INTRODUCTION
Most random analyses encompassing and modeling the wind effects on structures have assumed a Gaussian
process. This assumption was initially considered for convenience in analysis, since information on
statistics of Gaussian process is abundant. However, this assumption is no longer valid if the process is
non-Gaussian and; wind loads or response are characterized by skewness and kurtosis. For the wind
resistant design of a structure, the extreme value distribution plays an important role in estimating
appropriate wind loads, especially for the maximum response of the structure.
Current practice of estimating peak factor for calculating the peak wind loads on structures from
ASCE 7-05 Standard [1] AIJ standard [2] and other major codes is based on the assumption that they have
Gaussian distributions. S.O.Rice [3] discussed the statistical distribution of maxima, which depends only on
the standard deviation and the spectral width. He also showed the expected number of up-crossings per unit
time for Gaussian stationary stochastic processes. Cartwright and Longuet-Higgins [4] developed an
equation for the bandwidth parameter and applied it to water wave problems. They showed that the
bandwidth parameter can have values of 0 < < 1, where zero corresponds to narrowband process and one
corresponds to wide band processes. However, non-Gaussianity is not considered. Davenport [5, 6] showed
the distributions of largest values and derived a peak factor based on Gaussian and narrowband
assumptions. This peak factor shows good approximation when considering the Gaussian and narrowband
process. However, it under-predicts when the distribution is non-Gaussian or non-narrowband. Full
probability distributions have been estimated from non-Gaussian response moments with Gram Charlier
Edgeworth series, and Longuet-Higgins series distributions are discussed by Ochi [7]. Ochi gives the
probability density function in terms of skewness and kurtosis and good predictions for non-Gaussian
distributions are also shown. The method based on the expansion of a non-Gaussian variable in terms of
hermite polynomials involving a standard Gaussian process is discussed by Winterstein [8, 9]. For the
softening process, the relationship between standardized Gaussian and non-Gaussian relation is given by
Winterstein.
Kareem and Zhao [10] give the derivation of gust loading factor for tension leg platforms under
wind loads. This shows good prediction for any non-Gaussian response, but in some cases it may overpredict when considering the non-narrow bandwidth, for example normal stresses in columns in low- and
medium-rise building models, which tend to have wide bandwidth. Gurley et al [11] show wide application
of the analysis and simulation tools in wind engineering, and explain the non-Gaussian behavior and
derivations. Gurley and Kareem [12] show the non-Gaussian conditional simulation technique for velocity
and pressure fields. Wirsching and Light [13] show an improved model that can be used for design
purposes to predict metal fatigue under wide-band random stress processes. By this method, using a closed
form expression, fatigue damage under wide-band random stresses can be predicted directly from the
knowledge of the moments of the stress spectral density. Petrucci and Zuccarello [14] explain the fatigue
life prediction under wideband random loading. It shows the method of fatigue life prediction for Gaussian,
stationary and wide band random processes. Tamura et al [15] show the experimental results of peak
normal stress and a peak factor of more than 4.5, which cannot be calculated theoretically using available
procedures.
The
n = 0,1, 2,3.......
(1)
For non-Gaussian distribution, considering the skewness and kurtosis, Gram- Charlier probability
density function can be given as
N
P(u ) = ( x0 ) 1 + hn H n ( x0 )
(2)
n =1
where ( x0 ) is the standard Gaussian probability density function, ( x0 ) = (2 )1/ 2 exp( x02 / 2) ,
and the hermite polynomials are H n ( x0 ) . It can be shown that this method is applicable to any random
process with non-Gaussian behavior by considering the corresponding skewness and kurtosis coefficients.
For any non-Gaussian process, u(t) can be expressed in terms of the standard Gaussian process x(t) through
a monotonic function. Considering the softening process, the relation between the standardized nonGaussian and standard Gaussian process is given by
u = x + h3 x 2 1 + h4 x3 2 x
where
)}
= (1 + 2h32 + 6h42 ) 2 , h3 =
(3)
3
4 + 2 1 + 1.5 4
and h4 =
1 + 1.5 4 1
18
There are numerous efficient solutions for various problems based on the assumption that the PDF
is Gaussian. The Gram-Charlier series (2) is also good from this point of view, making it the most
frequently used non-Gaussian PDF.
The equation once solved using the hermite moments can give the equation for the non-Gaussian
probability density function with the skewness and kurtosis, as,
P( x) =
1 ( x x ) 2 3
(x x ) 4
( x x )
exp
H4
1 + H 3
+
2
x 4!
x
2 x
3!
(4)
where, 3 , 4 are the skewness and the kurtosis of the distribution minus three respectively. x is the
standard deviation. H3 and H4 are the hermite moments. For further explanation, the hermite moment of
softening process, 4 >3, is considered. An example for the non-Gaussian distribution is shown in Fig.1.
Bandwidth parameter
Cartwright & Longuet-Higgins [4] derived the bandwidth parameter based on the spectral
moments of random process. It is given by
m 2
(5)
= 1 2
m m
0 4
Here mr = n r S (n)dn , S ( n) is the power spectrum of the random function at the frequency n. For narrow
0
band process = 0 and it increases as the bandwidth increases, reaching 1 for the wide band process.
Examples for the narrow band and the non narrow band processes are shown in Fig.2.
GENERALIZED PEAK FACTOR
The expected maximum value for any random process x(t ) can be given by xmax = x + g . x
g = (2 log N ) +
(6)
2 log N
where g is the peak factor for the Gaussian and narrow band process, derived by Davenport (1964), and
1/ 2
N =T
m2
m0
exp x0 2
P( x0 ) =
(7)
2
2
q( x0 ) = P( x0 ) dx0
(8)
The cumulative probability distribution function Q( x0 ) , is defined as the probability of the function
exceeding some value
x0 , such that,
x0 (
1
1 2
1 2
2
q( x0 ) =
exp( 2 x )dx + ( 1 ) exp( 2 x0 )
2 x0
1 2 /
1 2
exp( x )dx
2
(9)
approximated as
Consider a sample of N maxima, the probability Pmax ( x0 ) that the largest of them has a value x0 , is the
probability that one of the maxima has this value and the rest are smaller, i.e.
Pmax ( x0 ) = N [1 q( x0 )]N 1 p( x0 )
(11)
q ( x0 ) = / N , where 0 N
(12)
For large values of N, we can write in the limit, the asymptotic form
Pmax ( x0 ) = d (1 / N ) N 1
= d exp( )
(13)
= d exp[ Nq ( x0 )]
Then, the maxima during the period when N is large is given by S.O.Rice [3] as
1/ 2
m
N = 4 T
m2
where T is the time period. Then, Equations (10) , (11) and (14) give the Equation
1
= N (1 2 )1/ 2 exp( x0 2 )
x0 max =
xP
0 max
( x0 )dx0
(14)
(15)
(16)
(17)
exp( )d
(18)
log
= (2 log(1 2 )1/ 2 N ) 1
1/ 2
log(1
)
N
log
= (2 log(1 2 )1/ 2 N )
2 log(1 2 )1/ 2 N
(19)
1
log 2
( 2 log(1
2 1/ 2
+ ..........
log exp( )d =
(20)
1/ 2
+ 2 log(1 2 )1/ 2 N
1/ 2
(21)
1/ 2
gb 2 log(1 2 )1/ 2 N +
1/ 2
2 1/ 2
2 log(1 ) N
g = b +
(22)
where b = (2 log(1 2 )1/ 2 N ) . Eqn (22) can be used for any bandwidth process in the frequency domain
to calculate the peak factor. It is not applicable to non-Gaussian processes. The following derivation
pertains to non-Gaussian processes, with the bandwidth parameter.
The effect of the non-Gaussian terms is included by employing the moment-based hermite
transformation approach. For the poisson model, the extreme distribution for the time period T is given as
x 2
(23)
Pmax ( x) = exp N exp 0
2
where x 0 (t ) = [ x x ] / x is the standardized Gaussian variable. The mean value of the positive maximum
values over the time T is given as
x0 max =
xP
0 max
( x0 )dx0
exp( )d
where
g is the peak factor and x is the standard deviation of the fluctuations of x and
Now, considering Eqn (23) and applying it to Eqn (24), we can write
= N (1 2 )1/ 2 exp( x0 2 )
(24)
(25)
Substituting in the Eqn (3) and (24) in Eqn (25) we can write
3 2
2
+
g ngb = b + + h3 b 2 + 2 1 + h4 b3 + 3b ( 1) +
b
b 12
2
(26)
where g ngb is the peak factor for Gaussian or non-Gaussian random processes with any bandwidth
parameter . Here b = (2 log
(1 2 ) N ) , = 0.5772 .
However Kareem & Zhao shows the non-Gaussian peak factor by considering the narrowband
approximation can be given as
32
2
g ng = + + h3 2 + 2 1 + h4 3 + 3 ( 1) +
+
12
2
(27)
where g ng is the non-Gaussian peak factor without considering the bandwidth parameter
. Here
= (2 log N ) , = 0.5772 .
PEAK FACTOR COMPARISON
Values of the calculated peak factor from the experimental results and the various theoretical
solutions are compared. Equation (6) shows the Davenports peak factor based on the assumption that the
random data is Gaussian and follows a narrowband process. Also the Cartwright & Longuet-Higgins peak
factor is for the Gaussian process and can have the bandwidth parameter shown in eqn (22). But these two
peak factors do not give a good comparison with the experimental results. As the non-Gaussianity is
dominant in these cases the non-Gaussian peak factor based on Kareem & Zhou gives better predictions.
But in this peak factor the bandwidth parameter is not considered. So the present derived peak factor for
non-Gaussianity and with the bandwidth parameter shown in eqn (26) shows good predictions with the
experimental results.
Fig.1 Example for the non-Gaussian distribution (Normal stress in the building column)
X(t)
(a)
X(t)
(b)
Fig.2 Example for the (a) narrow band ( 0 ) and (b) non narrow band processes ( 0 < <1 )
(a)
(b)
3.6
0.4
3.4
Skewne
ss
0.5
0.2
0.98
0.1
0.97
2.8
20
40
60
80
Sample number
2.6
0
100
0.99
3.2
Kurtos
is
0.3
Bandwidth parameter
Fig.3 (a) Pressure taps shown on the medium rise building model (B: D: H =
200:200:200m m, Scale = 1/250) (b) Structural model
20
40
60
80
Sample number
100
0.96
0
20
40
60
80
Sample number
100
Fig.4 (a) skewness (b) kurtosis (c) bandwidth parameter for the different samples shown (sorted
based on peak factor)
5.5
Davenport
Experimental
Peak factor
Peak factor
5.5
4.5
4
3.5
3
4.5
4
0
20
40
60
80
100
Sample numbers (sorted based one experimental results)
(b)
5.5
5.5
4.5
4
20
40
60
80
0
20
40
60
80
100
Sample numbers (sorted based one experimental results)
Present derivation
Experimental
4.5
4
3.5
Kareem-zhao
Experimental
3.5
(c)
Cartwright-Longuet Higgins
Experimental
3.5
Peak factor
Peak factor
(a)
100
(d)
0
20
40
60
80
100
Sample numbers (sorted based one experimental results)
Fig.5 Peak factor comparison between the experimental results and the various theoretical
solution for random pressures.
9
Davenport
Experimental
Peak factor
Peak factor
8
7
6
5
4
(a)
9
Peak factor
0
5
10
15
20
Sample numbers (sorted based on experimental results)
25
(b)
6
5
4
(c)
0
5
10
15
20
Sample numbers (sorted based on experimental results)
25
Kareem -Zhou
Experimental
Experimental
Peak factor
Present derivation
Experimental
6
5
4
0
5
10
15
20
Sample numbers (sorted based on experimental results)
25
(d)
0
5
10
15
20
Sample numbers (sorted based on experimental results)
Fig.6 Peak factor comparison between the experimental results and the various theoretical
solution for random normal stress acting on the corner columns of the building.
25