Prediction of Flood Frequency Factor For Gumbel Distribution PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Arab J Sci Eng (2017) 42:3895–3906

DOI 10.1007/s13369-017-2507-1

RESEARCH ARTICLE - CIVIL ENGINEERING

Prediction of Flood Frequency Factor for Gumbel Distribution


Using Regression and GEP Model
Fevzi Onen1 · Tamer Bagatur1

Received: 5 September 2015 / Accepted: 14 March 2017 / Published online: 22 March 2017
© King Fahd University of Petroleum & Minerals 2017

Abstract Floods are the most common natural disasters Keywords Flood · Frequency factor · Gumbel distribution ·
that affect societies around the world. One of the major Regression · Gene expression programming (GEP)
problems in water resources engineering design is the esti-
mation of maximum flood discharges. These estimations are
determined to assign hydrological and hydraulic dimensions 1 Introduction
to bridges, sewers, dam, spillway, protection embankments,
weirs, detention ponds and diversion canals. Accurate esti- Floods are natural hazards causing loss of life, injury, dam-
mation of flood frequency discharge increases safety of the age to agricultural lands, and major property losses [1].
hydraulic structures. In probability theory and statistics, flood Flood frequency analysis (FFA) is the estimation of how
frequency analysis is used to obtain the probability distribu- often specified flood events will occur. Before the estima-
tion of floods. The distribution models can be summarized tion can be done, analyzing of the stream or river flow data
the generalized extreme value, Gumbel or extreme value type is important in order to obtain the probability distribution
1, Log-Normal and the Log Pearson type III distributions. of flood [2]. Because, one of the greatest challenges fac-
The Gumbel distribution provides the best fit according to ing the hydrology is to gain a better understanding of flood
the extreme value analysis studies. This study concentrates regimes. To do this, flood frequency analysis is most com-
on prediction of flood frequency factor (K) for the Gum- monly used by civil engineers and hydrologists and consists
bel distribution using gene expression programming (GEP) of estimating flood peak quantities for non-exceedance prob-
and regression model. Some prediction models are presented abilities. The validity of the results in the application of FFA
for determining of flood frequency factor (K). The proposed is theoretically subject to the hypothesis that the series are
regression model (Model 4) and GEP model (Model 7) give independent and identically distributed [3,4]. Although stud-
a fast and practical way of estimating the flood frequency ies have employed several statistical distributions to quantify
factor. Thus, Gumbel’s method has been simplified in such the likelihood and intensity of floods, none had gained world-
a predictive model that one can obtain the magnitude of a wide acceptance and is specific to any country [5].
given return period for flood discharges without recourse to Estimated values of flood discharge are needed in various
looking at a table. The performance of the prediction models risks of exceedance for a wide range of engineering prob-
was evaluated with an illustrative example for 2, 5, 10, 20, lems such as culvert and bridge design. In a long record of
50, 100, 200, 250, 500 and 1000 years flood. measured floods, these flood estimates may be derived by
statistical analysis of the flow series. Alternatively the storm
magnitude for an appropriate duration, areal coverage and
return period may be converted into the flood of a given return
period using a rainfall-runoff model [4,5].
B Fevzi Onen In order to derive the risk of occurrence of any flood
fonen@dicle.edu.tr
event, the frequency distribution, which can best describe
1 Civil Engineering Department, Engineering Faculty, Dicle the past characteristics on the magnitude and the probability
University, 21280 Diyarbakir, Turkey of occurrence of such floods, must be known. This involves

123
3896 Arab J Sci Eng (2017) 42:3895–3906

the determination of the best flood frequency model, which Equation (3) may be developed as the following from Eqs.
can be fitted to the available historical record. (1) and (2):
To derive the probability of flood events, the frequency
1 −y
distributions were tested. These studies are to involve the = 1 − e−e (3)
determination of the best flood frequency models which T
could be fitted to the available recorded historical data. The Here, the reduced variable (y) may be obtained as the follow-
distribution models can be included the Generalized Extreme ing form:
Value, Gumbel or Extreme Value type 1 (EV1), Log-Normal   
and the Log Pearson type III distributions. The extreme value T
y = − ln ln (4)
analysis studies showed that the Gumbel distribution can be (T − 1)
provided the best fit. Gumbel distribution is a statistical tech-
nique for fitting frequency distribution data to predict the The statistical variate is presented with the following expres-
flood for a river. Also, Gumbel distribution is effective for sion:
smaller sample sizes than 50 record data [6]. Rasmussen and y = yn + K · σn (5)
Gautam [7] used probability weighted moments (PWM) for
estimating parameters of statistical distributions, including where yn = Gumbel’s reduced mean variable, σn = standard
the Gumbel distribution. Haktanir [8] compared various flood deviation, K = frequency factor.
frequency distributions for rivers in Anatolia. The Gumbel Equation (6) may be developed as the following form from
distributions were found more preferable than the other dis- Eqs. (4) and (5):
tributions. 

The objective of this study is to predict flood frequency − ln ln T
− yn
y − yn (T −1)
factor (K) for the Gumbel distribution using gene expression K = = (6)
programming (GEP) and regression. For this aim, the table σn σn
values of Gumbel frequency distribution are used as data col-
lection. Moreover, this paper presents GEP model, which is
an extension to genetic programming, and multiple nonlin- 3 Estimation and Predicted Alternatives of Flood
ear regression models as an alternative approach to modeling Frequency Factor (K )
of flood frequency factor. The performance of the prediction
models was evaluated with illustrative example. In this part, the estimation and predicted models are presented
for determining of flood frequency factor1 (K ). Therefore,
some approaches are evaluated.
2 Gumbel Distribution
3.1 Approach: Determination Using Gumbel’s Table
(Model 1)
In probability theory and statistics, the Gumbel distribution
(Extreme Value Type I) is used to model the distribution of The frequency factor is calculated based on Eq. (6). More-
the maximum (or the minimum) of a number of samples of over, Gumbel’s reduced mean variable (yn ) and standard
various distributions. Such a distribution might be used to deviation (σn ) can be selected using Table 1. But, the use
represent the distribution of the maximum level of a river in of Gumbel’s method is rather time consuming. Therefore,
a particular year if there was a list of maximum values for to use the method, one has to refer the values of table for
the past 10 years. determining of yn and σn .
The cumulative distribution function (cdf) of the Gumbel
distribution (maximum) is given as following expression [9]: 3.2 Approach: Determination Using Combined
Gumbel’s Table for K (Model 2)
  −y
F (x) = exp −exp (−y) = e−e (1)
The frequency factor values of Gumbel’s distribution can be
where F(x) is the probability distribution function of random directly selected using Table 2 based on data number and
variable x. Also, y is a reduced variable. return period.
The probability of exceedance is related to the return
period (T ). The probability of exceedance of the excluded 3.3 Approach: Estimation of Flood Frequency Factor
events is presented by: Using Powell Method for n = ∝ (Model 3)

1 For determining the value of K without referring to a table,


F1 (x) = 1 − F (x) = (2) Powell defined the equation as follows:
T

123
Arab J Sci Eng (2017) 42:3895–3906 3897

Table 1 Gumbel’s reduced


n yn σn n yn σn n yn σn
mean variable and standard
deviation based on data number 8 0.4843 0.9043 35 0.5403 1.1285 64 0.5533 1.1793
9 0.4902 0.9288 36 0.5410 1.1313 66 0.5538 1.1814
10 0.4952 0.9497 37 0.5418 1.1339 68 0.5543 1.1834
11 0.4996 0.9676 38 0.5424 1.1363 70 0.5548 1.1854
12 0.5035 0.9833 39 0.5430 1.1388 72 0.5552 1.1873
13 0.5070 0.9972 40 0.5436 1.1413 74 0.5557 1.1890
14 0.5100 1.0095 41 0.5442 1.1436 76 0.5561 1.1906
15 0.5128 1.0206 42 0.5448 1.1458 78 0.5565 1.1923
16 0.5157 1.0316 43 0.5453 1.1480 80 0.5569 1.1938
17 0.5181 1.0411 44 0.5458 1.1499 82 0.5572 1.1953
18 0.5202 1.0493 45 0.5463 1.1519 84 0.5576 1.1967
19 0.5220 1.0566 46 0.5468 1.1538 86 0.5580 1.1980
20 0.5236 1.0628 47 0.5473 1.1557 88 0.5583 1.1994
21 0.5252 1.0696 48 0.5477 1.1574 90 0.5586 1.2007
22 0.5268 1.0754 49 0.5481 1.1590 92 0.5589 1.2020
23 0.5283 1.0811 50 0.5485 1.1607 94 0.5592 1.2032
24 0.5296 1.0864 51 0.5489 1.1623 96 0.5595 1.2044
25 0.5309 1.0915 52 0.5493 1.1638 98 0.5598 1.2055
26 0.5320 1.0961 53 0.5497 1.1653 100 0.5600 1.2065
27 0.5332 1.1004 54 0.5501 1.1667 150 0.5646 1.2253
28 0.5343 1.1047 55 0.5504 1.1681 200 0.5672 1.2360
29 0.5353 1.1086 56 0.5508 1.1696 250 0.5688 1.2429
30 0.5362 1.1124 57 0.5511 1.1708 300 0.5699 1.2479
31 0.5371 1.1159 58 0.5515 1.1721 400 0.5714 1.2545
32 0.5380 1.1193 59 0.5518 1.1734 500 0.5724 1.2588
33 0.5388 1.1226 60 0.5521 1.1747 750 0.5738 1.2651
34 0.5396 1.1255 62 0.5527 1.177 1000 0.5745 1.2685

√   
6 T In present study, the regression equations derived of
K =− γ + ln ln (7)
π (T − 1) Table 1 for Gumbel distribution can be developed as fol-
lowing:
where γ is Euler’s constant (0.5772), 
   − 0.66
T yn = 0.5775 · n n
R 2 = 0.999 (9)

K = −0.78 0.5772 + ln ln (8) − 1.268
(T − 1) σn = 1.2811 · n n
R 2 = 0.998 (10)
What is different from Gumbel is that the value of K now
The comparison between predicted and measured Gum-
does not depend upon the number of years of record. The
bel’s reduced mean variable values for regression model is
values of K for the Powell method are given in Table 3.
presented in Fig. 2. Also, relation between Gumbel’s reduced
standard deviation and data number can be obtained in Fig. 3.
3.4 Approach: Prediction Using Regression Equations
In Fig. 4, Comparison between predicted and measured Gum-
Derived of Table 1 (Model 4)
bel’s reduced standard deviation and data number values
according to regression model is presented.
The number of record years and the return period are a strong
In practical, the frequency factor (K ) can be determined
function determining of the value of K . Relation between
using Eqs. (4) and (11) depends on data number and return
Gumbel’s reduced mean variable and data number is given
period years with determination coefficient, R 2 = 0.999.
in Fig. 1.

123
3898 Arab J Sci Eng (2017) 42:3895–3906

Table 2 Frequency factor values depend on data number and return 0.58
period 0.57

n T (return periods) (year) 0.56


0.55
2 5 10 20 50 100
0.54
K 0.53
10 −0.1355 1.0580 1.8483 2.6063 3.5874 4.3227 0.52
15 −0.1434 0.9672 1.7025 2.4078 3.3208 4.0049 0.51
20 −0.1478 0.9187 1.6248 2.3020 3.1787 3.8356 0.50
25 −0.1506 0.8879 1.5754 2.2350 3.0886 3.7284 0.49
30 −0.1526 0.8664 1.5410 2.1881 3.0257 3.6534
0.48
35 −0.1540 0.8504 1.5154 2.1532 2.9789 3.5976
0.47
40 −0.1552 0.8379 1.4954 2.1261 2.9425 3.5543 0 200 400 600 800 1000 1200
45 −0.1561 0.8279 1.4794 2.1044 2.9133 3.5195 Data number, n
50 −0.1568 0.8197 1.4663 2.0865 2.8892 3.4908
Fig. 1 Relation between Gumbel’s reduced mean variable and data
55 −0.1574 0.8128 1.4552 2.0714 2.8690 3.4667 number
60 −0.1580 0.8069 1.4458 2.0586 2.8518 3.4461
65 −0.1584 0.8018 1.4376 2.0475 2.8368 3.4284 0.58
70 −0.1588 0.7974 1.4305 2.0377 2.8238 3.4128
0.57
75 −0.1592 0.7934 1.4242 2.0291 2.8122 3.3991
0.56
80 −0.1595 0.7900 1.4185 2.0215 2.8020 3.3868
0.55
85 −0.1597 0.7868 1.4135 2.0146 2.7928 3.3758
90 −0.1600 0.7840 1.4090 2.0084 2.7844 3.3659 0.54
Predicted

95 −0.1602 0.7814 1.4048 2.0028 2.7769 3.3569 0.53

100 −0.1604 0.7791 1.4011 1.9977 2.7700 3.3487 0.52

0.51

0.50 y = 0,9981x + 0,0005


Table 3 Variation of K factor R² = 0,999
T (year) K (from Eq. 8) 0.49
based on return period and
n =∝ 2 −0.1642 0.48

2.33 0.0000 0.47


0.47 0.48 0.49 0.50 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58
5 +0.7195
Measured
10 +1.3046
20 +1.8659 Fig. 2 Comparison between predicted and measured Gumbel’s
50 +2.5924 reduced mean variable values according to regression model
100 +3.1368
200 +3.6792 1.30

250 +3.8536
1.20
500 +4.3949
1000 +4.9357 1.10

1.00

 0.90
− 0.66
y − 0.5775 · n n
K =  0.80
− 1.268
1.2811 · n n

0 200 400 600 800 1000 1200

− 0.66 Data number, n
− ln ln (T −1)
T
− 0.5775.n n

=  (11) Fig. 3 Relation between Gumbel’s reduced standard deviation and


− 1.268
1.2811 · n n data number

123
Arab J Sci Eng (2017) 42:3895–3906 3899

1.30
Table 4 Variation of K factor based on return period for n = 10
1.25 T (year) K (from Table 3 or Eq. 13)
1.20 2 −0.1355
1.15 5 +1.0580
10 +1.8483
Predicted

1.10
20 +2.6063
1.05 50 +3.5874
y = 0,9966x + 0,0039 100 +4.3227
1.00
R² = 0,998 200 +5.0548
0.95 250 +5.2903
0.90 500 +6.0213
1000 +6.7516
0.85
0.85 0.90 0.95 1.00 1.05 1.10 1.15 1.20 1.25 1.30
Measured 5
y = 0.9993x - 0.0022
Fig. 4 Comparison between predicted and measured Gumbel’s R² = 0.9997
reduced standard deviation and data number values for regression model 4

4.0 3
K (Predicted)
y = 0,9514x + 0,134
R² = 0,956
2
3.0

1
K (Predicted)

2.0
0
-1 0 1 2 3 4 5

1.0 K (Measured)
-1

Fig. 6 Comparison between predicted and measured frequency factor


values according to GEP model
0.0
0.0 1.0 2.0 3.0 4.0
K (Measured) The comparison between predicted and measured frequency
factor values for regression model is given Fig. 5.
Fig. 5 Comparison between predicted and measured frequency factor
values according to regression model
3.6 Approach: Estimation of Frequency Factor Using
n = 10 as Minimum Data Number
3.5 Approach: Prediction Using Regression Equation
Derived of Table 2 In this paper, the estimation of frequency factor using n = 10
as minimum data number is suggested as alternative approach
Also, another regression equation derived of Table 1 for due to higher Gumbel’s reduced mean variable and standard
Gumbel distribution was developed in present study. The deviation values opposite to Powell method (Model 3). The
frequency factor can be directly estimated using Eq. (12) variation of K factor based on return period and n = 10
depends on data number and return period. This equation is is given in Table 4. From Table 1 for n = 10, yn and σn
valid for return period’s 5 ≤ T ≤ 100 years with determina- can be selected as 0.4952 and 0.9497, respectively. Thus, the
tion coefficient, R 2 = 0.956. obtained equation can be developed as following:


 y − yn − ln ln (T T−1) − 0.4952
− 2.37 K = = (13)
K = 3.66.n T
for 5 ≤ T ≤ 100, R = 0.956 2
(12) σn 0.9497

123
3900

Table 5 Values of flood frequency factor (K ) for prediction models based on n and T
T Flood frequency factor (K )

123
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
from Eq. (6), from Table 3 (for n = ∝) regression regression (for n = 10) GEP Model
Table 2 from Eq. (8) model from model from (from Table 3) from Eq. (14)
Eq. (11) Eq. (12) (for
5 ≤ T ≤ 100
years)

2 −0.1355 for n = 10 −0.1355 for n = 10 −0.1642 −0.1350 for n = 10 – −0.1355 −0.1195 for n = 10
−0.1478 for n = 20 −0.1478 for n = 20 −0.1477 for n = 20 −0.1609 for n = 20
−0.1526 for n = 30 −0.1526 for n = 30 −0.1522 for n = 30 −0.1721 for n = 30
−0.1552 for n = 40 −0.1552 for n = 40 −0.1548 for n = 40 −0.1770 for n = 40
−0.1568 for n = 50 −0.1568 for n = 50 −0.1564 for n = 50 −0.1798 for n = 50
−0.1604 for n = 100 −0.1604 for n = 100 −0.1600 for n = 100 −0.1847 for n = 100
5 1.0580 for n = 10 1.0580 for n = 10 0.7195 1.0497 for n = 10 1.2288 for n = 10 1.0580 1.0663 for n = 10
0.9187 for n = 20 0.9187 for n = 20 0.9240 for n = 20 0.8847 for n = 20 0.9141 for n = 20
0.8664 for n = 30 0.8664 for n = 30 0.8693 for n = 30 0.7300 for n = 30 0.8617 for n = 30
0.8379 for n = 40 0.8379 for n = 40 0.8397 for n = 40 0.6369 for n = 40 0.8349 for n = 40
0.8197 for n = 50 0.8197 for n = 50 0.8206 for n = 50 0.5730 for n = 50 0.8187 for n = 50
0.7791 for n = 100 0.7791 for n = 100 0.7780 for n = 100 0.4126 for n = 100 0.7857 for n = 100
10 1.8483 for n = 10 1.8483 for n = 10 1.3046 1.8341 for n = 10 2.1207 for n = 10 1.8483 1.8384 for n = 10
1.6248 for n = 20 1.6248 for n = 20 1.6307 for n = 20 1.7994 for n = 20 1.6111 for n = 20
1.5410 for n = 30 1.5410 for n = 30 1.5456 for n = 30 1.6346 for n = 30 1.5309 for n = 30
1.4954 for n = 40 1.4954 for n = 40 1.4981 for n = 40 1.5268 for n = 40 1.4896 for n = 40
1.4663 for n = 50 1.4663 for n = 50 1.4675 for n = 50 1.4482 for n = 50 1.46433 for n = 50
1.4011 for n = 100 1.4011 for n = 100 1.3990 for n = 100 1.2288 for n = 100 1.4127 for n = 100
20 2.6063 for n = 10 2.6063 for n = 10 1.8659 2.5865 for n = 10 2.7860 for n = 10 2.6063 2.5523 for n = 10
2.3020 for n = 20 2.3020 for n = 20 2.3101 for n = 20 2.5663 for n = 20 2.2522 for n = 20
2.1881 for n = 30 2.1881 for n = 30 2.1944 for n = 30 2.4459 for n = 30 2.1454 for n = 30
2.1261 for n = 40 2.1261 for n = 40 2.1297 for n = 40 2.3639 for n = 40 2.0901 for n = 40
2.0865 for n = 50 2.0865 for n = 50 2.0879 for n = 50 2.3023 for n = 50 2.0563 for n = 50
1.9977 for n = 100 1.9977 for n = 100 1.9947 for n = 100 2.1207 for n = 100 1.9867 for n = 100
Arab J Sci Eng (2017) 42:3895–3906
Table 5 continued
T Flood frequency factor (K )

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7


from Eq. (6), from Table 3 (for n = ∝) regression regression (for n = 10) GEP Model
Table 2 from Eq. (8) model from model from (from Table 3) from Eq. (14)
Eq. (11) Eq. (12) (for
5 ≤ T ≤ 100
Arab J Sci Eng (2017) 42:3895–3906

years)

50 3.5874 for n = 10 3.5874 for n = 10 2.5924 3.5604 for n = 10 3.2816 for n = 10 3.5874 3.5051 for n = 10
3.1787 for n = 20 3.1787 for n = 20 3.1895 for n = 20 3.1755 for n = 20 3.1034 for n = 20
3.0257 for n = 30 3.0257 for n = 30 3.0341 for n = 30 3.1151 for n = 30 2.9597 for n = 30
2.9425 for n = 40 2.9425 for n = 40 2.9472 for n = 40 3.0729 for n = 40 2.8852 for n = 40
2.8892 for n = 50 2.8892 for n = 50 2.8911 for n = 50 3.0405 for n = 50 2.8395 for n = 50
2.7700 for n = 100 2.7700 for n = 100 2.7657 for n = 100 2.9423 for n = 100 2.7456 for n = 100
100 4.3227 for n = 10 4.3227 for n = 10 3.1368 4.2902 for n = 10 3.4656 for n = 10 4.3227 4.2941 for n = 10
3.8356 for n = 20 3.8356 for n = 20 3.8485 for n = 20 3.4092 for n = 20 3.8059 for n = 20
3.6534 for n = 30 3.6534 for n = 30 3.6634 for n = 30 3.3766 for n = 30 3.6309 for n = 30
3.5543 for n = 40 3.5543 for n = 40 3.5598 for n = 40 3.3536 for n = 40 3.5402 for n = 40
3.4908 for n = 50 3.4908 for n = 50 3.4929 for n = 50 3.3359 for n = 50 3.4846 for n = 50
3.3487 for n = 100 3.3487 for n = 100 3.3435 for n = 100 3.2816 for n = 100 3.3700 for n = 100
200 5.0548 for n = 10 5.0548 for n = 10 3.6792 5.0173 for n = 10 – 5.0548 5.0548 for n = 10
4.4902 for n = 20 4.4902 for n = 20 4.5051 for n = 20 4.5990 for n = 20
4.2787 for n = 30 4.2787 for n = 30 4.2903 for n = 30 4.3884 for n = 30
4.1638 for n = 40 4.1638 for n = 40 4.1702 for n = 40 4.2791 for n = 40
4.0900 for n = 50 4.0900 for n = 50 4.0906 for n = 50 4.2121 for n = 50
3.9252 for n = 100 3.9252 for n = 100 3.9191 for n = 100 4.0741 for n = 100
250 5.2903 for n = 10 5.2903 for n = 10 3.8536 5.2511 for n = 10 – 5.2903 5.5017 for n = 10
4.0076 for n = 20 4.0076 for n = 20 4.7162 for n = 20 4.8791 for n = 20
4.4797 for n = 30 4.4797 for n = 30 4.4919 for n = 30 4.6557 for n = 30
4.3598 for n = 40 4.3598 for n = 40 4.3664 for n = 40 4.5399 for n = 40
4.2827 for n = 50 4.2827 for n = 50 4.2854 for n = 50 4.4688 for n = 50
4.1106 for n = 100 4.1106 for n = 100 4.1042 for n = 100 4.3225 for n = 100

123
3901
3902

Table 5 continued
T Flood frequency factor (K )

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7

123
from Eq. (6), from Table 3 (for n = ∝) regression regression (for n = 10) GEP Model
Table 2 from Eq. (8) model from model from (from Table 3) from Eq. (14)
Eq. (11) Eq. (12) (for
5 ≤ T ≤ 100
years)

500 6.0213 for n = 10 6.0213 for n = 10 4.3949 6.0000 for n = 10 – 6.0213 6.5870 for n = 10
5.3538 for n = 20 5.3538 for n = 20 5.3714 for n = 20 5.8426 for n = 20
5.1037 for n = 30 5.1037 for n = 30 5.1175 for n = 30 5.5756 for n = 30
4.9680 for n = 40 4.9680 for n = 40 4.9755 for n = 40 5.4370 for n = 40
4.8807 for n = 50 4.8807 for n = 50 4.8837 for n = 50 5.3520 for n = 50
4.6859 for n = 100 4.6859 for n = 100 4.6786 for n = 100 5.1770 for n = 100
1000 6.7516 for n = 10 6.7516 for n = 10 4.9357 6.7017 for n = 10 – 6.7516 7.8619 for n = 10
6.0064 for n = 20 6.0064 for n = 20 6.0261 for n = 20 6.9741 for n = 20
5.7273 for n = 30 5.7273 for n = 30 5.7417 for n = 30 6.6556 for n = 30
5.5757 for n = 40 5.5757 for n = 40 5.5841 for n = 40 6.4903 for n = 40
5.4783 for n = 50 5.4783 for n = 50 5.4817 for n = 50 6.3889 for n = 50
5.2608 for n = 100 5.2608 for n = 100 5.2527 for n = 100 6.1802 for n = 100
Arab J Sci Eng (2017) 42:3895–3906
Arab J Sci Eng (2017) 42:3895–3906 3903

Fig. 7 Study area

3.7 Approach: Estimation of Frequency Factor Using Table 6 Observed annual peak discharges data in river gauge station
GEP Model Sequence number Years Flood discharge (m3 /s)

Gene expression programming (GEP) is a new evolutionary 1 1972 243


artificial intelligence (AI) technique developed by Ferreira 2 1973 46
[10]. GEP is an algorithm based on Genetic Algorithms (GA) 3 1974 104
and Genetic Programming (GP). This algorithm develops a 4 1975 193
computer program encoded in linear chromosomes of fixed 5 1976 334
length. The main aim of GEP is to develop a mathemati- 6 1977 137
cal function using a set of data presented to GEP model. 7 1978 163
For the mathematical equation, the GEP process performs 8 1979 202
the symbolic regression by means of the most of the genetic 9 1980 158
operators of GA. Successful applications of genetic program- 10 1981 202
ming (GP), a branch of soft computing, recorded in water 11 1982 357
resources engineering [11–14] has drawn the hydrologists 12 1983 109
in investigating the use of GP in estimating the river flow 13 1984 304
data. Most recently, Kisi and Shiri [15] forecasted precipi- 14 1985 458
tation using wavelet-genetic programming; Azamathulla et
15 1986 129
al. [16] developed stage-discharge rating curves of Pahang
16 1987 590
River by using GEP. Seckin and Guven [17] estimated peak
17 1988 430
flood discharges at ungauged sites across Turkey with GEP.
18 1989 125
Onen [18] predicted penetration depth in a plunging water
19 1990 210
jet using soft computing approaches. Bagatur and Onen [19]
20 1991 285
studied a predictive model on air entrainment by plunging
21 1992 445
water jets using GEP and ANN.
The powerful soft computing software package GeneX- 22 1993 293
proTools 4.0 [20] was used to develop GEP-based models 23 1994 385
for prediction of frequency factor. After two parameters 24 1995 300
(n, T ) are defined, the model was simulated. In the results of

123
3904 Arab J Sci Eng (2017) 42:3895–3906

Table 6 continued charge prediction. This model can be successfully applied in


Sequence number Years Flood discharge (m3 /s) estimating the flood peak discharges.

25 1996 215
26 1997 235 5 Gumbel Distribution for Observed Flood Series
27 1998 320
28 1999 332 Gumbel distribution is a statistical method often used for
29 2000 64 predicting extreme hydrological events such as floods [21].
30 2001 179 In this study it has been applied for flood frequency analysis
31 2002 251 because (a) peak flow data are homogeneous and independent
32 2003 238 hence lack long-term trends; (b) the river is less regulated,
33 2004 579 hence is not significantly affected by reservoir operations,
34 2005 294 diversions or urbanization; and (c) flow data cover a relatively
35 2006 142 long record (more than 10 years) and is of good quality. The
36 2007 196 equation for fitting the Gumbel distribution to observed series
37 2008 47 of flood flows at different return periods T is given with the
38 2009 159 following mathematical expression:
39 2010 344
Q T = Q̄ + K σ (15)
40 2011 225
Q̄ (mean value) = 250.55 m3 /s,
σ (standard deviaion) = 130.33 where Q T denotes the magnitude of the T -year flood event,
K is the frequency factor, Q̄ and σ are the mean and the
standard deviation of the maximum instantaneous flows,
respectively.
GEP models, Eq. (14) was effectively obtained as following
expression with determination coefficient, R 2 = 0.999:
     6 Illustrative Example
9.97 0.25 2.62 1
K = T − T 0.33 0.88 + 1.1 − +
n T n
This part includes a case study as an illustrative example. The
(14)
study area comprises of an upland watershed and a major
tributary of Kezer River in the Tigris basin, Siirt, Turkey.
The comparison between predicted and measured frequency The province of Kezer river is located in the northeast of
factor values for GEP model is given in Fig. 6. Southeastern Anatolia Region, 41◦ 51 24 —east longitude
and 37◦ 57 42 —north latitude and located on the Sirnak and
Van, east, north and Bitlis, Batman from the west, south,
4 Comparison of Prediction Models surrounded by the provinces of Mardin and Sirnak. Drainage
area of Kezer River is 1069.6 km2 at 530 m elevation (Fig. 7).
The obtained values of flood frequency factor (K ) for predic- The main objective of this case study is to perform flood
tion models are presented based on n and T in Table 5. The frequency analysis of the river catchment using annual peak
flood frequency factor values derived from Powell’s method flow or maximum discharge data obtained in the river in the
(Model 3) and Model 5 consistently gives lower values than water years 1972–2011 (Table 6; Fig. 7). Thus, peak flood
Gumbel’s method (Models 1 and 2). But, the purposed Model magnitudes for the 2, 5, 10, 20, 50, 100, 200, 250, 500 and
6 conversely gives higher values than Gumbel’s method. 1000 years recurrence intervals are estimated using the Gum-
The model 5 is validated from 5 to 100 years. Because, the bel distribution.
regression studies could be only achieved with high determi- Gumbel distribution was analyzed for 2, 5, 10, 20, 50, 100,
nation coefficients for these years. 200, 250, 500 and 1000 years flood of Kezer River, Turkey.
Gumbel’s method can be simplified as in Eq. 11 (Model The peak discharges were estimated according to proposed
4) and 14 (Model 7), and the flood frequency factor using models. Table 7 gives determined and estimated values of
these equations can be predicted without looking up Gum- flood frequency factor (K ) with prediction models (from
bel’s table. The proposed regression model (Model 4) and Table 5). Table 8 presents the model results of predicted dis-
GEP model (Model 7) give a fast and practical way of estimat- charges for return periods. Also, Table 9 shows relative errors
ing the flood frequency factor. The effective model (Model of prediction model results versus Gumbel’s method for peak
4) takes into account n and T as the two parameters for dis- discharges. The regression model (Model 4) gives close val-

123
Arab J Sci Eng (2017) 42:3895–3906 3905

Table 7 Determined and


For n = 40 K
estimated values of flood
frequency factor with prediction T Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
models
2 −0.1552 −0.1552 −0.1642 −0.1548 – −0.1355 −0.1770
5 0.8379 0.8379 0.7195 0.8397 0.6369 1.0580 0.8349
10 1.4954 1.4954 1.3046 1.4981 1.5268 1.8483 1.4896
20 2.1261 2.1261 1.8659 2.1297 2.3639 2.6063 2.0901
50 2.9425 2.9425 2.5924 2.9472 3.0729 3.5874 2.8852
100 3.5543 3.5543 3.1368 3.5598 3.3536 4.3227 3.5402
200 4.1638 4.1638 3.6792 4.1702 – 5.0548 4.2791
250 4.3598 4.3598 3.8536 4.3664 – 5.2903 4.5399
500 4.9680 4.9680 4.3949 4.9755 – 6.0213 5.4370
1000 5.5757 5.5757 4.9357 5.5841 – 6.7516 6.4903

Table 8 Model results of


QT Determined Gumbel Predicted discharge values of alternative models
predicted discharges for return
values from Models 1
periods
and 2 Model 3 Model 4 Model 5 Model 6 Model 7

Q2 230.33 229.15 230.37 – 232.89 227.48


Q5 359.75 344.32 359.98 333.55 388.43 359.36
Q 10 445.44 420.57 445.79 449.53 491.43 444.68
Q 20 527.64 493.73 528.11 558.63 590.22 522.95
Q 50 634.04 588.41 634.65 651.04 718.09 626.57
Q 100 713.78 659.36 714.49 687.62 813.92 711.94
Q 200 793.21 730.06 794.05 707.16 909.35 808.24
Q 250 818.76 752.78 819.62 711.16 940.03 842.23
Q 500 898.02 823.33 899.01 719.29 1035.3 959.15
Q 1000 977.23 893.82 978.32 723.40 1130.5 1096.4

Table 9 Relative errors of (Observed−Calculated)


prediction Models for peak QT Relative error (Re%), Re% = Observed × 100
discharges Model 3 Model 4 Model 5 Model 6 Model 7

Q2 0.51 0.01 – 1.11 1.23


Q5 4.28 0.06 7.28 7.97 0.10
Q 10 5.58 0.07 0.91 10.32 0.17
Q 20 6.42 0.08 5.87 11.86 0.88
Q 50 7.19 0.09 2.68 13.25 1.17
Q 100 7.62 0.10 3.66 14.02 0.25
Q 200 7.96 0.10 10.84 14.64 1.89
Q 250 8.05 0.10 13.14 14.81 2.86
Q 500 8.31 0.11 19.90 15.28 6.80
Q 1000 8.53 0.11 25.97 15.68 12.19

123
3906 Arab J Sci Eng (2017) 42:3895–3906

1400
References
1200 1. Fill, H.D.; Stedinger, J.R.: Homogeneity tests based upon Gumbel
distribution and a critical appraisal of Dalrymple’s test. J. Hydrol.
1000 166, 81–105 (1995)
2. Ahmad, U.N.; Shabri, A.; Zakaria, Z.A.: Flood frequency anal-
ysis of annual maximum stream flows using L-moments and
Q (Predicted)

Model 3
800 TL-moments. Appl. Math. Sci. 5, 243–253 (2011)
Model 4 3. Stedinger, J.R.; Vogel, R.M.: Frequency analysis of extreme events,
600 chapter 18. In: Maidment, D. (ed.) Handbook of Hydrology.
Model 5 McGraw-Hill, New York (1993)
4. Khaliq, M.; Ouarda, T.; Ondo, J.; Gachon, P.; Bobée, B.: Frequency
400 Model 6
analysis of a sequence of dependent and/or non-stationary hydro-
Model 7 meteorological observations: a review. J. Hydrol. 329(3–4), 534–
200 552 (2006)
Best fit 5. Law, G.S.; Tasker G.D.: Flood-frequency prediction methods for
unregulated streams of Tennessee. Water Resources Investigations
0 Report 03-4176. Nashville, Tennessee (2003)
0 200 400 600 800 1000 1200 1400
6. Cunnane, C.: Statistical distributions for flood frequency analy-
Q (Measured) sis. Operational Hydrology Report No. 33. World Meteorological
Organization (1989)
Fig. 8 Comparison between predicted and measured peak discharges 7. Rasmussen, P.F.; Gautam, N.: Alternative PWM-estimators of the
according to Models Gumbel distribution. J. Hydrol. 280, 265–271 (2003)
8. Haktanir, T.: Comparison of various flood frequency distributions
using annual flood peaks data of rivers in Anatolia. J. Hydrol. 136,
1–31 (1992)
ues to Gumbel’s method with maximum 0.11% relative error. 9. Hosking, J.R.M.; Wallis, J.R.: Regional Frequency Analysis. Cam-
Also, Fig. 8 shows the comparison between predicted and bridge University Press, Cambridge (1997)
measured peak discharges according to models. 10. Ferreira, C.: Gene expression programming: a new adaptive algo-
rithm for solving problems. Complex Syst. 13, 87–129 (2001)
11. Rabunal, J.R.; Puertas, J.; Suarez, J.; Rivero, D.: Determination
of the unit hydrograph of a typical urban basin using genetic pro-
gramming and artificial neural networks. Hydrol. Process. 27(4),
7 Conclusions 476–485 (2007)
12. Guven, A.; Aytek, A.; Yuce, M.I.; Aksoy, H.: Genetic
In this paper, the following results can be derived from pre- programming-based empirical model for daily reference evapotran-
spiration estimation. CLEAN Soil Air Water 36(10–11), 905–912
diction studies of flood frequency factor and estimating the (2008)
peak discharges: 13. Guven, A.; Gunal, M.: A genetic programming approach for pre-
diction of local scour downstream hydraulic structures. J. Irrig.
Drain. Eng. 132(4), 241–249 (2008)
1. Powell’s method (Model 3) consistently gives lower val- 14. Guven, A.; Kisi, O.: Estimation of suspended sediment yield in
natural rivers using machine-coded linear genetic programming.
ues than Gumbel’s method. Water Resour. Manag. 25(2), 691–704 (2011)
2. Gumbel’s method can be simplified as in Eqs. 11 and 15. Kisi, O.; Shiri, J.: Precipitation forecasting using wavelet-genetic
14, and the flood frequency factor using these equations programming and wavelet-neuro-fuzzy conjunction models. Water
can be predicted without looking up Gumbel’s table. The Resour. Manag. 25(13), 3135–3152 (2011)
16. Azamathulla, H.M.D.; Ghani, A.A.B.; Leow, C.S.; Chang, C.K.;
proposed regression model (Model 4) and GEP model Zakaria, A.Z.: Gene-expression programming for the development
(Model 7) give a fast and practical way of estimating the of a stage-discharge curve of the Pahang River. Water Resour.
flood frequency factor. Manag. 25(11), 2901–2916 (2011)
3. The effective model (Model 4) takes into account n and 17. Seckin, N.; Guven, A.: Estimation of peak flood discharges at
ungauged sites across Turkey. Water Resour. Manag. 26, 2569–
T as the two parameters for discharge prediction. This 2581 (2012)
model can be successfully applied in estimating the peak 18. Onen, F.: Prediction of penetration depth in a plunging water jet
discharges of flood events. using soft computing approaches. Neural Comput. Appl. 25(1),
4. Gumbel distribution was used in analysis to extract 2, 5, 217–227 (2013)
19. Bagatur, T.; Onen, F.: A predictive model on air entrainment by
10, 20, 50, 100, 200, 250, 500 and 1000 years flood of plunging water jets using GEP and ANN. KSCE J. Civ. Eng. 18(1),
Kezer River, Turkey. The peak discharges were estimated 304–314 (2014)
according to proposed models. The regression model 20. Ferreira, C.: Gene-Expression Programming: Mathematical Mod-
(Model 4) gives close values to Gumbel’s method with eling by an Artificial Intelligence. Springer, Berlin (2006)
21. Haan, C.T.: Statistical Methods in Hydrology. Iowa State Univer-
0.1% relative error. sity Press, Ames (1977)

123

You might also like