The Seventh Asia-Pacific Conference on

Wind Engineering, November 8-12, 2009,

Taipei, Taiwan

A PEAK FACTOR FOR PREDICTING NON-GAUSSIAN PEAK

RESULTANT RESPONSE OF WIND-EXCITED TALL BUILDINGS
M.F. Huang1, C.M. Chan2, K.C.S. Kwok3 and Wenjuan Lou4
1
Postdoctoral Researcher, Institute of Structural Engineering, Zhejiang University,
Hangzhou 310027, [Link]. E-mail: hmfust@[Link]
2
Associate Professor, Dept. of Civil and Environmental Engineering, The Hong Kong Univ. of
Science and Technology, Clear Water Bay, Kowloon, Hong Kong. E-mail: cecmchan@[Link]
3
Professor, School of Engineering, University of Western Sydney, NSW, Australia, E-mail:
[Link]@[Link]
4
Professor, Institute of Structural Engineering, Zhejiang University, Hangzhou 310027,
[Link]. E-mail: Louwj@[Link]
ABSTRACT
In the structural design of tall buildings, peak factors have been widely used to predict mean extreme responses
of tall buildings under wind excitations. Using the out-crossing theory for the first-passage problem, a
probabilistic peak factor, which is directly related to an explicit measure of structural reliability against a
Gaussian response process, can be obtained for structural design. Based on the asymptotic theory of statistical
extremes, a new closed-form peak factor, the so-called Gamma peak factor, can be also obtained for a non-
Gaussian resultant response, characterized by a Rayleigh distribution process. Using the Gamma peak factor, a
combined peak factor method has been developed for predicting the expected maximum resultant responses of a
building undergoing lateral-torsional motion. Utilizing the wind tunnel data derived from synchronous multi-
pressure measurements, the time history response analysis has been carried out for the 45-story CAARC
standard tall building. The Gamma peak factor has been applied to predict the peak resultant acceleration of the
building. Results of the building example indicate that the Gamma peak factor provides accurate prediction of
the mean extreme resultant acceleration response for the dynamic serviceability performance design of modern
tall buildings.
KEYWORDS: LEVEL-CROSSING RATE (LCR); TIME-VARIANT RELIABILITY; MEAN EXTREME
RESPONSE; PEAK FACTOR METHOD; DYNAMIC SERVICEABILITY
Introduction
Considering inherent random characteristics of wind and uncertainties in the material
properties and structural performance, much research work has also been carried out on the
probabilistic evaluation of wind-induced response of structures with uncertain parameters
[Kareem (1987), Solari (1997), Hong et al. 2001]. These studies about wind effects on
structures employed the classical work of Davenport to estimate the expected or mean value
of the largest peak response. Davenport (1964) has shown that, if the underlying parent
distribution of a response process is Gaussian, then the extreme values of the process will
asymptotically follow a Gumbel distribution. For a zero-mean response process, the so-called
peak factor, can be defined as the ratio of the largest peak response to the standard deviation
value of the response. In general, the Davenport’s peak factor provides satisfactory estimates
of the maximum peak response for wide-band response processes; but it may yield
conservative estimates for narrow-band response processes [Kareem (1987), Gurley et al.
(1997)].
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

The approach using the Davenport’s peak factor for estimating the expected maximum
peak responses of tall buildings to wind is based on the assumption that the underlying
stochastic response is Gaussian. Such an assumption is valid for many general wind
engineering applications. However, non-Gaussian wind effects may arise from specific but
important situations, such as responses of a non-linear building system, turbulence-induced
local pressure fluctuations on building surfaces, and combined resultant acceleration
responses of a tall building for dynamic serviceability design [Melbourne and Palmer 1992,
Isyumov et al. (1992), Chan et al. (2009)]. Specific efforts have been made to modify the
Davenport’s peak factor for predicting non-Gaussian gust and extreme effects either using
mathematical series representation of distribution functions or using statistical approach
[Gurley et al. (1997), Sadek and Simiu (2002), Holmes and Cochran (2003), Tieleman et al.
(2007)].
This paper firstly presents an analytical expression of the probabilistic peak factor,
which is directly related to an explicit measure of reliability (or the probability of random
peak responses without exceeding a certain threshold value). Secondly, based on the
asymptotic theory of statistical extremes, the so-called Gamma peak factor can be analytically
obtained for a non-Gaussian resultant response, characterized by a Rayleigh distribution
process. Using the Gamma peak factor, a combined peak factor method can then be developed
for predicting the expected maximum resultant acceleration responses of tall buildings under
wind excitations. Finally, the peak resultant acceleration responses of the CAARC building
were calculated using the Gamma peak factor, and compared with the time history response
data derived from wind tunnel based synchronous multi-pressure measurements.
Probabilistic Peak factor
The largest peak response over a given time duration τ can be defined as a new
random variable
Yτ = max { Y (t ) ; 0 ≤ t ≤ τ } (1)
Using the time-variant reliability R ( t ) from the Poisson model [Vanmarcke (1975)], for a
fixed time duration τ , the cumulative distribution function (CDF) of Yτ can be expressed as
  b2  
FYτ ( b ) = R (τ ) = exp  −v0τ exp  − 2   (2)
  2σ Y  
where b indicates a chosen response magnitude; v0 = σ Y& / (πσ Y ) , representing the mean zero-
crossing rate of the process Y (t ) . For a prescribed probability p of the largest peak response
being within the specific threshold over the time period τ , the corresponding response
threshold denoted as bτ , p , can be obtained by solving the following equation
  bτ2, p  
( )
P Yτ ≤ bτ , p = p = exp  −v0τ exp  − 2  
  2σ Y  
(3)
 
By taking logarithm transform twice at both sides of Eq. (3) and rearranging terms, one
obtains
bτ , p v0τ
= 2 ln (4)
σY ln (1/ p )
The ratio of bτ , p / σ Y given in Eq. (4) can be regarded as a peak factor. Based on the fact that
the largest extreme values of a Gaussian process asymptotically follow the Gumbel
distribution, Davenport (1964) developed the following peak factor for practical use
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

g f = 2ln v0τ + γ / 2 ln v0τ (5)

where the Euler’s constant γ = 0.5772 . For a narrow-band resonant response, the mean zero-
crossing rate v0 can be simply approximated by the natural frequency of a building, and the
observation time duration τ may be normally taken as 600s or 3600s in wind engineering
practice. It is worth noting that the Davenport’s peak factor is independent of spectral
bandwidth parameter. Based on extreme value theory, the probability of the largest peak
response not exceeding the expected maximum peak response can be evaluated by using the
Type I extreme value distribution (or the Gumbel distribution) as
−γ
P(Yτ ≤ g f σ Y ) = e− e = 0.5704 (6)
The Davenport’s peak factor was developed under the assumption that the out-
crossings constitute a Poisson model, which has been found to be too conservative when the
response Y (t ) is a narrowband process and the threshold level b is not high enough with
respect to the standard deviation value of the response. Furthermore, the consecutive out-
crossings of the response Y (t ) , cannot be realistically assumed as independent events, as they
tend to occur generally in clumps. Vanmarcke (1975) developed a corrected mean out-
crossing rate based on a modified Poisson model accounting for the dependence among the
crossing events as
 π 1.2 b 
1 − exp  − q 
 2 σY 
ηb = vb (7)
 b2 
1 − exp  − 2 
 2σ Y 
where q = 1 − λ12 / ( λ0λ2 ) =shape factor that characterizes the bandwidth of the process, in
which the spectral moments λm can be defined as
∞
λm = ∫ ω mGY (ω )dω ; m = 0,1, 2, 4 (8)
0
where GY (ω ) =one-sided power spectral density function of the process and one can show
that λ0 = σ Y2 , λ2 = σ Y2& and λ4 = σ Y&&2 .
Using the Vanmarcke’s corrected mean out-crossing rate of level b in Eq. (7), the CDF
of Yτ can be rewritten as
  π 1.2 b  
 1 − exp  − q 
  2 σY  
FYτ ( b ) = exp  −v0τ  (9)
 b2 
 exp  2  − 1 
  2σ Y  
Given a specific probability of no exceedance p of the largest peak response being within the
specific threshold bτ , p over the time period τ , Eq. (9) can be expressed as
  π 1.2  
 1 − exp  − q gp 
  2 
p = exp  −v0τ  (10)
  gp 
2

 exp   −1 
 2 
   
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

where g p is the so called probabilistic peak factor, which can be defined as the ratio of
bτ , p / σ Y . Taking logarithm at both sides of (10) and rearranging terms, one obtains
v0τ  
 π 1.2    g 2p
1 − exp  − q g p   = exp 
 −1 (11)
ln(1/ p )   2     2
 
( )
In light of the fact that exp g 2p / 2 >> 1 , Eq. (11) can be first rewritten as

v0τ  
 π 1.2    g 2p
1 − exp  −  q g 
p ≈ exp  (12)
ln(1/ p )   
2   2
 
Using Eq. (4) as a first approximation of g p and substituting it into the left side of Eq. (12),
the probabilistic peak factor g p can be explicitly expressed in terms of p, v0 , τ and q as
 v τ   v0τ   
g p ≈ 2 ln  0 1 − exp  − q1.2 π ln   (13)
ln(1/ p )   ln (1/ p ) 
     
The peak factors, calculated according to the Davenport’s peak factor given in Eq. (5)
and the probabilistic peak factor of Eq. (13), are plotted as a function of v0 within the range
of 0.1-1.1 Hz for typical multi-story building structures with several chosen values of
bandwidth shape factor q as shown in Figure 1. For the sake of comparison with the
conventional Davenport’s peak factor, the probabilistic peak factor has been given based on
Eq. (13) with the probability of no exceedance p=0.5704 and an excitation duration
time τ =3600s. It is evidently shown in Figure 1 that the Davenport’s peak factor is
independent of the spectral bandwidth parameter q, and always gives more conservative
results, particularly for a narrowband process with a smaller value of q. For a wide-band
process with the value of q approaching to 1, the probabilistic peak factor approaches to the
value of the Davenport’s peak factor.
4.3

4.1
Peak factors

3.9

Davenport's peak factor (p=57%)

3.7
Probabilistic peak factor (p=57%; q=0.7)

3.5 Probabilistic peak factor (p=57%; q=0.4)

Probabilistic peak factor (p=57%; q=0.2)

3.3
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
mean zero-crossing rate ν 0 (Hz)
Figure 1: Peak factors with spectral bandwidth parameter q
Peak distribution for a non-Gaussian combined random process
Under wind excitation, a building may vibrate in a lateral-torsional manner such that
the maximum resultant response may involve several component responses in a 3-D manner.
Assuming that the corner of a building experiences two perpendicular translational component
responses, X(t) and Y(t). Then the combined resultant process can be written as
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

A(t ) = X 2 ( t ) + Y 2 ( t ) (14)
Assuming that X(t) and Y(t) follow a Gaussian distribution with a zero-mean and a common
standard deviation of σ X = σ Y , then the resultant process A(t) follows a Rayleigh distribution,
which can be given as
a  a2 
f A ( a ) = 2 exp  − 2  (15)
σA  2σ A 
where σ A denotes the mode value of A(t). Since it is assumed that the two component
processes have the same standard deviation, the mode value of A(t) coincides with the
component standard deviation such that σ A = σ X = σ Y . As can be observed from full-scale
measurement data and recently demonstrated by means of a numerical study [Chen and
Huang (2009)], the actual peak resultant response attains the most deviation from the SRSS
combination of two individual peak components when σ X = σ Y . Since the most significant
joint action between two component responses X and Y occurs when σ X = σ Y , it becomes
necessary and meaningful to first investigate the joint action effects under the condition of
two random component processes with equal fluctuating variation (i.e. σ X = σ Y ) and then
later to extend to more general cases in which σ X ≠ σ Y .
The peak distribution of the non-Gaussian combined random process A(t) can be
determined by its mean level-crossing rate (LCR). For a stationary combined process A(t), the
probabilistic distribution function (PDF) of peaks of the combined process can be obtained as
the derivative of the frequentist definition of probability
1 dVb+
f Am (b) = − + (16)
V−∞ db
where Am denotes a random variable, which represents the peak values arising from a
combined process A(t).
For a narrowband process, each upcrossing event can possibly lead to a corresponding
peak. The expected number of peaks above the given threshold level b per second can then be
approximated by the level upcrossing rate vb+ . If the desired peaks are counted by the peak-
+
over-threshold approach, the expected number of total peaks per second V−∞ can be well
estimated from the mean upcrossing rate of a sufficiently small threshold level b such as mode
+
value σ A , i.e., V−∞ ≈ vb+ . The probability peak distribution of a non-Gaussian combined
b =σ A
process given in Eq. (16) can then be related to the mean LCR as
1 dvb+
f Am (b) ≈ − + (17)
vσ A db
Considering the Rayleigh resultant process A(t) with the known mean LCR vb+ [Huang
(2008)], the probability peak distribution of the non-Gaussian combined process A(t) can be
obtained from Eq. (17) as
1  b2   1  b2 
f Am (b) ≈  2 − 1 exp  −  2 − 1  , b ≥ σ A (18)
σA σA   2  σ A  
By introducing the intermediate response threshold level c = b 2 / σ A2 − 1 associated with a
peak-dependent intermediate random variable C = Am2 / σ A2 − 1 , which is referred to as the
intermediate peak variable, the elementary probability of the event for an occurring peak with
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

{b ≤ Am ≤ b + db} is equal to the elementary probability of the event for the intermediate peak

{ }
variable with c ≤ Am2 / σ A2 − 1 ≤ c + dc as
c  1 
f Am (b)db = exp  − c  dc, c ≥ 0 (19)
2 c +1  2 
The PDF of the intermediate peak variable C can be further approximated from Eq. (19) into a
form of Gamma distribution, as
c  1 
fC (c) ≈ exp  − c  (20)
4  2 
In general, the Gamma probability model of Eq. (20) results in a higher PDF value than that
of the original PDF model at the tail range of intermediate peak variables, indicating a more
conservative estimation of the crossing failure probability using the Gamma probability
model given in Eq. (20).
Gamma peak factor and combined peak factor method
A closed-form expression of the peak factor for a combined resultant response process
A(t ) = X 2 ( t ) + Y 2 ( t ) with σ X = σ Y is developed in this section. Based on the PDF of the
intermediate peak variable C = Am2 / σ A2 − 1 in the form of the Gamma distribution given in Eq.
(20), the corresponding CDF of C can be analytically obtained as follows
 1   1 
FC (c) = P ( C ≤ c ) = 1 − 1 + c  exp  − c  , c ≥ 0 (21)
 2   2 
Since the extreme value arising from the Gamma PDF of C (i.e., Eq. (21)), converges
asymptotically to the Type I Gumbel distribution, the location parameter un and the scale
parameter β n of the corresponding extreme value of the intermediate peak variable can be
determined, respectively as follows,
 1   1  1
1 + un  exp  − un  = (22)
 2   2  n
−1
−1  u  1 
β n = [ nfC (un ) ] =  n n exp  − un   (23)
 4  2 
Taking logarithm at both sides of Eq. (22), one obtains
un = 2 ln n + 2 ln(1 + 0.5un ) (24)
When the sample size n is sufficiently large (e.g. when n>100), the characteristic largest value
for the Gamma peak distribution can thus be approximated as
un ≈ 2 ln n + 2 ln ln n (25)
The scale parameter or dispersion of the largest peak value from the Gamma
distribution can then be reduced from Eq. (23) to the following form as
β n ≈ 2 ln n /(ln n + ln ln n) (26)
Based on Eqs. (25) and (26), the mean µCn of the extreme value of the intermediate peak
variable C can be given as follows
2γ ln n
µCn = un + γβ n = 2 ln n + 2 ln ln n + (27)
ln n + ln ln n
Using the relationship of the intermediate peak variable C to the peaks of a Rayleigh
process Am , i.e., C = Am2 / σ A2 − 1 , the expected peak factor of a combined resultant process
with Rayleigh distribution can be written as
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

1/ 2
 2γ 2 ln n 
gG = µCn + 1 =  2 ln n + 2 ln ln n + + 1 (28)
 ln n + ln ln n 

The above equation gives a closed-form formula for estimating the mean extreme response of
a combined resultant process following the Rayleigh distribution. The expected peak factor
determined by Eq. (28) is herein called as the Gamma peak factor, which is derived from the
Gamma distribution given in Eq. (20). Since the peak distribution of the resultant process
according to Eq. (17) is defined for those peaks over the threshold of mode value σ A , the
expected maximum resultant response can then be estimated in terms of the Gamma peak
factor and the mode value of the resultant process A(t) as
µ An = gGσ A (29)
If two components do not have the same value of standard deviation, i.e., σ X < σ Y , the mean
extreme resultant response of the combined process A(t) can then be approximated as [Huang
(2008)]
µ An ≈ (g 2
G )
− g 2f σ X2 + g 2f σ Y2 (30)
The use of Eq. (30) for predicting the peak resultant acceleration response can be referred to
as the combined peak factor (CPF) method.
Illustrated Example: The CAARC building
The prototype CAARC building has an overall height of 180m and a rectangular floor
plan dimension of 30 m by 45 m [Melbourne (1980)]. Aerodynamic wind forces acting on the
45-story CAARC building were measured by the synchronous multi-pressure measurement
technique in the wind tunnel using a 1:400 scale rigid model. A 10-year return period hourly
mean wind speed of 34.7m/s at the reference height of 90 m was used for calculating the
acceleration responses of the building. The 0-degree wind perpendicular to the wide face
acting in the short direction of the building was considered in the example. Time history
analysis was carried out by applying the measured time history wind forces in a computer-
based finite element model of the CAARC building.
Table 1 presents the peak resultant acceleration results at the top corner of the
CAARC building. The mean extreme resultant acceleration was calculated using the CPF
method of Eq. (30) with two component standard deviation acceleration, which can be
obtained from either frequency-domain approach or time-domain approach. The proposed
combined peak factor method resulted in a peak resultant acceleration of 19.1 milli-g, slightly
less than (-3%) that of the statistically evaluated value of 19.7 milli-g, which were calculated
by averaging the 24 extreme resultant peaks over six samples of 10-minute resultant
acceleration time history.

Table 1: Peak resultant accelerations in the CAARC building under 0-dgree wind
Peak resultant acceleration (milli-g) Used peak factors
CPF method gf
Time history analysis gG
Eq. (30)
19.7 19.1 (-3%) 3.276 3.751

Conclusion
In this paper, the probabilistic peak factor and the Gamma peak factor are developed
analytically by investigating the mean level-crossing rate and peak distribution of the resultant
response process. The probabilistic peak factor is explicitly expressed in terms of the random
excitation duration, the non-exceedance probability and the spectral bandwidth shape
parameter of the random acceleration response process. For a narrowband response process,
The Seventh Asia-Pacific Conference on Wind Engineering, November 8-12, 2009, Taipei, Taiwan

the use of Davenport peak factor without accounting spectral bandwidth effects yields a
conservative estimate of the mean extreme component response. The probabilistic peak factor
has the advantage to predict extreme response at different levels of time-variant reliability or
non-exceedance probability, which may be useful in the context of reliability performance-
based design of tall buildings.
Based on the asymptotic theory of statistical extremes, the Gamma peak factor has
been obtained for a non-Gaussian combined resultant response, characterized by a Rayleigh
distribution process. Using the Gamma peak factor, the combined peak factor method has
been proposed for predicting the mean extreme resultant acceleration response of wind-
sensitive tall buildings. The 45-story CAARC building tested in the wind tunnel was used to
demonstrate the applicability of the combined peak factor method to evaluate the mean
extreme resultant response. The wind tunnel derived acceleration time history results of the
building verify that the combined peak factor method gives reasonably accurate prediction of
the mean extreme resultant acceleration response of wind-excited tall buildings.
References
Chan C.M., Huang M.F., Kwok K.C.S. (2009). “Stiffness optimization for wind-Induced dynamic serviceability
design of tall buildings.” Journal of Structural Engineering, ASCE, 135(8), 985-997.
Chen X, Huang G. (2009). “Evaluation of peak resultant response for wind-excited tall buildings.” Engineering
Structures, 31, 858-868.
Davenport A.G. (1964). “Note on the distribution of the largest value of a random function with application to
gust loading.” Proceedings, Institution of Civil Engineering, 28, 187-196.
Gurley K.R., Tognarelli M.A., and Kareem A. (1997). “Analysis and simulation tools for wind engineering.”
Probabilistic Engineering Mechanics, 12(1), 9-31.
Holmes J.D, and Cochran L.S. (2003). “Probability distribution of extreme pressure coefficients.” Journal of
Wind Engineering and Industrial Aerodynamics, 91, 893-901.
Hong H.P., Beadle S., and Escobar J.A. (2001). “Probabilistic assessment of wind-sensitive structures with
uncertain parameters.” Journal of Wind Engineering and Industrial Aerodynamics, 89, 893-910.
Huang, M.F. (2008). “Performance-based serviceability design optimization of wind sensitive tall buildings.”
Ph.D. thesis, Hong Kong Univ. of Science and Technology, Hong Kong.
Isyumov N., Fediw A.A., Colaco J. and Banavalkar P.V. (1992). “Performance of a tall building under wind
action.” Journal of Wind Engineering and Industrial Aerodynamics, 41-44, 1053-1064.
Kareem A. (1987). “Wind effects on structures: a probabilistic viewpoint.” Probabilistic Engineering Mechanics,
2, 166-200.
Melbourne, W.H. (1980). “Comparison of measurements of the CAARC standard tall building model in
simulated model wind flows.” Journal of Wind Engineering and Industrial Aerodynamics, 6, 78–88.
Melbourne W.H., and Palmer T.R. (1992). “Accelerations and comfort criteria for buildings undergong complex
motions.” Journal of Wind Engineering and Industrial Aerodynamics, 41-44: 105-116.
Sadek F. and Simiu E. (2002). “Peak non-Gaussian wind effects for database-assisted low-rise building design.”
Journal of Engineering Mechanics ASCE, 128(5), 530-539.
Solari G. (1997). “Wind-excited response of structures with uncertain parameters.” Probabilistic Engineering
Mechanics, 12(2), 75-87.
Vanmarcke E.H. (1975). “On the distribution of the First-passage time for normal stationary random processes.”
Journal of Applied Mechanics, 42, 215-220.
Tieleman H.W, Ge Z, and Hajj M.R. (2007). “Theoretically estimated peak wind loads.” Journal of Wind
Engineering and Industrial Aerodynamics, 95, 113-132.