Attribution Non-Commercial (BY-NC)

236 tayangan

Attribution Non-Commercial (BY-NC)

- Final Draft of the Climate Science Special Report
- tif_ch07
- Combining Priors and Noisy Visual Cues in a Rapid Pointing Task
- Broersen2000_FiniteSampleCriteriaforAutoregressive
- ExamC Sample Questions
- Final Draft of the Climate Science Special Report
- Statistical Arbitrage
- Tutorial on PLLs_ Part 1
- Measurement in Variance
- Model Fit Almost
- 2004-11-29
- Asa266 Test Output
- pr_l4
- UNITEN ICCBT 08 Determination of Basic Wind Speed for Building Structures In
- 13-Subjective Probability and Expert Elicitation
- An Adventure of Epic Porpoises
- Non Lin Filt Neuro Morph Networks
- Chap05 Probability
- E 141 - 91 R03 _RTE0MQ__
- DEC 2016

Anda di halaman 1dari 18

and Industrial Aerodynamics 95 (2007) 165–182

www.elsevier.com/locate/jweia

speed estimation

Ying An1, M.D. Pandey

Department of Civil Engineering, University of Waterloo, Waterloo, Ont., Canada N2L 3G1

Received 14 November 2005; received in revised form 6 April 2006; accepted 22 May 2006

Available online 28 July 2006

Abstract

The paper presents the statistical estimation of extreme wind speed using annually r largest order

statistics (r-LOS) extracted from the time series of wind data. The method is based on a joint

generalized extreme value distribution of r-LOS derived from the theory of Poisson process. The

parameter estimation is based on the method of maximum likelihood. The hourly wind speed data

collected at 30 stations in Ontario, Canada, are analyzed in the paper. The results of r-LOS method

are compared with those obtained from the method of independent storms (MIS) and speciﬁcations

of the Canadian National Building Code (CNBC-1995). The CNBC estimates are apparently

conservative upper bound due to large sampling error associated with annual maxima analysis. Using

the r-LOS method, the paper shows that the wind pressure data can be suitably modelled by the

Gumbel distribution.

r 2006 Elsevier Ltd. All rights reserved.

Keywords: Wind speed; Extreme value estimation; Generalized extreme value distribution; Order statistics;

Annual maxima; Maximum likelihood method; Method of independent storm

1. Introduction

The estimation of design wind speed corresponding to a long return period is generally

based on the extreme value theory, which derives the three asymptotic domains of attraction,

namely, the Gumbel, Frechet and Weibull distributions [1]. These three distributions can be

written in a uniﬁed form, referred to as the Generalized Extreme Value distribution (GEV).

Corresponding author. Tel.: +1 519 888 4567x5858; fax: +1 519 888 4349.

E-mail address: mdpandey@uwaterloo.ca (M.D. Pandey).

1

Graduate student.

0167-6105/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.

doi:10.1016/j.jweia.2006.05.008

ARTICLE IN PRESS

166 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

Traditionally, a sample of annual maximum wind speed is ﬁtted with the Gumbel

distribution using the methods of moments or least squares. However, the statistical

extrapolation to estimate wind speed corresponding to 500–1000 year return period is

seriously contaminated by sampling and model uncertainty, if data are available for a

limited period (20–30 years). This has motivated the development of approaches to enlarge

the sample extreme values beyond the annual maxima.

The method of independent storms (MIS), proposed by Cook [2] and reﬁned by Harris

[3,4], considers several wind storm maxima, rather than just annual maxima. The extremes

of storm maxima are ﬁtted with the Gumbel distribution. The MIS method is limited to the

Gumbel distribution, and it discounts the possibility of GEV model representing the data.

The Peaks-Over-Threshold (POT) method is another alternative that models the peaks of

wind speed time series exceeding a threshold by the Generalized Pareto Distribution

(GPD), which is shown to be the domain of attraction of the peaks [5,6]. However, the

application of POT is confounded by an erratic variation of a quantile estimate with

respect to the threshold used in creating the sample of peaks [7].

The paper presents an alternate extreme value analysis of the Canadian wind speed data

that is based on estimation of the joint distribution of annually r largest order statistics

(r-LOS) of data. Assuming that r-LOS are generated by an underlying inhomogeneous

Poisson process, they can be modelled by a joint GEV distribution [8,9]. The paper

shows that the r-LOS method provides a systematic approach to (1) ascertain whether data

belong to the Gumbel or the GEV distribution, and (2) estimate the sampling error

associated with quantile estimates. Although the theoretical basis of the r-LOS method is

well established, the paper illustrates its versatility in the estimation of extreme wind speed.

The paper is organized as follows. A brief review of extreme value theory and MIS is

presented in Section 2. The proposed r-LOS method is described in Section 3. Section 4

presents a detailed analysis of wind data collected at 30 sites in Ontario (Canada) using

r-LOS and MIS methods. The wind speed quantile estimates are compared with the design

values speciﬁed in the Canadian National Building Code (CNBC) 1995. Section 5

summarizes the ﬁnding of this paper.

2.1. Background

Consider a sample of iid random variables, (X1, X2,y, Xn), with a cumulative

distribution (CDF) FX(x) and denote the maximum value in the sample as Mn ¼ max

(X1,yXn). The CDF of Mn can be obtained as

PrfM n oxg ¼ ½F X ðxÞn . (1)

According to extreme value theory, if there exist some constants an and bn, then the

distribution of extremes converges to a non-trivial result as

Pr ðM n bn Þ=an px ¼ ½F X ðan x þ bn Þn ! G X ðxÞ as n ! 1. (2)

The asymptotic distribution, GX(x), must converge to one of the three types of

distributions, namely, the Gumbel, Frechet and Weibull forms [1]. The three distributions

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 167

( )

xðx mÞ 1=x

G X ðxÞ ¼ exp 1 þ . (3)

s

where m, s and x are the location, scale and shape parameters, respectively. If x40, the

GEV is known as Type II (Frechet) distribution with an unbounded upper tail (m-s/

xoxoN). The case of xo0 is called the Type III (the reverse Weibull) distribution with a

ﬁnite upper tail (Noxom–s/x). As x-0, the Type I (Gumbel) distribution is obtained:

G X ðxÞ ¼ exp exp ðx mÞ=s . (4)

Since the GEV in a theoretical sense encompasses all the three types of extreme value

distributions, it has become a popular choice for extreme value analysis without any

presumption about the Gumbel distribution.

It should be noted that the GEV converges to the Gumbel distribution only in an

asymptotic sense, because Eq. (3) is not applicable to the case of x ¼ 0 due to a singularity.

This has an important practical consequence that a non-zero value of the shape parameter

obtained during the distribution ﬁtting cannot be accepted as it is. In fact, a statistical

signiﬁcance test is required to test whether or not a non-zero shape parameter is indicative

of the GEV or Gumbel distribution.

A key issue related to the application of extreme value distribution is the convergence, or

lack of it, of [FX(x)]n to the asymptotic distribution GX(x).

Cook [2] showed that the exponential parent distribution most rapidly converges to the

Gumbel form when nE100. However, the rate of convergence can be quite slow for other

parents. For example, there is a signiﬁcant departure between the tails of the Rayleigh

parent and Gumbel distributions for n as high as 10,000. For this reason, prior to the

extreme value analysis a suitable transformation of original data is desirable to bring it

closer to an exponential form. Assuming that the parent wind speed distribution

approximately follows the Rayleigh distribution, the dynamic pressure, which involves a

square of the wind speed, would be closer to the exponential distribution. Hence, extremes

of pressure should be ﬁtted with the Gumbel distribution, due to their much faster rate of

convergence than the wind speed data [2–4].

The empirical data analysis typically shows that the three parameter GEV distribution

provides a superior ﬁt to the data that follow a curve on the Gumbel plot due to added

ﬂexibility associated with an additional parameter. However, in many cases the GEV shape

parameter tends to be negative implying a bounded upper tail. It has sparked a debate due

to possible underestimation of extreme values by a bounded distribution. The proponents

of GEV justify its use due to its general form, better quality of empirical ﬁt and a possible

physical upper limit to wind speed. The opponents of the GEV argue that a lack of

convergence to the Gumbel distribution is manifested by the curvature of sample data (i.e.,

empirical distribution) on the Gumbel plot, and propose data transformation as a way to

accelerate the convergence.

The extreme value theory is based on solid mathematical foundation, though practical

applications are confounded by limited data that do not satisfy the conditions of

ARTICLE IN PRESS

168 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

asymptotic convergence. The r-LOS method will be applied to examine this issue in the

context of the Canadian wind speed data analysis.

It is a simple and straightforward method, adopted by many national design codes world

wide, in which the Gumbel distribution is used to ﬁt a sample of annual maximum wind

speed. The annual maxima are plotted on the Gumbel probability paper and parameters

are estimated from the method of least squares.

The design wind speeds speciﬁed in the CNBC 1995 is based on the Gumbel analysis of

annual maxima of wind speed data [11]. Given a sample of n year maxima, the method of

moments is used to estimate the Gumbel parameters [12]:

pﬃﬃﬃ.

s ¼ sx 6 p,

m ¼ x̄ 0:5772s, ð5Þ

where x̄ is the mean and sx is the standard deviation of the annual maxima data, and m and

s are location and scale parameter, respectively. In CNBC, the 30-year wind speed quantile

is chosen as a reference speed. The mean 30-year wind speed (V̄ 30 ) is ﬁrst estimated for

every station in Canada from the Gumbel analysis of annual maxima data, varying from

15 to 25 years in duration [12]. The data uncertainty is added to the mean estimate to

obtain the ﬁnal design value. pﬃﬃﬃ

The sampling error associated with the 30-year quantile was estimated as es ¼ 2:96sx = n

[12]. This formula is based on sampling error associated with the Gumbel quantile written

as a function of sample moments. The additional sources of error are climate variability,

siting uncertainty, anemometer height uncertainty and uncertainty associated with log law

for height correction, which were assumed to be 25%, 20%, 15% and 15% of the sampling

error (es), respectively. The ﬁnal data uncertainty (ed) was obtained by adding variance of

all these components, i.e, ed ¼ 1.0712 es. The 30-year wind speed and the associated error

were plotted on a map, and the ﬁnal extreme wind contour map was prepared using expert

judgement. The design speed appears to be speciﬁed as V30 ¼ V̄ 30 +ed, i.e., it corresponds

to 68% upper bound conﬁdence interval.

The design speed corresponding to any other T-year return period can be calculated in

terms of V30 as

T

V T ¼ V 30 0:7797sx 3:3843 þ lnðln . (6)

T 1

The CNBC provides an alternate formula in which a T- year quantile is expressed in

terms of 10 and 30-year quantiles as

" #

x10 x30 0:0339

xT ¼ x30 þ ln . (7)

1:1339 ln 1 1=T

This method enlarges the sample for extreme value analysis by including the wind storm

maxima, typically 100 storms per year. Through the examination of continuous records of

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 169

wind speed, independent wind storms are identiﬁed between each pair of lulls, and

maximum value within each storm is extracted to form a sample of storm maxima. The

wind speed data are converted into dynamic pressure. The top r order statistics of storm

maxima data are plotted on the Gumbel probability paper and extreme quantiles are

obtained by extrapolation of the straight line ﬁtted to the data. The main concepts of the

improved MIS, as discussed by Harris [3,4], are summarized below:

Suppose N independent storm maxima are extracted from S years of records, and FP(x)

denotes their CDF. The probability distribution of annual maxima can be given as

FA(x) ¼ [FP(x)]m, assuming that it is generated by independent storms with annual

occurrence rate of m ¼ N/S. Arranging the storm maxima in a decreasing order and

denoting a kth order statistic as Yk, such that Y1 being the largest and the YN being the

smallest value. Its probability density function (PDF) is given as [4]

N!

f Y k ðyk Þ ¼ ½F P ðyk ÞNk ½1 F P ðyk Þk1 f P ðyk Þ. (8)

ðk 1Þ!ðn kÞ!

Using the probability integral transformation, zk ¼ FP(yk), the PDF of the cumulative

frequency or the plotting position, zk, of Yk can be derived as

dyk

f Zk ðzk Þ ¼ f Y k fF 1

P ðzk Þg and dzk ¼ f P ðyk Þ dyk . (9)

dzk

N!

f Zk ðzk Þ ¼ ½zk Nk ½1 zk Þk1 . (10)

ðk 1Þ!ðn kÞ!

R1

E½zk ¼ 0 zk f Zk ðzk Þ dzk ¼ (N-k+1)/(N+1), where E[ ] denotes the expectation operation.

Since FA(x) ¼ [FP(x)]m ¼ zm is assumed to converge to the Gumbel distribution, Harris [4]

recommended to plot the expected value of the

n o Gumbel plotting positions

m

R1

E½ lnf lnðzk Þ g ¼ 0 ½ lnð ln zk Þ ln mf Zk ðzk Þ dzk against the order statistics,

Yk, of the storm maxima. The annual storm rate, m, has fairly small effect in analysis,

because ln(m) is involved in the plotting position formula and its value doe not change

much even with large variations in the value of m.

Harris also presented a computer program for accurate computation of the

mean and variance of the Gumbel plotting positions for various values of k. Another

reﬁnement that Harris proposed was to use the weighted least-squares method to ﬁt a

straight line to the data on the Gumbel plot, instead of the classical least squares. The

weight associated with kth plotting position is taken as inversely proportional to its

standard deviation [4].

In the MIS method, only r top-order statistics out of N storm maxima are plotted on the

Gumbel paper. The value of r is recommended so that approximately 2.5 observations per

year can be included in the plot. In summary, MIS method uses a larger data set than

annual maxima, employs preconditioning of data (i.e., square transformation) and plots

them consistently on the Gumbel paper.

ARTICLE IN PRESS

170 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

3.1. Analysis

This method selects r largest observation in each year of the data collected and derives

their joint distribution based on the theory of the Poisson process [13]. As seen from

Eqs. (2) and (3) the sample extreme value distribution, [FX(x)]n, asymptotically converge to

the GEV as

n 1=x o

½F X ðuÞn ¼ G X ðxÞ ¼ exp 1 þ xðu mÞ=s (11)

n 1=x o

n log½F X ðuÞ ¼ 1 þ xðu mÞ=s . (12)

The probability that wind speed exceeds a threshold, u, is p ¼ [1FX(u)] and the number

of exceedances in n trials follows the binomial distribution with parameters n and p. As

n ! 1 and p-0, then np-L, a constant. The distribution of number of exceedances

converges to the Poisson distribution with L as the intensity measure. Using a ﬁrst-order

approximation, log½F X ðuÞ ½1 F X ðuÞ, the Poisson intensity measure can be obtained

from Eq. (12) as

1=x

n log½F X ðuÞ n½1 F X ðuÞ3np ! LðuÞ ¼ 1 þ xðu mÞ=s ðn ! 1; p ! 0Þ.

(13)

Denoting the rth LOS in a sample of n as M ðrÞ

n , its distribution, Gr(y), can be related to the

Poisson distribution as

X

r1

½LðyÞk

n oy ¼ G r ðyÞ ¼

P½M ðrÞ exp Lð yÞ . (14)

k¼0

k!

(r1) order statistics (OS) of the sample. The increased information is expected to reduce

the uncertainty from estimation. In a special case of r ¼ 1, Eq. (14) reduces to the GEV

distribution.

The joint distribution of an r-LOS vector (Xr5Xr1,y,oX1) can be derived using the

theory of Poisson process as

Y

r

1

f 1r ðx1 ; . . . ; xr Þ ¼ exp½Lðxr Þ ½Lðxk Þ1þx . (15)

k¼1

s

For details of the derivation of this distribution, readers are referred to [9, Chapter 7].

The joint distribution (16) is the basis for inference, and is referred to as the r-LOS model.

As x-0, the Poisson rate parameter in limit converges to

h
y mi

Lð yÞ ¼ exp ðx ! 0Þ (16)

s

which is the Gumbel analog of the r-LOS model.

The r-LOS model has emerged as a versatile method for extreme value analysis in many

areas of science and engineering. It has been effectively applied for the estimation of

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 171

extremes of sea levels [5], wave heights [14] and rainfall [15]. A monograph by Coles

illustrates several interesting applications of this method [9]. This paper presents for the

ﬁrst time a comprehensive application of r-LOS model to wind speed estimation problem.

The joint distribution of r-LOS given by Eq. (15) provides a basis for maximum

likelihood (ML) method. Suppose j years of data are available and r-LOS are extracted

from each year of data. It is reasonable to assume that wind speed data from separate years

are independent, which makes a vector of r-LOS as independent and identically distributed

across j years. Thus, the likelihood function can be constructed as a product of the j

densities (15) corresponding to the observed sample:

( )

Y

j Y

r

1þx

Lðm; s; xÞ ¼ exp½Lðxr;i Þ ½Lðxk;i Þ . (17)

i¼1 k¼1

Here, xk,i denotes the kth largest OS in ith year of wind data. The distribution parameters

are obtained by maximizing the log-likelihood function. In a special case of r ¼ 1, it

reduces to the GEV model for annual maxima. A quantile estimate corresponding to a T

year return period is obtained as

sn x o

xT ¼ m 1 log 1 1=T . (18)

x

A given data set can be ﬁtted with either joint GEV (i.e., x6¼0) or Gumbel (i.e., x ¼ 0)

version of the r-LOS model. A formal statistical test, referred to as the likelihood ratio test,

can be applied to discriminate between the Gumbel or GEV model [15]. Let L1 denote the

ML associated with the GEV assumption (i.e., x6¼0) and ML associated with the Gumbel

model, obtained by setting x ¼ 0, denoted as L2. The likelihood ratio is deﬁned as

D ¼ 2 log L2 =L1 . (19)

It is distributed as a w2 variable with one degree of freedom. Thus, at the 5% signiﬁcance

level, the Gumbel model would be preferred if Dow21;0:95 3:841 [15]. This criterion is used

in the paper to select the appropriate distribution.

Smith [8] showed that the inverse of the information matrix is a good approximation of

the variance–covariance matrix of the ML estimators. The information matrix is the

Hessian of the negative log-likelihood evaluated at the maximum of the likelihood

function. The variance of a quantile value is calculated in terms of the variance of the

distribution parameters based on the delta method [16]. The estimated variance is used to

calculate the conﬁdence intervals for a quantile [9].

There is a bias-variance trade-off associated with the number of order statistics, r, used

in the analysis. A small value of r can result in large variance, but a large r is likely to cause

ARTICLE IN PRESS

172 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

the bias and violate the assumption of Poisson process generating the extreme values [8].

A practical criterion is that r should be selected such that it minimizes the variance

associated with a required quantile estimate. Tawn [17] also concluded that the results for

r ¼ 3–7 are very stable, showing that provided r is not too large this method leads to

consistent results. The analysis presented in Section 4 shows that r ¼ 5 is sufﬁcient to

provide minimum variance quantile estimates. Thus, r ¼ 5 is considered in the rest of the

analysis.

This section presents the analysis of time series of the aviation wind (HLY 01) recorded

by the Environment Canada at 30 stations in the Province of Ontario. The location and

other details of these stations are given in Table 1 and Fig. 1. The data for a site contain

Table 1

Information about wind stations in Ontario Canada

No. Name of sites Site ID Begin year End year Total years

2 Red Lake A 6016975 1953 2004 52

3 Kenora A 6034075 1953 2004 52

4 Sioux Lookout A 6037775 1953 2004 52

5 Geraldton A 6042716 1981 2004 24

6 Thunder Bay A 6048261 1954 2004 51

7 Sault Ste Marie A 6057592 1961 2004 44

8 Sudbury A 6068150 1954 2004 51

9 Earlton A 6072225 1953 2004 52

10 Kapuskasing A 6073975 1953 2004 52

11 Timmins 6078285 1955 2004 50

12 North Bay A 6085700 1953 2004 52

13 Gore Bay A 6092925 1953 2004 52

14 Kingston A 6104146 1967 2004 38

15 Ottawa M-C Int’l A 6106000 1954 2004 51

16 Petawawa A 6106398 1969 2004 36

17 Muskoka A 6115525 1953 2004 52

18 Wiarton 6119500 1953 2004 52

19 Sarnia A 6127514 1967 2004 38

20 St. Catharines A 6137287 1971 2004 34

21 Simcoe 6137730 1962 1986 25

22 Windsor 6139525 1953 2004 52

23 London A 6144475 1954 2004 51

24 Mount Forest 6145503 1962 1986 25

25 Waterloo Wellington A 6149387 1966 2002 37

26 Hamilton A 6153194 1970 2004 35

27 Toronto Island Airport 6158665 1957 2004 48

28 Toronto Pearson Airport 6158733 1953 2004 52

29 Trenton 6158875 1953 2004 52

30 Peterborough A 6166418 1969 2004 36

Note: The letter ‘A’ in the names of a site means that data were recorded at the local airport ﬁeld.

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 173

hourly wind speed, which is the 2-min average of the wind speed recorded just before the

hour. The largest daily wind speed is the maximum value of these hourly wind speeds

(24 values) within a day. The design wind estimates given in CNBC were derived from

these data [11].

The main assumption in the extreme value analysis is that the wind speed maxima are

independent. To reduce mutual dependence in the data, the annual time series of daily

maximum speed are partitioned into blocks that are equal to or larger than the duration of

typical storms in days (4–8 days). Using a procedure described by Simiu and Heckert [5],

new time series of 4-day independent maxima were created for each station. The wind

speed (V) time series were converted into dynamic pressure series using the relation

0.5rV2, where r ¼ 1.29 kg/m3 is the air density in Ontario [12].

The processed data sets were analyzed using the following three methods:

This method, as described in Section 3, was implemented using the algorithms provided

by Cole [9].

(2) Method of Independent Storms (MIS)

The computer code described by Harris [4] was programmed to implement this

method. The purpose is to compare its performance with r-LOS and CNBC estimates.

(3) Annual maxima Gumbel method (AMG)

The Gumbel analysis of annual maxima data using the method of moments was

performed to conﬁrm the estimates of CNBC [11].

ARTICLE IN PRESS

174 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

The steps involved in the extreme value estimation methods are explained through a

detailed analysis of Kingston station data. The 38 years of daily maximum hourly speed

data (13,750 values) were processed to create a time series of 3466 independent values of

4-day maximum wind pressure.

In the r-LOS method, a sample of size 38r was extracted from the data, and the analysis

was repeated for r 1–10. ML estimates of the three GEV parameters and the associated

standard errors (SEs) were calculated. The SEs associated with the shape parameters (x)

for various values of r are shown in Fig. 2. The variability decreases with an increase in r

up to 5, but there is no appreciable change in SE for r between 5 and 10. The SEs

associated with 50 and 500 year quantiles are shown in Fig. 3. A deceasing trend in SE is

seen with a modest increase in r, which levels off for r45. Therefore, an optimum choice of

r is expected to be close to 5. Similar to the work of Smith [5], we have chosen r ¼ 5 as a

ﬁxed value in the subsequent analysis. Thus, the r LOS sample for Kingston includes 190

values of wind pressure.

To apply the likelihood ratio test, the ML analysis was repeated under the assumption of

the Gumbel distribution (i.e., x ¼ 0). As shown in Table 2, the likelihood ratio test discerns

the Gumbel model as a statistically representative distribution of the wind pressure.

An interesting observation is made when the likelihood test was reapplied to wind speed

data, which showed a preference for the GEV distribution with x ¼ 0.0715 (Table 2). The

analysis of other stations, presented in the next section, also shows that the wind pressure

distribution is much closer to the Gumbel form than the wind speed data.

The MIS analysis included 88 top-order statistics out of 3466 storm maxima of the

dynamic pressure series, i.e., it uses approximately 2.3 maxima/year. The annual maxima

data (38 values) were also ﬁtted with the Gumbel distribution, and its parameters were

estimated from the method of moments. The Gumbel parameters estimated from the three

methods are compared in Table 3.

0.4

0.3

Shape Parameter of GEV

0.2

0.1

0.0

-0.1

-0.2

-0.3

0 1 2 3 4 5 6 7 8 9 10

r

Fig. 2. The GEV shape parameter with 95% conﬁdence interval (Kingston, Ont.).

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 175

105 105

50-year Design Speed (km/h)

100 100

95 95

90 90

85 85

80 80

75 75

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

r r

Fig. 3. 50 and 500 year wind speed with 95% conﬁdence interval (Kingston, Ont.).

Table 2

Likelihood ratio test applied to r-LOS data (Kingdom, Ont.)

r-LOS Gumbel r-LOS GEV test of model

Wind speed (km/h) 453.0 447.4 D ¼ 11.2443.841 GEV

Table 3

Parameters of the Gumbel distribution (Kingdom, Ont.)

MIS (Gumbel) 86.11 2.98

r-LOS (Gumbel) 85.86 1.97

Fig. 4 compares the wind speed quantiles for return periods ranging from 50 to 1000

years. It is interesting that the MIS and r-LOS curves are in close agreement in this case.

The design wind speeds speciﬁed in the CNBC are higher than those obtained from the

r-LOS and MIS methods. The reason, as discussed in Section 2.1, is the higher uncertainty

associated with estimates obtained from annual maxima data.

The data from other remaining 29 stations in Ontario were analyzed as described in the

previous section. The quantity of data used by each of the three methods varies, as shown

in Fig. 5. Some of the important trends observed from the results are discussed below.

ARTICLE IN PRESS

176 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

125

120

115

Quantile (Km/h)

110

105

100

95

90

NBCC 95

85

MIS

80 r-LOS

75

10 100 1000

Return Period (Year)

300 r-LOS

MIS

AMG

250

Number of Samples

200

150

100

50

0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 5. Comparison of the sample size used by different methods in Ontario data analysis.

The effect of considering r41 order statistics is the reduction in sampling uncertainty.

To illustrate this point, the SE associated with the r-LOS estimates of 500-year wind speed

are shown in Fig. 6. It is noteworthy that the SE on average is reduced by 50% as r is

increased from 1 to 5. It is a remarkable advantage of using the r-LOS method over the

classical annual maxima method.

For all 30 stations, the mean value and 95% conﬁdence intervals of the GEV shape

parameters estimated from the r-LOS (r ¼ 5) analysis of the wind pressure data are plotted

in Fig. 7. The mean shape parameter is mostly close to zero or in the positive range which

implies unbounded upper tail of the distribution. The results of the likelihood ratio test, as

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 177

5

r=1

Standard Error (km/h)

r=3

3

2

r=5

1

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 6. The number of order statistics (r) versus the standard error associated with 500-year speed.

0.4

0.3

0.2

Shape Parameter

0.1

0.0

-0.1

-0.2

-0.3

-0.4

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 7. The GEV shape parameter with 95% conﬁdence interval: wind pressure data (Ont., Canada).

shown in Fig. 8, conﬁrm that the Gumbel model is preferred for 27 stations. The GEV can

be used only for three stations numbered 1 (Wawa), 19 (Sarnia) and 28 (Mount Forest).

The estimated shape parameters for these stations are highly negative: 0.11, 0.12 and

0.19, respectively.

In contrast, the shape parameter (mean and 95% conﬁdence intervals) estimated from

the wind speed data is mostly negative (Fig. 9). The likelihood ratio test conﬁrms that 22

stations follow the GEV model (Fig. 10). The negative shape parameter means that the

upper tail is bounded at (ms/x), and it corresponds to the reverse Weibull (Type III)

domains of attraction. These results are in line with conclusions of the POT results

ARTICLE IN PRESS

178 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

20

18

16

14

Likelihood Ratio

12

10

6 GEV

Criteria Line = 3.841

4

2 Gumbel

0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 8. Results of likelihood ratio test: wind pressure data (Ont., Canada).

0.3

0.2

0.1

Shape Parameter

0.0

-0.1

-0.2

-0.3

-0.4

-0.5

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 9. The GEV shape parameter with 95% conﬁdence interval: wind speed data (Ont., Canada).

reported in the literature [5], in which the shape parameter for the generalized Pareto

distribution is predominantly negative.

On the other hand, the comparison of Figs. 8 and 10 appears to support the notion that

a square transformation of wind speed (to wind pressure) accelerates the convergence to

the Gumbel distribution [2,3,18].

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 179

20

18

16

14

Likelihood Ratio

12

10

6 GEV

Criteria Line =3.841

4

2 Gumbel

0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 10. Results of likelihood ratio test: Wind speed data (Ont., Canada).

110 CNBC 95

r-LOS

MIS

100

50-year Speed (km/h)

90

80

70

60

50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

The quantile estimates obtained from r-LOS analysis of wind pressure data are

compared with those obtained from MIS and CNBC speciﬁcations. The comparisons of

50-year and 500-year quantiles are given in Figs. 11 and 12, respectively.

It is interesting to note that mean estimates of 50-year wind speed obtained from MIS

and r–LOS are in close agreement, as the two methods are conceptually similar. However,

ARTICLE IN PRESS

180 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

130

CNBC95

r-LOS

MIS

500-year Speed (km / h)

110

90

70

50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

130

120

110

Quantile (km/h)

100

90

80

70 NBCC 95

r-LOS+1.07*Std.Err.

Mean of r-LOS

60

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Site No.

Fig. 13. Comparison of 500-year wind speed estimated from RLOS and NBC 95 methods.

r-LOS estimates of 500-year wind speed are slightly higher than the corresponding MIS

values.

For the sake of a consistent comparison, mean r-LOS estimates are converted into the

design values by adding the data uncertainty (ed ¼ 1.0712 es) similar to CNBC approach.

Fig. 13 shows that r-LOS estimates of 500-year design speed are much lower than CNBC

speciﬁcations. It suggests that the conservatism-associated CNBC can be reduced by

improved extreme value analysis of the wind speed data.

ARTICLE IN PRESS

Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182 181

5. Conclusions

The paper presents the statistical estimation of extreme wind speed using annually

r-LOS extracted from the time series of wind data. The method is based on a joint

generalized extreme value distribution of r largest order statistics derived from the theory

of Poisson process. The parameter estimation is based on the method of ML . A formal

likelihood ratio test is applied to discern between the GEV and Gumbel distribution. The

data collected at 30 stations in Ontario, Canada, are analyzed in the paper. The results

of r-LOS method are compared with those obtained from the MIS and speciﬁcations of the

CNBC.

The r-LOS analysis of the Ontario data shows that the Gumbel distribution is

statistically more preferable than the GEV distribution with a bounded tail due to negative

shape parameter. In this sense, the r-LOS results support the basis of the MIS method

which adopts the Gumbel distribution for wind pressure. However, the r-LOS method is

more versatile than the MIS method, as it also retains the ﬂexibility of adopting the GEV

distribution.

Acknowledgements

The authors are grateful to the Sciences and Engineering Research of Canada (NSERC)

and the University Network of Excellence for Nuclear Engineering (UNENE) for

providing the ﬁnancial support for this study. The authors are also thankful to the

Environment Canada for providing the wind speed data. The authors gratefully

acknowledge the use of computer programs provided by Coles for r-LOS model and

Simiu for ﬁltering of the data.

References

[1] E.J. Gumbel, Statistics of Extremes, Columbia University Press, Columbia, 1958.

[2] N.J. Cook, Towards better estimation of extreme winds, J. Wind Eng. Ind. Aerodyn. 9 (1982)

295–323.

[3] R.I. Harris, Gumbel re-visited—a new look at extreme value statistics applied to wind speeds, J. Wind Eng.

Ind. Aerodyn. 59 (1996) 1–22.

[4] R.I. Harris, Improvements to the method of independent storms, J. Wind Eng. Ind. Aerodyn. 80 (1999)

1–30.

[5] E. Simiu, N.A. Heckert, Extreme wind distribution tails: a peaks over threshold approach, J. Struct. Eng. 122

(1996) 539–547.

[6] J.I. Pickands, Statistical inference using extreme order statistics, Ann. Stat. 3 (1975) 119–131.

[7] M.D. Pandey, An adaptive exponential model for extreme wind speed estimation, J. Wind Eng. Ind.

Aerodyn. 90 (2002) 839–866.

[8] R.L. Smith, Extreme value theory based on the r largest annual events, J. Hydrol. 86 (1986) 27–43.

[9] S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer, Berlin, 2001.

[10] A.F. Jenkinson, The frequency distribution of the annual maximum (or minimum) of meteorological

elements, Quart. J. R. Meteorol. Soc. 81 (1955) 158–171.

[11] National Building Code of Canada, 1995, National Research Council.

[12] Yip, T.C., Auld, H., 1993. Updating the 1995 National Building Code of Canada wind pressures. In: Paper

presented to Canadian Electricity Association, pp. 1–9.

[13] I. Weissman, Estimation of parameters and large quantiles based on the k largest observations, J. Am. Stat.

Assoc. 73 (1978) 812–815.

ARTICLE IN PRESS

182 Y. An, M.D. Pandey / J. Wind Eng. Ind. Aerodyn. 95 (2007) 165–182

[14] C.G. Soares, M.G. Scotto, Application of the r largest-order statistics for long-term predictions of signiﬁcant

wave height, Coast. Eng. 51 (2004) 387–394.

[15] S. Nadarajah, Extremes of daily rainfall in west central Florida, Clim. Change 69 (2005) 325–342.

[16] G.W. Oehlert, A note on the delta method, Am. Statist. 46 (1992) 27–29.

[17] J.A. Tawn, An extreme-value theory model for dependent observations, J. Hydrol. 101 (1988)

227–250.

[18] A. Naess, Estimation of long return period design values for wind speeds, J. Eng. Mech. 124 (1998)

252–259.

- Final Draft of the Climate Science Special ReportDiunggah olehBenjamin Spillman
- tif_ch07Diunggah olehJDoe198811_219263216
- Combining Priors and Noisy Visual Cues in a Rapid Pointing TaskDiunggah olehdinomite
- Broersen2000_FiniteSampleCriteriaforAutoregressiveDiunggah olehAndrés Felipe Gómez
- ExamC Sample QuestionsDiunggah olehJessica Min Yea Sul
- Final Draft of the Climate Science Special ReportDiunggah olehBayAreaNewsGroup
- Statistical ArbitrageDiunggah olehhkless2
- Tutorial on PLLs_ Part 1Diunggah olehRoshdy AbdelRassoul
- Measurement in VarianceDiunggah olehfastford14
- Model Fit AlmostDiunggah olehAhsan Sheikh
- 2004-11-29Diunggah olehmounabs
- Asa266 Test OutputDiunggah olehmnbvcxz0987123
- pr_l4Diunggah olehengineer86
- UNITEN ICCBT 08 Determination of Basic Wind Speed for Building Structures InDiunggah olehlim_samuel2696
- 13-Subjective Probability and Expert ElicitationDiunggah olehgirellibruni
- An Adventure of Epic PorpoisesDiunggah olehbrkdance2003935
- Non Lin Filt Neuro Morph NetworksDiunggah olehAlim Sheikh
- Chap05 ProbabilityDiunggah olehIman Nurakhmad Fajar
- E 141 - 91 R03 _RTE0MQ__Diunggah olehGiorgos Siorentas
- DEC 2016Diunggah olehaakash mehta
- A Feasibility Study On The Prospects Of Establishing A Piggery Farm In Enugu-Ezike.pdfDiunggah olehJosette Ripalda
- Windata.pdfDiunggah olehSajith Chembekkara
- Macro Support Vector MachineDiunggah olehDwi Ratnasari
- MIT18_05S14_class10slides.pdfDiunggah olehIslamSharaf
- ECO 4000 R AssignmentDiunggah olehjl123123
- r2012-Thesis-Inference for a Gamma Step-stress Model Under Censoring_(Alkhalfan,BALA)Diunggah olehsss
- XYZ A Case-Study.pptxDiunggah olehSandipan Chowdhury
- Lecture 1 Introduction Use and Misuse of StatisticsDiunggah olehGurwinder Singh
- Accident Prediction Model and Mitigative Measures(FINAL)Diunggah olehIftiyaz Topinkatti
- Sampling PptDiunggah olehPulkit Sharma

- RingDiunggah olehShumayl Iman
- AIATS JEE Main 2016 ScheduleDiunggah olehAvinash Kanaujiya
- Darling - Differential FormsDiunggah olehPatrik Jakobsson
- 離散數學第一次作業1.1～1.4Diunggah olehyrugreen26
- chapter1(01-08-07)Diunggah olehJessie L. Labiste Jr.
- MF26Diunggah olehNg Shu Qing
- NguyenXuanHung ThesisDiunggah olehAndrés Mercado
- 4 LP Simplex Big MDiunggah olehRizal Ahmad
- lessonplanbooktemplate 2014 math 8Diunggah olehapi-260696748
- Section_3_7 (1).pdfDiunggah olehUmer Younas
- Orthogonal MatricesDiunggah olehHemant
- AsymptoticNotations.pptDiunggah olehmnkshgupta
- 1.5and2.1notesDiunggah olehg1191059
- Calculus - Early TranscendentalsDiunggah olehKofi Appiah-Danquah
- Non-linear Buckling and Post-buckling Analysis of Cylindrical Shells Subjected to Axial Compressive Loads: A Study on Imperfection Sensitivityl Shells Axial Load 2013 (JB Gunda)Diunggah olehShell Buckling
- Stability Assignment DynamicDiunggah olehtengyan
- GATE CSE Syllabus 2018 With Weightage (Topic Wise), Download PDF!Diunggah olehNag Raju
- Adjust CDiunggah olehFelipe Carvajal Rodríguez
- ExercisesDiunggah olehJimmy Granja
- Mathematics 1a Study Material 1Diunggah olehGautam Varma
- tyler nadeau resume 2018Diunggah olehapi-245089911
- MATH2061 PRACTICE SESSION 1Diunggah olehddddddda
- Project 2Diunggah olehA K
- Spiral TrainingDiunggah olehNoy
- Life_Assurance_Mathematics_W.F._ScottDiunggah olehMoses Njenga
- GannDiunggah olehDouglas J James
- Formula Sheet Final ExamDiunggah olehAli Amran
- Texto - Descartes and the Bird of Analytic GeometryDiunggah olehAna Carolina Pereira
- 1-s2.0-S0301679X11000752-main(1)Diunggah olehanandr50
- B.Tech (III-IV Semesters) in COMPUTER SCIENCE AND ENGINEERINGDiunggah olehJinu Madhavan