Académique Documents
Professionnel Documents
Culture Documents
Research Article
Estimation of Extreme Values by the Average Conditional
Exceedance Rate Method
A. Naess,1 O. Gaidai,2 and O. Karpa3
1
Department of Mathematical Sciences and CeSOS, Norwegian University of Science and Technology, 7491 Trondheim, Norway
Norwegian Marine Technology Research Institute, 7491 Trondheim, Norway
3
Centre for Ships and Ocean Structures (CeSOS), Norwegian University of Science and Technology, 7491 Trondheim, Norway
2
1. Introduction
Extreme value statistics, even in applications, are generally
based on asymptotic results. This is done either by assuming
that the epochal extremes, for example, yearly extreme wind
speeds at a given location, are distributed according to the
generalized (asymptotic) extreme value distribution with
unknown parameters to be estimated on the basis of the observed data [1, 2]. Or it is assumed that the exceedances above
high thresholds follow a generalized (asymptotic) Pareto
distribution with parameters that are estimated from the data
[14]. The major problem with both of these approaches is
that the asymptotic extreme value theory itself cannot be used
in practice to decide to what extent it is applicable for the
observed data. And since the statistical tests to decide this
issue are rarely precise enough to completely settle this problem, the assumption that a specific asymptotic extreme value
distribution is the appropriate distribution for the observed
data is based more or less on faith or convenience.
On the other hand, one can reasonably assume that in
most cases long time series obtained from practical measurements do contain values that are large enough to provide
where := means by definition, the following one-step memory approximation will, to a certain extent, account for the
dependence between the s,
Prob { | 1 , . . . , 1 }
Prob { | 1 } ,
(3)
() 2 ()
:= Prob { | 1 } Prob (1 ) .
=2
(4)
By conditioning on one more data point, the one-step memory approximation is extended to
Prob { | 1 , . . . , 1 }
Prob { | 1 , 2 } ,
(5)
() 3 () := Prob { | 1 , 2 }
=3
Prob {2 | 1 } Prob (1 ) .
(6)
For a general , 2 , it is obtained that
() ()
:= Prob { | 1 , . . . , +1 }
=
() = Prob ( ) = Prob { , . . . , 1 }
= Prob { | 1 , . . . , 1 }
Prob {1 , . . . , 1 }
Prob { | 1 . . . , 1 }
=2
(1)
= Prob { | 1 , . . . , 1 }
=2
Prob (1 ) .
In general, the variables are statistically dependent.
Hence, instead of assuming that all the are statistically
independent, which leads to the classical approximation
() 1 () := Prob ( ) ,
=1
(7)
Prob (1 ) ,
where () = ().
It should be noted that the one-step memory approximation adopted above is not a Markov chain approximation
[911], nor do the -step memory approximations lead to
th-order Markov chains [12, 13]. An effort to relinquish the
Markov chain assumption to obtain an approximate distribution of clusters of extremes is reported in [14].
It is of interest to have a closer look at the values for ()
obtained by using (7) as compared to (2). Now, (2) can be
rewritten in the form
(2)
() 1 () = (1 1 ()) ,
=1
(8)
where 1 () = Prob{ > }, = 1, . . . , . Then the approximation based on assuming independent data can be written
as
() 1 () := exp ( 1 ()) ,
(9)
=1
() () = (1 ()) (1 ()) ,
(10)
=1
=1
() () := exp ( () ()) ,
(11)
,
and () () as with () = () for .
For the cascade of approximations () to have practical
significance, it is implicitly assumed that there is a cut-off
value satisfying such that effectively () =
(). It may be noted that for -dependent stationary data
sequences, that is, for data where and are independent
whenever | | > , then () = +1 () exactly, and, under
rather mild conditions on the joint distributions of the data,
lim 1 () = lim () [15]. In fact, it can be shown
that lim 1 () = lim () is true for weaker conditions than -dependence [16]. However, for finite values of
the picture is much more complex, and purely asymptotic
results should be used with some caution. Cartwright [17]
used the notion of -dependence to investigate the effect on
extremes of correlation in sea wave data time series.
Returning to (11), extreme value prediction by the conditioning approach described above reduces to estimation of
(combinations) of the () functions. In accordance with
the previous assumption about a cut-off value , for all values of interest, , so that 1
=1 () is effectively
negligible compared to
= (). Hence, for simplicity, the
following approximation is adopted, which is applicable to
both stationary and nonstationary data,
() = exp ( ()) ,
1.
(12)
1
() ,
+ 1 =
= 1, 2, . . . .
(13)
(14)
= , . . . , , = 2, . . . ,
where 1{A} denotes the indicator function of some event A.
Then,
() =
E [ ()]
E [ ()]
= , . . . , , = 2, . . . ,
(15)
= ()
()
(16)
= E [ ()]
+1
(17)
E [ ()]
E [ ()]
)
(18)
exp ( E [ ()]) .
no. , = 1, . . . , , then
= E[ ()] = (). Hence,
(19)
where
() =
1
() .
+ 1 =
() =
2
1 ()
( () ()) ,
1 =1
(20)
(21)
where () () denotes the ACER function estimate from realization no. , and () = =1 () ()/.
Assuming that realizations are independent, for a suitable
number , for example, 20, (21) leads to a good approximation of the 95% confidence interval CI = ( (), + ())
for the value (), where
() = ()
() exp ( ( + 1) ()) ,
1.96 ()
.
(22)
Alternatively, and which also applies to the non-stationary case, it is consistent with the adopted approach to assume
that the stream of conditional exceedances over a threshold
constitute a Poisson process, possibly non-homogeneous.
Hence, the variance of the estimator () of (), where
() =
= ()
(23)
+1
is Var[ ()] = (). Therefore, for high levels , the approximate limits of a 95% confidence interval of (), and also
(), can be written as
() = () (1
1.96
( + 1) ()
).
(24)
() = () exp { ( ) } ,
1 ,
(25)
()
) = log ( ) log ( ) .
log log (
()
(26)
Therefore, under the assumptions made, a plot of
log | log( ()/ ())| versus log( ) will exhibit a
perfectly linear tail behaviour.
It is realized that if the function () could be replaced
by a constant value, say , one would immediately be in a
position to apply a linear extrapolation strategy for deep tail
prediction problems. In general, () is not constant, but its
variation in the tail region is often sufficiently slow to allow
for its replacement by a constant, possibly by adjusting the tail
marker 1 . The proposed statistical approach to the prediction
of extreme values is therefore based on the assumption that
we can write,
() = exp { ( ) } ,
1 ,
(27)
2
(, , , ) = log ( ) log + ( ) , (28)
=1
where 1 < < denotes the levels where the ACER function has been estimated, denotes a weight factor that
log ( ) 0,
0 < < +,
min < 1 ,
(29)
0 < < +,
0 < < 5.
=1 ( ) ( )
2
=1 ( )
,
(30)
log (, ) = + (, ) ,
1/
1/
() = () [1 + ( ( ) )]
where =
=1 / =1 , with a similar definition of .
To calculate the final optimal set of parameters, one
may use the Levenberg-Marquardt method on the function
) = ( (, ), , , (, )) to find the optimal values
(,
and , and then use (30) to calculate the corresponding
and .
For a simple construction of a confidence interval for
the predicted, deep tail extreme value given by a particular
ACER function as provided by the fitted parametric curve, the
empirical confidence band is reanchored to the fitted curve
by centering the individual confidence intervals CI0.95 for the
point estimates of the ACER function on the fitted curve.
Under the premise that the specified class of parametric
curves fully describes the behaviour of the ACER functions in
the tail, parametric curves are fitted as described above to the
boundaries of the reanchored confidence band. These curves
are used to determine a first estimate of a 95% confidence
interval of the predicted extreme value. To obtain a more
precise estimate of the confidence interval, a bootstrapping
method would be recommended. A comparison of estimated
confidence intervals by both these methods will be presented
in the section on extreme value prediction for synthetic data.
As a final point, it has been observed that the predicted value
is not very sensitive to the choice of 1 , provided it is chosen
with some care. This property is easily recognized by looking
at the way the optimized fitting is done. If the tail marker is
in the appropriate domain of the ACER function, the optimal
fitted curve does not change appreciably by moving the tail
marker.
1 () [1 + ( ( ))]
(31)
1 , (32)
() = [1 + ( ( ) + ())]
1 ,
(33)
() = [1 + ( ) ] ,
1 ,
(34)
where = 1/, = .
For the analysis of data, first the tail marker 1 is
provisionally identified from visual inspection of the log
plot (, ln ()). The value chosen for 1 corresponds to the
beginning of regular tail behaviour in a sense to be discussed
below.
The optimization process to estimate the parameters is
done relative to the log plot, as for the Gumbel case. The mean
square error function to be minimized is in the general case
written as
(
, , , , ) = log ( ) log
(35)
=1
2
+ log [1 + ( ) ] ,
(36)
method [40, 41]. Note that the assumption that the initial
extreme values are actually generated with good approximation from a Gumbel distribution cannot easily be verified with
any accuracy in general, which is a drawback of this method.
Compared with the POT method, the Gumbel method would
also seem to use much less of the information available in
the data. This may explain why the POT method has become
increasingly popular over the past years, but the Gumbel
method is still widely used in practice.
(38)
() = (; , 0) = exp ( ) .
Since the recorded data in practice are rarely independent, a declustering technique is commonly used to filter the
data to achieve approximate independence [1, 2].
7.2. Return Periods. The return period of a given value
of in terms of a specified length of time , for example,
a year, is defined as the inverse of the probability that the
specified value will be exceeded in any time interval of length
. If denotes the mean exceedance rate of the threshold
per length of time (i.e., the average number of data points
above the threshold per ), then the return period of the
value of corresponding to the level = + is given by
the relation
1
1
=
.
=
(39)
Prob ( > ) Prob ( > )
1
.
()
(40)
[1 () ]
.
(41)
(42)
2
)} ,
2
(43)
2
)} ,
2
(44)
Table 1: 100-year return level estimates and 95% CI (BCI = CI by bootstrap) for A = ACER, G = Gumbel, and P = POT. Exact value = 4.80.
A 100
5.07
4.65
4.86
4.75
4.54
4.80
4.84
5.02
4.59
4.84
4.62
4.82
A CI
(4.67, 5.21)
(4.27, 4.94)
(4.49, 5.06)
(4.22, 5.01)
(4.20, 4.74)
(4.35, 5.05)
(4.36, 5.20)
(4.47, 5.31)
(4.33, 4.81)
(4.49, 5.11)
(4.29, 5.05)
(4.41, 5.09)
A BCI
(4.69, 5.42)
(4.37, 5.03)
(4.44, 5.19)
(4.33, 5.02)
(4.27, 4.88)
(4.42, 5.14)
(4.48, 5.19)
(4.62, 5.36)
(4.38, 4.98)
(4.60, 5.30)
(4.45, 5.09)
(4.48, 5.18)
G 100
4.41
4.92
5.04
4.75
4.80
4.91
4.85
4.96
4.76
4.77
4.79
4.84
P 100
4.29
4.88
5.04
4.69
4.73
4.79
4.71
4.97
4.68
4.41
4.53
4.72
G BCI
(4.14, 4.73)
(4.40, 5.58)
(4.54, 5.63)
(4.27, 5.32)
(4.31, 5.39)
(4.41, 5.50)
(4.36, 5.43)
(4.47, 5.53)
(4.31, 5.31)
(4.34, 5.32)
(4.31, 5.41)
(4.37, 5.40)
P BCI
(4.13, 4.52)
(4.42, 5.40)
(4.48, 5.74)
(4.24, 5.26)
(4.19, 5.31)
(4.31, 5.34)
(4.32, 5.23)
(4.47, 5.71)
(4.15, 5.27)
(4.23, 4.64)
(4.05, 4.88)
(4.27, 5.23)
101
102
ACER1 ()
Sim. No.
1
10
20
30
40
50
60
70
80
90
100
Av. 100
103
104
105
4.85
2.5
3.5
4.5
5.5
10
4.72
4.7
100 yr
4.7
4.68
4.66
4.64
4.62
4.6
100
130
160
190 204
230
250
3
2
102
ACER ()
ln(ln(( + 1)/))
1
0
103
4.75
1
3.6
3.8
4.2
4.4
4.6
4.8
104
3.5
4.5
5.5
6.5
/
=1
=2
=4
= 24
= 48
= 72
= 96
for the prediction of a 100-year value, one may use the first
ACER function. The reason for this is that Figure 5 shows that
all the ACER functions coalesce in the far tail. Hence, we may
use any of the ACER functions for the prediction. Then, the
obvious choice is to use the first ACER function, which allows
us to use all the data in its estimation and thereby increase
accuracy.
In Figure 6 is shown the results of parametric estimation
of the return value and its 95% CI for 13 years of hourly
11
5
101
4
3
ln(ln(( + 1)/))
ACER1 ()
102
103
104
105
2
1
0
106
8.62
3
6
/
6.5
7.5
/
8.5
9.23
9
9.5
8.1
8
7.95
100 yr /
8.56
Method
7.9
ACER, various
7.8
7.7
7.6
Annual maxima
100
120
140
161
180
200
Figure 7: The point estimate 100 yr of the 100-year return level based
on 13 years hourly data as a function of the number of data points
above threshold. = 6.01 m/s.
POT
Spec
1
2
4
24
48
72
96
MM
GL
100 yr , m/s
51.85
51.48
52.56
52.90
54.62
53.81
54.97
51.5
55.5
47.8
95% CI (
100 yr ), m/s
(48.4, 53.1)
(46.1, 54.1)
(46.7, 55.7)
(47.0, 56.2)
(47.7, 57.6)
(46.9, 58.3)
(47.5, 60.5)
(45.2, 59.3)
(48.0, 64.9)
(44.8, 52.7)
12
102
ACER ()
Method
10
Spec
100 yr , m/s
95% CI (
100 yr ), m/s
60.47
(53.1, 64.9)
62.23
(53.3, 70.0)
63.03
(53.0, 74.5)
24
60.63
(51.3, 70.7)
48
60.44
(51.3, 77.0)
72
58.06
(51.2, 66.4)
96
59.19
(52.0, 68.3)
MM
58.10
(50.8, 67.3)
GL
60.63
(53.0, 70.1)
53.48
(48.9, 57.0)
ACER, various
104
4.5
5.5
6.5
7.5
8.5
Annual maxima
= 48
= 72
= 96
=1
=2
=4
= 24
POT
9.42
9.38
9.35
100 yr /
102
9.34
ACER1 ()
103
9.3
104
9.26
105
100
140
185
220
260
300
106
10.6
5
10
11
12
Figure 11: The point estimate 100 yr of the 100-year return level
based on 14 years hourly data as a function of the number of data
points above threshold. = 5.72 m/s.
13
2.5
2
1.5
1
0.5
()
ln(ln(( + 1)/))
0.5
1
1.5
0
10.2 10.6
1
7
7.5
8.5
9.5
10
10.5
2
2.5
60
70
80
90
100
110
120
Time (s)
Figure 13: Part of the narrow-band response time series of the linear
oscillator with fully sampled and peak values indicated.
14
ACER ()
102
103
104
105
1.5
2.5
3.5
/
= 32
= 64
=1
=2
= 25
ACER ()
Acknowledgment
101
102
References
103
1.5
2.5
3.5
/
=1
=2
=3
=4
=5
15
[32] A. Naess, Estimation of Extreme Values of Time Series with
Heavy Tails, Preprint Statistics No. 14/2010, Department of
Mathematical Sciences, Norwegian University of Science and
Technology, Trondheim, Norway, 2010.
[33] E. F. Eastoe and J. A. Tawn, Modelling the distribution of the
cluster maxima of exceedances of subasymptotic thresholds,
Biometrika, vol. 99, no. 1, pp. 4355, 2012.
[34] J. A. Tawn, Discussion of paper by A. C. Davison and R. L.
Smith, Journal of the Royal Statistical Society. Series B, vol. 52,
no. 3, pp. 393442, 1990.
[35] A. W. Ledford and J. A. Tawn, Statistics for near independence
in multivariate extreme values, Biometrika, vol. 83, no. 1, pp.
169187, 1996.
[36] J. E. Heffernan and J. A. Tawn, A conditional approach for
multivariate extreme values, Journal of the Royal Statistical
Society. Series B, vol. 66, no. 3, pp. 497546, 2004.
[37] Numerical Algorithms Group, NAG Toolbox for Matlab, NAG,
Oxford, UK, 2010.
[38] E. J. Gumbel, Statistics of Extremes, Columbia University Press,
New York, NY, USA, 1958.
[39] K. V. Bury, Statistical Models in Applied Science, Wiley Series
in Probability and Mathematical Statistics, John Wiley & Sons,
New York, NY, USA, 1975.
[40] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap,
vol. 57 of Monographs on Statistics and Applied Probability,
Chapman and Hall, New York, NY, USA, 1993.
[41] A. C. Davison and D. V. Hinkley, Bootstrap Methods and Their
Application, vol. 1 of Cambridge Series in Statistical and Probabilistic Mathematics, Cambridge University Press, Cambridge,
UK, 1997.
[42] J. Pickands, III, Statistical inference using extreme order
statistics, The Annals of Statistics, vol. 3, pp. 119131, 1975.
[43] A. Naess and P. H. Clausen, Combination of the peaks-overthreshold and bootstrapping methods for extreme value prediction, Structural Safety, vol. 23, no. 4, pp. 315330, 2001.
[44] N. J. Cook, The Designers Guide to Wind Loading of Building
Structures, Butterworths, London, UK, 1985.
[45] N. J. Cook, Towards better estimation of extreme winds,
Journal of Wind Engineering and Industrial Aerodynamics, vol.
9, no. 3, pp. 295323, 1982.
[46] A. Naess, Estimation of long return period design values for
wind speeds, Journal of Engineering Mechanics, vol. 124, no. 3,
pp. 252259, 1998.
[47] J. P. Palutikof, B. B. Brabson, D. H. Lister, and S. T. Adcock, A
review of methods to calculate extreme wind speeds, Meteorological Applications, vol. 6, no. 2, pp. 119132, 1999.
[48] O. Perrin, H. Rootzen, and R. Taesler, A discussion of statistical
methods used to estimate extreme wind speeds, Theoretical and
Applied Climatology, vol. 85, no. 3-4, pp. 203215, 2006.
[49] M. E. Robinson and J. A. Tawn, Extremal analysis of processes
sampled at different frequencies, Journal of the Royal Statistical
Society. Series B, vol. 62, no. 1, pp. 117135, 2000.
[50] A. Naess, A Note on the Bivariate ACER Method, Preprint Statistics No. 01/2011, Department of Mathematical Sciences, Norwegian University of Science and Technology, Trondheim,
Norway, 2011.
Advances in
Operations Research
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Advances in
Decision Sciences
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Mathematical Problems
in Engineering
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Journal of
Algebra
Hindawi Publishing Corporation
http://www.hindawi.com
The Scientific
World Journal
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
International Journal of
Differential Equations
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Volume 2014
Advances in
Combinatorics
Hindawi Publishing Corporation
http://www.hindawi.com
Mathematical Physics
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Journal of
Complex Analysis
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
International
Journal of
Mathematics and
Mathematical
Sciences
Journal of
Stochastic Analysis
Abstract and
Applied Analysis
International Journal of
Mathematics
Volume 2014
Volume 2014
Discrete Dynamics in
Nature and Society
Volume 2014
Volume 2014
Journal of
Journal of
Discrete Mathematics
Journal of
Volume 2014
Applied Mathematics
Journal of
Function Spaces
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Volume 2014
Volume 2014
Optimization
Hindawi Publishing Corporation
http://www.hindawi.com
Volume 2014
Volume 2014
Numerical methods for calculating the crossing rate of high and extreme
response levels of compliant offshore structures subjected to random waves
A. Naess a,b, , H.C. Karlsen b , P.S. Teigen c
a Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, A. Getz vei 1, NO-7491, Trondheim, Norway
b Department of Mathematical Sciences, Norwegian University of Science and Technology, A. Getz vei 1, NO-7491, Trondheim, Norway
c Statoil Research Centre, Rotvoll, Trondheim, Norway
Received 16 December 2005; received in revised form 3 April 2006; accepted 8 April 2006
Available online 5 June 2006
Abstract
The focus of the paper is on methods for calculating the mean upcrossing rate of stationary stochastic processes that can be represented as
second order stochastic Volterra series. This is the current state-of-the-art representation of the horizontal motion response of e.g. a tension leg
platform in random seas. Until recently, there has been no method available for accurately calculating the mean level upcrossing rate of such
response processes. Since the mean upcrossing rate is a key parameter for estimating the large and extreme responses it is clearly of importance
to develop methods for its calculation. The paper describes in some detail a numerical method for calculating the mean level upcrossing rate
of a stochastic response process of the type considered. Since no approximations are made, the only source of inaccuracy is in the numerical
calculation, which can be controlled. In addition to this exact method, two approximate methods are also discussed.
c 2006 Elsevier Ltd. All rights reserved.
Keywords: Second order stochastic Volterra model; Mean crossing rate; Extreme response; Slow drift response; Method of steepest descent
1. Introduction
The problem of calculating the extreme response of
compliant offshore structures like tension leg platforms or
moored spar buoys in random seas, has been a challenge for
many years, and, in fact, it still represents a challenge. Starting
with the state-of-the-art representation of the horizontal
excursions of moored, floating offshore structures in random
seas as a second order stochastic Volterra series, we shall in
this paper develop a general method for estimating the extreme
response of such structures. Even if the Volterra series model
was formulated more than 30 years ago, it is not until quite
recently that general numerical methods have become available
that allow accurate calculation of the probability distribution
and, perhaps more importantly, the mean upcrossing rate of
the total response process. This last quantity is the crucial
parameter for estimating extreme responses.
During the 1980s significant efforts were directed towards
developing methods for calculating the response statistics of
Corresponding author. Tel.: +47 73 59 70 53; fax: +47 73 59 35 24.
Z 2 (t) =
(1)
j=1
(2)
(3)
(i = 1)
W2 j1 (t) + i W2 j (t) =
u j (k )Bk eik t
(4)
k=1
Z 1 (t) =
j W j (t).
(5)
j=1
S X (k ) u j (k )
(6)
s f Z Z (, s)ds
(7)
(8)
1
(2 )2
M(u, v)
(9)
By substituting from Eq. (9) back into Eq. (7), the mean
crossing rate is formally expressed in terms of the characteristic
function, but this is not a very practical expression.
The solution to this is obtained by considering the
characteristic function as a function of two complex variables.
It can then often be shown that this new function becomes
holomorphic in suitable regions of C2 , where C denotes the
complex plane. As shown in detail in [14], under suitable
conditions, the use of complex function theory allows the
derivation of two alternative expressions for the crossing rate.
Here we shall focus on one of these alternatives, viz.
Z+ ( ) =
2 j1 + i 2 j =
Z+ ( ) =
1
(2 )2
ia
ib
ia
ib
1
M(z, w)eiz dzdw (10)
w2
where 0 < a < a1 for some positive constant a1 , and b0 < b <
b1 for some constants b0 < 0 and b1 > 0.
12
22
(11)
(ik )u i (k )u j (k ) .
(12)
k=1
(13)
while
r2i1,2 j = r2i,2 j1 = (Ri j )
(14)
where R(z) denotes the real part and (z) the imaginary part of
z. Similarly, let
N
k2 u i (k )u j (k ) .
(15)
(16)
Si j =
k=1
Then
while
s2i1,2 j = s2i,2 j1 = (Si j ).
(17)
1
(2 )2
ia
ia
1
I (, w)dw
w2
(22)
where
4. Numerical calculation
1
1
1
ln(det(A)) v 2 V + t A1 t
2
2
2
(18)
ib
ib
ib
(23)
ib
where
M(z, w) eiz dz
I = I (, w) =
z1 = zs + h
(25)
A = I 2 i u 2 i v ( 21 + 12 ) + 4 v 2 V
(19)
V = 22 21 12
(20)
t = i u I + i v 12 2 v V
2
(21)
zj =
g (z j )
h,
|g (z j )|
z j+1 = z j +
z j,
j = 1, 2, . . .
j = 1, 2, . . .
(26)
(27)
z 1 = z s h
zj =
g (z j )
h,
|g (z j )|
z j1 = z j +
z j,
j = 1, 2, . . .
j = 1, 2, . . . .
(29)
(30)
(31)
+
+
Z ( ) = Z (ref )
where
h
I+ = exp{g(z s )} +
2
z j exp{g(z j )}
(32)
j=1
and
h
I = exp{g(z s )}
2
z j exp{g(z j )}
(33)
j=1
1
R
(2 )2
L
j=L
1
I (, w j ) w j
w 2j
(34)
2
2
s ;w)
g(z
x2
exp{g(z s ; w)}
(35)
(36)
f Z ( )
f Z (ref )
(n)
(m)
(37)
n=1 m=1
(n)
(m)
(38)
= exp T
0
Z+ (
| h, t) f Hs Ts (h, t) dh dt .
(39)
Haver and Nyhus [22] have proposed a model for the joint PDF
f Hs Ts (h, t).
Table 1
Main particulars of the MDF
Draught (m)
Column diameter (m)
Natural period surge/sway (s)
Natural period yaw (s)
80.0
10.0
133.5
121
Fig. 1. Computer mesh of the submerged part of the moored deep floater.
g 2
5
exp
5 p
4
+ ln exp
6. Numerical examples
[2
1
.
+ 2i 0 + 02 ]
1
p
(41)
L()
=
1
2 2
(40)
Hs
T p2
(1 0.287 ln )
(42)
= /1
h = 1.0 103
h = 1.0 102
2.0
5.0
10.0
15.0
20.0
25.0
8.38 103
3.93 103
5.53 104
5.70 105
5.34 106
4.81 107
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
Table 3
Calculated upcrossing rates for 10 eigenvalues
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
7.41 103
3.59 103
5.23 104
5.59 105
5.36 106
4.92 107
Table 4
Calculated upcrossing rates for 20 eigenvalues
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
8.09 103
4.19 103
6.51 104
6.95 105
6.56 106
5.91 107
8.10 103
4.19 103
6.51 104
6.95 105
6.56 106
5.91 107
7.27 103
3.82 103
6.06 104
6.74 105
6.58 106
6.11 107
Table 5
Calculated upcrossing rates for 30 eigenvalues
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
7.21 103
3.58 103
5.48 104
5.85 105
5.54 106
5.00 107
7.21 103
3.58 103
5.48 104
5.85 105
5.54 106
5.00 107
6.51 103
3.27 103
5.11 104
5.67 105
5.55 106
5.17 107
Table 7
Calculated upcrossing rates for 100 eigenvalues
= /1
+
Z ()
+
Z ()
+
Z ()
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
6.55 103
3.25 103
5.04 104
5.44 105
5.19 106
4.71 107
6.55 103
3.25 103
5.04 104
5.44 105
5.19 106
4.71 107
5.93 103
2.98 103
4.70 104
5.28 105
5.20 106
4.86 107
2.0
5.0
10.0
15.0
20.0
25.0
6.17 103
3.03 103
4.71 104
5.11 105
4.88 106
4.44 107
5.59 103
2.78 103
4.40 104
4.96 105
4.90 106
4.63 107
6.03 103
2.74 103
3.65 104
3.44 105
2.94 106
2.44 107
7. Concluding remarks
In this paper we have described a general numerical method
for calculating the mean level upcrossing rate of a second order
stochastic Volterra series. The specific model considered was
the surge response of a moored deep floater in random seas.
The method has been successfully implemented, and results
have been calculated. However, it has been pointed out that
a numerical calculation for the full problem by the general
method is in many cases too demanding computationally
for practical use. Therefore, two approximations have been
investigated, and it has been shown that the approximation
obtained by using what is known as a Laplace approximation
for integrals, appears to provide estimates that are sufficiently
accurate for practical purposes. The second approximation,
which was based on a widely used simplification, did not in
general perform that well.
The main accomplishment of the present work, is that there
is now available a general numerical method for calculating the
mean upcrossing rate of any second order stochastic Volterra
system with a stationary Gaussian process as input. Since this is
the state-of-the-art representation of the motion response of any
large volume floating offshore structure subjected to random
seas, this makes it possible to calculate accurately the response
statistics of such structures.
Acknowledgements
The financial support from the Statoil Research Centre
and the Research Council of Norway (NFR) through the
Centre for Ships and Ocean Structures (CeSOS) at the
Norwegian University of Science and Technology is gratefully
acknowledged.
References
[1] Vinje T. On the statistical distribution of second-order forces and motions.
International Shipbuilding Progress 1983;30:5868.
[2] Naess A. Statistical analysis of second-order response of marine
structures. Journal of Ship Research 1985;29(4):27084.
[3] Naess A. The statistical distribution of second-order slowly-varying forces
and motions. Applied Ocean Research 1986;8(2):1108.
[4] Langley RS. A statistical analysis of low frequency second-order forces
and motions. Applied Ocean Research 1987;9(3):16370.
[5] Kato S, Ando S, Kinoshita T. On the statistical theory of total second order
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
Mori
CH219.tex
5/6/2007
18: 24
Page 1
Simulation methods
Mori
CH219.tex
5/6/2007
18: 24
Page 2
Mori
CH219.tex
5/6/2007
18: 24
Page 3
Applications of Statistics and Probability in Civil Engineering Kanda, Takada & Furuta (eds)
2007 Taylor & Francis Group, London, ISBN 978-0-415-45134-5
O. Gaidai
Centre for Ships and Ocean Structures
Norwegian University of Science and Technology, Trondheim, Norway
ABSTRACT: The development of simple and efficient methods for estimation of the extreme response of
dynamical systems subjected to random excitations is discussed in the present paper. The key parameter for
calculating the statistical distribution of extreme response is the mean level upcrossing rate function. Exploiting
the regularity of the tail behaviour of this function, an efficient simulation based methodology for estimating the
extreme response distribution function is developed.
Keywords:
Monte Carlo simulation, mean crossing rate, extreme response, nonlinear dynamic systems.
INTRODUCTION
During the last decade it has become increasingly feasible to estimate the response statistics of dynamical
systems driven by stationary stochastic loading processes by standard Monte Carlo methods. However,
the problem of estimating the extreme response is still
largely considered to be too costly computationally if
done by a standard Monte Carlo technique. This is due
to the necessity of providing estimates of events with
a very low probability of occurring, thereby requiring
very long simulation times. Recent years have seen the
appearance of importance sampling techniques also
for dynamical systems (Tanaka 1998; A. Naess 2000;
Au and Beck 2003; Macke and Bucher 2003; Olsen and
Naess 2005), which can reduce substantially the computational cost. But these methods are fairly involved
and are not likely to become widely used in practical
applications.
Recognizing the huge increase in computational
power over the last 510 years, which is not showing
any signs of weakening, we consider it of some importance to develop methods based on simple, standard
Monte Carlo simulation methods that have the capability to offer robust and accurate estimates of extreme
response of quite general nonlinear dynamical systems
subjected to stationary stochastic inputs. To simplify
matters slightly, it is expedient to adopt the simplification that the statistical distribution of extreme values
Mori
CH219.tex
5/6/2007
18: 24
Page 4
where
where n+
j (; 0, T ) denotes the counted number of
upcrossings of the level for time history no. j.
For a suitable number k, e.g. k 20 30, a good
approximation of the 95% confidence interval for the
value + () can be obtained as
which brings out the crucial role of the mean upcrossing rate X+ (; t) in determining the extreme value
distribution. To calculate the mean upcrossing rate of
the stochastic process X (t), it is necessary to find the
joint PDF fX X (, ). However, for most nonlinear systems, this function cannot be calculated analytically, or
even numerically. The following question can then be
Mori
CH219.tex
5/6/2007
18: 24
Page 5
From the previous comments concerning the function (x), it may be assumed that [1 (y)]/y will
be negligible for relevant extreme values. This leads
to the conclusion that the transformed process Y (t)
has a mean upcrossing rate that is very close to an
exponential function.
Mori
CH219.tex
5/6/2007
18: 24
Page 6
several decimal orders of magnitude of + (x). Obviously, one should neglect data points from the very tail
where the confidence band width is greater than the
mean value itself (9), that is,
Figure 1. Monte Carlo simulated upcrossing rates versus analytical results for the mean upcrossing rate on the
loglog-log scale, Duffing oscillator: Monte Carlo (); linear
fit (); confidence bands (- -); range of linearity (--).
NUMERICAL EXAMPLES
Mori
CH219.tex
5/6/2007
18: 24
Page 7
Figure 4. Monte Carlo results for PDF versus mean upcrossing rate, 0-1st floor relative displacement, Benchmark problem: Monte Carlo (); linear fit ( ).
Mori
CH219.tex
5/6/2007
18: 24
Page 8
Mori
CH219.tex
5/6/2007
18: 24
Page 9
Figure 8. Monte Carlo results for PDF versus mean upcrossing rate, 01st floor relative displacement, Benchmark problem: Monte Carlo (); linear fit ( ). Random structural
parameters.
reduced by hundreds of times, compared to straightforward Monte Carlo simulations. This should make
the proposed method of considerable practical interest.
As a last comment, it is easily realized that there
are nonlinear systems for which the proposed simulation procedure might not work very well. For instance,
if the system stiffness characteristics suddenly change
significantly when the response levels reach very high
values that are rarely captured for short simulation
times, then the crossing rate function at extreme values could be difficult to estimate from more moderate
response values. Fortunately, such structural systems
are rarely seen in engineering practice.
Figure 9. Monte Carlo results for the mean upcrossing
rate, 910th floor relative displacement, Benchmark problem: Monte Carlo (); linear fit (); confidence bands (- -);
range of linearity (--). Random structural parameters.
ACKNOWLEDGEMENTS
The financial support from the Research Council of
Norway (NFR) through the Centre for Ships and Ocean
Structures (CeSOS) at the Norwegian University of
Science and Technology is gratefully acknowledged.
REFERENCES
A. Naess, C. S. (2000). Importance sampling for dynamical systems. In ICASP Applications of Statistics and
Probability, Volume 8, pp. 749755. A.A. Balkema.
Au, S. K. and J. L. Beck (2003). Importance sampling in high
dimensions. Structural Safety 25, 139163.
IfM (1994). Benchmark study on reliability estimation in
higher dimensions of structural systems. Technical Report
(see www.uibk.ac.at/mechanik/Benchmark Reliability/),
Institute of Engineering Mechanics, Leopold- Franzens
University, Innsbruch, Austria.
Macke, M. and C. Bucher (2003). Importance sampling for
randomly excited dynamical systems. Journal of Sound
and Vibration 268, 269290.
CONCLUDING REMARKS
Mori
CH219.tex
5/6/2007
18: 24
Page 10
10
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 1/35
Introduction
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 2/35
Introduction
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 2/35
Introduction
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 2/35
Introduction
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 2/35
Introduction
NTNU
NTNU
M X(t)
t = F(t),
A Monte Carlo approach for efficient estimation of extreme response statistics p. 3/35
NTNU
M X(t)
t = F(t),
M denotes a generalized n n mass matrix,
X = X(t) = (X1 (t), . . . , Xn (t))T = the system response vector,
H a nonlinear vector function,
F(t) denotes a stochastic loading process.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 3/35
NTNU
M X(t)
t = F(t),
M denotes a generalized n n mass matrix,
X = X(t) = (X1 (t), . . . , Xn (t))T = the system response vector,
H a nonlinear vector function,
F(t) denotes a stochastic loading process.
Hence the solution X(t) is also a stochastic vector process.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 3/35
NTNU
M X(t)
t = F(t),
M denotes a generalized n n mass matrix,
X = X(t) = (X1 (t), . . . , Xn (t))T = the system response vector,
H a nonlinear vector function,
F(t) denotes a stochastic loading process.
Hence the solution X(t) is also a stochastic vector process.
For specific prediction purposes, it is usually the extreme values
of one, or possibly a combination of several, of the component
processes of X(t) that is sought. For simplicity, denote it by
X(t).
A Monte Carlo approach for efficient estimation of extreme response statistics p. 3/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 4/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 4/35
NTNU
s fX(t)X(t)
(, s) ds
where fX(t)X(t)
(, ) denotes the joint PDF of X(t) and
X(t)
= dX(t)/dt.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 4/35
NTNU
s fX(t)X(t)
(, s) ds
where fX(t)X(t)
(, ) denotes the joint PDF of X(t) and
X(t)
= dX(t)/dt.
+ (; t) =
0
0).
where X + = max(X,
A Monte Carlo approach for efficient estimation of extreme response statistics p. 4/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 5/35
NTNU
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
A Monte Carlo approach for efficient estimation of extreme response statistics p. 5/35
NTNU
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
+
(; t) in
This brings out the crucial role of the mean upcrossing rate X
determining the extreme value distribution.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 5/35
NTNU
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
+
(; t) in
This brings out the crucial role of the mean upcrossing rate X
determining the extreme value distribution.
Note that the parameter of the Poisson distribution is
T
E[N + (; 0, T )] = 0 + (; t) dt.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 5/35
NTNU
+ (; t) dt.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 6/35
NTNU
+ (; t) dt.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 6/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 7/35
NTNU
() =
kT
+
n+
j (; 0, T )
j=1
A Monte Carlo approach for efficient estimation of extreme response statistics p. 7/35
NTNU
() =
kT
+
n+
j (; 0, T )
j=1
+
conf. band() = () 1.96 s()/ k
A Monte Carlo approach for efficient estimation of extreme response statistics p. 7/35
NTNU
() =
kT
+
n+
j (; 0, T )
j=1
+
conf. band() = () 1.96 s()/ k
The empirical standard deviation s() is given as
1
s() =
k1
j=1
n+
j (; 0, T )
T
()
A Monte Carlo approach for efficient estimation of extreme response statistics p. 7/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 8/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 8/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 8/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 9/35
NTNU
= (x) + ln q + (x)
A Monte Carlo approach for efficient estimation of extreme response statistics p. 9/35
NTNU
= (x) + ln q + (x)
+
(x) versus ln fX (x) will then clearly show to what
Plotting ln X
extent |(x)| is dominated by (x) as x .
A Monte Carlo approach for efficient estimation of extreme response statistics p. 9/35
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
where a, b and c are suitable constants, and d(x) is a function of
much slower increase than (x).
A Monte Carlo approach for efficient estimation of extreme response statistics p. 10/35
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
where a, b and c are suitable constants, and d(x) is a function of
much slower increase than (x).
Hence, we assume that
+
X
(x) = q(x) exp{a(x b)c } , x x0 ,
A Monte Carlo approach for efficient estimation of extreme response statistics p. 10/35
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
where a, b and c are suitable constants, and d(x) is a function of
much slower increase than (x).
Hence, we assume that
+
X
(x) = q(x) exp{a(x b)c } , x x0 ,
A Monte Carlo approach for efficient estimation of extreme response statistics p. 10/35
NTNU
It follows that
+
X
(x)
log log
q(x)
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
It follows that
+
X
(x)
log log
q(x)
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
It follows that
+
X
(x)
log log
q(x)
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
It follows that
+
X
(x)
log log
q(x)
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
It follows that
+
X
(x)
log log
q(x)
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
It follows that
+
X
(x)
log log
q(x)
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
A Monte Carlo approach for efficient estimation of extreme response statistics p. 11/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 12/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 12/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 12/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 12/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 12/35
NTNU
+
0.5
1010
0.5
102
1.5
103
6
2.5
10
3
0.9 1
1.1 2
1.2
31.3
1.4 4
1.5
/
A Monte Carlo approach for efficient estimation of extreme response statistics p. 13/35
NTNU
+0
1
101
2
102
3
103
104
4
5
105
6
106
6 6
10
5
105
4 4
10
3
103
2
102
1
101
A Monte Carlo approach for efficient estimation of extreme response statistics p. 14/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 15/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 15/35
NTNU
Z |Z|1 X|Z|
Z = |X|
+ AX
A Monte Carlo approach for efficient estimation of extreme response statistics p. 15/35
NTNU
Z |Z|1 X|Z|
Z = |X|
+ AX
A Monte Carlo approach for efficient estimation of extreme response statistics p. 15/35
NTNU
Z |Z|1 X|Z|
Z = |X|
+ AX
NTNU
+
0.8
1
1.2
2
1.4
10
1.6
1.8
103
2
4
10
2.2
2.4
6
102.6
2.8
2.8
2.9
3
1
3.1
3.2
2
3.33
3.4 4
3.5
5
3.6
A Monte Carlo approach for efficient estimation of extreme response statistics p. 16/35
NTNU
+2
2.5
3
103
3.5
4
104
4.5
105
5
5.5
106
6
6 105
5 104
4 103
3 102
2 101
A Monte Carlo approach for efficient estimation of extreme response statistics p. 17/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 18/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 19/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 19/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 19/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 20/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 20/35
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 20/35
3
P0 level
ln(ln((n+1)/k))
NTNU
2
0.25
0.3
0.35
0.4
0.45 L
G
M
0.5
0.55
0.6
0.65
A Monte Carlo approach for efficient estimation of extreme response statistics p. 21/35
3
P0 level
2
ln(ln((n+1)/k))
NTNU
2
0.35
0.4
0.45
0.5
0.55
M
0.6 L 0.65
G
0.7
0.75
A Monte Carlo approach for efficient estimation of extreme response statistics p. 22/35
10
NTNU
0
0.3
0.35
0.4
0.45
0.5
0.55
Level estimate
0.6
0.65
0.7
A Monte Carlo approach for efficient estimation of extreme response statistics p. 23/35
10
NTNU
0
0.45
0.5
0.55
0.6
0.65
0.7
Level estimate (m)
0.75
0.8
0.85
A Monte Carlo approach for efficient estimation of extreme response statistics p. 24/35
2
log10+()
NTNU
6
15
10
0
/
10
15
A Monte Carlo approach for efficient estimation of extreme response statistics p. 25/35
2
log10+()
NTNU
6
15
10
0
/
10
15
A Monte Carlo approach for efficient estimation of extreme response statistics p. 26/35
NTNU
2
10
0.4
0.6
0.8
1
1.2
103
1.4
1.6
4
1.8
10
2
105 P level
2.2 0
6
2.4
10
/
0.5
3
14
1.56
8 2
10
2.5
12
A Monte Carlo approach for efficient estimation of extreme response statistics p. 27/35
NTNU
2
10
0.4
0.6
0.8
1
1.2
103
1.4
1.6
4
1.8
10
2
5
10
P level
2.2 0
6
2.4
10
/
0.5
3
14
1.56
8 2
10
12 2.5
A Monte Carlo approach for efficient estimation of extreme response statistics p. 28/35
NTNU
0
5
10
15
20
25
30
20
20
10
10
0
0
10
10
20
20
A Monte Carlo approach for efficient estimation of extreme response statistics p. 29/35
NTNU
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
A Monte Carlo approach for efficient estimation of extreme response statistics p. 30/35
NTNU
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
A Monte Carlo approach for efficient estimation of extreme response statistics p. 30/35
NTNU
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
A simplified model for the surge response is adopted here:
A Monte Carlo approach for efficient estimation of extreme response statistics p. 30/35
NTNU
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
A simplified model for the surge response is adopted here:
1
2
3
Z + 2e (0 + cF2 (t))Z + e (Z + Z ) =
F1 (t) + F2 (t)
M
A Monte Carlo approach for efficient estimation of extreme response statistics p. 30/35
NTNU
+1.8
3
101.9
2
2.1
2.2
104
2.3
2.4
5
102.5
2.6
106
2.7
7
102.8
/
3.7
3 3.8
3.9
4.1
5
4.2
4.3
4.4
7
4.5
94.6
A Monte Carlo approach for efficient estimation of extreme response statistics p. 31/35
NTNU
+1.6
1.8
103
2.2
104
2.4
105
2.6
106
7
102.8
3.4
/
23.5
3.6
3 3.7
3.8
3.95
4.1
7
4.2
A Monte Carlo approach for efficient estimation of extreme response statistics p. 32/35
NTNU
+0.8
1
1.2
3
101.4
1.6
4
101.8
2
5
102.2
6
102.4
7
10
2.6
0.9
1.1
1.2
1.3
1.4
1.5
1.6
5 1.7
1.8
A Monte Carlo approach for efficient estimation of extreme response statistics p. 33/35
NTNU
1.9
10
2
2.1
4
102.2
2.3
2.4
5
10
2.5
6
102.6
2.7
7
10
0.8
1.2 4
1.4
1.6
1.8
A Monte Carlo approach for efficient estimation of extreme response statistics p. 34/35
Conclusions
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 35/35
Conclusions
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 35/35
Conclusions
NTNU
A Monte Carlo approach for efficient estimation of extreme response statistics p. 35/35
ARTICLE IN PRESS
Department of Mathematical Sciences, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway
Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway
c
Statoil ASA, NO-4035 Stavanger, Norway
Abstract
The paper describes a novel approach to the problem of estimating the extreme response statistics of a drag-dominated offshore
structure exhibiting a pronounced dynamic behaviour when subjected to harsh weather conditions. It is shown that the key quantity for
extreme response prediction is the mean upcrossing rate function, which can be simply extracted from simulated response time histories.
A commonly adopted practice for obtaining adequate extremes for design purposes requires the execution of 20 or more 3-h time domain
analyses for several extreme sea states. For early phase considerations, it would be convenient if extremes of a reasonable accuracy could
be obtained based on shorter and fewer simulations. The aim of the work reported in the present paper has therefore been to develop
specic methods which make it possible to extract the necessary information about the extreme response from relatively short time
histories.
The method proposed in this paper opens up the possibility to predict simply and efciently both short-term and long-term extreme
response statistics. The results presented are based on extensive simulation results for the Kvitebjrn jacket structure, in operation on the
Norwegian Continental Shelf. Specically, deck response time histories for different sea states simulated from an MDOF model were
used as the basis for our analyses.
r 2007 Elsevier Ltd. All rights reserved.
Keywords: Monte Carlo simulation; Mean crossing rate; Extreme response prediction; Jacket platform
1. Introduction
For drag-dominated offshore structures like jacket and
jack-up platforms that exhibit pronounced dynamic
behaviour in severe seas, there has not been available a
simple, accurate and fast method for estimating the
extreme response values. There are in particular two
different aspects of this problem that have caused this lack
of easily accessible methods. Firstly, there is the nonlinear
characteristics of the drag force itself. Secondly, due to the
fact that the structures considered are bottom-xed, there
is the added complexity of the inundation effect. These two
complications make simple approximations very hard to
achieve, if accurate results are required.
Corresponding author.
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2189
(1)
(2)
(3)
kd C d rD=2,
(4)
ARTICLE IN PRESS
2190
(6)
3. Dynamic response
In agreement with the lumped parameter model referred
in the previous section (Karunakaran et al., 2001), the
following dynamic model is adopted
CX
_ KX Q.
MX
(5)
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
n x lim
(9)
k
1 X
n x; T 0 ,
kT 0 j1 j
(10)
where n
j x; T 0 denotes the counted number of upcrossings of the level x by time history no. j. This will be the
approach to the estimation of the mean upcrossing rate
adopted in this paper.
For a suitable number k, e.g. kX20 30, and provided
that T 0 is sufciently large, a fair approximation of the
95% condence interval (CI0:95 ) for the value n x can be
obtained as
s^x
s^x
2
n^ x .
(12)
s^x
k 1 j1
T0
Note that k and T 0 may not necessarily be the number and
length of the actually simulated response time series.
Rather, they have been chosen to optimize the estimate
of Eq. (12). If initially, k~ time series of length T~ are
~ 0 and T~ k0 T 0 . That is, each initial
simulated, then k kk
time series of length T~ has been divided into k0 time series
of length T 0 , assuming, of course, that T~ is large enough to
allow for this in an acceptable way. The consistency of the
estimates obtained by Eq. (12) can be checked by the
observation that VarN x; t n xt since N x; t is a
Poisson random variable by assumption. This leads to the
equation
"
#
k N x; T
X
1
n x
0
j
2
,
(13)
sx Var
k
T0
T0
j1
where fN
1 x; T 0 ; . . . ; N k x; T 0 g denotes a random sample
2
s^x =k % n^ x=kT 0 . Since this last relation is consistent
with the adopted assumptions, it could have been used as
the empirical estimate of the sample variance in the rst
place. It is also insensitive to the blocking of data discussed
~ However, the advantage of Eq. (12)
above since kT 0 k~T.
is that it does not rely on any specic assumptions about
the statistical distributions involved.
2191
xXx0 ,
(14)
ARTICLE IN PRESS
2192
Table 1
Representative sea states
H s (m)
T p (s)
12.0
14.7
14.7
12.0
15.0
16.5
1
0
1
2
0.25
0.3
0.35
0.4
0.45 LG 0.5
Mk
0.55
0.6
0.65
Fig. 3. Empirical Gumbel plot of the 20 simulated 3 h extremes for the sea
state H s 12 m and T p 12 s together with the tted Gumbel distribution ( ).
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2193
Table 2
90% Fractile values of tted Gumbel distributions
LG (m)
CI0:95
0.47
0.618
0.632
(0.396, 0.540)
(0.548, 0.695)
(0.558, 0.726)
1
Bootstrapped PDF, Hs=12m, Tp=12s
0
12
10
2
0.35
0.4
0.45
0.5
0.55
Mk
0.6 LG 0.65
0.7
0.75
Fig. 4. Empirical Gumbel plot of the 20 simulated 3 h extremes for the sea
state H s 14:7 m and T p 15 s together with the tted Gumbel
distribution ( ).
ln(ln((n+1)/k))
P0 level
2
0
0.3
4
3
0.4
0.45
0.5
0.55
Level estimate
0.6
0.65
0.7
Fig. 6. Thee empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 12 m, T p 12 s. The
indicates the limits of CI0:95 .
P0 level
2
ln(ln((n+1)/k))
0.35
0
10
8
0.4
0.5
0.6 LG
Mk
0.7
0.8
Fig. 5. Empirical Gumbel plot of the 20 simulated 3 h extremes for the sea
state H s 14:7 m and T p 16:5 s together with the tted Gumbel
distribution ( ).
0
0.45
0.5
0.55
0.6
0.65
0.7
Level estimate (m)
0.75
0.8
0.85
Fig. 7. The empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 14:7 m, T p 15 s. The
indicates the limits of CI0:95 .
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2194
11
10
9
8
2
log10+()
7
6
5
3
4
4
3
2
1
0
0.4
0.5
0.6
0.7
Level estimate
0.8
0.9
Fig. 8. The empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 14:7 m, T p 16:5 s. The
indicates the limits of CI0:95 .
log10+()
2
3
4
5
6
15
10
0
/
10
15
Fig. 9. Mean upcrossing rate statistics along with 95% condence bands
() for the sea state with H s 12 m, T p 12 s, s 0:047 m. : Monte
Carlo; : linear t.
6
15
10
10
15
/
Fig. 10. Mean upcrossing rate statistics along with 95% condence bands
() for the sea state with H s 14:7 m, T p 15 s, s 0:066 m. : Monte
Carlo; : linear t.
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2195
102
log10+()
103
104
105
P0 level
106
6
15
10
10
15
/
Fig. 11. Mean upcrossing rate statistics along with 95% condence bands
() for the sea state with H s 14:7 m, T p 16:5 s, s 0:068 m. :
Monte Carlo; : linear t.
10
12
102
Fig. 13. Transformed plot along with 95% condence bands () for the
sea state with H s 14:7 m, T p 15 s, s 0:066 m. : Monte Carlo; :
linear t, q 0:06, b 0:9s.
103
103
104
105
6
/
104
P0 level
105
106
P0 level
106
3
10
12
/
Fig. 12. Transformed plot along with 95% condence bands () for the
sea state with H s 12 m, T p 12 s, s 0:047 m. : Monte Carlo; :
linear t, q 0:04, b 1:4s.
10
12
/
Fig. 14. Transformed plot along with 95% condence bands () for the
sea state with H s 14:7 m, T p 16:5 s, s 0:068 m. : Monte Carlo; :
linear t, q 0:06, b 1:1s.
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2196
Table 3
90% Fractile values by the new method
LCR (m)
PI0:95
0.489
0.618
0.623
(0.464, 0.515)
(0.607, 0.630)
(0.608, 0.645)
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
Naess, A., 1984b. On the long-term statistics of extremes. Applied Ocean
Research 6 (4), 227228.
Naess, A., Gaidai, O., 2006. Monte Carlo methods for estimating the
extreme response of dynamical systems. Technical Report, Norwegian
University of Science and Technology, Trondheim, submitted for
publication.
Naess, A., Gaidai, O., Haver, S., 2007. Estimating extreme response of
drag dominated offshore structures from simulated time series of
structural response. In: Proceedings 26th International Conference on
Offshore Mechanics and Arctic Engineering. ASME, New York, Paper
no. OMAE200729119.
Naess, A., Moan, T., 2005. Probabilistic design of offshore structures. In:
Chakrabarti, S.K. (Ed.), Handbook of Offshore Engineering, vol. 1.
Elsevier, Amsterdam, pp. 197277 (Chapter 5).
2197
PMC2000-151
Abstract
The paper investigates the application of the peaks over threshold (POT) method (Smith 1989; Davison and
Smith 1990) in combination with the bootstrapping method (Davison and Hinkley 1997) for estimating long
return period characteristic values of environmental loads, e.g. wind loads, on the basis of observed data.
In particular, attention is focused on the effect that different statistical estimators of the parameters inherent
in the POT method have on the predicted characteristic values. The accuracy of the predicted long return
period characteristic values provided by the different methods will be evaluated by application to synthetic
data where the relevant characteristic values can be calculated. The bootstrapping method is advocated as
a practical tool for the estimation of confidence intervals on the provided point estimates of the long return
period values.
Introduction
The focus of this paper is the practical implementation of the peaks over threshold (POT)
method for the estimation of extreme values characterized by return periods that are much
longer than the periods of data acquisition. Typically, the goal is to estimate extremes of an
environmental process like for instance wind speed with a return period of say 50 or 100
years on the basis of 25 years of recorded data.
The traditional method adopted for estimating 50 or 100 year return period values is usually
based on the assumption that the distribution of yearly extreme values can be approximated
by a Gumbel or Type I extreme value distribution (Gumbel 1958; Castillo 1988). The
general procedure is then to plot the data on Gumbel probability paper and carry out some
form of linear regression on the data to fit a straight line. By restricting the database to
comprise only yearly extreme values, a great amount of potentially useful information is
discarded. This procedure can be improved as discussed by Naess (1998a). In recent years
the POT has also received considerable attention.
The POT method is based on utilizing all peak events of a given time series exceeding
a specified threshold. The word event is used deliberately, because in many practical
applications, like wind speed time series, one would consider peak values with a time
separation less than a certain value to belong to the same peak event. For example, the
passage of a low pressure system with a duration of a few days is seen as an event in this
context. By considering peak events instead of yearly maxima, the number of datapoints
for statistical processing may be increased considerably. However, in this connection it is
Naess, Clausen
appropriate to recall the basic assumption of the POT method that the threshold is high
relative to typical values. Excessive lowering of the threshold to obtain more data may lead
to substantial bias. Gross et al. (1994) have performed numerical simulations that indicate
that in samples taken from extreme value populations, optimal results are obtained by a
choice of threshold corresponding to a number of exceedances of the order of ten per year
on the average.
Naess (1998b) used a POT analysis based on estimators proposed by de Haan (1994) to
analyse wind speed data recorded at 44 US weather stations. It was shown that the predicted
50 year values of wind speed obtained by the POT method were in general significantly
below the corresponding predictions obtained by using the traditional Gumbel method.
Naess and Clausen (1999) investigated the performance of the POT method using several
different estimators. The data used for the analyses were synthetic data sampled from
a stationary Gaussian process. For such processes the 50 year values can be accurately
given. It was established that the POT method in combination with the chosen estimators
significantly underestimated the correct values.
In the present paper a new estimator for use with the POT method will be introduced. It
will be demonstrated that this estimator provides good agreement with the target values.
The problem of estimating confidence intervals for the obtained long return period values
is also discussed, and it is pointed out that the bootstrapping technique offers a practical
solution.
The Peaks Over Threshold Method
The POT method is based on what is called the generalized Pareto (GP) distribution (defined below) in the following manner: It has been shown (Pickands 1975) that asymptotically, the excess values above a high level will follow a GP distribution if and only if the
parent distribution belongs to the domain of attraction of one of the extreme value distributions. The assumption of a Poisson process model for the exceedance times combined
with GP distributed excesses can be shown to lead to the generalized extreme value (GEV)
distribution for corresponding extremes, see below. The expression for the GP distribution
is
(1)
The asymptotic result referred to above implies that equation (1) can be used to represent
the conditional cumulative distribution function of the excess
of the observed
variate over the threshold u, given that
for sufficiently large (Pickands 1975).
,
and
correspond to Frechet (Type II), Gumbel (Type I), and
The cases
reverse Weibull (Type III) domains of attraction, respectively. For
, the expression
between the parentheses in equation (1) is understood in a limiting sense as the exponential
.
Naess, Clausen
The return period of a given wind speed, in years, is defined as the inverse of the probability that the specified wind speed will be exceeded in any one year. In this section we give
expressions that allow the estimation from the GP distribution of the excess value above
a level corresponding to any percentage point , where is the mean crossing
rate of the threshold per year (i.e., the average number of data points above the threshold
per year). That is,
. Invoking equation (1) for
leads to
the result
. Similarly, for
, it is found that
. The
, where is the threshold
required value with the return period is then
used in the estimation of and .
Statistical Estimators
The de Haan Estimators
Let denote the total number of data points, while the number of observations above the
threshold value is denoted by . The threshold then represents the -th highest
data point(s). An estimate for is
, where denotes the length of the record
in years. The highest, second highest, ... , -th highest, -th highest variates are
denoted by , , ..., ,
, respectively.
The parameter estimators proposed by de Haan (1994) are based on the following two
quantities
Estimators for
and
and
(2)
(3)
where
if
, while
if
. de Haan (1994) showed that under general
conditions on the underlying probability law,
and
as
(in prob.).
The Moment Estimators
In terms of the the mean value and the standard deviation of the exceedance
variate , it can be shown that (Hosking and Wallis 1987)
and
(4)
Naess, Clausen
(5)
3
, where = 1 year, which would be typical for the wind speed process. The
distribution of the yearly extreme value of is then calculated by the formula
. The 50-year return period value is then calculated from
the relation , which gives
.
The corresponding data to be used for the POT analyses, are generated by observing that the
peak events extracted from measurements of the wind speed process are usually peak events
that are separated by three to four days. This is done to obtain approximately independent
data, cf. Naess (1998b). In accordance with this, the peak event data are generated from
the extreme value distribution
, where
,
which implies that corresponds to approximately 3.65 days.
For the POT analyses, 5000 independent time series of 2000 peak events in each time series
were generated from . Each time series then corresponds to 20 years of data.
The POT analyses were carried out as follows: First, a threshold level was chosen so that
about 40 peak events exceeded this threshold. The subsequent threshold levels were chosen
so that approximately 10 new peak events were added for each new threshold. A typical
feature of the resulting estimates is a significant variation as function of threshold. To
reduce the variability and facilitate the practical choice of a suitable threshold to be used
for the prediction of , the results are smoothed by a running
filter. Specifically, in
and denotes
Figure 1 are plotted the results for , where =
the estimate for data points above the threshold. For comparison, the result obtained by
the Gumbel method for yearly extremes combined with transformation of data was 4.64,
cf. Naess (1998a). The target value is 4.65.
Figure 2 shows the mean values of the calculated estimates from the 5000 independent samples as a function of the number above the threshold. The target value of 4.65 is
Naess, Clausen
4.9
Dehaan
c0
c0 mom
mom
4.9
Dehaan
c0
c0 mom
mom
4.8
4.8
50yr
x50yr
4.7
4.7
4.6
4.6
4.5
4.5
50
100
150
200
250
50
100
150
200
250
Figure 2. The mean value of the calculated estimates from 5000 independent
samples.
accurately predicted by the Gumbel method combined with transformation of data giving
4.66. Judged from the mean values, Figure 2 indicates that the POT analysis in combination with the de Haan and the de Haan estimators lead to varying degrees of significant
underestimation of the target value. The best performance is clearly offered by the moment
and the moment estimators. On the average they are also more stable as a function of
threshold level expressed through the parameter . While the moment estimator leads on
the average to a slight underestimation, the moment estimator leads to a slight overestimation of the target value.
As mentioned in the Introduction, another topic that we want to discuss is the problem of
estimating confidence intervals associated with the predicted long return period values. In
particular, our concern is the performance of the bootstrapping method for this purpose.
To elucidate on this, for the specific choice of = 145, two empirical PDFs have been
calculated for each one of the estimators for discussed above. One is obtained from
the 5000 independent samples, providing an accurate approximation to the target PDF,
from which the confidence interval can be derived. The other is obtained on the basis of
one arbitrary sample for which 5000 nonparametric bootstrap samples have been generated.
The conclusion reached from a comparison of these PDFs is that the bootstrap method will
provide reasonably accurate estimates of the confidence intervals associated with the various estimators. In particular, there is excellent agreement between the target and bootstrap
PDFs for the two estimators. A feature of significant practical interest is the observation
that the estimators have much better efficiency than the other two estimators. This is
clearly connected to the fact that effectively only one estimated parameter enters the
estimator, viz. , while estimates of both and enter the other two estimators.
Conclusions
The main focus of this paper has been the performance of the POT method in combination
with various estimators. The goal has been the prediction of characteristic values associated
with long return periods, typically 50 or 100 years, on the basis of recorded data for shorter
Naess, Clausen
periods of time, e.g. 20 years. In particular, the attention has been directed toward the de
Haan and moment estimators applied to the data as recorded, that is, to untransformed data.
In addition, two new estimators have been proposed, which are related to transformed data.
The transformation is derived from the underlying distribution of the instantaneous value
of the process under study.
An extensive set of synthetic data have been used to test the performance of the various estimators. The overall conclusion is that the de Haan estimator for untransformed data and the
de Haan estimator for transformed data on the average lead to significant underestimation of the target values. On the other hand, the moment estimator for untransformed data
and the proposed moment estimator for transformed data, provide on the average much
better agreement with the target values. It has also been pointed out that the estimators
are superior in terms of efficiency.
References
Castillo, E. (1988). Extreme Value Theory in Engineering. San Diego: Academic Press.
Davison, A. C. and D. V. Hinkley (1997). Bootstrap Methods and their Applications. London: Cambridge
University Press.
Davison, A. C. and R. L. Smith (1990). Models of Exceedances Over High Thresholds. Journal of the
Royal Statistical Society 52(Ser. B), 393442.
de Haan, L. (1994). Extreme Value Statistics. In J. Galambos, J. A. Lechner, and E. Simiu (eds.), Extreme
Value Theory and Applications. Dordrecht: Kluwer Academic Publishers.
Gross, J. L. et al. (1994). Novel Extreme Value Procedures: Application to Extreme Wind Data. In J. Galambos, J. A. Lechner, and E. Simiu (eds.), Extreme Value Theory and Applications. Dordrecht: Kluwer Academic Publishers.
Gumbel, E. (1958). Statistics of Extremes. New York: Columbia University Press.
Hosking, J. R. M. and J. R. Wallis (1987). Parameter and Quantile Estimation for the Generalized Pareto
Distribution. Technometrics 29, 339349.
Naess, A. (1998a). Estimation of Long Return Period Design Values for Wind Speeds. Journal of Engineering Mechanics, ASCE 124(3), 252259.
Naess, A. (1998b). Statistical Extrapolation of Extreme Value Data Based on the Peaks Over Threshold
Method. Journal of Offshore Mechanics and Arctic Engineering, ASME 120, 9196.
Naess, A. and P. H. Clausen (1999). Statistical Extrapolation and the Peaks Over Threshold Method.
OMAE996422. In Proceedings 18th International Conference on Offshore Mechanics and Arctic Engineering. New York: ASME.
Pickands, J. (1975). Statistical Interference Using Order Statistics. Annals of Statistics 3, 119131.
Smith, R. L. (1989). Extreme Value Theory. In W. Ledermann, E. Lloyd, S. Vajda, and C. Alexander (eds.),
Handbook of Applicable Mathematics, pp. 437472. New York: John Wiley and Sons.
Naess, Clausen
Received 4 October 2006; received in revised form 27 August 2007; accepted 30 August 2007
Available online 12 September 2007
Abstract
The paper presents a detailed study of the structure and asymptotic behaviour of a second-order stochastic Volterra series model of the slow
drift response of large volume compliant offshore structures subjected to random seas. A long standing challenge has been to develop efficient
and accurate methods for calculating the response statistics of compliant offshore structures to random seas. Recent work has revealed that
the statistical properties of the response process, which consist of a linear, first-order component and a nonlinear, second-order component, is
surprisingly complex. The goal of the research work presented here is to complement efforts to develop numerical procedures to calculate the
statistics of the response process.
c 2007 Elsevier Ltd. All rights reserved.
Keywords: Stochastic Volterra series; Slow drift response; Offshore structures; Response statistics
1. Introduction
The response of compliant large volume offshore structures
to random seas is typically characterized by two timescales.
In the case of the motion response of for example a moored
drilling rig or a tension leg platform, a common feature
is the presence of slowly varying response at frequencies
well below those of the wave energy spectrum. This motion
behaviour occurs in the lateral motion modes (surge, sway,
yaw) due to the existence of the mooring system, which often
leads to resonance periods in these motion modes of the
order of minutes. The common model for the hydrodynamic
excitation forces at these periods is a nonlinear, second-order
transformation from the wave elevation process. Since the wave
elevation in this context is usually modelled as a Gaussian
process, this leads to a non-Gaussian excitation and therefore
in general to a non-Gaussian slow drift response. In such cases
the total response is consequently also non-Gaussian. As will be
shown, when the equations of motion for the floating structure
Corresponding address: CeSOS/NTNU, Otto Nielsens vei 10, NO-7491
Trondheim, Norway. Tel.: +47 73 59 46 31; fax: +47 73 59 35 24.
E-mail address: arvidn@math.ntnu.no (A. Naess).
0266-8920/$ - see front matter c 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.probengmech.2007.08.003
344
Z 1 (t) =
0
F1 (t) =
0
k1 ( ; ) d X (t ) d,
(1)
k2 (, ; , )
d X (t ) d X (t ) d d ,
(2)
Z (t) =
l( ) F(t ) d,
0
h 2 (, ; , )
(5)
l(t) k1 ( t; ) dt,
h 1 ( ; ) =
(6)
d X (t ) d X (t ) d d .
Z 2 (t) =
(4)
h 1 ( ; ) d X (t ) d,
(3)
and
h 2 (, ; , ) =
K 1 (; ) =
k1 ( ; ) ei d,
(8)
and
K 2 (, ; , ) =
k2 (, ; , )
0
ei( + ) d d .
(9)
X (t) =
j=L k=1
1
S X (| j |; k ) j k B jk ei j t .
2
(10)
345
H 1 (; ) = L()
K 1 (; ) and H 2 (, ; , ) = L(
+
L()
= [2 M() + i C() + K ]1 ,
(11)
Z 1 (t) =
q jk B jk ei j t ,
(12)
j=L k=1
where
q jk = H 1 ( j ; k )
1
S X (| j |; k ) j k ,
2
(13)
and
L
Z 2 (t) =
(14)
where
Q i jkl = H 2 (i , j ; k , l )
1
2
S X (|i |; k ) S X (| j |; l ) i j k l . (15)
for < 0.
(16)
This condition clearly implies that all sum-frequency components of the second-order excitation forces are neglected, retaining only the difference frequency components, which are
the only terms that can give rise to the slowly varying response.
Introducing the L M L M-matrix G = (G ), where
G = Q i jkl ,
= (k 1)L + i
= (l 1)L + j
i, j = 1, . . . , L
(17)
k, l = 1, . . . , M,
(18)
Z 2 (t) =
(19)
=1
W (t) =
w ( j , k ) B jk ei j t ,
(20)
j=L k=1
( j , k )
2
w21 ( j , k ) =
1
( j , k )
j > 0
j < 0,
(21)
and
( j , k )
2
w2 ( j , k ) =
i
( j , k )
2
j > 0
j < 0,
(22)
Z 1 (t) =
(23)
=1
H 1 ( j ; k )
d21 i d2 =
j=1 k=1
S X (| j |; k ) j k ( j , k ) .
(24)
346
Z (t) =
( W (t)2 + d W (t)),
(25)
=1
Z+ ( ) =
(31)
s f Z Z (, s) ds.
(26)
f Z Z (z, z ) =
f W2 ...Wn Z Z
f W2 ...Wn Z Z (w2 , . . . , wn , z, z )
w1+
z
+ f W1 W2 ...Wn Z (w1 , w2 , . . . , wn , z )
w1
z
11
21
(33)
= E[ Z |W = w] = p 21 w,
(34)
1
2
exp
(z )
2
.
2 2
(36)
(37)
j=1
(28)
(29)
j=2
(30)
1
(2)n/2
w1
1
= [1 (z + c12 /(41 ) q(w))]1/2 .
z
2
12
22 ,
( j w 2j + c j w j ),
f W (w) =
(32)
j=1
(27)
= f W1 W2 ...Wn Z (w1+ , w2 , . . . , wn , z )
2 j W j W j + c j W j ,
Z =
...
1
1
exp
+
1
1
exp
2
s ds dw2 . . . dwn .
1
(w1+ )2
2
1
(w1 )2
2
(38)
347
1
2
exp
1
2
x dx
(39)
exp
21
(2 )n/2
n
exp 21
...
2[1 (
=+,
c1
21
ik w j (k , l ) Bk l eik t
k =L l =1
L
ik wi (k , l ) w j (k , l )
(l j w 2j c j w j /1 )
j=3
ik wi (k , l ) w j (k , l ) .
c1
c2
w2
[ + c12 /(41 ) q(w)]1/2 +
21
21 1
dw2 . . . dwn ,
(40)
where l j = 1 j /1 and = / . Eq. (40) will
now be used to study the asymptotic behaviour of Z+ ( ) as
. To this end it is necessary to first establish the
asymptotic behaviour of and . The following parameters
are introduced
r jk
E[W j W k ]
= qk j .
s j sk
(41)
i, j = 1, . . . , n.
j = 1, . . . , N .
(42)
ik w2i (k , l ) w2 j (k , l )
s2 j r2i,2 j =
k=L l=1
L
and
=
n
= Sk j ,
j = k.
(43)
l=1
k=L l=1
L
ik w2i1 (k , l ) w2 j1 (k , l )
k=L l=1
= s2 j1 r2i1,2 j1 ,
ik w2i1 (k , l ) w2 j (k , l )
k=L l=1
k2 | w2 j1 (k , l ) |2 = s22 j1 .
k=L l=1
(50)
s2 j r2i1,2 j =
k=L l=1
k2 |(k ) w2 j1 (k , l )|2
L
ik (k ) w2i1 (k , l )
k=L l=1
(k ) w2 j1 (k , l )
k2 |w2 j (k , l )|2
(49)
k=1
(48)
and
rl j rlk
r2i,2 j = r2i1,2 j1
S jk = s j sk q jk
(47)
q2 j1,2 j = 0 ,
rk2j ,
(46)
k=L l=1
i, j = 1, . . . , N ,
ik w j (k , l ) wi (k , l )
k=L l=1
G( )
E[W j W k ]
=
,
sk
si r ji = E[W j W i ] =
s 2j = Var[W j ] = E[W j2 ],
(45)
k=L l=1
exp
q jk =
wi (k , l ) Bkl eik t
k=L l=1
x
(t) dt.
Z+ ( ) =
=E
(44)
ik w2i1 (k , l ) (k ) w2 j1 (k , l )
=
k=L l=1
348
ik (k ) w2i1 (k , l ) w2 j1 (k , l )
S j j (2 j w j + c j )2
+
j=2
k=L l=1
L
Si j (2i wi + ci )(2 j w j + c j ).
+2
ik w2i (k , l ) w2 j1 (k , l )
(58)
2i< jn
k=L l=1
= s2 j1 r2i,2 j1 ,
(51)
i, j = 1, . . . , N ,
(52)
s2 j1r2 j1,2 j c2 j w2 j1 c2 j1 w2 j
and
j=2
2
2
2
2
r2i1,2
j1 + r2i,2 j1 = r2i1,2 j + r2i,2 j ,
i, j = 1, . . . , N .
(53)
s2 j1r2 j1,2 j c2 j w2 j1 c2 j1 w2 j
+ 21
j=2
j = 1, . . . , N ,
ri,2 j1 ri,2 j = 0,
(54)
i=1
and
c1 (w2 + c2 ) ,
n
2
(ri,2
j1
2
ri,2
j)
= 0,
j = 1, . . . , N ,
(55)
i=1
k2 w2 j1 (k , l ) w2 j (k , l )
s22 j1 q2 j1,2 j =
k=L l=1
L
k2 (k ) |w2 j1 (k , l )|2 = 0,
(56)
k=L l=1
+ s1 r12 c2 w1 c1 w2 + c2 s1
r j2 w j
(2k wk + ck ) sk r jk w j ,
j=3 k=3
where l j = 1 j /1 .
2
( ) = ( p ) S p =
+ 2(21 w1
(21 w1
n
+ c1 )
+ c1 ) S11
2
S1 j (2 j w j + c j )
j=2
+ c22
(61)
n
2
l=1 rlk .
+ c12
c1 (w2 + c2 ) ,
j=3
n
12
j=3
(60)
= ( p ) 21 w
= 21 s1
(59)
(57)
s1 + c12 + c22
(2)n/2
+c12 +c22
+c12 +c22
dw3 . . .
du
dwn
349
exp
1
2
exp c1 [
[
(l j w 2j 2c j w j )
j=3
=+,
+ c2 u
G( )
Prob (M Z (T ) ) = exp{ Z+ ( ) T },
(62)
expression (u =
Z+ ( )
n
s1 + c12 + c22
2
exp{c2j /(2l j )}
j=3
= exp + eb
1 =+,
+ =
dx.
(63)
Z+ ( )
+ c12 + c22
c12 + c22
+ c12 + c22
c12 + c22
, (64)
1
kj
exp
y
1
+ b y + ln y
2
4
j=2
(65)
(68)
y dPY (y) y0 +
E[Y ] =
0
,
(y0 )
(69)
E[Y ] = Q + 2b Q +
1/4
2
s1
n+
Z 2 ( )
2
s1
2 b
kj
j=2
T exp
where
G( )
+ c12 + c22 c1 1 x 2 + c2 x
1 x2
2 /2
(67)
exp
where the expression for Z+ ( ) given by the rhs of Eq. (64) will
be used.
Let Y = 1 M Z (T ) + b2 , where b2 = c12 + c22 . Then the
probability distribution of Y is given as
PY (y) = Prob (Y y)
lj
(66)
1
ln Q
2
b
1+
Q
b
+ 2 + 2b2 + (1 + b2 2 ) + O
Q
1
Q
, (70)
E[M Z (T )] = 21 ln( + T ) + 1 2b Q +
b
1+
Q
1
ln Q
2
b
+ 2 + (1 + b2 2 )
+O
Q
1
Q
.
(71)
+ ln
21
2
1
(72)
where
n+ =
2s1
1
.
k
j=2 j
(73)
350
Table 1
Particulars of the TLP
Column diameter D (m)
Eigenperiod surge Te (s)
Relative damping
Total mass (incl. added mass) M (kg)
E[M Z 2 (T )] = 1 Q 2 + ln Q 2 + 2
+O
1
3/2
Q2
10.0
128.8
0.05
1.5 107
1
1
Q2
(74)
| L()|
is equal to or higher than half of the resonance peak
5 p
4
g 2
exp
5
exp
1
2 2
+ ln
1
p
(75)
Hs
T p2
(1 0.287 ln ) ,
(76)
(77)
351
Z 2 , mod.
sea
Z 1 + Z 2 , mod.
sea
Z 2 , sev.
sea
Z 1 + Z 2 , sev. sea
0.70
0.75
0.80
0.85
0.90
0.95
0.98
0.98
0.98
0.98
0.98
0.99
0.89
0.89
0.90
0.90
0.90
0.91
0.99
0.99
0.99
0.99
0.99
0.99
0.97
0.97
0.98
0.98
0.99
0.99
(A)
[1] Kac M, Siegert AJF. On the theory of noise in radio receivers with square
law detectors. Journal of Applied Physics 1947;18:38397.
[2] Neal E. Second order hydrodynamic forces due to stochastic excitation.
in: Proc. 10th ONR symposium. 1974.
[3] Vinje T. On the statistical distribution of second-order forces and motions.
International Shipbuilding Progress 1983;30:5868.
[4] Naess A. Statistical analysis of second-order response of marine
structures. Journal of Ship Research 1985;29(4):27084.
[5] Langley RS. A statistical analysis of low frequency second-order forces
and motions. Applied Ocean Research 1987;9(3).
[6] Kato S, Ando S, Kinoshita T. On the statistical theory of total second order
response of moored floating structures. in: Proc. 19th annual offshore
technology conference. Paper no. 5579. 1987.
[7] Donley MG, Spanos PD. Dynamic analysis of non-linear structures by the
method of statistical quadratization. Lecture notes in engineering, vol. 57.
Berlin: Springer-Verlag; 1990.
[8] Naess A. Statistical analysis of nonlinear, second-order forces and
motions of offshore structures in short-crested random seas. Probabilistic
Engineering Mechanics 1990;5(4):192203.
[9] Langley RS, McWilliam S. A statistical analysis of first and second
order vessel motions induced by waves and wind gusts. Applied Ocean
Research 1993;15(1):1323.
[10] McWilliam S, Langley RS. Extreme values of first and second order wave
induced vessel motions. Applied Ocean Research 1993;15(3):16981.
[11] Kareem A, Zhao J, Tognarelli MA, Gurley KR. Dynamics of nonlinear
stochastic systems: A frequency domain approach. In: Haldar A,
Ayyub BM, Guran A, editors. Uncertainty modelling in vibration, control
and fuzzy analysis of structural systems. Singapore: World Scientific
Publishing Co.; 1997. p. 10146.
[12] Naess A. The response statistics of non-linear second-order transformations to Gaussian loads. Journal of Sound and Vibration 1987;115(1):
10329.
352
[13] Naess A, Karlsen HC. Numerical calculation of the level crossing rate
of second order stochastic Volterra systems. Probabilistic Engineering
Mechanics 2004;19(2):15560.
[14] Naess A, Karlsen HC, Teigen PS. Numerical methods for calculating the
crossing rate of high and extreme response levels of compliant offshore
structures subjected to random waves. Applied Ocean Research 2006;
28(1):18.
[15] Benfratello S, Di Paola M, Spanos PD. Stochastic response of MDOF
wind excited structures by means of Volterra series approach. Journal of