Professional Documents
Culture Documents
1982 Identification of Independent Storms
1982 Identification of Independent Storms
1982 Identification of Independent Storms
net/publication/223127221
CITATIONS READS
172 337
2 authors, including:
Pedro J. Restrepo
Consultant
58 PUBLICATIONS 861 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 Pedro J. Restrepo on 13 November 2017.
[1)
ABSTRACT
Restrepo-Posada, P.J. and Eagleson, P.S., 1982. Identification of independent rainstorms.
J. Hydrol., 55: 303-319.
An easily-applied approximate criterion is proposed for the separation of point-
precipitation records into statistically-independent storms. The criterion is tested success-
fully in parameter estimation for Poisson models of annual rainfall frequency in an arid
climate. Analysis of rainfall records from a wide range of climates produces an empirical
relation for estimating the minimum time between independent rainstorms, given only
the mean annual rainfall and the mean length of the rainy season.
INTRODUCTION
DON002571
304
BACKGROUND
and
tx =(storm interarrival time) = t, + t, (2)
in general, a ramstorm is a realization of a non-homogeneous multivariate
random generating process. The heterogeneity arises primarily from the dif-
ferent cooling mechanisms responsible for producing storms at a given lo-
cation. These mechanisms produce storm types with different characteristic
distributions of the constituent random variables: storm depth, storm dur-
ation, and time since the last storm. When the storm types are well-segregated
seasonally (e.g., convective in the summer andcyclonic inthe winter) model-
lers of point-rainfall sequences may account for the heterogeneity by using
distributions of the constituent random variables that have seasonally-varying
parameters (Todorovic and Yevjevich, 1969). When the cooling mechanisms
are well-mixed temporally, the modeller has a choice of adding the outputs
of homogeneous generating processes for each storm type, or of defining a
single generating process having a lower order of homogeneity in .which
storm type (i.e. cooling mechanism) is one of the constituent random vari-
ables.
In what follows, we assume that the homogeneity issue has been dealt
with, and we address only the identification of independent events from
homogeneous populations.
It seems intuitively clear that statistical independence of samples of a
particular constituent variable will increase with the time between the events
from which those samples are drawn. There should, theref4re, be a minimum
DON002572
305
L
© Idealized InUepenaent Events
I' er^
^ tOI
separation time between samples that is necessary in order to gain any pre-
scribed degree of independence. Unfortunately, however, this is a multi-
variate process and we have no assurance that the minimum time between
samples of storm depth (for example) will bring the same degree of indepen-
dence between the associated values of storm duration, or that the arrivals
of the events from which the samples of depth were drawn are themselves
independent to at least this degree.
To avoid separate investigations of independence for each constituent ran-
dom variable, it would be useful to work only with the single variable for
which the actual generating process has the longest "memory", that is, for
DON002573
306
NOTATION
List of symbols used
CV[coefficient of variation
E[' ] expected value
fx(x) probability density function (p.d.f. of x)
F'g(x) cumulative distribution function (c.d.f. of x)
h storm depth
i storm intensity
k value of probability density of rainy season length
mp, average seasonal precipitation
mtr average storm duration
m" seasonal average number of storms
mr average length of rainy season
r square root of coefficient of determination
t„ storm interarrival time
time between storms
minimum tim between independent storms
tb^_a minimum time between storms
lr storm duration
inverse of the mean time between storms
a shape parameter in the gamma distribution
a[ • ] standard deviation of [' I
w inverse of the mean storm interarrival time = mean storm arrival rate
r length of rainy season
rc minimum lengtb of rainy season
p[' ] Pearson's incomplete gamma function of [•]
y[ •] incomplete gamma function of [' ]
DON002574
307
In this application, the arrivals have finite duration, t„ which raises the prob-
lem of overlapping events. With:
mtr = (mean storm duration) and w t = (mean time between
storm arrivals)
we can expect the Poisson process to adequately describe the arrivals of
independent rainstorms, provided:
wmir G<1 (4)
Since the origin of the exponentially-distributed arrival time is arbitrary
for the true ( i.e. instantaneous event) Poisson process (
Cox and Lewis, 1966,
p. 22), we may expect the time between storms, tE„ in our application to be
exponentially distributed with the same goodness-of-fit as ta. Once again, the
only interference with precision-of-fit is the magnitude of the product, wmtr.
We thus expect:
DON002575
308
these events to vanish and the probability of rainfall in the next Atb to de-
cline toward the small, but constant, value associated with a successful
set of naturally-occurring rainless periods will thus
Bernoulli trial. The full
not follow the exponential distribution due to the inclusion therein of small
tb's which separate dependent events. To demonstrate this, we will examine
the distribution of the complete set of observed tb's for the six primary (i.e.
long, hourly record) stations of Table I. These distributions are presented in
Fig. 2 where they are compared with the exponential as fitted by setting the
mean value of the exponential distribution equal to the sample mean. The
poor fit is as expected.
To obtain exponentiality and, therefore, an independent set of events, for
these series, we will first successively redefine the storms, as shown in Fig. lc,
incorporating within the appropriate new event all rainless periods tb < tbmin -
We continue our assumption that wm1t << 1.
Values of this parameter are
tabulated in Table I. When the rainless period separating two rainy periods
has a duration less than this value, tbm^n, called the minimum time between
storms the two rainy periods belong to the same storm. Otherwise, they
form parts of different storms. Tests at successively larger values of tbm+n
should ultimately bring sufficient exponentiality to the distribution, jTb (tb).
The value of tbm,n for which this occurs is called tbo and will establish the
set of independent storms.
DON002576
309
^ µ^ .+ M t` N W O u^ N N H^ c0 0^0 0 0
^ M NODh0u9c01TtD
NN^C1Q>C^ ^LOm d^Om Nd^O
.-^.+NN^d'+-^ Mm d' d'M m NtNN^
0000^+N00tiNOttl'OOT V'N
d^m ^^1NONN MeMMN C-OCi rit^
N M cO ri tD N n 6^ ^J 1' fO O Vi Oi .i W O
^y .yNC^d^ d'mN^tl'mT 0^.9
^i
.i t^ u)00OOQ+000 oOOOOO
m t^d^ u'JOiO T u,1CiQ^<p.O^ N^^^
<N+]ti000iNt cmOe^O<OG^^MC+Ml^OiCO ri
r N ti
O
N
^t1 MON^0^6^N^N^0^f1NNN0
_T _T_T _TTTT
4=°
s°.4°.°
A^ ^
^ c00
zzzcSoo3
^ g
H . ^ o_
F I ga• 0
^y¢Gtrx^"Ax4awA ^sca^c^^
«
0 ..^cvM^ru><atiaorno-,cva.>v^^mm^
z ti ti ti ti.y ti ti ti
DON002577
310
rescaled exponential, we are implicitly assuming that the population of inde-
pendent events actually contains values in the range 0 < tp < tb even though
chosen criterion for independen ce does not permit their identification.
our
To identify the independent fraction of these events separated by tb < tb ,
we would need a criterion which examined individual events.
4-
^ Ir .,^ 1 1+
+a
^
^ }I {I^ II
DON002578
311
assuming
Fig. 2. Fitting of observed time between storms to an exponential distaibution,
all events independent.
DON002579
312
For small ten y", we have seen in Fig. 2(where tb mln = 0) that the obser-
vations follow a distribution having a standard deviation much larger than
the mean. For small tbtherefore, the observations give CV[t„ )> 1.
For increasing tbmin, the observed distribution of tb is truncated and re-
scaled. This will nierease the ,nean of tb while maintaining all central mo-
ments (including a) almost the same (as the distribution Of tb is approxi-
mately exponential in this range). The coefficient of variation Of t(, will
thus be driven toward zero as tbmi. increases.
We will choose the tbmi„ at which the observed CV [tb ] passes through the
value for the exponential distribution (eq. 8) as the criterion, tf,o , for mini-
mum separation of independent events.
It is demonstrated in the Appendix how the CV(tb) can be calculated for
different values of tbmfn in a single processing of the original data. These cal-
culations lead to the values presented in Fig. 3 for three of the stations of
Table I. The interpolated intersections of these CV(tb) with the line,
CV(th )= 1, yield the values Of tb, shown in the figure and tabulated for all
seventeen stations in Table I.
DON002580
313
(9)
= e-'"v { 1+^ (m^)°Plvtc, m xz] f
Prob IPA
mPA ^- z
l l „=i v 1
where
pA = annual precipitation
m,.A = mean annual precipitation
= mean number of storms per rainy season
= shape factor of gamma distribution of storm depth
T = length of rainy season
To = minimum length of rainy season
mz = mean length of rainy season
and by definition:
(12)
(average storm arrival rate) = m„/mT
and
(13)
wJk = 2(m. - wTo]
Analysis of five years of daily rainfall at Al Ta'if, Saudi Arabia, shows the
season length to be relatively constant with mT = 235 days. Successive com-
putations of CV (t„ ] in the manner described above leads to tb = 5.5 days,
as shown in Fig. 3. Also available at this station are nine years of annual rain-
fall and four years of 10-min. intensity-iiuration data. The former data are
plotted in Fig. 4 using the Thomas (1948) plotting position and they form a
base against which to test the Poisson model of eq. 9. The latter data are
used to give a first estimate of the Poisson parameters, ic = 0.43 and m„ =
34. Since the intensity-duration data are not continuous, the effective mini-
mum time between storms is unknown but from the data it is estimated to
be - 1 hr. or 0.04 days. Eq. 9 is plotted in Fig. 4, using these parameters and
is labeled tbmin ^ 1 hr.
The continuous daily data from Al Ta'if are analyzed successively for
tbmin = 2, 4 and 6 days. The corresponding values of ic and m„ are estimated
and eq. 9 is plotted for each case in Fig. 4A. If the annual series is actually
DON002581
314
Poisson, we expect a c.d.f. for tb(n = t6 = 5.5 days to best represent the
population distribution. Certainly, the Poisson agreement with the few plot-
ted observations improves with increasing t,,min , at least in the better-defined
high-frequency (low-recurrence interval) end of the distribution.
The same test is conducted in Fig. 4B and C for Dhahran and Riyadh,
Saudi Arabia - locations at which eq. 10 adequately describes the season
F ^tbmin ^ m. ^ K mr I
(days) (d0Y5)
0.04 34 0.43 235
2 {276 0.65 235
4 10.44 0.53 235 I \
6 9.11 0.65 235 4
1.25 2 5 10 20 50 100
1.01
RECURRENCE INTERVAL ( 9eors)
UHAHRAN
SAUOI ARABIA
2
o OOservations
-Poisson Model
tb ' mv K ^ m T
m
d0YS1 Itlays) (Gays)
L - 37 0.40 115 ^ 44
.. ^.__.y...
01 1,25 2 5 10
RECURRENCE INTERVAL ( yearsl
DON002582
315
006servafione
-Poiseon Mudel
m, mi
(bmin
(day0 tdays)
01 1 2^ 2 5 10 20 50 100
RECURRENCE INTERVAL ( years)
Fig. 4. Frequency of annual precipitation at: Al Ta'if, Saudi Arabia; Dhahran, Saudi
Arabia; and Riyadh,8audi Arabia.
length distribution. Once again, the daily data are analyzed repetitively both
to obtain tbe , as shown in Fig. 3, and to obtain the Poisson parameters of
eq. 11: (c, m,,, Ta and mT. The agreement between tbe and the best-fit tbml„
is remarkable in each case.
In Fig. 4C, the sensitivity to season length distribution is shown by also
presenting eq. 9, using the parameters for tbmla = 2 days.
It must be remembered that the purpose of isolating independent storms
has been the convenience which independence brings in solving a class of pre-
cipitation modeling problems. For other problems concerning the dynamic
response of hydrologic systems to precipitation inputs, the "interior" of the
individual events must be modeled. In such cases, the independent events
may not be the most appropriate since they can involve, particularly in arid
climates, long internal rainless periods. In particular, the idealization of
storm rainfall as a random sequence of rectangular intensity-pulses that is so
convenient in studying the statistics of event hydrology (Eagleson, 1972,
1978a), can become physically meaningless. That is, the equivalent uniform
storm intensity, i = h jtr, will have little physical significance as tbe gets
large and all shorter rainless periods become internal to the independent
events. For dynamic purposes, the distributions of the constituent storm
variables are more appropriately and conveniently determined from the
raw storm data without concern for independence.
DON002583
316
^ -1
m,
I ( mm/month)
DON002584
317
The finding that tba is larger for arid climates than for humid ones results
from the clustering of aridelimate rainfall and the consequent dependence
of multiple storms upon the same large-scale weather system.
Exponentiality of the time between storms, tb, has been used as a suf-
ficient condition for the statistical independence of storm arrivals within the
restriction that overlapping is negligible (i.e. wmtr <41). A sample coefficient
of variation of unity, CV [tb ]= 1 was accepted as a sufficiently unambiguous
criterion for independence of arrivals on the basis of observations of expon-
tial-like distributions of tb from numerous series of raw storms. Causality
arguments have been used to justify using arrival independence also as an
approximate condition for independence with respect to the separate storm
variables such as depth and duration.
A convenient, simple algorithm is presented for obtaining CV[tb] from a
given storm sequence for successively larger minimum storm separation. Esti-
mates of tb, so obtained are shown to produce Poisson distributions of
annual precipitation which agree well with observed distributions at three
arid-climate sites. Values of t,,a from a wide range of climates are shown to
correlate with the rainy-season monthly precipitation thus providing a means
for estimating tb. without record analysis.
ACKNOWLEDGEMENT
Support for this work was provided by the National Science Foundation
under NSF Grant No. ENG 76-11236.
APPENDIX
Let br, i = 1, ..., No, be the duration of the non-rainy period i, and No =
total number of non-rainy periods. Then, for tbmin = 0:
where
1 N^ 1 N (A-2), (A-3)
= N Y_ b3 and Si.o = No ^l bi
S2,0
o i=1
DON002585
318
Calculating a histogram for b; with a width of class equal unity, such that:
and
1)}]
f{S ^-(Si^ )Z}{N,l(Ni v2 (A-4)
CV(1) = Si
where
and
(A-7)
N, = No'hi
In general,
CV(K) [{S2 K_"' (St rc )2 }{^ l(NK - 1)}l'^2 (A-8)
S i,K
where
- hK.K2 } (A-9)
S2, K NK-1 S2.x-I
1 (A-10)
hx Kl
S,x - NK[Nx tSi.xi
(A-1 1)
NA = Nx-1 ' hx
Finally,
tbm.n K, if CV(K) = 1
or by interpolation, when CV(K - 1) > 1 and CV(K) < 1:
tno = (K-1)4,[(CVx-i-1)/(CVx_i-CVrc)I (A-12)
REFERENCES
DON002586
319
Eagleson, P.S., 1978b. Climate, soil and vegetation, 7. A derived distribution of annual
water yield. Water Resour, Res., 14(5): 765-776.
Eagleson, P.S., 1981. Some limiting forms of the Poisson distribution of annual station
precipitation. Water Resour. Res. (in press).
Gnedenko, B.V., Belyayev, Y.K. and Solovyev, A.D., 1969. Mathematical Methods of
Reliability Theory. Academic Press, New York, N.Y.
Grace, R.A. and Eagleson, P.S., 1967. A model for generating synthetic sequences of
short-time-interval rainfall depths. Proc. Int. Hydrol. Symp., Fort Collins, Colo., pp.
26"76.
Grayman, W.M. and Eagleson, P.S., 1971. Evaluation of radar and raingage systems for
flood forecasting. Ralph M. Parsons Lab,, Dep. Civ. Eng., Mass. Inst. Technol., Cam-
bridge, Mass., Tech. Rep. No. 138.
Sariahmed, A. and Kisiel, C.C., 1968. Synthesis of sequences of summer thunderstorm
volumes for the Atterbury watershed in the Tucson area. Proc. Int. Assoc. Hydrol. Sci.
Symp, on Use of Analog and Digital Computers in Hydrology, 2: 439-447.
Thomas, H.A., 1948. Frequency of minor floods. J. Boston Soc. Civ. Eng., 35: 425-442.
Todorovi¢, P. and Yevjevich, V., 1969. Stochastic processes of precipitation. Colo. State
Univ., Fort Collins, Colo., Hydrol. Pap. No. 35, pp. 1-61.
DON002587