Simulation of Stochastic Wind Velocity Field On Long-Span Bridges
Simulation of Stochastic Wind Velocity Field On Long-Span Bridges
Simulation of Stochastic Wind Velocity Field On Long-Span Bridges
net/publication/245286240
CITATIONS READS
113 354
3 authors, including:
Yinghong Cao
Independent Researcher
21 PUBLICATIONS 216 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Yinghong Cao on 23 July 2016.
ABSTRACT: An improved algorithm is introduced in this paper for digital simulation of the stochastic wind
velocity field on long-span bridges, when the cross-spectral density matrix of the field is given. The target wind
velocity field is assumed to be a one-dimensional, multivariate, homogeneous stochastic process. The basic
method of simulation used is the spectral representation method. It is improved by explicitly expressing Cho-
lesky’s decomposition of the cross-spectral density matrix in the form of algebraic formulas, then cutting off as
many as possible of the cosine terms, so long as the accuracy of results is not affected. The fast Fourier transform
technique is used to enhance the efficiency of computation. A numerical example of simulation for buffeting
analysis is included in this paper to illustrate the improved method introduced. It is demonstrated that deviations
between the simulated correlation functions and the target are sufficiently small and that the simulated power
spectra are close to the target.
Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
SIMULATION OF ONE-DIMENSIONAL, MULTIVARIATE The period of the simulated function expressed by (2) can be
STOCHASTIC PROCESS determined by the formula
Let us consider a one-dimensional, multivariate stochastic 2pn 2pnN
T0 = = (10)
process { f (t)}, which has n components f 1(t), f 2(t), . . . , fn(t), Dv vup
with zero as their mean value. The cross-spectral density ma-
The ergodicity of the results of (2) has been proved by Deo-
F G
trix is given by
datis (1996).
S 011(v) 0
S 12(v) ??? 0
S 1n (v) It is clear that the one-dimensional, multivariate, stationary
S 021(v) 0
S 22(v) ??? 0
S 2n (v) stochastic Gaussian process can be simulated quite well by
S0(v) = (1)
??? ??? ??? ??? means of expression (2), when the cross-spectral density ma-
S 0n1(v) 0
S n2(v) ??? S 0nn(v) trix has been given and values of the parameters N, vup , and
Dt have been properly chosen.
According to Shinozuka and Jan (1972), the typical com-
ponent fj (t) of { f (t)} can be simulated by the sequence CALCULATION OF H(v)
OO
j N
The cross-spectral density matrix S0(v) is usually a complex
fj (t) = Ï2(Dv) uHjm (vml)ucos(vml t 2 ujm (vml) 1 Fml),
m=1 l=1
matrix; hence, the calculation of its Cholesky’s decomposition
is tedious and often turns to recursion formulas. Since H(v)
j = 1, 2, . . . n (2) is a function of v, it can be seen from the structure of (2) that
Cholesky’s decomposition has to be calculated separately for
where N is a sufficiently large number; vup = upper cutoff every frequency vml . Moreover, it must be calculated once
frequency, with the condition that, when v > vup , the value of more when another cosine term is added. The labor and time
S0(v) is less than ε1 and can be ignored; Dv = (vup /N) is the consumed in computation thus become enormous. When long-
frequency increment; F1l, F2l, . . . , Fjl = sequences of indepen- span bridges are considered, the writers found that some ex-
dent random phase angles, distributed uniformly over the in- pedient simplification could be introduced so that Cholesky’s
terval [0, 2p]; and Hjm (vml) is a typical element of matrix decomposition can be expressed in an explicit form with sev-
H(v), which is defined by Cholesky’s decomposition of cross- eral algebraic formulas, and the time for computation can thus
spectral density matrix S0(v); thus be remarkably reduced. Such simplification will be explained
in the following.
F G
S0(v) = H(v)HT*(v) (3)
The orthogonal spectrum of wind velocity is very small in
H11(v) 0 ??? 0 atmosphere and can be omitted. Hence, S 0ij (v) can usually be
H21(v) H22(v) ??? 0 treated as real, and S0(v) can be considered a real matrix.
H(v) = (4)
??? ??? ??? ??? When the horizontal wind velocity field acting on a long-span
Hn1(v) Hn2(v) ??? Hnn(v) bridge is to be simulated, the velocity field can be treated as
a one-dimensional, multivariate Gaussian process. Usually we
As S0(v) is usually a complex matrix, H(v) is generally can approximately take the spectra of horizontal wind veloc-
also a complex matrix whose diagonal elements are real, non- ities as not varying along the length of the bridge. Then the
negative functions of v, while the off-diagonal elements (ex- spectra of all points on the bridge deck are the same. We then
cept zeros) are complex functions of v. For the diagonal ele- have
ments, there exists the relation
S 011(v) = S 022(v) = ? ? ? = S 0nn (v) = S(v) (11)
Hjj (v) = Hjj (2v), j = 1, 2, . . . n (5)
and
while for the off-diagonal elements, there exists the relation
S 0jm (v) = ÏS 0jj (v)S 0mm (v)Coh(Djm , v) = S(v)Coh(Djm , v)
Hjm (v) = uH *
jm (2v)ue
iujm(v)
, j = 1, 2, . . . n, m = 1, 2, . . . j 2 1
(6) j, m = 1, 2, . . . n, j≠m (12)
where Djm = horizontal distance from point j to m; and
where ujm (v) = complex angle of Hjm (v) and is given by
Coh(Djm , v) = coherence function between points j and m.
S D
where Im[Hjm (v)] and Re[Hjm (v)] are, respectively, the imag-
lvDjm
inary and real parts of the complex function Hjm (v). Coh(Djm , v) = exp 2 (13)
According to the work of Shinozuka et al. (1989), the period 2pU (z)
of the simulated sample function will be sufficiently large, if Considering the case of n uniformly distributed points on
we let the bridge deck with a distance interval D between successive
m points, we have Djm = D( j 2 m) when j > m, or Djm = D(m 2
vml = (l 2 1)Dv 1 Dv, l = 1, 2, . . . , N (8) j) when m > j. Thus, Djm = Du j 2 mu. It can be proved that
n
Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
C = exp S2
lvD
2pU(z) D (15) Sjm (v) = S(v) O
m
k=1
Gjk (v)GTkm (v) (22)
Utilizing the relations (11)–(14), the cross-spectrum matrix It should be noticed that the value of Gjm (v) in (18b) will
F G
S0(v) can be rewritten as follows: decrease rapidly when the variables v, D, and u j 2 mu increase
in value. If Gjk (v0) < 1023, and v > v0, it is clear that Gjk (v)
1 C C2 ??? C n21 will be still smaller; hence, the contribution of Gjk (v) to Sjm (v)
C 1 C ??? C n22 is very small and can be omitted without affecting accuracy.
S0(v) = S(v) C2 C 1 ??? C n23 (16) It is easy to determine some small number ε0, and then cut off
??? ??? ??? ??? ??? those cosine terms in (21), when their multiplier Gjk (v) is less
C n21 C n22 C n23 ??? 1 than ε0. In a numerical example, 1023 is used for ε0 and 83.6%
of the cosine terms can be cut off in this way; hence, time of
S0(v) in (16) is expressed in the form of an exponential de- computation is remarkably reduced.
cayed matrix. It can be proved with the method of mathemat-
ical induction that Cholesky’s decomposition H(v) of (16) can APPLICATION OF FAST FOURIER TRANSFORM
be expressed in explicit form; thus TECHNIQUE
H(v) = ÏS(v)G(v) (17) It has been stated in some papers that the efficiency of sim-
ulation can be enhanced greatly by utilizing the FFT tech-
where nique. In works by Brigham (1988) and Deodatis (1996), it
1 0 has been proved that (21) can be rewritten as follows:
HO F S D GJ
j
C Ï1 2 C 2 mDv
fj ( pDt) = Re hjm (qDt)exp i ( pDt) (23a)
C2 CÏ1 2 C 2 Ï1 2 C 2 m=1 n
G(v) =
C3 C 2 Ï1 2 C 2 CÏ1 2 C 2 Ï1 2 C 2 p = 0, 1, . . . 2N 3 n 2 1, j = 1, 2, . . . n (23b)
??? ??? ??? ??? ???
C n21 C n22Ï1 2 C 2 C n23Ï1 2 C 2 C n24Ï1 2 C 2 ? ? ? Ï1 2C 2 where q is the remainder of p/(2N), q = 0, 1, 2, . . . n 2 1 and
(18a) hjm (qDt) is given by
O S D
2N21
Gjm (v) = H 0,
C u j2mu,
C u j2muÏ(1 2 C 2),
when 1 # j < m # n
when m = 1, m # j # n
when 2 # m # j # n
(18b)
where, in the current paper, Bjm (lDv) is expressed with the
following equations:
Bjm (lDv)
H S D
Since 0 < C < 1, Ï(1 2 C 2) must be a positive real num- mDv
ber, H(v) is a real matrix, and it can be proved that Ï2(Dv)S(v)Gjm lDv 1 exp(ifml), when 0 # l < N
= n
T*
H (v) = H (v) T
(19) 0, when N #l < 2N
OO
j N
pension bridge to be constructed across the Yangtze River in
fj (t) = Ï2(Dv) ÏS(vml)Gjm (vml)cos(vml t Jiangsu Province, P.R. China. The main span of Jiangyin
m=1 l=1
Bridge is 1,385 meters. The wind velocity field on the bridge
1 Fml), j = 1, 2, . . . n (21) deck is assumed to be composed of 50 wind velocity waves
at 50 different points uniformly distributed along the bridge.
Obviously, the simulation of the wind velocity field on long- The distance between successive points is (1,385/49) = 28.27
span bridges is greatly simplified, when we use (21) together m. The bridge, with 50 uniformly distributed points, is shown
with (15), (18a), and (18b). in Fig. 1.
The main data of the bridge and of simulating conditions
CUTTING OFF PART OF COSINE FUNCTIONS are as follows:
When the distance D is large enough (say, 300 m), there • Span: L = 1,385 m
will be in reality little coherence between the wind velocities
(especially their high frequency part) at two points with dis-
tance D apart. Hence, it is practicable to enhance the efficiency
of computation without affecting its accuracy by omitting the
coherence between velocities of points at large distances apart.
Utilizing the relations (3), (17), (18a), and (18b), it can be
proved that FIG. 1. Position of 50 Points on Bridge Deck
Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
• Height of the deck above ground: z = 50.0 m
• Ground roughness: z0 = 0.03 m
• Average wind velocity on the deck: U(z) = 40.0 m/s
• Number of simulated points: n = 50
• Interval between points: D = 28.27 m
• Upper cutoff frequency: vup = 4p rad/s
• Dividing number of frequency: N = 1,024
• Time interval: dt = 0.25 s
• Period: T0 = 25600 s
• Target wind spectrum: Kaimal’s spectrum
SD
f= , U = , K = 0.4
U (z) * z
ln
z0
After substituting the above variables with values of this
bridge, S(v) is computed by the formula
1163
S(v) = m2/s (27)
(1 1 9.95v)5/3
Davenport’s coherence function is computed with the formula
Coh(D, v) = exp S
2
lvD
2pU (z)
D = exp(21.125v) (28)
Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
(from Kaimal’s spectrum) to the Panofsky-McCormick spec- It can be seen from (32) that for n = 1 and n = 2, the values
trum (or some other vertical spectrum). of G(v) agree with (18a) and (18b).
We can suppose that for n = p 2 1, the values of G(v) can
CONCLUDING REMARKS be computed according to (30).
The new value of the first column of G(v) is
Fast and effective simulation of the natural wind velocity
field is of great importance in the nonlinear buffeting analysis Sp 1(v)
Gp 1(v) = = Sp 1(v) = C p21 (33)
of long-span bridges. In this paper, an improved simulation G11(v)
algorithm is proposed to generate the sample functions of a Obviously, the value of Gp1(v) agrees with (18a) and (18b).
stationary, one-dimensional, multivariate stochastic wind ve- The new value of the second column of G(v) is
locity process, when the cross-spectral density matrix of the
wind velocity field is given. Sp 2(v) 2 Gp 1(v)G21(v) C p22 2 C p21C
A numerical example is included in this paper to demon- Gp 2(v) = =
G22(v) Ï1 2 C 2
strate the capabilities and efficiency of the proposed algorithm.
When the problem of buffeting analysis of long-span bridges = C p22Ï1 2 C 2 (34)
under turbulent wind excitations is studied, the proposed
method can perform satisfactory simulations with high effi- Again, the value of Gp2(v) agrees with (18a) and (18b).
ciency. Suppose that the new value of column s 2 1 (s < p 2 1)
Application of the proposed algorithm is limited to analysis agrees with (18a) and (18b); then the new value of column s
of a one-dimensional (considering the horizontal or vertical is
O
component only), multivariate wind velocity field at uniformly s21
be strongly correlated even for points far apart). There are, of Gss
O
course, approximations and simplifications. However, for most s21
dynamic analyses in structural wind engineering, including
C p2s 2 C p21C s21 2 C p2rÏ1 2 C 2 ? C s2rÏ1 2 C 2
buffeting analysis, such approximations are usually good r=2
enough. =
A very efficient simulation of the full vector wind field, with Ï1 2 C 2
wind shear and proper coherence considered, has been pro- C p2s 2 C p1s22 2 (C p2s12 2 C p1s22)
posed by Mann (1998). Thus it should be considered when = = C p2sÏ1 2 C 2
more advanced problems are studied. Ï1 2 C 2 (35)
The new value of column p is
S O D
APPENDIX I. PROOF OF (18a) AND (18b) p 21 1/2
F G
Eq. (16) is rewritten here as (29) for convenience. r=1
S O D
p 21 1/2
1 C C2 ??? C n21
= 12C 2( p21)
2 (1 2 C ) 2
C 2( p2r)
S D
C 2( p22) 2 C 2 ? 1/2
C n21 C n22 C n23 ??? 1 C2
= 12C 2p22
2 (1 2 C ) 2
1
According to the standard formulas of Cholesky’s decom- 12
position, G(v) is computed as C2
= (1 2 C 2p22 1 C 2p22 2 C 2)1/2 = Ï1 2 C 2 (36)
Si1(v)
G11(v) = ÏS11(v); Gi1(v) = ; It can be seen from (35) and (36) that for n = p, all new values
G11(v)
of G(v) (row p) agree with (18a) and (18b).
S O D
k21 1/2
So, for any arbitrary order of matrix S0(v), the Cholesky’s
Gkk (v) = Skk (v) 2 G 2kr (v) ; decomposition agrees with (18a) and (18b).
r=1
O ACKNOWLEDGMENTS
k=1
Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
motion using stochastic waves.’’ J. Engrg. Mech., ASCE, 115(12), Shinozuka, M. (1977). Analyses on random vibration. Seismic Press, Bei-
2723–2737. jing.
Kovacs, I., Svensson, H. S., and Jordet, E. (1992). ‘‘Analytical aerody- Shinozuka, M., and Jan, C. M. (1972). ‘‘Digital simulation of random
namic investigation of cable-stayed Helgeland Bridge.’’ J. Struct. processes and its applications.’’ J. Sound and Vibration, 25(10), 111–
Engrg., ASCE, 118(1), 147–168. 128.
Li, Y., and Kareem, A. (1993). ‘‘Simulation of multivariate random pro- Shinozuka, M., Kamata, M., and Yun, C. (1989). ‘‘Simulation of earth-
cesses: hybrid DFT and digital filtering approach.’’ J. Engrg. Mech., quake ground motion as multi-variate stochastic process.’’ Technique
ASCE, 119(5), 1078–1098. Rep. 1989.5, Dept. of Civ. Engrg. and Operations Res., Princeton Uni-
Liu, C. H. (1994). ‘‘Simulation of Divers correlative stochastic processes versity, Princeton, N.J.
and its applications.’’ J. Tongji Univ., Shanghai, 22 (May), 61–67. Simiu, E., and Scanlan, R. H. (1986). Wind effects on structures. Wiley,
Mann, J. (1998). ‘‘Wind field simulation.’’ Probabilistic Engrg. Mech.,
New York.
13, 269–282.
Spanos, P. D., and Zeldin, B. A. (1998). ‘‘Monte Carlo treatment of ran-
Niemann, H. J., and Hoffer, R. (1995). ‘‘Nonlinear effects in the buffeting
dom fields: a broad perspective.’’ Appl. Mech. Rev., 51(3), 219–237.
program.’’ State of the Art in Wind Engrg., 9th Int. Conf. on Wind
Engrg., Wiley Eastern, New Delhi, India, 320–341. Yamazaki, F., and Shinozuka, M. (1988). ‘‘Digital generation of non-
Ramadan, O., and Novak, M. (1993). ‘‘Simulation of spatially incoherent Gaussian stochastic fields.’’ J. Engrg. Mech., ASCE, 114(7), 1183–
random ground motions.’’ J. Engrg. Mech., ASCE, 119(5), 997–1016. 1197.
Shinozuka, M. (1974). ‘‘Digital simulation of random processes in en- Yang, J. (1972). ‘‘Simulation of random envelope processes.’’ J. Sound
gineering mechanics with the aid of FFT technique.’’ Stochastic prob- and Vibration, 25(1), 73–85.
lems in mechanics, S. T. Ariaratnam and H. H. E. Leipholz, eds., Uni- Yang, J. (1973). ‘‘On the normality and accuracy of simulated random
versity of Waterloo Press, Waterloo, Canada, 277–286. processes.’’ J. Sound and Vibration, 26(3), 417–428.
View publication stats Downloaded 13 Nov 2010 to 129.10.107.106. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org