Professional Documents
Culture Documents
1.1 Introduction
IDF curves (Intensity Duration Frequency curves), which comes from the term curved intensity
duration return period rainfall, consist a statistical hydrology tool mainly used in flood defenses design.
They are analytical or graphical expressions of the maximum rainfall intensity i as a function of the
duration of the rainfall event d and the return period T. More details on the use of IDF curves, one can
find in the majority of hydrology textbooks (e.g. Koutsoyiannis and Xanthopoulos, 1999).
The development of IDF curves from rain gauge measurements may be carried out into two stages, which
in Hydrognomon come in the form of two distinct subsystems. The first step is to derive the time series
of annual maximum mean intensities ijl for a broad range of aggregation durations such as dj = {5min,
10min, 30 min, 1h, 2h, 6h, 12h, 24h, 48h, which are obtained from the time series of the rainfall
measurement (Extremes Evaluation subsystem in Hydrognomon). These samples ijl are used in the
second stage of the procedure in order to develop the final IDF curves (Ombros subsystem in
Hydrognomon).
The implementation of a coherent methodology that enables a single mathematical expression for the IDF
curves i(d, T) is suggested. Additionally, it is also proposed to use a statistical distribution with EV2
characteristics and the Lmoments method for the estimation of its parameters (in contrast to the classical
approach of EV1/Gumbel fit and the common method of moments for the parameter estimation).
However, the conventional methodology for estimating IDF curves in specific return periods is also
given.
The following remarks should be underlined concerning the use of the above correction coefficient:
If samples with both high and low resolution are simultaneously used, then the coefficients
should be applied with the appropriate criteria in order to remove the heterogeneity, i.e. higher
weights should be used for the lower resolution samples.
If the maximum rainfall time series have been derived from the aggregation of high-resolution
time series (e.g. 5 or 10 minutes), then application of the coefficients provides better estimates
(weights) of the IDF curves in the small durations.
If the IDF curves are derived from rain gauge sample with daily time step, then a coefficient with
value 1.13 should be used; in the opposite case there would be underestimation of the rainfall
intensity throughout the whole range of time durations.
The temporal resolution can be individually set for each time series, which is suggested if the time series
are obtained from instruments with different time steps. Alternatively, if the resolution is unknown it may
not be introduced. In any case, its effect may or may not be taken into account on the following step, i.e.
the development of the IDF curves (note that the default system option is not to be taken into account).
Relationship between annual maxima time series and time series above threshold
If the return period is small (T<10) then it has been shown that IDF curves derived from time series above
threshold are more representative (Koutsoyiannis, 1997, p 287). Most often, however, only the annual
maxima time series are available, and thus a relationship is needed in order to approximate the IDF curves
derived from series above threshold. If T is the return period corresponding to the time series of annual
maxima and T' the return period corresponding to the time series of maxima above threshold, then the
following relation links T to T:
1 1
T T ' 1.1
1
1
1 e T' ln 1
T
The relationship above can be approximated to precision of two decimal places by the following, simpler
one:
= +0.5 1.2
Both relationships are meaningful only for 2<T<10 and they can be used directly in the IDF curves
equations, by replacing T with T. This transformation may simplify the expressions of the IDF curves,
especially in the case of EV distributions (see Section Error! Reference source not found.). In
Hydrognomon the transformation is automatically applied if the corresponding option is selected, which
results to slightly higher intensity values for the given durations and return periods.
Moments method
L-Moments method
Unified equation of the i IDF
Quantile function in relation curve (intensity) in relation to
s/n
Distribution
to the return period T: x(T) the return period T and the
duration d
ln( T ) ln( T )
d
1 Expotential
1
1 G 1 1 , ,
2 Gamma G 1 1 , , T
T
d
1
G 1 1 , ,
Log- 1 1 T
3 G 1 , , e
T
Pearson III e
d
1
1 ln ln 1
4 EV1-Max ln ln 1 T
T
d
1
1 ln 1
5 EV2-Max ln 1 T
T
d
6 GEV-Max
1
1 ln 1 1
GEV-Max ( ln 1 1 T
7 T
predifened)
d
1
1 1
8 Pareto 1 T
T
d
In Table 1.3 there are demonstrated the transformed quantile equations for the EV distributions, when the
return period T is used which corresponds to an above-threshold maxima time series. The transformation
is made by introducing the expression of T in relation to (Relationship 1.1). The resulting relationships
may be alternatively used for small return periods (see Paragraph 1.3.2), which in Hydrognomon may be
achieved by y=the simple numerical transformation T = T+0.5.
Table 1.3: EV distributions used in the development of the IDF curves for return period T that
corresponds to the above-threshold time series of threshold T.
ln T '
ln T '
1 EV1-Max
d
T '
2 EV2-Max T'
d
T ' 1
3 GEV-Max
GEV-Max (
T ' 1
4
predifened) d
To merge the samples, each of the time series ijl is multiplied by b(dj)=(dj+): yjl=ijlb(dj), and then all of
them are used as the new sample. This sample should be sorted in descending order, but in this step it is
suggested to use only a subset of the data because in this manner the function b(d) is better fitted in the
area of the higher intensities. Thus, initially each time series is sorted separately and only a partition of its
highest values is taken to create the merged sample (the default value in Hydrognomon is 1/3). In this
manner m may become smaller depending on the percentage drawn from the original samples which may
vary from 10 to 100%.
The influence of this percentage in the final results is not significant and therefore it is suggested to use
the default value.
For smaller samples a reduction of this percentage is made: if the percentage chosen is and the
maximum number of values from the times series is nmax = max(nj), then the final percentage q is:
, nmax 10
q 10 / nmax , nmax 10 nmax 10 1.9
1, 10 nmax 10
nmax
When the q nj maximum values of each value have been chosen the size of the merged sample should be:
k
m q nj 1.10
j 1
Once the merged sample is sorted in descending order, a single rank is assigned to each value in
ascending order, i.e. 1 to the highest value, 2 to the next, etc. If two sample values are identical then the
mean of the corresponding ranks is used. Finally, the mean of each time series ranks is estimated
separately:
k
1
rj
nj
rj 1
jl 1.11
where a percentage of q nj is used instead of nj. If all the individual samples had the same distribution,
then each r j would have been very close to the value (m+1)/2. To achieve this goal, and eventually
estimate the optimal values of and , the statistical parameter h of the Kruskal-Wallis test is used:
m 1
2
12 k
h n j rj
mm 1 j 1 2
1.12
This statistical test examines the mean ranks of all the individual samples and the minimization of h is
achieved after an exhaustive testing process. In Hydrognomon a 31x31 matrix is used for the estimation
of the values and , which range in the space (0,1). When the optimal combination , is determined
then a new 31x31 matrix is formed centered on this pair of values. This new matrix dimensions are
1/32x1/32 in order to have a better approximation of the optimal solution (third decimal place). Hence, a
total of 1922 trials is required to derive the final values.
As an example for the algorithms efficiency, a comparison between the values of and determined by
Hydrognomon software (using the 1/3 of maxima) and Koutsoyiannis (1997, p. 280) for the station of
Ellinikon (sample data demonstrated in Table 1.4). There is only a slight difference of magnitude about
1% in the findings of the two approaches, as Hydrognomon estimates =0.792 and =0.186, while in the
textbook =0.796 and =0.189.
1.4: Measured annual maximum rainfall at Ellinikon meteorological station (Source: Hellenic
National Meteorological Service).
Hyd. Maximum rain intensity i (mm/h) for duration d=
year 5min 10min 30min 1h 2h 6h 12h 24h
1957-58 81.60 66.00 53.20 35.00 26.90 8.97 6.27
1958-59 58.80 48.00 33.00 21.50 11.20 6.75 3.98
1959-60 39.60 34.80 20.00 11.60 6.85 2.40 1.82
1960-61 54.00 34.80 18.40 11.00 6.65 3.62 2.28
1961-62 120.00 85.80 41.80 24.80 19.30 7.32 3.73
1962-63 67.20 60.00 23.60 13.80 7.20 3.03 1.94
1963-64 78.00 48.00 27.80 14.30 8.50 3.52 2.72
1964-65 96.00 63.00 28.00 15.50 10.65 4.28 2.17
1965-66 38.40 36.00 23.00 12.00 6.55 2.45 1.69
1966-67 74.40 63.60 28.40 15.10 7.55 4.88 2.46
1967-68 36.00 24.60 16.60 10.20 6.65 3.65 2.75 1.58
1968-69 126.00 69.00 43.20 26.80 15.15 5.93 2.97 1.48
1969-70 82.80 64.20 41.60 24.50 12.45 5.45 2.75 1.76
1970-71 42.00 42.00 25.20 17.70 8.95 3.70 3.09 1.55
1971-72 117.60 85.20 65.20 35.90 19.75 10.02 5.01 2.92
1972-73 68.40 49.80 39.40 33.50 17.75 6.78 5.27 2.68
1973-74 60.00 42.00 28.60 15.20 9.85 4.20 3.47 2.00
1974-75 48.00 48.00 30.60 15.90 8.30 4.27 2.60 1.30
1975-76 120.00 120.00 74.00 40.90 21.50 7.38 4.54 2.27
1976-77 115.20 87.60 41.40 23.20 14.90 6.12 3.30 1.65
1977-78 56.40 46.20 38.60 32.70 20.15 6.73 3.37 1.68
1978-79 78.00 66.60 47.60 30.00 19.55 11.93 6.12 3.37
1979-80 67.20 40.80 17.20 13.30 8.60 4.22 2.81 1.62
1980-81 58.80 56.40 30.40 19.40 11.10 5.58 3.27 1.95
1981-82 67.20 64.80 40.60 24.70 13.05 4.35 2.28 1.14
1982-83 141.60 79.80 49.60 36.20 22.90 7.63 4.52 2.29
1983-84 102.00 69.00 50.40 29.00 17.70 7.03 3.63 1.82
1984-85 40.80 31.80 16.40 12.90 12.15 9.87 6.00 3.40
1985-86 74.40 66.00 29.20 15.60 9.40 3.13 1.57 0.83
1986-87 32.20 29.10 18.55 9.50 7.24 3.85
Finally, there is the option to determine if each time series shall be used separately for the estimation of
and , or for the distribution parameters, or for both estimations. For example, and may be extracted
from 10-min time series, whereas distribution parameters may be derived from longer duration time series
(i.e. 12, 24 or 48 h). There is even the option to set explicit values, if there is some prior knowledge for
and from other empirical studies (see also Koutsoyiannis 1997, p. 284).
1.6.2 Determination of the distribution parameters
The final step in the development of IDF curves is to estimate the parameters of the statistical distribution
adopted. The final merged sample yjl results from the multiplication of ijl with b(dj) = (dj+), i.e. yjl = ijl
b(dj), and its statistical characteristics are then used in the determination of the distribution parameters.
Again, the paradigm from Koutsoyiannis (1997, p. 282) comes handy in order to check the algorithms
efficiency. The values demonstrate only a slight difference for the EV1-Max (Weibull) distribution:
=7.962 and =2.652 versus =7.95 and =2.64 (which accounts for rounding errors). Implementation of
GEV-Max (L-moments) with =0.15, which is the suggested (default) option, gives respectively: =7.04
and =2.88.
IDF Curves - Distribution: EV1-Max (Gumbel)
1,000
10000
500
50
5
100
i (mm/h)
10
1
0.01 0.1 1 10 100
d (h)
Figure 1.1: IDF curves for return periods T=5, 10, 50, 500, 5000 at the Ellinikon meteorological station,
derived by a robust unified equation and EV1-Max statistical distribution
1.05
1.11
1.25
1.43
1.67
3.33
100
200
2.5
10
20
50
2
180
170
160
150
140
130
120
110
i (mm/h)
100
90
80
70
60
50
40
30
20
10
0
Figure 1.2: Goodness-of-fit check by depicting simultaneously both the distribution functions Gumbel-
Max (straight lines) and Weibull (empirical distribution points).
n
j 1
j
nm 1.14
k
In Figure 10.3 an application based on the example shown above is illustrated for return period T = 50. It
is evident that there is almost no difference between the confidence intervals of the sample and the
population; this is strongly affected by the use of logarithmic axes which mitigate the actual difference.
1,000
100
i (mm/h)
10
1
0.01 0.1 1 10 100
d (h)
Figure 1.3: The 95% confidence intervals of the IDF curve for return period T=50. The EV1-Max
(Gumbel) distribution was used in the sample from Ellinikon station.