Proceedings of the 2006 IASME/WSEAS Int. Conf.

on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

ROLE OF SYNTHETIC STORMS ON PEAK FLOW ESTIMATION

BETÜL SAF
Department of Hydrology&Hydraulic&Water Resources
Pamukkale University
Kınıklı Campüs, Denizli
TURKEY
Abstract

In this study, synthetic annual maximum storms distributed as Gumbel with random effective storm durations and
specific time distribution of precipitation for a given population mean and variance are generated by Monte Carlo
method. Population means and variances are defined by a three parameter nonlinear regression model and by a cubic
logarithmic model, respectively. Effective durations of the synthetic annual maximum storms are related to basin
characteristics, length and harmonic slope of the main course through Kirpich’s time of concentration relationship.
Annual peak flow series of the hypothetical basins with physiographic characteristics are generated through a
deterministic rainfall-runoff model, namely the Soil Conservation Service synthetic unit hydrograph method and
discrete convolution technique. Sample statistics and frequency distributions of the generated annual maximum
storms of random effective durations and of the resulting peak flow series are investigated. Eight well-known
probability distribution models with moment and maximum likelihood parameters are tested by chi-square and
probability plot correlation goodness of fit tests.

Key-words: synthetic rainfall, monte carlo, annual peak flow, SCS abstraction, frequency distribution, goodness-of-fit.

1. Introduction
The design of water resources structures for use in ungaged basins requires some estimate of flood flows and their
frequency of occurence. If there is no historical streamflow data for these watershed areas, regional flood frequency
analysis, or parametric rainfall-runoff event simulation can be used to estimate floods. Available precipitation data
with maximum value are used as input variables in parametric rainfall-runoff models. If both rainfall and runoff data
are not available, random input variables are generated by the Monte Carlo method and frequency analysis techniques
are used to analyze output variables.

Sometimes rainfall data recorded at meteorological stations may be much longer than corresponding streamflow
observations. Hence, frequency models fitted to extreme rainfall data are more reliable than the frequency models of
peak flows. This advantage can be transferred into improving or performing decisions on the appropriate frequency
distribution model for the peak flows. Of course, occurrences and frequencies of peak flows are also dependent on the
time distribution or pattern of the rainfall, basin physiography, vegetation, land management, and antecedent moisture
conditions of the watershed.

Methods of generating design storms are available and in wide use, but they are general in nature and assume storms
occur with the same temporal distribution over a wide area. Because of extreme climatic differences between the
areas, design curves are not likely to be representative of the actual time distribution of storms in semi-arid regions.
For example, [12] decided to develop a new design storm generation procedure applicable to the State of Wyoming
based on observed storm rainfall in Wyoming. Synthetic storm series, with similar statistical properties to observed
series, were developed by [17]. Data of the Alcorta raingauge (with a 5-year series of data) were analyzed by means
of five variables: duration of the rain, time between events, average and maximum intensity of the rain, and storm
advance coefficient. The variables were classified as independent (the first three variables) and dependent (the last
two variables). Probability distribution functions were fitted to the independent variables. Multiplicative relationships
were proposed for dependent variables and their coefficients were adjusted. The statistical characteristics of the
synthetic series were calculated and compared with the observed data series. A good agreement between calculated
and observed series was obtained [17].
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

2. Scope of the Study

The major objectives of this study were to find satisfactory answers in the following questions and to give an insight
into the problem of rainfall-runoff transformation on a probabilistic basis.

(1) Provided that the storm events of different durations have the same type of probability distribution (say, Gumbel),
does the type of the probability distribution of mixed storm events remain unchanged?

(2) Does, or how does, a rainfall-runoff transformation process transfer the distributional characteristics of rainfall
events into the output (peak flows)?

(3) Are there significant and physically meaningfull relations between the statistics of input and output?

(4) The storm events which are responsible for the annual peak flows may be drawn from different populations.
Therefore, return periods of input and output may or may not be the same.

In order to find answers to the above given questions, a Monte Carlo simulation is followed in the study. Twenty-five
sets of synthetic storms of different durations each N=100 size are generated. Kirpich’s empirical relationship
between the time of concentration of the small watershed and the length and slope of the main course is used for
deciding the critical rainfall durations over the hypothetical watersheds. Synthetic storm inputs are transformed into
peak flows by following Soil Conservation Service (SCS) triangular unit graph and discrete convolution procedures.

Statistical descriptors of both input (rainfall) and output (peak flow) series and parameters of various probability
distribution models by the methods of moments and maximum likelihood are estimated. Tests for the goodness of fit
of those distribution models by the chi-square test, as well as the probability plot correlation test, are performed and
frequency of acceptance of each model is found. Relations between the types of probability distributions of the input
and the output are investigated on the basis of relative acceptance frequencies of the postulated probability
distribution models.

3. Methodology
3.1 Probability Models, Parameter Estimations and Goodness of Fit

In this study; the normal (NOR), lognormal with two- and three parameter (LN2 and LN3), gumbel (GUM), log-
gumbel (LGUM), gamma with two-and three parameters (G2 and G3), and the log-pearson III distributions (LP3) are
used for synthetically generated storms and peak flows which are generated from these storms by using a rainfall-
runoff models. Numerous methods of estimation are available and two most common approaches that method of
moments (MOM) and maximum likelihood (ML) methods are used in this study. Detailed information about these
probability models and parameter estimation methods are given by [3] and [8].

The selection of a distribution type is crucial and propounds a fundamental challenge to hydrologists of the flood
frequency approach. For many cases, custom and convenience often play a significant role in the selection of a flood
distribution. In this study, chi-square (χ2) and probability plot correlation (PPC) test were used for goodness of fit.

3.2 Synthetic Unit Hydrograph

Determination of the unit hydrograph of a basin is very important for the design of water structures, because it gives
information about peak discharge, time of the peak discharge, and duration of excess runoff. When it is necessary to
determine a unit hydrograph for a ungaged basin, therefore, one of the synthetic unit hydrograph determination
methods is used. Synthetic unit hydrographs can be estimated for ungauged drainage basins by means of relationships
with parameters of characteristics of the drainage basin. The most commonly used methods are the Snyder (1938), the
Mockus (1957), and U.S. Soil Conservation Service (SCS) (1972) methods. In this study the Mockus synthetic unit
hydrograph method was used.
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

4. Computational Algorithm for Generation Synthetic Storms and Peak Flows

4.1 Generate Synthetic Storms

4.1.1 Time of concentration, effective storm duration and unit duration

A very simple event-based rainfall-runoff model is applied to a 100 km2 hypothetical watershed area without
baseflow. Durations of synthetic storms conforming basin lag are generated by considering that the length (L) and
harmonic slope (S) of the main channel are dominating factors on the time of concentration. The effective
storm duration, De, is given as [2] and [10];

De=2(tc)0.5 (1)

tc=0.00032(L)0.77/S0.385 (2)

where L is the length of the main course in meters, S is the harmonic slope, tc is the time of concentration of the
watershed and De is the effective storm duration, both in hours.

Time of concentration (tc) according to Kirpich’s formula Eq.2 for various lengths of main course (L) and harmonic
slopes (S) are given in Table 1. For tc≥4 hour, effective storm durations are assumed to be equal to time of
concentration (De=tc).

Table 1 Time of Concentration (tc; hour ) According to Kirpich’s Formula for Various Lengths of Main Course (L)
and Harmonic Slopes (S)
S (Harmonic Slope)
L(m) 0.001 0.005 0.01 0.015 0.02
8000 4.63 2.49 1.91 1.63 1.46
12000 6.32 3.40 2.60 2.23 2.00
16000 7.89 4.25 3.25 2.78 2.49
20000 9.37 5.00 3.86 3.30 2.95
25000 11.13 5.99 4.58 3.92 3.51

A typical rainfall hyetograph is a composition of M number of discrete pulses of ∆t time increments. Therefore the
critical rainfall duration, D, generated in a random manner from a uniform distribution in the range De≤D≤2De is
rounded off as multiples of ∆t:

D = M. ∆t (3)

In order to simplify calculations of composite hydrograph M is assumed as 5, 6, or 7. This assumption is in

accordance with unit graph durations used in hydrologic practice, 0.15tc≤∆t ≤0.20tc.

4.1.2 Type-I Extremal Distribution of Annual Maximum Storms with Asymptotic Moment Parameters

In this study, the mean of rainfall depth-rainfall duration and standard deviation of rainfall depth-rainfall duration
relationships of Uşak Meteorological Station developed by Benzeden (2001) are used as population statistics of the
synthetic storms. The mean of rainfall depth-rainfall duration and the standart deviation of rainfall depth-rainfall
duration relationships are as follows:

M D = 4.154[ln(D 1.1763)]
1.0263
(4)
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

[
S D = exp − 1.46047 + 1.79839 ln(D ) − 0.3155 ln 2 (D ) + 0.01944 ln 3 (D ) ] (5)

where D is in minutes and MD and SD both are in mm. According to these relationships, synthetic rainfall depths, YD,
distributed as the Extreme Value Type-I (Gumbel) with asymptotic parameters are generated through Eq.6.

YD,T = MD +SD.KT (6)

where

KT = - {0.45 + 0.7797 ln[-ln(1-1/T)]} (7)

is the frequency factor corresponding to the return period T or probability of nonexceedance PT which can be
calculated from generated uniform random numbers 0 <PT <1, that is T=1/(1-PT) (Kite, 1977).

4.1.3 Time Distribution of the Generated Storms

The SCS dimensionless cumulative rainfall curves were developed for various storm types, storm durations and
regions in the United States [10]. In this study, SCS 6-hour time distributions are used in order to calculate rainfall
hyetographs of given synthetic storms.

A design storm, YD, is divided into increments ∆Ym using an appropriate time distribution curve for the project site. A
time distribution curve represents the cumulative percentage of the precipitated rainfall f m = Ym YD , during the
percentage time X m = tm D , where D=M.∆t is the rainfall duration and tm=m∆t (m=1,2,...,M). Having fm values, the
cumulative rainfall amount (Ym) precipitated during period 0 to tm, and incremental rainfall amount (∆Ym) can be
computed as:

Ym = f m D , m=1,2,...,M (8)
∆Ym = Ym − Ym −1 , m=1,2,...,M (9)

where y0=0 for m=1.

4.2 Genaration Synthetic Peak Flows According to Synthetic Annual Maximum Storms

The SCS-curve number method with CN=90 is applied in order to account for the initial abstractions and its role on
the excess rainfall hyetograph and composite hydrographs. For the sake of simplicity, it is assumed that the Mockus
unit hydrograph is a sufficient tool to simulate the rainfall-runoff transformation process. The triangular unit graph
[10] of a unit duration ∆t includes the size of the basin drainage area, A, the length of the main channel L, and slope,
S because of the relation between time to peak, tp, and time of concentration, tc.

t p = ∆t 2 + 0.6t c = f1 ( L, S , ∆t ) (10)
q p = KA t p = KA f1 ( s ) = f 2 ( A, L, S , ∆t ) (11)

Ordinates of ∆t-hour triangular unit hydrograph at times ∆t, 2∆t, 3∆t, .... are used in calculating superposed output
resulting from given ERH of M number of incremental excess rainfall pulses, R1, R2, .... RM.

n≤ M
Qn = ∑R U
m =1
m n − m +1 (12)
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

5. Results

5.1 Probability Distribution of Mixed-Storm Durations

When the results of χ2 tests for 25 synthetic storm series are evaluated the most suitable model for the mixed-duration
storm series cannot be determined clearly (Table 3). Though LN2 distribution for the case with MOM parameters has
the highest acceptance frequency, GUM, G2, LN3 and G3 distributions are also compatible distributions. For the
maximum likelihood parameters, LN2, G2 and GUM distributions are the most suitable distributions. On the other
hand the PPC test results reveal that the most suitable models are LN2, G2, GUM, LN3 and G3 either with MOM or
ML parameters.

5.2 Relations Between the Types of Probability Distribution Functions

In order to compare the probability distribution models most frequently accepted (that is, the goodness of fit test is
passed) for the synthetic storm samples (input) and for the peak flows (output), relative acceptance frequencies of
each model, fI and fO, computed from Eq.13 and 14, at a significance level α=5% for the χ2 test are presented
comparatively with and maximum likelihood parameters. Relative acceptance frequency of a specific model (fI) is
defined as:

fI= 100 (TNCH)/25 (13)

where TNCH is total number of series that passed the χ 2 test for a specific distribution. Similarly, using parameters
estimated by method of moments, relative acceptance frequencies of each model at a significance level α=5% are
presented comparatively. The results of the PPC tests are evaluated in the same way except that the relative
acceptance frequencies of the output series are calculated now from Eq.14.

fO = 100 (TNPP)/25 (14)

where TNPP is total number of storm series that have a PPC coefficient greater than rc=0.95. The PPC results are
shown in Fig.5 and 6. Relative acceptance frequencies according to the chi- test results with MOM parameters LN2
is the most appropriate model. Similarly, with ML parameters; LN3 is the most appropriate model. For PPC test, it
LN2 and LN3 models are most suitable both of parameter estimation methods.

Table 3 Results of godness-of fit tests of synthetic storms (input) and peak flow series (output) with MOM and ML
estimators (GOF:Goodness-of-fit, PEM:Parameter estimation method)
GOF methods PEM Data NOR LN2 G2 GUM LGUM LN3 G3 LP3

Input 48 92 84 88 76 84 84 32
MOM
Output 0 80 40 20 68 52 72 68
Chi-square
Input 48 96 96 96 28 92 88 28
ML
Output 0 76 56 48 72 84 56 36
Input 84 100 100 100 96 100 100 52
MOM
Output 12 100 96 84 96 100 96 0
PPC
Input 84 100 100 100 96 100 100 28
ML
Output 12 100 84 84 96 100 96 40
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

5.3 Relations Between Distributional Characteristics of the Input and Output

Relation between the types of the probability distribution of input and output according to the chi-square tests are
investigated on the basis of correlations (r) between the relative acceptance frequencies and are calculated with
maximum likelihood parameters according. The correlations between the acceptance frequencies of input and output
are calculated. The input-output relationships are 0.197 and 0.350 for chi-square goodnes-of fit test with MOM and
ML, respectively. When PPCC model is used, input-output relationships are 0.350 and 0.625 for MOM and ML,
respectively. As it can be seen, result of ML parameter estimation method is better than the MOM. But, the
relationship between input and output distribution is not good in both of case.

6. Conclusions and Recommendations

As can be seen from Table 3, when the twenty-five synthetic storm series were evaluated with chi-square goodness of
fit test, the most suitable distribution with MOM parameters is LN2; and LN2, G2 and GUM with ML parameters.
This means that when the generated synthetic storms distributed as Gumbel are put into a mixed duration series the
type of the appropriate distribution may change. Similarly, according to the results of probability plot correlation test
given in Table 3, LN2, G2, GUM, LN3 and G3 distributions are the most suitable distributions either with MOM or
ML parameters. These results of the study reveals that the probability distribution of the rainfall input may even
diverge from their parent (Type-I Extremal) distributions because of the sampling, and since the generated input series
is a mixture of rainfall events of variable durations. Provided that the storm events of different durations have the
same type of probability distribution, the probability distribution types of mixed storm events have not been seen to
change drastically.

The answer of the second question may be given as, a rainfall-runoff transformation process doesn’t transfer the
distributional characteristics of rainfall events into the peak flows. The answer of third question is there aren’t
significant and physically meaningfull relations between the probability distribution of storm and peak flow. The last
answer is the storm events which are responsible for the annual peak flows may be drawn from different populations.
Therefore, return periods of input and output may not be the same.

This study is based on various assumptions and simplifications and therefore the conclusions should be evaluated
carefully. A rather complicated and comprehensive research should be conducted by taking into account for the
complicated mechanism of rainfall-runoff transformation process which must include the storage effects, the
nonlinearities in the watershed system, and alternative probability distribution models for the input. It is hoped that
these conclusions of this study will be beneficial in small basins where significant information on precipitations is
available in case of limited information on peak flows.

References

[1] Benzeden E., Standart Süreli Maksimum Yağışların Frekans Analizinde Karşılaşılan Sorunlar, DSI Teknoloji D.
Bşk. Basım Şb. Md., III. Ulusal Hidroloji Kongresi Bildirileri, Izmir, p.11-18, 2001.

[2] Chow, V. T., Handbook of Applied Hydrology, McGraw-Hill, 1964.

[3] Kite, G. W., Frequency and Risk Analysis in Hydrology, Water Resources Publications, Fort Collins, Colorado,
USA, 1977.

[4] Lettenmaier, D. P., & Burges, S. J., Gumbel’s Extreme Value I Distribution: A New Look, Journal of Hydraulic
Division, 502-514, 1982.

[5] Linsley, R. K., Kohler, M. A., & Paulhus, J. L. H., Applied Hydrology, McGraw-Hill, New York, 1949.
 Proceedings of the 2006 IASME/WSEAS Int. Conf. on Water Resources, Hydraulics & Hydrology, Chalkida, Greece, May 11-13, 2006 (pp13-20)

[6] Marien, J. L., & Vandewiele, G. L., A Point Rainfall Generator with Internal Storm Structure, Water Resources
Research, 22, 475-482, 1986.

[7] Mockus, V., Use of Storm and Watershed Characteristics in Synthetic Hydrograph Analysis and Application,
Amer. Geophys., Union, Southwest Region Meeting, Sacramento, Calif., 1957.

[8] Rao, A. R. and Hamed K. H., Flood Frequency Analysis, CRC Pres LLC, Boca Raton, Florida, 2000.

[9] Saf, B.,Role of Rainfall Pattern and Basin Physiography on the Probability Distribution of Floods, Ph. D.
Dissertation, Dokuz Eylül University, Turkey, 2003.

[10] SCS, National Engineering Handbook, Section 4, Hydrology Soil Conservation Services, U.S., Department of
Agriculture, Washington, D. C., 1972.

[11] Snyder, F. F., Synthetic Unit-Graphs, Transactions of the American Geophysical Union,19, 447-454, 1938.

[12] Tyrell, Patrick T., Development of Design Rainfall Distribution for the State of Wyoming, Msc. Thesis,
University of Wyoming, 71 pp., 1982.

[13] Vogel, R. M., The Probability Plot Correlation Coefficient Test for the Normal, Lognormal and Gumbel
Distributional Hypthesis, Water Resources Research, 22, 587-590, 1986.

[14] Vogel, R. M., & McMartin, D. E., Probability Plot Goodness-of-fit and Skewness Estimation Procedures for the
Pearson Type 3 Distribution, Water Resources Research, 27, 3149- 3158, 1991.

[15] Wanieliesta, M., Kersten, R., & Eaglin, R., Hydrology: Water Quantity & Quality Control, John Wiley & Sons,
1997.

[16] Wei, Tseng C., and C. L. Larson, Effects of Areal and Time Distribution of Rainfall on Small Watershed Runoff
Hydrographs, Minnesota Univ., Minneapolis, Water Resources Research Center, Bulletin No. 30, 130 pp., 1971.

[17] Yen, Ben Chie, and Ven Te Chow, Design Hyetographs for Small Drainage Structures, Journal of the Hydraulics
Division, ASCE, Vol. 106.,No. HY6., 1980.

[18] Zimmermann, Erik D., Synthetic Storm Generatıon in a Flatland Region, Santa Fe, Argentina, Journal of the
International Association for Environmental Hydrology, 6, 2, 1998.