INTRODUCTION
Principle
The model shown is a stochastic model for generating chronological rainfall series on
an hourly time step. This modelling is based on a statistical study of descriptive
variables used to create rainfall time series. There are different types of models to
generate hourly rainfalls:
- daily rainfall disaggregation models (Woolhiser & Osborn, 1985; Econopouly et al.,
1990; Koutsoyiannis, 1994) which transform daily rainfall into low step rainfalls;
- models based on a cluster process which result from LeCam's works (1961). These
models are usually based on a combination of two random processes: occurrence
and position of rainy cells, and cells depth and duration processes (Waymire &
Gupta, 1981; Rodriguez-Iturbe et al., 1987; Acreman, 1990; Cowpertwait, 1991;
Onofetai, 1995);
- other stochastic models (Tourasse, 1981; Lebel, 1984), including the model
designed by Cemagref (Cernesson, 1993; Arnaud, 1997), which are based on a
geometric description of hyetographs by random variables;
- scaling models based on the dimensionless description of hyetographs (Huff,
1967) which are capable of generating storms (Garcia-Gazman & Aranda-Olivier,
1993; Koutsoyiannis & Foufoula-Georgiou, 1993).
Here, the model studied is based on the theory that rainfall can be linked to a
random and intermittent process whose evolution is described by stochastic laws. It is
also based on the hypothesis of independence between variables describing
hyetographs and on the hypothesis of the stationary nature of the phenomenon studied.
Generating a rainfall time series involves two steps:
- Descriptive study of the phenomenon: choosing the variables which define the
phenomenon. By hypothesis, each descriptive variable should be independent of
the others and will be defined by a theoretical law of probability fitted to the
observations.
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 435
The analysis focuses on hyetographs of rainfall episodes selected from daily criteria
(all daily rainfalls of the episode greater than 4 mm and presence of one daily rainfall
of at least 20 mm). From these selected daily episodes, hourly hyetographs were
generated based on nine descriptive variables (Cernesson et al, 1996): average number
of episodes per year (NE); time of onset of storm (TSE); number of rainy periods
(group of storms without dry period) within an episode (NG); number of storms
(elementary hyetograph which have only one maximum) in a rainy period (A/A) and
duration of each storm (DA); average intensity (HMA); dry period (DIA) between
successive storms (zero for grouped storms); relative maximum position (RPX); and
ratio between maximum and average intensity (RX). [It should be noted that a
difference is made between average intensities and durations of two types of storms:
the main storm which provides the largest amount of water (maximum product HMA x
DA) during an episode (denoted by HMAP and DAP) and other storms which are
considered to be ordinary storms (denoted by HMAO and DAO).]
A study of 50 rain record stations made it possible to define the theoretical laws of
probability which best characterized the different descriptive values: Poisson's law (for
the variables NE, DAP and DAO), geometric law (for the variables NG, NA, TSE and
DIA), normal law truncated between 0 and 1 (RPX variable) and the exponential law
(variables RX, HMAP and HMAO). Fitting these laws of probability required using 19
parameters for each season.
Particular features
An additional parameter was defined for each storm type. This parameter is used to
model the dependence observed between mean intensity rainfall values and rainfall
duration values (Arnaud, 1997; Arnaud et al., in press). The generated model has 21
parameters defined for the two seasons.
Hourly hyetographs were created storm after storm from general descriptive
variables chosen through random selection within their cumulative distribution law
following a precise order. It should be recalled that the hypothesis of independence
of the different descriptive variables was made, which allowed to choose the
variables independently of each other as needed. However, this was not true for mean
rainfall intensity and duration variables as they have a mutually dependent
probability.
The persistence of rainstorms within a single rainy episode generates high
cumulative rainfall and was also modelled. Basically, it allows to group together
ordinary storms around one main storm. This point will be developed and explained in
later, while analysing model sensitivity.
436 Patrick Arnaud & Jacques Lavabre
To study the model's asymptotic behaviour (modelling extreme episodes), the mean
cumulative frequency distribution of simulated maximum precipitation over a long
time period (1000 years) was compared to observed cumulative frequency distribution
fitted by an EV type I or Gumbel law. Frequencies were then expressed as return
periods (7) to define 5-year ( 7 = 5 years), 10-year (7 = 10 years), 100-year
(7= 100 years) or 1000-year (7 = 1000 years) quantiles of maximum precipitation
over different durations. [Rainfall simulation over a given period (here 1000 years),
causes sampling problems due to the extreme values (like for the observations). To
solve this problem, a hundred simulations over a chosen period were performed to
obtain a sample of 100 cumulative frequency distributions. Mean distribution was
calculated for these 100 distributions to represent simulated distribution. Nevertheless,
this distribution was still subject to the uncertainty of the model's parameters for these
observation samples.]
The graph in Fig. 1 shows maximum precipitation distributions over 1 and 24 h at
the Néoulous station based on return period. The distribution of observed values
(Observations), simulated values (Simulation) and the fitting of a Gumbel law
(Fitting), for values which return period was greater than 1 year, are shown.
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 437
250
• Observations
Simulation 1000 • Observations
Fitting Simulation /
200 i
If 900
Fitting /
F
800i
500-
400-
300-
y /
*T ! "fitted" 100-year
fc 200- rainfall
X
CO
2
Return period (year)
* -fl- . , , ,
1 000 10 000 0.1 10 100 1 000 10 000
Fig. 1 Cumulative frequency distributions for maximum rainfall over 1 and 24 h at
Néoulous station (eastern Pyrenees).
300 Maximum rainfall over 24-hours (mm) * 600 ftlaximilm raintatl over i4-nours (mm) y
5-YEAR VALUES / 100-YEAR VALUES /
250 500
m
__
• » 0 » /
MOL
SIMULATION
_i
400
è •
150 -
m *yA• 300
£*.
• j
^ fm/
200
• *V^
m/ • j
100
i
FITTING FITTl NG
0 0
0 50 100 150 200 250 300 0 100 200 300 400 500 600
Fig. 2 Five-year and 100-year values of maximum rainfall over 24 h for 50 stations:
comparison between quantiles from simulation and from fitting of a Gumbel law.
438 Patrick Arnaud & Jacques Lavabre
estimates. Five-year values from simulation and from fitting were very similar (the
same result was obtained for short-duration maximum precipitation). Theoretical
probability laws always correctly represented cumulative distribution of observations
(up to a 10 years return period). Five-year quantiles from fitting are representative of
observation sample. The good agreement between these quantiles and those from
simulation helps to validate the model. It can be concluded that the model correctly
restores observed maximum precipitation distributions (with respect to mean and
5-year or 10-year values for maximum precipitation).
For 100-year quantiles, a clear divergence can be seen for 24 h maximum
precipitation (this result is less noticeable for short duration maximum precipitation).
This confirms the observations made for the graphs in Fig. 1 for the Néoulous station.
Beyond the observation period, the fitted quantiles are estimated by extrapolation
of the theoretical law. Simulated quantiles are always estimated empirically, based on
mean distribution of modelled maximum precipitation. Divergence between the two
approaches appeared during extrapolation of the distribution beyond the observed
values, where no direct proof is possible. Without entering the argument about the
better approach, we have attempted to analyse the specificities of modelling which
determine this behaviour.
This section is focused on examining the model's features which can cause over-
exponential behaviour.
It was first verified that shape variables (like TSE, RPX) and some other variables,
used to generate hyetographs, had only a slight influence on depth of generated
rainfalls. However, the law of mean rainfall intensities (HMA variable) clearly
differentiates the stations. As this law does not have a strict exponential behaviour, its
influence at low frequencies was studied.
Choosing a law: modelling breakage The search for a law which was not very
sensitive to sampling, to describe storm intensities led us to choose a simple
exponential law fitted to values four times lower than the mean of the variable
(\xHMA), and a slope breakage was then introduced to generate all the values beyond
this limit (Arnaud, 1997; Arnaud et al, 1998). The breakage at the value four times the
mean of the variable and modelling this breakage were based on a study of
so-called "regional" distributions of HMA variables. These distributions were designed
by clustering the variable's homogenized values for all 50 studied stations. "Regional"
distributions provide a larger sample of particularly high intensity values for study
(Fig. 3). [Homogenizing values was performed by dividing the values of each series by
its local mean. The hypothesis was made that the variables reduced by their mean have
the same standard deviation (or same coefficient of variation for untreated values)
which makes it possible to cluster homogenized values.]
A break can be seen in the "regional" laws of HMA variables which defined the
relationship between the slopes (a and a') of exponential laws. The second slope (a')
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 439
HMAP/u(HMAP) HMAO/ii(HMAO)
~ 0 10 15 20 25 0 A
10 15 20 2
4~
ë 0 •*—i r^ ^ A
41
§• -2
4>
fc
-4
41
^'W
\ l
\/ " V i .
8
I " \ ^•\» N.
^H^ I *
V -io r a ' = a/3 *
1-12 Va «j^-
T B —^a'- a/2
Winter season Summer season
Intensity of main storms Intensity of normal storms
Fig. 3 Distributions of "regional" frequency of storm intensities for all 50 stations
reduced by local mean.
of the intensities law (fitted locally to values four times less than mean) was defined by
this ratio. Even, if the break in slope in the exponential law only occurs at very low
frequencies, it could explain the model's over-exponential behaviour.
1000 -
Néoulous station (P6601) Néoulous station (P6601)
900 -
E 800 -
fci
s»
3 700 •
•= 600 •
C4
<U
S 500 •
dmum ntinfa
400-
increase of more than 20% in 1000-year maximum rainfall over 24 h per break in
slope, and 80% of the stations showed a less than 10% increase in 1000-year
maximum rainfall over 24 h. The difference obtained through this break in slope does
not explain the model's over-exponential behaviour.
Another factor might explain the rainfall's over-exponential behaviour: modelling
storm persistence.
Role of persistence
Modelling storm persistence within a single rainfall episode originated from the
problems encountered within the initial model for generating high rainfall
accumulation over 24 h and from observations made on grouping storms together
around the main storm.
Illustration of cluster phenomenon for heavy storms The graphs in Fig. 5 show
the episode at the Néoulous station when rainfall observed over 24 h was the heaviest
(322 mm) on record. On the left side, ordinary storms are presented (in grey) with the
main storm (in black) which had the highest volume during the episode. On the right
side are mean intensities of storms during the same episode. Within this rainy episode
mean intensity of the main storm was less than the intensity of the ordinary storm
which followed.
Looking qualitatively at other rainy episodes with high daily accumulation
(as observed at the Col de Bavella in Corsica: nearly 800 mm over 24 h), it can be seen
that they were primarily caused by a succession of several storms all clustered together
which individually had strong, but not exceptional, mean intensities. The high
accumulated rainfall over several dozen hours was caused by a cluster of high intensity
storms, rather than by the exceptional depth of the storms themselves.
Quantitatively, the probability of ordinary storm intensities occurring according to
their position in relation to the main storm within an episode was found (denoted F(i)
where / is the position of the ordinary storm compared to the position of the main
storm and F a cumulative frequency). The mean of F(i) was calculated for each
/-position value and denoted uF(/). The graph on the left in Fig. 6 illustrates the change
in mean occurrence probability for ordinary storms at the Néoulous station according
1
30 J 30 • S £
20 ~ 20
10
'." 1m*. A w i l f f i = ioMJlat i 0
Time (hours) 21 26 31 36 41 46 51 56 61 66 71 Time (hours) 21 26 31 36 41 46 51 56 61 66 71
Fig. S Episode with the highest daily rainfall at Néoulous station: graph of hourly
rainfall and mean storm intensities.
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 441
-5 -4 -3 -2 -1 +1 +2 +3 +4 +5
Fig. 6 Changes in uF(() and number of /-position storms according to position of
ordinary storm compared to the main storm: Néoulous station (P6601).
to their relative position (/) in relation to the main storm. The right hand graph shows
the number of ordinary storms used to make the calculation.
Theoretically, if the storm position is random within an episode, mean occurrence
probability of storm intensities will be equal to 0.5, no matter when the storms occur
within the episode. The left hand graph in Fig. 6 illustrates that this theory does not
hold true. Ordinary storms close to the main storm (i = ±1) have an average occurrence
probability (|xF(0) greater than 0.5. The further away from the main storm (\i\ > 2), the
lower the uF(0 value. Calculations were stopped at uF(z) when i = 5 as too few storms
were available for study (right hand graph in Fig. 6).
The (xF(/) values were calculated in the same way for i = -2, - 1 , +1 and +2 for six
stations representative of the pluviométrie variability within the study zone (Table 1).
As in the left hand graph of Fig. 6, a clear tendency can be observed for ordinary
storms to occur around a main storm whose intensities have a higher than normal
occurrence probability (average 0.51-0.59). The storm intensities which are the most
remote from the main storm have a lower than normal occurrence probability (with an
average up to 0.42).
This illustrates the persistence phenomenon of heavy storms within a single
episode around the heaviest rainstorm.
Table 1 Mean occurrence probability of ordinary storm intensities located near the main storm.
Station number P1310 P2091 P2098 P6601 P8354 P8369
(Station name) (Aix-en -P.) (Felce) (Bavella) (Néoulous) (Bonnaux) (Rimbaud)
mean F(storm P-2) 0.42 0.53 0.49 0.49 0.47 0.46
mean F(storm P-1 ) 0.59 0.55 0.51 0.56 0.51 0.51
mean F( storm P+1 ) 0.58 0.58 0.59 0.54 0.56 0.58
mean F(storm P+2) 0.51 0.46 0.47 0.49 0.53 0.51
0.8
i = -2 i = -1 i = +1 i = +2
0.7
|0.6
c 0.5 4
EO.4
~
"a
cO0.3
Frequency classes of HMAP (FHMAP) occurrence probability
S" 0- 0.2- 0.4- 0.6- 0.8- 0- 0.2- 0.4- 0.6- 0.8- 0- 0.2- 0.4- 0.6- 0.8- 0- 0.2- 0.4- 0.6- 0.8-
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
Fig. 7 Mean occurrence probability of ordinary storms according to the occurrence
probability of the main storm: example at Néoulous station.
By examining strong ordinary storms along with the main storm in an episode, an
association can be observed between the heaviest ordinary storms and the heaviest
main storms.
These persistence phenomena were modelled as follows: the strongest ordinary
storms were grouped around the main storm, which had been placed within the longest
rainy period in the episode. Persistence was parameterized by a frequency defining the
HMAP frequency threshold beyond which persistence modelling was used. This
allowed to associate clustering of heavy ordinary storms with a high intensity main
storm. This modelling frequency was determined by a simple ratio using the variation
coefficient for maximum daily precipitation. High rainfall accumulation can occur at
high altitude stations where a disturbance can be blocked for several hours or even a
whole day. These stations have a high variation coefficient for maximum daily
precipitation and are well-adapted to persistence modelling.
The graphs in Fig. 8 show the influence of persistence modelling on model behaviour
for hourly rainfall generation at the Néoulous station. Maximum hourly rainfalls are
slightly influenced by storm persistence, which is normal as it is not dependent on the
position of the storm within the episode. The maximum rainfall, which was the most
influenced by persistence, had a duration between 6 and 48 h. It should be noted that
including persistence is necessary to improve the restitution of maximum precipitation
within the observation range (right hand graph in Fig. 8).
To expand these observations to all the studied stations, maximum 5-year
precipitation over 24 h, shown in the graphs in Fig. 9, was compared first with, then
without persistence modelling.
A clear improvement can be seen for maximum 5-year precipitation over 24 h at
high values which was underestimated during modelling without persistence. The
model could not easily generate high accumulations over 24 h and tended to
underestimate long rainfall episodes. The graphs in Fig. 9 demonstrate the value of the
persistence model as they correctly reproduce these observations.
If 1000-year values are compared and modelled with and then without persistence,
the deviation is even greater. It can represent as much as a factor of two, clearly
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 443
300 Maximum rainfall e ver 24-hours (m »)—71 300 - Maximum rainfall over 24-hours (mm ' -A
5- YEARN 'ALUES 5-YEAR VALUES /
250
• •
WITH PERSISTENCE
200 i 200 -
SIMULATION
OUT PL RSISTE
SIMVL ÂTION
150
•/•
1
A»
• •
150 -
£
1/4
100 100 -
m/
50 50 • i ••
F ITTING FITTING
i 0 - 1 1
50 100 150 200 250 300 50 100 150 200 250 300
Fig. 9 Five-year maximum rainfall values over 24 h obtained with and without
persistence modelling for 50 study stations.
DISCUSSION
Persistence modelling is entirely justified by the fact that ordinary storms cluster
together around the main storm. When the study of this phenomenon is extended, it
can be observed that there is a certain positive dependency between occurrence
probability of the main storm and occurrence probability of storms which come before
or after it. Two combined effects occur:
- within one rainy episode, the strongest ordinary storms are preferentially clustered
together around the main storm,
- considering the number of ordinary storms throughout all the episodes, the
strongest storms close to the main storm are preferentially associated with the
strongest main storms and vice versa.
The arrangement of storms within a rainy episode was modelled to respect the
phenomenon observed for all the episodes. During virtual long period simulation, the
rainy episodes were statistically reproduced at an hourly time step whose
characteristics were similar to the observed rainy episodes. Extrapolating these
cumulative frequency distributions beyond the observed frequencies was achieved
only by virtual long period simulation as the rainfall modelling; it was the same as that
which would have been obtained for short periods. A good concurrence between
distributions from the simulations and observed distributions made it possible to
evaluate modelling quality for hourly rainfall episodes. Respecting rainfall
characteristics within the range of observations allowed the validation of distribution
extrapolation towards rare frequencies.
Model behaviour towards rare frequencies was achieved by extrapolating the
internal characteristics of observed rainfall episodes. The temporal structure of the
hyetographs was considered through analysis of the different storm characteristics and
their occurrence and arrangement which yielded a model of strong rainfalls.
It is believed that the over-exponential behaviour simulated for some pluviométrie
stations merely reflects observational reality.
It should be noted that there is no numerical limit to the duration of a simulation
period (apart from calculation time). Depths of simulated rainfall did not approach a
limit which does not concur with the concept of limit rainfall defined by some research
school (Hershfield, 1961; WMO, 1973, Chow et al., 1988). However, the concept of
limit rainfall is valid for frequencies over 104 (return period 10 000 years) within a
range where modelling is not generally used. It is obvious that these very long
simulations are only performed to understand very low occurrence probability
phenomena and are not estimation devices.
CONCLUSIONS
Rainfall estimation for short time steps can be performed by simulating hourly rainfall
time series using a stochastic model. The descriptive analysis of the internal structure
of rainfall episodes observed over an hourly period makes it possible to define random
variables used to statistically reproduce the observed phenomenon. Simulating
hyetographs for long virtual periods yields a number of possible outcomes which
reflect rainfall diversity and temporal data. A range of frequencies is obtained linked to
the simulation period for a multitude of episodes which provide a new approach to
rainfall asymptotic behaviour.
There is a good similarity between simulated rainfall quantiles and observed
rainfall quantiles in the observable frequency range. Extrapolating cumulative
Using a stochastic model for generating hourly hyetographs to study extreme rainfalls 445
REFERENCES
Acreman, M. C. (1990) A simple stochastic model of hourly rainfall for Farnborough. Hydrol. Sci. J. 35(2), 119-148.
Arnaud, P. (1997) Modèle de prédétermination de crues basé sur la simulation—extension de sa zone de validité,
paramétrisation horaire par l'information journalière et couplage des deux pas de temps. Thesis Doct., Univ.
Montpellier II, France.
Arnaud, P., Lavabre, J. & Masson, J. M. (In press) Amélioration des performances d'un modèle stochastique de génération
de hyétogrammes horaires: application au pourtour méditerranéen français. Rev. Sci. Eau.
Cernesson, F. (1993) Modèle simple de prédétermination des crues de fréquences courantes à rares sur petits bassins
versants méditerranéens. Thesis Doct., Univ. Montpellier II, France.
Cernesson, F., Lavabre, J. & Masson. J. M. (1996) Stochastic model for generating hourly hyetographs. Atmos. Res.
42(1-4), 149-161.
Chow, V. T., Maidment, D. R. & Mays, L. W. (1988) Applied Hydrology. McGraw-Hill, London.
CIGB (1992) Choix de la crue de projet. Méthodes actuelles. Bulletin 82. Commission Internationale des Grands Barrages,
Paris.
Cowpertwait, P. S. P. (1991) Further developments of the Neyman-Scott clustered point process for modelling rainfall.
Wat. Resour. Res. 27(7), 1431-1438.
Duband, D. & Guillot, P. (1968) La méthode du Gradex pour le calcul de la probabilité des crues à partir des pluies. SHF,
Xème journées de l'hydraulique, question l, rapport 7.
Econopouly, T. W., Davis, D. R. & Woolhiser, D. A. (1990) Parameter transferability for a daily rainfall disaggregation
model. J. Hydrol. 118, 209-228.
Garcia-Gazman, A. & Aranda-Olivier, F. (1993) A stochastic model of dimensionless hyetograph. Wat. Resour. Res.
29(7), 2363-2370.
Hershfield, D. M. (1961) Estimating the probable maximum precipitation.. J. Hydraul. Div. ASCE 87,99-106.
Huff, F. A. (1967) Time distribution of rainfall in heavy storms. Wat. Resour. Res. 3(4), 1007-1019.
Koutsoyiannis, D. & Foufoula-Georgiou, E. (1993) A scaling model of a storm hyetograph. Wat. Resour. Res. 29(7),
2345-2361.
Koutsoyiannis, D. (1994) A stochastic disaggregation method for design storm and flood synthesis. J. Hydrol. 156, 193—
225.
Lebel, T. ( 1984) Moyenne spatiale de la pluie sur un bassin versant: estimation optimale, génération stochastique et gradex
des valeurs extrêmes. Thesis, Institut National Polytechnique de Grenoble, France.
LeCam, L. (1961) A stochastic description of precipitation. Preprints: Fourth Symp. on Mathematical Statistics and
Probability (Univ. of California, Berkeley), vol. 3, 165-186.
Onof, C , Chandler, R., Kakou, A. & Northrop, P. (1995) Rainfall modelling using Poisson-cluster process. In:
International Conference in honour of Jacques Bernier (11-13 September 1995), Chapter II. UNESCO, Paris.
Rodriguez-Iturbe, I., Cox, D. R. & Isham, V. (1987) Some models for rainfall based on stochastic point processes. Proc.
Roy. Sot: London A410, 269-288.
Tourasse, P. (1981) Analyses spatiales et temporelles des précipitations et utilisation opérationnelle dans un système de
prévision de crues. Application aux régions cévenoles. Thesis, IMG Univ. Scientifique et Médical, Inst. National
Polytechnique de Grenoble, France.
Waymire, E. & Gupta, V. K. (1981) The mathematical structure of rainfall representations: a review of the stochastic
rainfall models. Wat. Resour. Res. 17(5), 1261-1272.
446 Patrick Arnaud & Jacques Lavabre
WMO (1973) Manual for estimation of probable maximum precipitation. Operational Hydrology Report no. I, World
Meteorological Organization, Geneva, Switzerland.
Woolhiser, D. A. & Osborn, H. B. (1985) A stochastic model of dimensionless thunderstorm rainfall. Wat. Resour. Res.
21(4), 511-522.