Escolar Documentos
Profissional Documentos
Cultura Documentos
715
716
2.
C. Onof et al.
(a)
RAINFALL DATA
The data examined are a subset of data recorded on daily charts and then
digitized by the GLC and the UK Meteorological Office, using a Precision
Encoder and Pattern Recognition (PEPR) system and are specified in Table 1.
The data are comprehensively analysed in Moore et al. (1993). More recent
digitized data from three raingauges were also available (Table 1), and a record
of 39 years of Heathrow hourly rainfall from 1949 to 1987 has been used to
derive DDF statistics. The gauges used are shown on Fig. 1.
The analysis shows that rainfall depths of return periods up to 50 years
are within 2 mm of those found in the FSR. There is a greater discrepancy for
100 year depths, but this appears to be a consequence of the way in which the
EV-II distribution was fitted to the data. Therefore the extreme-value statistics
obtained from the FSR were used in the following analyses.
111
718
C. Onof et al.
Period
Quality
Format
PEPR:
Kew Observatory
good
raw digitized
Brent reservoir
Clapton pond
hourly
Hampstead
good
hourly
Recent:
Kew sewerage treatment works
good
tip times
good
tip times
Brent reservoir
good
15-min totals
WATTORD
T Brent_
s
SLOUGH
Y
THa
LON^M|^=^
y^
Heathrow .
STORES
Hampstead)
X
x, .
V
T Clapton i,
KeV V=J
Cheam )
WOKING)
"^~s
10 km
Overview
The library model selects from historical data those profiles which are
realizations of storms with a depth and duration of a given return period. The
aims in developing the method, addressing some of the problems with the
current FRQSIM procedures, were:
to use observed storm profiles, not averaged or idealized ones;
to reproduce the frequency of occurrence of different profile shapes;
to use FSR statistics for depth-duration-frequency; and
to avoid scaling of profile ordinates where possible, because profile
shape was found to depend on total depth.
The name of the method arose from the large "library" of storm profiles which
was compiled from sub-hourly rainfall data from seven gauges in London, UK.
The records run from 1928 to 1994, with a few breaks, providing 59 years of
data in total.
The library
This is a collection of profiles at half-hourly discretization which correspond
to the most intense parts, over several durations, of every major storm in the
record. Profiles may include dry periods, as long as their total amount does not
exceed one sixth of the duration.
The library is divided into 79 duration classes (every half-hour from 1
to 40 hours), each of which is subdivided into six classes based on the total
rainfall over the duration. There are 3922 profiles in total.
Description
The storm simulation method has the following steps:
(1) Decide on a storm return period, T;
(2) Select a duration randomly from a uniform distribution over a range of
durations suitable for the catchment;
(3) Find the FSR rainfall corresponding to the return period and duration;
(4) Select a profile in a uniformly random fashion from the library class
corresponding to the depth and duration (or from the nearest class);
(5) Scale the profile slightly if necessary, so that depth matches the FSR
figures exactly;
(6) Repeat steps 2 to 5 several times to generate several realisations of a
T-year storm; and
(7) Apply the storms to a FRQSIM catchment model, using appropriate
catchment parameters for estimating the T-year flood.
719
720
C. Onofet al.
E[Y[hn = ^
variance:
var[y;(/!>] =
(i)
[hr, -1 + e ^ ]
(2)
= Jll[ehr>+e-hl>-2]e^h
(3)
i? 3
721
The way the model is used provides a way of generating storms having
the required duration. By simulating a large number of such storms, the
extrema of a given return period can be identified.
The parameters for the model are fitted by using moments of the rainfall
depth within storms for the extreme storms so as to reproduce features of the
storm profile and by using moments of the total storm depth so as to reproduce
the statistics of the FSR.
For such a calibration to be possible, analytical expressions must be
derived for the moments of the total storm depth as a function of the
parameters and the total storm duration. If the instantaneous cell depth is X
with a probability distribution function (pdf):
f(x) = exp[-x//u x ]
and the cell duration is Y with pdf:
g(y) = jexpt-^y]
the total storm depth is therefore:
Z = lXpYp
(4)
where Xp and Yp are the depth and duration of the pth cell in a storm of JV cells
and are distributed as Zand F respectively. The properties
of the distribution of Z are best obtained using the characteristic function. The
details of the calculation are given in Appendix 2.
The moments used in the calibration are the mean and the variance of the
storm depth:
E[D] = Ll (fiL + 1)
V
r
var[>] =
-I
Mx
(5)
(4/3L + 3)
(6)
exp{z(T)/a}-b
w h e r e z(T)
__ _ l n ( _ l n ( 1 _ h
Therefore, by using n different return periods Th for which the FSR gives the
depth as xt, parameters a, a and b can be estimated by minimizing the sum of
722
C. Onof et al.
the squares of the deviations between x(7}) and xh i.e. Y [xl; -x(7})]2. In this
way, the best possible EV-II fit to the FSR statistics is found. Setting the partial
derivatives to 0 yields equations which can be solved by using the Powell
hybrid method. The moments of the EV-II distribution can then be obtained,
for example as in equation 1.2.4.12 in NERC (1975).
12
18
24
[)]
varUJ
covK, Yi+1]
2.90
17.4
2.21
2.14
12.8
2.62
1.58
9.10
2.37
1.26
7.23
2.12
0.93
5.11
1.62
0.74
3.73
1.24
4
4.90
4.28
0.158
23.6
78
6
4.89
4.04
0.148
25.3
86
9
5.01
3.6
0.155
28.0
103
12
5.03
3.46
0.138
29.5
116
18
5.11
3.23
0.095
32.9
120
24
5.24
2.92
0.137
34.7
139
(h"11)
0
n Or )
4
2.66
9.76
19.8
6
1.61
6.40
15.2
9
1.12
5.11
12.9
12
18
0.840
4.53
12.1
0.642
4.16
10.9
24
0.546
3.96
9.74
Parameters were found for certain durations L and for the other
durations, regression relationships were estimated for each parameter:
723
jS = 8.075ZTU-S/4
ij = 16.67 -2.359L + 0.144L 2 - 2.8 x 10"3L3
0 381
Mx = 31.41L- -
(7)
with values of the percentage of explained variance, R2 equal to 0.80, 0.88 and
0.89 respectively. All three parameters showed an overall decrease as L
increased. Equations (7) indicate that longer storms are simulated by longer and
less intense cells which arrive less frequently. With these parameters, storms
of all durations L can be simulated and are considered to be the annual
maxima. The parameters X and 7 were not required.
Table 5 shows the storm depths for different return periods and for a
number of simulated storms (sample size) of 1000 or 2000. In this Table, the
T-year ranking was found by ranking the total depths of the annual maximum
storms and using Weibull plotting positions.
1000 storms
2000 storms
20
100
57
93
57
89
44
61
Instead of the combination used above, equations (3), (5) and (6) were
next used, so that the variance of the rainfall depths within storms from
equation (2) was replaced by the variance of the total storm depths
(equation (6)). With this calibration method therefore:
(a)
DDF curves given in the FSR were used so that the model reproduced
the mean and variance of the total storm depth; and
(b) equation (3) was used so that the correlation structure of the rainfall
process within the storm was reproduced. The parameters obtained are
given in Table 6.
The regression relationships for all durations L were:
13 = 22.5ZT0-875
7] = 1741-1.43
\xx = 128Z.-1.33
(9)
12
18
_ _
__
29.3
24.4
12.6
10.8
7.20
6.68
3.90
3.86
2.50
2.30
2.50
2.52
r, (h' )
^(mmh-1)
24
724
C. Onof et al.
Table 7 shows the storm depths for different return periods for 6-hour
and 12-hour storms and a sample size of 1000 storms.
Table 7 Storm statistics for the second calibration method
Return period
(years)
5
20
50
100
Depth (mm)
6-h storms
12-h storms
6-h storms
12-h storms
34
47
54
62
40
54
64
73
31
44
52
61
36
49
59
69
(a)
725
(b)
60-
30-
reduced variate y
Fig. 2 Frequency of (a) simulated 6-hour rainfall and (b) simulated 12-hour
rainfall.
regular profiles, as in Fig. 3(a) (early peak), which the stochastic model seldom
reproduced.
Estimation of design floods
The Nutfield Brook catchment The Nutfield Brook is a small tributary
of a river (Redhill Brook) whichflowsnorth, joining the Thames at the western
edge of London. The catchment has an area of 7.6 km2. It is predominantly
rural, although there is some impermeable area in a village and on a motorway
which crosses the catchment. Further details are given in a report on an
investigation (NRA, 1993) into flooding problems in the catchment. The study
involved constructing a FRQSIM model, calibrating it with rather limited data
and estimating design flows for several points in the catchment. For the current
study, design flows were estimated at the Mid Street Bridge in the village of
South Nutfield. A constant proportional loss model was used, with a catchment
wetness index of 120 mm.
Application of the Poisson model Sets of 1000 simulated annual maximum storms were applied to the catchment model, and the resulting sets of
flow peaks were ranked to enable the selection of flows with various return
periods, neglecting for the time being the effect of antecedent conditions. The
influence of storm duration on design flow was expected to be relatively small:
it was found that storms lasting between 10 and 15 hours gave slightly higher
design flows than 5 to 10 hour storms. Figure 4 is a plot of one realization of
the annual maximum floods produced by storms which were generated by the
Poisson model when fitted using method 2. It is clear from that Figure that
fitting a Gumbel distribution to this series, as is usually done with FRQSIM,
would not have adequately represented the results.
726
C. Onof et al.
llllllll.-.l
(b)
earlypeak
LAL
1
^m
double peak
-HLIl-l-l lllllllllll
Fig. 3 (a) Historic 6-hour profiles from the library and (b) simulated 6-hour
profiles.
Application of the library model Four sets of 250 storm profiles were
used to run the model, each set consisting of different storm profile shapes and
durations (between 4 and 36 hours), corresponding to storms with return
periods 5, 20, 50 and 100 years. Each set of storms produced a distribution of
flow peaks. Examples are displayed in Figs 5(a) and 5(b). From these Figures,
it is clear that a storm of a given return period can give rise to a wide range
of floods depending only on the storm event rainfall profile. A way round this
indeterminate relationship is to examine the peak intensities of storms, for
727
I 6-
0-
reduced variate y
4
5
6
peak How /cumecs
Fig. 5 Distribution of floods (a) from 250 5-year storms and (b) from 250
100-year storms.
728
C, Onof et al.
example the 100-year storms which gave rise to the largest floods in the distribution had peak intensities close to the FSR 100-year rainfall over half an hour
for the catchment. This indicates that it would make sense to choose the 100year flood from the top end of the distribution and it is the maximum which is
chosen for comparison below.
Comparison of the results Design floods estimated using the two rainfall models are compared in Table 8 with those from the original FRQSIM
method (NRA, 1993). The two models presented in this paper gave remarkably
similar estimates, differing by a maximum of 9%, for the 20-year flood. The
original method overestimated floods up to 50 years, but gave a lower 100-year
flood; this could be due to a split regression technique used to fit Gumbel
distributions to flows simulated by FRQSIM, i.e. several Gumbel lines were
fitted piecewise to the data points.
CONCLUSION
This paper has presented two models for the temporal distribution of storm
event rainfall, the first a library model which improves upon existing methods
for the generation of design rainfall for flood simulation, and the second a
stochastic model based upon a Poisson process for rainfall cell structure, which
generates random storms which are transformed into floods when input into
FRQSIM.
It has been demonstrated that, at least for the catchment considered, peak
discharge is dependent on storm interior distribution.
These results of applying methods which include variability of storm
interior distribution are encouraging and research is currently being undertaken
to extend them to other catchment areas.
The two rainfall models were seen to produce very similar flood estimates. The Poisson model was found to generate design rainfalls which both
satisfied the FSR DDF statistics and presented a diversity of profiles similar to
those found in the library of historical storms. The Poisson model, however,
presents several advantages:
it does not require that so much data be available, since one statistic
from the actual storm profiles, namely the lag-one autocorrelation of
729
730
C.Onofetal.
APPENDIX 1
Analytical expression of the moments of aggregated precipitation depths
within a storm
The total intensity Y(t) of rainfall is the sum of all the active contributions at
time t:
Y(t) =
lXt_u(u)dN(t-u)
H=0
where Xu(v) is the rainfall intensity at a time v after the beginning of a pulse
which arrived at time u.
The first- and second-order moments of Y(t) are given in RodriguezIturbe et al. (1987):
E[Y(t)] = nx., var[F(f)] = 2/ix2. and c^r) = 2y.x2^T,
V
V
V
autocorrelation function of lagr of the process Y(t).
These moments give:
(Al)
(A2)
-h
Thus, for k = 0:
h
var[Y,.(/!)] = 2 f (h-v)c^v)v
(A3)
fXx
^[e^
kr fl
+ e-^-2]e~
for k > 1
(A4)
V
and:
4W 2jS
var[Y[ ] = -L-[hv
h)
iff
- 1 +e~f"!]
(A5)
731
APPENDIX 2
Analytical expression of the total storm depth for a Bartlett-Lewis
Rectangular Pulse Model
The charateristic generating function of the random variable Z is given by:
(j>^t) = EJ?xp{izt}]
and from (4):
N
= E-XY EN[exp{itZXY}}
Downloaded by [212.12.187.60] at 06:36 08 August 2016
P'l
ZEXJ[Ilexp{iXpYpt}]P{N=n}
(A6)
, _ (ffLf-iexpl-ffL}
where L is the duration of activity of the storm, i.e. the time during which new
cells can be born. This is because the cells arrive according to a Poisson
process within the storm and that there is one cell associated with the beginning
of the storm.
Now calculate:
EXY[exp{iXpYpt}]
= EXY
g axpYpty
y=o
J1-
ZEx[X/}ET[Y/\
M J1
where Ex[XJp] and Ey[Yp] are given by the pdf of X and Y:
1
1
1/Mr
4>x(t) = i e x p { 0 - f - _ )x}dx =
_L_
V ^x
It..
nx
V/i-it
and since:
W) = E(ityEixq
y=o
j\
(A7)
732
C. Onofet al.
then:
f-
EVU] =
(A8)
(1W
Similarly:
0y(O = f i?exp{(if ij)x}<k
if it
and since:
then:
Wl =
I
rf
^[exp^y}] =
l^f-^-
j-o j \ (Vfy.xy
and by substitution into (A6):
4>M = S
n=\
(j3L)n~lexp{-PL}
(-!)!
J=o(rj/ay
fi-i
00
4>Sf) = exp-|8( ^ ^ - 1 )
Will id
U\
(A10)
j=0(r,/nx)>
(<t>M)lf){0) nr
^
= Ll(pL + l)
i
[Z 2 ] =
(All)
ij
(d^O/d^KO)
i
* V ,o2i-2
= (Id)
(j3/Z/ + 6|3L + 4)
=*var[Z] = ( l i ) (4/3L + 3)
(A12)
E[Z3] =
733
;3
l x^tl^T^^,r
i i R2T2_
= (Liy(p
L
+15 $lLl + 30 0L +6)
so that:
E[(Z-Emn
^3^ _
(Px,
= (li)(6/3L-4)
6*2-4
=> 5 =
(A13)
()/4BL + 3 )
Remarks
1.
2.
Note that the skewness will not necessarily be equal to 1.14 which is the
skewness of an EV-I Gumbel distribution.
Equations (All) and (A 12) used in the calibration are a function of the
duration of activity L. This is not known for real storms, but since the
average cell duration is small, i.e. 1/T/. The actual observed storm
duration, i.e. the period during which the storm is alive, can be approximated by the duration of activity L. Numerical simulations have shown
that this introduces a maximum error of the order of 5 % for the total
storm depth.