ARMA Algorithms for Ocean Wave

P-T. D. Spanos Modeling

The P.D. Henderson Teaching Fellow
Three different algorithms are presented for simulating a time series which is
in Engineering,
Associate Professor of Engineering
compatible with a given power spectrum of ocean waves. The first algorithm
Mechanics, generates the current value of the time series as the sum of a linear combination of
University of Texas, its past values and a white noise deviate. The second algorithm produces the values
Austin, Tex. 78712 of the time series as a linear combination of white noise deviates. The third
Mem.ASME algorithm is a combination of the first and second algorithms. These algorithms are
applied to the Pierson-Moskowitz (P-M) spectrum, exclusively. The third
algorithm is associated with simple analog filter approximations of the P-M
spectrum. The advantages and disadvantages of each of the three algorithms are
discussed in context with their applicability to offshore engineerngproblems.

Introduction
The problem of spectral analysis of ocean wave-induced as the output of an autoregressive (AR) digital filter to white
motions of floating or submerged, and flexible or rigid noise input.
structural systems has received considerable attention over a The AR algorithms are best suited for treating all-pole
period of recent years. One aspect of this problem deals with spectra. However, it can be proved, under quite general
the development of proper analytical models which describe conditions, that theoretically any pole-zero spectrum can be
adequately the statistical power versus frequency distribution approximated, as closely as is desired, by using a digital AR
of confused ocean states. For this purpose field and filter of an appropriately large order. Such an approximation
laboratory data are used to derive the so-called wave power was attempted in reference [6] in connection with the Pierson-
spectrum or spectral density. Another significant aspect of the Moskowitz (P-M) spectrum of wind-generated ocean waves.
same problem is the digital generation of a time series which is This spectrum is perhaps the most widely used in offshore
compatible with a target (specified) power spectrum of ocean engineering studies. Rather unexpectedly, it was found that
waves. This need has arisen primarily through nonlinear the spectrum of the time series generated by the (AR)
Monte Carlo dynamic studies of structural systems operating algorithm exhibited considerable narrowband fluctuations
in the ocean environment. Traditional methods of addressing with respect to the target P-M spectrum.
this problem can be found in references, such as [1-4]. These The present study aims to demonstrate the appropriateness
methods are founded on the representation of a stationary of autoregressive moving-average (ARMA) algorithms for
random process as an infinite sum of harmonic components digital synthesis of ocean wave records; for numerical
with random amplitudes or phases [5], This mathematical computations the P-M spectrum is used, exclusively.
form is versatile in the sense that it can account, if necessary, ARMA algorithms generate the current value of the simulated
for statistical spatial dependence and cross-correlation of time series as the sum of a linear combination of its previous
ocean waves [2, 4], However, due to spectral peakedness values and a linear combination of white noise deviates. AR
considerations, a large number of harmonic components must algorithms are considered herein as a special class of ARMA
be employed. Thus, these methods can become com- algorithms. In this regard this study elaborates on the findings
putationally costly, and they cannot be recommended of reference [6]. Another special class of ARMA algorithms is
unreservedly. the class of moving-average (MA) algorithms. These
An alternative technique of simulating ocean wave records algorithms synthesize the current value of the time series as a
has been proposed in reference [6]. Based on this technique linear combination of white noise deviates. Their applicability
the time series which represents the ocean wave is considered to ocean spectral analyses is examined, as well.
It is assumed that a two-sided target spectrum S(w) = S( —
o)) is given in the interval 0 < I co I < ub, where wb is designated
NOTE: Appropriate SI Conversion Units—1 ft = 0.3048 m, 1 knot = 0.515
m/s as the cut-off frequency. It is sought to approximate S(a>) by
the power spectrum P(co) of the output to white noise input of
Contributed by the Offshore Mechanics Committee of the Petroleum a digital filter having a pulse transfer function H(z), in terms
Division and presented at the Second International Symposium on Offshore of z-transform notation, and sampling time step
Mechanics and Arctic Engineering at the Energy Sources Technology Con-
ference and Exhibition, Houston, Texas, January 30-February 3, 1983, of THE T=r/u>„ (1)
AMERICAN SOCIETY OF MECHANICAL ENGINEERS. Manuscript received by the
Petroleum Division, September 13, 1982; revised manuscript received April 28, Three different algorithms associated with digital filters will
1983. be considered.

300/Vol. 105, SEPTEMBER 1983 Transactions of the ASME

Copyright © 1983 by ASME

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 Auto-Regressive Algorithms (AR) F u r t h e r m o r e , the following relationship is valid:
This class of numerical schemes has been examined in
reference [6], In general, the pulse transfer of a p - o r d e r A R ^ == ~ - HakRi
Rx l^uk^i\-k\;'x>p (11)
Ar=l
filter is given by the equation [7-10]
E q u a t i o n (9) can be viewed as predicting the current value if,,
G by convolving the coefficients ak of the filter with the
HAR(z)= (2) previous values of the process U(t). Similarly, equation (10)
can be thought of as providing an extrapolated (predicted)
value R^, \>p, in terms of the values R\, X < p , of the
*=1
autocorrelation function R. Furthermore, examining
T h e constants ak are determined so that equation (8) it can be proved that the criterion expressed by
equation (3) is the frequency domain equivalent of the least
S(u»)
•i? * ^AR(«)
du> = m i n i m u m (3) square, time d o m a i n , criterion
<(U„- U„)2 > = m i n i m u m (12)
r
where In equation (12) the symbol < • > denotes the operation of
PAR(o>) = HAK(el"T)HXR(ei«T),i = ^ \ (4) mathematical expectation, and t/„ is the stationary process
associated with the target spectrum P(w).
In equation (4) the symbol ( )* denotes complex conjugate. In connection with the P-M spectrum the preceding ap-
T h e scaling constant G is determined by requiring that proach was originally used in reference [6]. T h e algebraic
7T
expression of this spectrum, single sided herein, is
S(oj)do) P(0))d(D (5)
S(co)= — e x p ( - — ) (13)
co V or /
It can be proved that the minimization criterion introduced by where
equation (3) leads to the following Toeplitz matrix equation: A = 8 . 2 x 1 0 - 3 g 2 ; fi = 0.74(g/y) 4 (14)

Rn R, .R -\ R, Alternatively, the P-M spectrum can be expressed in the

P
dimensionless form
R, Rn R\ «2 R-, 5
* *, S(w) _a
S(o>) = ——r =e
Ri S(coJ ( " )

exp
R p-i
Ri
•R„ Rn [-{(£) 4];< (15)

(6) T h e symbols g and v represent the gravitational ac-

In equation (6) the symbol celeration, and the velocity of the wind generating the ocean
waves, respectively, while co0 = (2w)(4B/5),M is the peak
frequency of t h e spectrum [11-12]. In reference [6] the spectra
Rx S(w)cos(\uT)do)-,\ = 0,\,. (7)
\ of A R generated processes approximated reliably the target
spectrum in an average sense. However, they exhibited quite
denotes the value of the autocorrelation function of the target sharp fluctuations. Interestingly, these fluctuations were not
spectrum calculated at a time lag T which is equal to XT. U p o n eliminated by increasing the order of the A R filter. O n the
solving equation (6) and determining the constants ax . . . , contrary, by selecting larger values for p, the Toeplitz matrix
ap, the scaling constant G can be found by using the equation which appears in equation (6) became gradually ill-
conditioned [13]. In those cases either the constants ax, . . . ,

°2 = h lR°+ £fl***] (8) ap could not be determined or their computed values yielded
through equation (8) imaginary values for the constant G. It
E q u a t i o n (8) is equivalent to equation (5). W h e n the was speculated that these fluctuations were associated with
parameters ax, . . . , ak and G are determined, a time series the peakedness and the mathematical peculiarity of the P-M
{/„, can be generated digitally by utilizing the following spectrum. Specifically, examining equation (13) it becomes
recursive algorithm: clear that the P - M spectrum possesses a zero of infinite order
at co = 0. T h u s , experiencing numerical difficulties in ap-
U„ = U(nT)=Gw„- ~£ > * { > „ _ (9) proximating this spectrum by A R models which are best
suited for treating pole-dominated spectra should n o t be
In equation (9) the symbol w„ signifies a deviate of clipped unexpected. Consequently, further consideration has been
white noise with power spectrum equal to one over the in- given to the source of the A R spectral fluctuations and the
terval 0 < I co I < 7 7 T T . Denote by R the autocorrelation func- computational difficulties encountered in solving the Toeplitz
tion of an infinite collection of samples of the recursively equation in context with this spectrum. It has been found that
produced time series. T h e n , it can be proved that the values of an eight-term Taylor expansion of the exponential term of
R and R coincide at the p + 1 points of the time lag axis T, equation (13) leads to excellent approximations of the P - M
T=0,T=T,. . . ,r=pT. T h a t is, spectrum for the range of wind velocities commonly used in
/?X=/JX;X=O,I, (10) engineering practice. T h a t is,

A co2
(16)
P(<*)<
32 28 ,6 B $
12
B6 , B1 B*
co +Bco +- i+
TT" 2 0 + - 4! -CO 16 + — ~ C 0 co8 + ^ r ^ c o 4 + -
1 2
+ •
2! 5! 6! 7!

Journal of Energy Resources Technology SEPTEMBER 1983, Vol. 105/301

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Clearly, equation (16) reduces the order of the zero of the
P-M spectrum at o> = 0, from infinity to 27. It is understood PIERS0N-M0SK0WITZ SPECTRUM
that the quality of this approximation improves for large to. TAYLOR APPROXIMATION ORDER • 8
This is due to the fact that the exponent of the P-M spectrum

1.0
/ WINU VLLULI1 T 3 DO HHOI9
decreases with increasing frequency. The Taylor ap- v M u> b »27r/5,rad/sec
proximation reduced the level of the AR spectral fluctuation « / V-WIND VELOCITY' 20knots
but it did not eliminate it. In an effort to further lessen this o
\CuV«7r, rad/sSC
fluctuation, a thorough examination of the behavior of the Era,
556
autocorrelation function of the P-M spectrum was un- CO
dertaken. This was deemed logical since the AR modeling is
•£
equivalent to pointwise matching of the autocorrelation r> ID
functions of the target and the AR spectra, as is manifested by £6'
equation (10). This examination has demonstrated that o
UJ
knowledge of the autocorrelation function is important in Q.
CO ^
selecting intelligently the sampling time step T, and the order O
p of the filter. Based on these studies the following guidelines 1\ -THEORETICAL
\ EXPRESSION
have been developed for implementing AR fitting to ocean \ -AUTOREGRESSIVE MODEL
C\J
wave spectra.
o

J
! ORDERED FILTER = 100
I Calculate the autocorrelation function which is
associated with specified spectrum. Examining equation
(7) it becomes obvious that wb should be assigned the 0 0.5 1.0 1.5 2 0 2.5 3.0
minimum value beyond which the calculated value of R does OJ, rad/sec
not change appreciably. In this manner the upper bound -K/ab Fig. 1 Normalized P-M spectrum: wind velocity: v = 20,50 knots: AR
of the sampling time Twill be kept maximum. algorithm
II Determine the magnitude rmax of the time lag T beyond
which the value of the autocorrelation function is judged
negligible. Ut
III Select the sampling time step T compatible with P1ERS0N-M0SK0WITZ SPECTRUM
equation (1), and with integer multiples coinciding with as TAYLOR APPROXIMATION ORDER'8
many as possible abscissas which correspond to local peaks of ORDER OF FILTER • I00
the autocorrelation function.
IV Estimate the minimum order p m i n of the AR filter by
using the integer part of the number r m a x /T. Recognize that
the order p of the filter can not be selected unreservedly large. -WIND VELOCITY • 20knots
Otherwise, extremely small values of the autocorrelation TfME STEP' I sec.
function may be introduced in the Toeplitz matrix of equation WIND VELOCITY • SOknots
(6) and numerical difficulties can be encountered in solving TIME S T E P ' 2 . 5 sec
this equation. A measure of the ill conditioning of the
Toeplitz matrix can be provided by the ratio
Mmax
(17)

where /xmax and ^ min are its maximum and minimum eigen-
values, respectively. An a priori estimate of r can be obtained
by using the approximation

/ • « / • ' - ^ l«ls«6, (18)

•*min

where P m a x and P m i n are the maximum and the minimum

spectral values used, respectively. Thus, taking into con-
sideration the peakedness of the P-M spectrum, in fact of all
of the wave spectra, caution must be exercised to select a
frequency approximation interval which does not yield an
unduly small value for Pmia. 40 60
T,98C
Following the preceding guidelines the numerical data Fig. 2 Normalized autocorrelation function: P-M spectrum, wind
plotted in Figs. 1-6 have been obtained. It is seen that the velocity: v = 20,50 knots: AR algorithm
agreement between the target and the AR spectra is quite
good. Observe, as well, the matching between the target and
the AR autocorrelation functions at all of their peaks. In equation (19) the symbols b„ and q denote a constant and
an integer number, respectively. It is known, that any weakly
Moving Average Algorithms (MA) stationary process can be regarded as the output of an infinite
order filter which has pulse transfer function expressible by
Consider the class of algorithms of spectral analysis which equation (19), and white noise input. Thus, a time series can
corresponds to nonrecursive digital filters with pulse transfer be synthesized by using the recursive numerical scheme
functions of the form [14]
U„ = Un(nT)= J2 b„wn_p,q-~o° (20)
H*, dz)= £ b,z- (19)
which possesses a power spectrum P(oi) which is identical to

302/Vol. 105, SEPTEMBER 1983 Transactions of the ASME

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 N.
PIERSON-MOSKOWITZ SPECTRUM PIERSON-MOSKOWITZ SPECTRUM
TAYLOR APPROXIMATION ORDER • 8 TAYLOR APPROXIMATION ORDER "8

WIND VELOCITY »60 knots WIND VELOCITY-70 knots

t o » 7 r / 3 , rad/sec w.»27r/7,rad/88c
b
WIND VELOCITY- 30knots WIND VELOCITY'40 knots
0Ju»2TT/3,r<Hl/9«C UJL» 7T/2,rad/89C

— THEORETICAL EXPRESSION
~ AUTOREGRESSIVE MODEL
THEORETICAL EXPRESSION
.AUTOREGRESSIVE MODEL ORDER OF FILTER = 100

ORDER OF FILTER = 100

1.5 2.0 2.5 3.0 5 2.0 2.5 3.0

w, rod/sec OJ. rad/sec.
Fig. 3 Normalized P-M spectrum: wind velocity: v = 30,60 knots: AR
algorithm Fig. 5 Normalized P-M spectrum: wind velocity: v = 40,70 knots: AR
algorithm

PIERSON-MOSKOWITZ SPECTRUM
TAYLOR APPROXIMATION
ORDER - 8
ORDER OF FILTER - 100

WIND VELOCITY - 4 0 knots

TtME STEP -2sso.
-WIND VELOCITY - 7 0 knots
.TIME STEP- 3.5sac.

— THEORETICAL EXPRESSION
x AUTOREGRESSIVE MODEL

100 20 4 0 - 60 80 100
T ,s#c.
Fig. 4 Normalized autocorrelation function: P-M spectrum, wind Fig. 6 Normalized autocorrelation function: P-M spectrum, wind
velocity: v = 40,70 knots: AR algorithm
velocity: v = 30,60 knots: AR algorithm
7T

T
EMA = — f lV5 , (cj)-^ MA (e- /a,;r ) 11do> = minimum (21)
the P-M spectrum or any other ocean wave spectrum. 2ir J "
Equation (20) can be thought of as simulating a time series by
the weighed average of 2q + 1 white noise deviates which move That is, bv are the coefficients of the Fourier expansion of
in time. Hence, the algorithm represented by this equation is VS(o) in the interval, 0< I col <K/T; they are calculated by
commonly qualified by the name moving average (MA). In making use of the formula
practical applications a finite value of q is chosen. Upon
7r
selecting the value of q the constants bv can be determined by
ensuring that the succeeding optimization criterion is satisfied bv = 1 T *JS(ui)cos(vTw)d<j). (22)
T Jo

Journal of Energy Resources Technology SEPTEMBER 1983, Vol. 1 0 5 / 3 0 3

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 PERSON - MOSKOWITZ SPECTRUM
ORDER OF F I L T E R - 2 9
q.
-
X
o WIND VELOCITY • SOknots
E
CO WIND VELOCITY » 2 0 knots
-o
2 — THEORETICAL
K EXPRESSION
t-co • MOVING-AVERAGE
O .
UJO MODEL

0 0.5 1.0 15 3.0 0 05 1.0 15

GJ, rad/sec b j , rod/sec
Fig. 7 Normalized P-M spectrum: wind velocity: v 20,50 knots: MA Fig. 9 Normalized P-M spectrum: wind velocity: v = 40,70 knots: MA
algorithm algorithm

algorithm requires a smaller number of terms than the AR

— algorithm for the digital simulation.
PIERSON-MOSKOWITZ SPECTRUM
ORDER OF FILTER - 2 9 Autoregressive Moving-Average Algorithms (ARMA)
q.
WND VEL0CITY It is pointed out that AR algorithms are best suited for
11 / V" * SO knots treating all-pole spectra, and MA algorithms are best suited
SPECTRUM , S/S_

l / V - W I N D VELOCITY* 30 knots for treating all-zero spectra. However, there exist spectra
0.8

which involve both poles and zeros. This is true for the ap-
Y \ —THEORETICAL proximation of the P-M spectrum which is provided by
> \ EXPRESSION equation (16). In these situations it is appropriate to consider
0.6

\ • MOVING-AVERAGE what is commonly called an autoregressive moving-average

\ MODEL (ARMA) algorithm. An ARMA algorithm can be associated
with a digital filter which has the following pulse transfer
function
0.4
0.2

f=0
^ARMA(Z) — (25)

J —A—i —=»*f i"*~——«=—

1+I>^ f
0

0.5 1.0 1.5 2.0

The symbols c { and d( denote constants. Evidently AR and
U), rad/sec
MA algorithms can be produced as special cases of ARMA
Fig. 8 Normalized P-M spectrum: wind velocity: v = 30,60 knots: MA algorithms by setting in equation (22) c { = 0 for £ > 1, d^O,
algorithm
and Cj^O, d^=Q, respectively. For a particular spectral
fitting problem the constants c f and d^ must be determined so
that a certain optimization criterion is satisfied. For instance,
It can be shown that the minimum error is the criterion posed by equation (3) can be adopted. That is, c f
and rfj are determined so that
iMA.min Zj 6„I2<<X. (23) 7T
M><7 S(co)
rfco = min (26)
The magnitude of £ M A min can serve as an indicator of the J IT
IT D
D
U«)
quality of the approximation
S(co) = ^ M A ( e - r ) ^ A ( e ' w r ) . (24) in which
Based upon the preceding theoretical developments, filters -PARMA(u)=-WARMA( e "" )-^ARMA ( e ' " ) (27)
have been constructed for the P-M spectrum. The values of Equation (26) leads to the following equations:
the wind velocity have been taken identical to those which
have been employed in the AR modeling. Pertinent numerical O£ARMA
results are depicted in Figs. 7 through 9. It is observed that the dct
-2Dc/^(£-')=0,l</</ (28)
agreement between the target and the MA spectra is excellent.
Furthermore, attention is called to the fact that the MA and

304/Vol. 105, SEPTEMBER 1983 Transactions of the ASME

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 dE*
= 2 ^ t f / £ ( £ - / ) = 0, !<;/<H,efo = l (29) (37)
dd f 1=0 f=l
where In this case, the innovation of Ut requires the storage of
max(m,n) values at each time step.
1+I> Are"
Application of A R M A Fitting to the P-M Spectrum
cos(£-/)cortfco
Least-Squares Spectral Approximation. Evidently
c*e" equation (16) can be rewritten in the form of equation (32).
Thus, it is logical to seek to fit an ARMA spectrum to the P-M
(30) spectrum. In this regard it has been considered worthwhile to
and explore the existence of approximations which are simpler
than equation (16) and reasonably accurate. Specifically, a
least-squares approximation S(co) of the P-M spectrum has
7T been investigated. That is, it has been sought to determine
r S(co)
S(co)sothat
^(?-/)=J T - cos(£- /) uTdu (31)
c^e" £[S(u,)-S(co,)] 2 = minimum (38)

It is recognized that equations (28) and (29) are nonlinear in The symbol M denotes the number of frequency points at
ck and dk. Thus, for the determination of the constants c { and which the difference between .P(co) and S(co} is taken into
df. it is necessary to solve m + n nonlinear equations. This is a account. The succeeding algebraic forms of S(oi) have been
major shortcoming of ARMA algorithms constructed by considered:
using the criterion expressed by equation (26); pertinent
discussions can be found in references like [10]. However, the SD(co) (39)
2
need of solving nonlinear algebraic equations for developing (CO -k0y + (C C0)
o
2

ARMA algorithms can be circumvented in some special cases.

Assume that the target spectrum can be written, exactly or Gjco 4
S 0 («) = 2 (40)
with acceptable accuracy, in the form [(co -- £ i ) 2 + ( c u) 2 ] 2
A (co) and
S(co) = (32)
ri(co) G2co4
S2(co) = (41)
in which A and II are polynomials of co. Then, an MA filter [(co2 - k 2 f + (c2co)2][(co2 - k2f + (c2co)2]
with pulse transfer function
The scaling constants G0, Gu G2, the "springs" k0, klt k2,
k2, and the "dashpots" c0, c,, c 2 , c2 are to be determined so
(33) that the criterion specified by equation (38) is satisfied. It is
f=l clear that the spectra S0, S{ and S2 exhibit the following three
can be constructed for A (co). Furthermore, an AR filter with characteristics of the P-M spectrum. First, they exhibit
pulse transfer function peakedness for values of the "springs" and the "dashpots"
yielding "ratios of critical damping" much smaller than
unity. Second, they tend to zero, with zero slope, as co—0.
1
Third, they tend to zero, with zero slope, for co-—oo. Fur-
HT (z) = (34)
thermore, by including the terms co4 in the numerator of the
spectra Sx and S 2 another characteristic of the P-M spectrum
is maintained approximately. Specifically, examining
equation (16) it is seen that the difference of orders of the
can be constructed for the spectrum l/II(co). Clearly, this
polynomials A (co) and II(co) is equal to 5. In fact, this is true
methodology can be compatible with the criteria defined by
independently of the order of the Taylor expansion which is
equation (3) and equation (21) for the AR and MA spectra,
used to approximate the exponential term. The value of this
respectively. However, in general it does not ensure that the
difference which corresponds to the spectra Sl and S2 is equal
criterion defined by equation (26) is satisfied.
to 4.
Independently of the technique which has been utilized to
Using the criterion expressed by equation (38) with M =
determine the coefficients c5 and dj., sample records of the
200-250, a code based on the Levenberg-Marquardt
time series which is campatible with the spectrum P(co) can be
algorithm [15, 16] has-been used to determine the parameters
generated by employing the difference equation
of the spectra S0, S, and S2. Initial estimates of these
parameters have been obtained by selecting the "springs,"
(35) "dashpots," and scaling constants so that the "resonance
?=o f=i frequencies," the "half-power bands," and the peak values of
the model spectra are close to the corresponding values of the
Clearly, the innovation of U(t) necessitates the storage of P-M spectrum. This selection has proved quite suitable in
n + m values at each step. Alternatively, the MA and AR terms of the rate of convergence of the Levenberg-Marquardt
filters can be considered in cascade and the following algorithm. Pertinent numerical results for various wind
equations can be used: velocities, are shown in Table 1 and Figs. 10-15. Note that the
last row of the table involves values which correspond to the
dimensionless P-M spectrum given by equation (15). Thus,
U, = U(lT)=w,- ^d(U,. (36) these coefficients are uniformly valid for all values of the
f=i
wind velocity v.
and It is observed that the agreement between the target

Journal of Energy Resources Technology SEPTEMBER 1983, Vol. 105/305

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 1 Parameters of analog filter approximations of the Pierson-
Moskowitz spectrum; the last row pertains to approximations of the
dimensionless spectrum 5(to), equation (15)
Wind
speed
U
knots 'l c\ 'l C2 C2
20 0.639 0.422 0.759 1.032 0.740 0.767 0.848 0.338 0.586 0.946 1.500
25 1.235 0.336 0.485 1.292 0.592 0.491 1.068 0.270 0.375 0.760 0.961
30 2.126 0.278 0.337 1.551 0.494 0.341 1.282 0.226 0.261 0.634 0.668
35 3.368 0.238 0.248 1.809 0.424 0.251 1.497 0.194 0.191 0.544 0.490
40 5.022 0.208 0.190 2.065 0.370 0.192 1.711 0.170 0.147 0.476 0.375
45 7.144 0.186 0.150 2.324 0.330 0.152 1.924 0.150 0.116 0.422 0.297
50 9.959 0.168 0.121 2.579 0.296 0.123 2.126 0.136 0.094 0.380 0.240
55 13.198 0.152 0.100 2.839 0.270 0.101 2.346 0.122 0.078 0.346 0.199
60 17.082 0.140 0.084 3.098 0.246 0.085 2.563 0.112 0.065 0.316 0.167
65 21.669 0.128 0.072 3.356 0.228 0.073 2.778 0.104 0.056 0.292 0.142
70 27.020 0.120 0.062 3.614 0.212 0.063 2.994 0.096 0.048 0.272 0.123
75 33.366 0.112 0.054 3.873 0.099 0.055 3.207 0.090 0.042 0.254 0.107
eq.(15) 0.265 0.504 1.085 0.606 0.894 1.096 0.527 0.410 0.841 1.158 2.149

PIERSON-MOSKOWITZ SPECTRUM
- T H E O R E T I C A L EXPRESSION
— ' ANALOG FILTER APPROXIMATION
TRANSFER FUNCTION

WIND VELOCITY
I 7 0 knot!
II ' 6 0 knot!
III 'SOknoti

1.0 1.5
w ,rod/ H C
Fig. 10 Normalized P-M spectrum: wind velocity: v = 20,30,40 knots: Fig. 11 Normalized P-M spectrum: wind velocity: v 50,60,70 knots:
ARMA algorithm, filter H 0 ARMA algorithm, filter H0

spectrum and S 2 is quite good. Furthermore, the per- U2 + (c2 + c2)U2 + (k2 +k2+ c2c2)U2
formances of S0 and S, also reasonably satisfactory. These
results are remarkable considering the low orders of the + (c2k2 + c2k2)U2 + k2k2 U2 = VG^W (44)
analog filters, and the fact that the P-M spectrum represents In equations (42) through (44), a dot over a variable denotes
only an empirical deduction about what happens in nature. its derivative with respect to time; the symbol w represents
The simplicity of S0, S~i and S2 implies significant advantages clipped white noise with spectral density equal to unity.
in terms of simulation studies of structures exposed to ocean Clearly, ARMA algorithms for simulating U0, Ut, and U2
waves. digitally can be determined by using finite differences, for-
ward, central or backward, to approximate the derivatives
Simulation Based on Analog Filters. Equations (39) appearing in equations (42) through (44). Appropriate
through (41) can be thought of as the transmittancy functions formulas for this purpose can be found in references such as
[17] of analog filters. Therefore, records of time series which [18, 19]. It is understood that the time integration step f,
are compatible with these power spectra can be generated by which is required for a specified level of accuracy, can be
integrating numerically the following differential equations: made larger with increasing order of the selected finite dif-
U0+c0U0+k0U0=JG~0w (42) ference scheme.
2 2 Alternatively to the preceding procedure, ARMA
U{ + 2cj £/, + (c] +2ki)Ul +2clk[Ul + ki U1 =^G[w (43) algorithms associated with equation (42) through (44) can be
and derived by using the transfer functions of the analog filters

306/Vol. 105, SEPTEMBER 1983 Transactions of the ASME

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 PIERSON-MOSKOWITZ SPECTRUM
-•THEORETICAL EXPRESSION
-•ANALOG FILTER APPROXIMATION

ITJcjTTk,!2
WIND VELOCITY
4-0 knot!
II : 30 knot!
III • 20 knoll

1.0 1.9 1.0 1.5

u , rod/««c til , rod«/t«c
Fig. 12 Normalized P-M spectrum: wind velocity: v = 20,30,40 knots: Fig. 14 Normalized P-M spectrum: wind velocity: v = 20,30,40 knots:
ARMA algorithm, filter H 1 ARMA algorithm, filter H 2

0.50
050
u , rods/we
cu , rod/tec
Fig. 13 Normalized P-M spectrum: wind velocity: v = 50,60,70 knots:
ARMA algorithm, filter H 1 Fig. 15 Normalized P-M spectrum: wind velocity: v = 50,60,70 knots:
ARMA algorithm, filter H 2

associated with the spectra S0, §i and S2. Specifically, in VG,S2

terms of Laplace transform notation, these functions can be Hi (s) = s4+2c1.s3+(c2+2A:1),s2+2clA:1s + £2
written as
VG> 2
VG„ VG: (46)
H0(s)-- (45) (s-Pi)Hs-pt)2
2
s +c0s + k0 (s^Po)(s-pZ) and

Journal of Energy Resources Technology SEPTEMBER 1983, Vol. 105/307

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 ANALOG, EQ. (41)

DIGITAL, EQ. (51), T =T

DIGITAL, EQ. (51),T = T/2

DIGITAL, EQ. (51), T = T/4

i
»3

5 2.0 2.5

U) , DIMENSIONLESS DIMENSIONLESS
F
Fig. 16 nini<oi III.=r „ v i n , „ , i „ „ , _i c i \i6 i \- n i , k . .,„„*<<>.. '9- 1 7 Digital filter approximations of S 2 (u)/^2("o); P u l s e transfer
Digital filterbyapproximations of S2(»)lS2(a0); pulse transfer function determined by using equation (51)
..."ti-.HBtarminnrihvMainnBo.miinn/sn
function determined using equation (50)

H2(s) motion of a structure which is exposed to loads induced by

ocean waves possessing the P-M spectrum.

s4 + (c2 + c2)s3 + (k2 + k2+c2c2)s2 + (c2k2 + c2k2)s + k2kz Concluding Remarks

VG 2 5 2 Three algorithms have been discussed for simulating time

(47) series which are compatible with a target power spectrum of
(s~p2)(.s-pt)(s--p2)(s-pt) ocean waves. As far as numerical results are concerned, the P-
The symbols p, p signify the poles of S0, Si and S2. Then, M spectrum has been used exclusively.
approximation of the filters given by equations (45) through The first algorithm, AR, represents the current value of the
(47) can be constructed by using digital filters having pulse time series as a linear combination of its past values and of a
transfer functions determined either by relying on the white noise deviate. This algorithm is associated with a
following equations: recursive digital filter. It ensures that the autocorrelation
functions of the target and the simulated process match at a
. fVG^l-z-')
(48) number of equally spaced points along the time-lag axis. This
H0 (z) =
(l-e-"° z-1)(l-e-',°:rz-1)
7
number is equal to the number of constants of the digital
filter. This property can be the basis of a technique of
f2VGl(l-z-')(l-z-1) deducing analytical models for power spectra of existing wave
Hx (z) = (49) data [6, 20, 21]. This technique possesses a major advantage
(l-e-^fz-l)2(.\-e-»*Tz-x)2 over the traditional methods of spectral analysis. Specifically,
and it provides through equation (11) extrapolated values for the
H2(z) auto-correlation function for time lag greater than the
duration of the available record; the traditional methods
assume that these values are equal to zero. Another interesting
r 2 VG 2 (i-z- 1 )d-z- 1 ) feature of AR algorithms is that they are realizable physically
(1 - e - ^ - i ) ( i -e-n rz-i)(i - e - * 2 ^ - i ) ( 1 -e-nrz-i) by means of delay elements. Therefore, they can be used for
(50) developing devices generating random waves in laboratories.
However, when using an AR algorithm either for simulation
or by using the bilinear transformation or for spectral estimation of ocean waves their spectral
2 l-z-[ peakedness must be kept in perspective. In this manner the
(51) sampling time step and the order of the filter can be selected
T i+r judiciously to avoid ill conditioning of the Toeplitz matrix
in equations (45) through (47). Obviously these procedures which is used in the determination of the filter parameters.
lead to expressions for the transfer functions which can be put
in the form of equation (25). Then, simulations can be made The second algorithm, MA, is associated with a
by using either equation (35) or equation (36) and equation nonrecursive digital filter and generates the values of the time
(37). series as a linear combination of white noise deviates. These
It must be noted that for the approximations provided by deviates are weighted by suitable constants. Several criteria
equations (48) through (51) to be reliable, it is necessary to can be adopted for determining these constants. The par-
oversample, that is to select ticular one which has been used in the present study furnishes
constants which are associated with the Fourier coefficients of
T<<-w/wb (52) the square root of the target spectrum. The numerical
Pertinent numerical results are shown in Fig. (16) and Fig. calculation of the weighting constants which correspond to
(17). It is observed that for securing reliable spectral ap- the P-M spectrum is explicit. Furthermore, it is not sensitive
proximations of S2, the algorithm which is constructed by to the magnitude of the sampling time step, as the AR
using the bilinear transformation requires values of T which algorithms are. However, MA algorithms can not be used
are considerably larger than those required by the algorithm readily to fit ocean wave data prior to the determination of
based on equation (50). However, the latter algorithm is of a their power spectra by any of the standard signal processing
smaller order. In any case, it is noted that the magnitude of T techniques.
which corresponds to reliable spectral approximations shown The third algorithm, ARMA, produces the current value of
in Fig. (16) and Fig. (17) is comparable with the value of the the time series as a linear combination of its past values and
time step which should be used in integrating the equation of values of white noise deviates. Both the values of the process

308/Vol. 105, SEPTEMBER 1983 Transactions of the ASME

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 and the white noise deviates are weighted by appropriate Journal of Waterways and Harbors Division, American Society of Civil
constants. An approach to determining the weight constants is Engineers, Nov. 1969, pp. 557-583.
2 Shinozuka, M., Fang, S. L. S., and Nishitani, A., "Time Domain
to satisfy the criterion which is used for the AR algorithms. Structural Response Simulation in a Short Crested Sea," ASME JOURNAL OF
However, in this case a set of nonlinear algebraic equations ENERGY RESOURCES TECHNOLOGY, Vol. 101,1979, pp. 270-275.
must be solved. Alternatively, a least-squares approximation 3 Hudspeth, R. T., and Chen, M - C , "Digital Simulation of Nonlinear
can be used to model the target spectrum as the output of an Random Waves," Journal of the Waterway, Port, Coastal and Ocean Division,
American Society of Civil Engineers, Vol. WW1, 1979, pp. 67-85.
analog filter to white noise input. Three filters of this nature 4 Shinozuka, M., and Wai, P., "Digital Simulation of Short-Crested Sea
have been developed in connection with the P-M spectrum; Surface Elevation," Journal of Ship Research, Vol. 23, 1979, pp. 133-141.
the accuracy of the approximation is satisfactory. In fact, one 5 Rice, S. O., "Mathematical Analysis of Random Noise," Selected Papers
of the developed spectral approximations reproduces the P-M in Noise and Stochastic Process, ed. N. Wax, Dover, New York, 1954.
spectrum quite reliably. The analog filter approach to 6 Spanos, P-T. D., and Hansen, J. E., "Linear Prediction Theory for
Digital Simulation of Sea Waves," ASME JOURNAL OF ENERGY RESOURCES
modeling the P-M spectrum leads, through finite difference TECHNOLGOY, Vol. 103, 1981, pp. 243-249.
formulas for the derivatives, or through "^-domain to z- 7 Oppenheim, A. V., and Schafer, R. W., Digital Signal Processing,
domain" techniques to quite efficient numerical schemes of Prentice-Hall, Englewood Cliffs, N.J., 1975.
digital simulation. 8 Tretter, S. A., Introduction to Discrete Time Signal Processing, Wiley,
New York, 1976, pp. 102-116.
Considering the reliability, the simulation convenience, and 9 Papoulis, A., Signal Analysis, McGraw-Hill, New York, 1977.
the analytical simplicity of the analog filters approximations 10 Childers, D. G., Editor, Modern Spectrum Analysis, IEEE Press, New
of the P-M spectrum, the necessity of using its exact ex- York, 1978.
pression in future dynamic studies of offshore structures is 11 Pierson, W. J., and Moskowitz, L., " A Proposed Spectral Form for
Fully Developed Wind Seas Based on the Similarity Theory of S. A.
strongly questioned; especially when its empirical nature is Kitargorodskii," Journal of Geophysical Research, Vol. 69, No. 24, 1964, pp.
kept in perspective. 5181-5190.
It is believed that analog filter modeling of the P-M 12 Sarpakaya, T., and Isaacson, M., Mechanics of Wave Forces on
Structures, Van Nostrand Reinhold, New York, 1981.
spectrum, in fact of any other wave spectrum, can have a 13 Isaacson, E., and Keller, H. B., Analysis of Numerical Methods, Wiley,
significant impact on analytical studies of dynamic behavior New York, 1966.
of offshore structures. Furthermore, it can lead to the 14 Koopmans, L. H., The Spectral Analysis of Time Series, Academic Press,
development of quite efficient algorithms for simulating New York, 1974.
ocean wave samples in connection with Monte Carlo studies 15 Levenberg, K., " A Method for the Solution of Certain Non-linear
Problems in Least Squares," Quarterly of Applied Mathematics, Vol. 2, 1944,
of random wave induced vibrations. Finally, it could con- pp.164-168.
ceivably be useful in dealing with directional spectra [22] and 16 Marquardt, D. W., "An Algorithm for Least-Squares Estimation of
two dimensional arrays of wave data. Nonlinear Parameters," Journal of the Society of Industrial and Applied
Mathematics, Vol. 11, No. 2, 1963, pp. 431-441.
17 Lin, Y. K., Probabilistic Theory of Structural Dynamics, Krieger,
Acknowledgments Huntington, N.Y., 1976.
18 Dahlquist, G., and Bjorck, A., Numerical Methods, Prentice-Hall,
Appreciation is expressed to V. K. Agarwal, especially, and Englewood Cliffs, N.J., 1974.
T. A. Strouboulis for their assistance in the numerical 19 James, M. L., Smith, G. M., and Wolford, J. C , Applied Numerical
computations. The financial support of this work by the Methods for Digital Computation with FORTRAN and CSMP, Harper and
Row, New York, 1977.
Grant NSF-CME 7918570 from the National Science 20 Holm, S.,andHovem, J. M., "Estimation of Scalar Ocean Wave Spectra
Foundation is gratefully acknowledged. by the Maximum Entropy Method," IEEE Journal of Oceanic Engineering,
Vol. OE-4, July 1979, pp. 76-83.
21 Houmb, O. G., and Overvik, T., "Some Applications of Maximum
References Entropy Spectral Estimation to Ocean Waves and Linear Systems Response in
Waves," Applied Ocean Research, Vol. 3, No. 4, 1981, pp. 154-162.
1 Borgman, L. E., "Ocean Wave Simulation for Engineering Design," 22 Kinsman, B., Wind Waves, Prentice-Hall, Englewood Cliffs, N.J., 1965.

Journal of Energy Resources Technology SEPTEMBER 1983, Vol. 105/309

Downloaded From: http://energyresources.asmedigitalcollection.asme.org/ on 01/28/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use