is found to be an appropriate
method for estimation of design wind speed based on the
amount of variation on RMSE. The results showed that the
RMSE on the estimated wind speed given by Gumbel
distribution (using OSA) for Delhi and Visakhapatnam regions
are 2.71 km/ hr and 3.41 km/ hr respectively. The results also
showed that the CC values for Delhi and Visakhapatnam are
0.977 and 0.979 respectively. The study suggested that the
Mean+SE (where Mean denotes the estimated extreme wind
speed) values of about 147 km/ hr and 183 km/ hr related to
1000-yr return period may be adopted for design purposes of
buildings and structures at Delhi and Visakhapatnam regions.
ACKNOWLEDGEMENTS
The authors are grateful to the Director, Central Water and
ing
the study. The authors
s NPCIL, Mumbai, and India Meteorological Department,
Pune, for making available the wind speed data of Delhi and
Visakhapatnam regions.
REFERENCES
[1] International Atomic Energ
ISSN 2277 - 5048 | 2012 Bonfring
Bonfring International Journal of Data Mining, Vol. 2, No.4, December 2012 15
ISSN 2277 - 5048 | 2012 Bonfring
[2] J. Lieblein, Note on simp or Type I extreme value
in estimating
lified estimates f
distribution - NBSIR 75-647, National Bureau of Standards, U.S.
Department of Commerce, Washington D.C, 1974.
[3] J.M. Landwehr, N.C. Matalas, and J.R. Wallis, Probability weighted
moments compared with some traditional techniques
Gumbel parameters and quantiles, Water Resources Research, Vol. 15,
No. 5, Pp 1055-1064, 1979.
[4] J.A. Ranyal, and J.D. Salas, Estimation procedures for the type-1
extreme value distribution, Journal of Hydrology, Vol. 87, Nos. 3-4, Pp
315-336, 1986.
[5] K. Arora and V. P. Singh, On statistical intercomparison of EVI estimators
by Monte Carlo simulation, Advances in Water Resources, Vol. 10, No. 2,
Pp 87-107, 1987.
[6] H.N. Phien, A review of methods of parameter estimation for the
extreme value type1 distribution, Journal of Hydraulics, Vol. 90, Nos.
3-4, Pp 251-268, 1987.
[7] D. Manik and S.K. Datta, A comparative study of estimation of
extreme value, Journal of River Behaviour & Control, Vol. 25, No. 1,
Pp 41-47, 1998
[8] P.F. Rasmussen and N. Gautam, Alternative PWM-estimators of the
Gumbel distribution, Journal of Hydrology, Vol. 280, Nos. 1-4, Pp 265-
271, 2003.
[9] A.N. Celik, On the distributional parameters used in assessment of the
suitability of wind speed probability density functions, Energy
Conversion and Management, Vol. 45, No. 11-12, Pp 1735-1747, 2004.
[10] N. Vivekanandan, F.T. Mathew and Dr. S.K. Roy, Modelling of wind
speed data using probabilistic approach, Journal of Power and River
Valley Development, Vol. 62, No. 3-4, Pp 42-45, 2012.
[11] E.J. Gumbel, Statistic of Extremes,
2nd
Edition, Columbia University
Press, New York, USA, 1960.
Atomic Energy Regulatory Board (AERB), Extreme values of [12]
meteorological parameters AERB Safety Guide No. NF/SG/S-3,
2008.
J. Zhang, Powerful goodness-of-fit tests based on the li [13] kelihood ratio,
Journal of Royal Statistical Society, Vol. 64, No.2, Pp 281-294, 2002.
W. May, Variability and extr [14] emes of daily rainfall during the Indian
summer monsoon in the period 1901-1989, Global and Planetary Change,
Vol. 44, No. 1-2, Pp 83-105, 2004.
India [15] Meteorological Department (IMD), Report on availability of
meteorological data, 2010.