You are on page 1of 8

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/317756056

Runoff Projection under Climate Change Conditions with Data-Mining Methods

Article  in  Journal of Irrigation and Drainage Engineering · August 2017


DOI: 10.1061/(ASCE)IR.1943-4774.0001205

CITATIONS READS
6 338

4 authors:

Parisa Sarzaeim Omid Bozorg-Haddad


University of Nebraska at Lincoln University of Tehran
18 PUBLICATIONS   42 CITATIONS    362 PUBLICATIONS   5,702 CITATIONS   

SEE PROFILE SEE PROFILE

Atiyeh Bozorgi Hugo Loaiciga


University of Tehran University of California, Santa Barbara
9 PUBLICATIONS   13 CITATIONS    298 PUBLICATIONS   4,482 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

From Gene To Global Hydroclimatic Controls On Hybrids Performance Predictability View project

Hydropolitics, Water Conflicts and War, Water Power and Hydro-Hegemony, Geopolitics View project

All content following this page was uploaded by Hugo Loaiciga on 03 May 2018.

The user has requested enhancement of the downloaded file.


Runoff Projection under Climate Change
Conditions with Data-Mining Methods
Parisa Sarzaeim 1; Omid Bozorg-Haddad 2; Atiyeh Bozorgi 3; and Hugo A. Loáiciga, F.ASCE 4
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

Abstract: This work proposes data-mining algorithms for runoff projection under climate change conditions. Specifically, genetic program-
ming (GP), artificial neural network (ANN), and support vector machine (SVM) data-mining tools are applied for runoff projection
and their predictive skills are compared by means of several standard indicators of models’ performance. The approach herein implemented
predicts future regional precipitation and temperature with the Hadley Centre Coupled Atmosphere-Ocean General Circulation Model
version 3 (HadCM3) atmosphere-ocean general circulation model (AOGCM) followed by runoff prediction with GP, ANN, and SVM
in the Aidoghmoush Basin, Iran. This paper’s results demonstrate that SVM outperforms GP and ANN by 7 and 5%, respectively.
DOI: 10.1061/(ASCE)IR.1943-4774.0001205. © 2017 American Society of Civil Engineers.
Author keywords: Climate change; Genetic programming; Artificial neural network; Support vector machine; Runoff projection.

Introduction Wagesho et al. (2013) reported on the potential effects of climate


change on runoff in two basins in Ethiopia. Precipitation and tem-
The effects of climate change on runoff have been the subject of perature parameters were produced by statistical downscaling and
various studies, such as runoff prediction under different climate then the runoff was simulated with the soil and water assessment
change scenarios (Xie and Eheart 2003), the study of the current tool (SWAT). Their study also showed that the duration of the dry
status and future conditions of water resources in California under periods increased in dry months and remained stable in wet months.
climate change conditions (Hancock et al. 2004), hydrological Hong et al. (2013) developed a methodology for evaluating the po-
modeling to assess the effects of climate change on agricultural tential impacts of climate change on hydrological parameters and
and municipal water sectors in the Conchos Basin in Brazil water resources in the Han River Basin located in South Korea.
(Ingol-Blanco and McKinney 2009), and runoff simulation under Their investigation showed that average low flow would increase,
climate change conditions (Ashofteh et al. 2013). while average wet flow and normal flow would decrease. Zhang
Mimikou et al. (2001) investigated the effects of climate change et al. (2013) considered climate change as an environmental stress
on water quantity and quality in the Thessaly region in Greece. that will affect the management of water resources in the Huaihe
Their results indicated that average annual runoff would decrease Basin, China.
under climate change conditions. Sidhu et al. (2011) investigated There are various types of rainfall-runoff models. A conceptual
climate change effects on water resources, energy consumption, model such as the Hydrology Engineering Center’s Hydrologic
and agriculture production in Punjab, India. Their results showed Modeling System (HEC-HMS) estimates runoff from watershed
a decrease in precipitation and increase in temperature, which lead and a wide range of climatic and hydrologic data. Anderson et al.
to a fall in agriculture production. Nazif et al. (2012) researched (2002) predicted runoff in Nevada and California with HEC-HMS
climate change effects on runoff from snowmelt in northwestern for flood management. Another conceptual model is the identifica-
Tehran, Iran. Two greenhouse gas emissions scenarios (A2 and tion of unit hydrographs and component flows from rainfall, evapo-
B2) were employed, and after downscaling temperature was pro- ration, and streamflow (IHACRES), which has been used by many
jected to rise, whereas no trend in precipitation was discerned. authors for runoff estimation based on precipitation and tempera-
ture data. Dye and Croke (2003) applied IHACRES to runoff pre-
1
M.Sc. Student, Dept. of Irrigation and Reclamation Engineering, diction in two basins in South Africa.
Faculty of Agriculture Engineering and Technology, College of Agriculture Genetic programming (GP), artificial neural network (ANN),
and Natural Resources, Univ. of Tehran, Karaj, 14697-13491 Alborz, Iran. and support vector machine (SVM) are data-mining models whose
E-mail: Parisa.Sarzaeim@ut.ac.ir
2 effectiveness has been tested in several studies. Genetic program-
Professor, Dept. of Irrigation and Reclamation Engineering, Faculty
of Agriculture Engineering and Technology, College of Agriculture and ming applications in water resources include modeling of rainfall-
Natural Resources, Univ. of Tehran, Karaj, 14697-13491 Alborz, Iran runoff processes (Whigham and Crapper 2001), calculation of
(corresponding author). E-mail: OBHaddad@ut.ac.ir the optimal operation rules in an aquifer-dam system (Fallah-
3
M.Sc. Student, Dept. of Irrigation and Reclamation Engineering, Mehdipour et al. 2013), determination of irrigation allocation
Faculty of Agriculture Engineering and Technology, College of Agriculture policy under climate change (Ashofteh et al. 2014), and modeling
and Natural Resources, Univ. of Tehran, Karaj, 14697-13491 Alborz, Iran. of groundwater (Fallah-Mehdipour et al. 2014). Artificial neural
E-mail: Bozorgi.Atiyeh@ut.ac.ir network has been implemented for forecasting of streamflow of
4
Professor, Dept. of Geography, Univ. of California, Santa Barbara, the Apalachicola River (Huang 2001), rainfall-runoff modeling
CA 93106. E-mail: Hugo.Loaiciga@ucsb.edu
Note. This manuscript was submitted on October 21, 2016; approved on
in Portugal (Fernandes and Haie 2001), runoff simulation in the
February 27, 2017; published online on May 20, 2017. Discussion period Manitoba region, Canada (Ahmad and Simonovic 2001), predic-
open until October 20, 2017; separate discussions must be submitted for tion of water availability for basin management (Khalil et al.
individual papers. This paper is part of the Journal of Irrigation and Drai- 2005), modeling of the relation between runoff and large-scale
nage Engineering, © ASCE, ISSN 0733-9437. atmospheric processes (Maity and Kumar 2006), forecasting of

© ASCE 04017026-1 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


contaminants in distribution systems (Romero-Gomez et al. 2007),
prediction of quality parameters (Kalin and Isik 2010), the optimal
design of water distribution systems (Andrade et al. 2013), and
solar radiation estimation (Citakoglu 2015). Support vector ma-
chine applications in water resources include identification of op-
timal well locations to monitor aquifers (Asefa et al. 2004), spring
and fall flow forecasting in rivers in the United States (She and
Basketfield 2005), long-term prediction of lake water level (Khan
and Coulibaly 2006), statistical downscaling in tropical regions of
India (Tripathi et al. 2006), and prediction of monthly reservoir
storage by genetic algorithm (GA)–based SVM (Su et al. 2014).
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

Makkeasorn et al. (2008) employed GP and ANN to project


runoff under climate change conditions and compared their effi-
ciencies. Behzad et al. (2009) predicted daily Bakhtiyari River flow
in Iran with SVM, and compared SVM results with those of ANN
and ANN-GA. Their investigation found better accuracy of
SVM than those of ANN and ANN-GA. The comparison of GP,
ANN, and SVM for runoff projection under climate change con-
Fig. 1. Mathematical relation produced by GP
ditions has been rarely addressed in past studies. This paper applies
GP, ANN and SVM for runoff projection under climate change
conditions and reports a comparison of their predictive accuracies.
Atmosphere-ocean general circulation models (AOGCMs) are ap- and testing. After training and identifying the relation between in-
plied to project hydrologic variables such as precipitation and tem- puts and outputs, ANN predicts the corresponding output for a new
perature under climate change. These models provide large-scale input based on the mathematical function identified in the training
outputs that must be downscaled for regional applications, runoff step. Artificial neural networks are generally nonlinear mathemati-
prediction being a case in point. This is followed by implementa- cal learning algorithms that attempt to imitate some of the workings
tion of the GP, ANN, and SVM data-mining methods to project of the human brain. A schematic of the ANN algorithm is portrayed
runoff under climate change conditions. Monthly precipitation in Fig. 2. It is seen in Fig. 2 that inputs are employed in the ANN to
and temperature are inputs to the data-mining methods with which make predictions that minimize simulation errors. The ANN pre-
to project runoff. dictions are passed through a nonlinear transfer function and even-
tually the final ANN outputs are achieved. Several algorithms exist
for network training and calculating the predictions weights and
Genetic Programming bias. One of them is the Levenberg-Marquardt (LM) algorithm
Genetic programming is an evolutionary algorithm whose princi- (Marquardt 1963), which minimizes the sum of the squared errors
ples are based on the GA (Aboutalebi et al. 2016) and was proposed of predictions. The ANNs are particularly adept for solving prob-
by Cramer (1985). Genetic programming relies on tree-structured lems that do not exhibit clear mathematical relation between inputs
search for solutions of various types of problems. The tree structure and outputs.
of chromosomes in GP renders its search algorithm robust for solv- This paper predicts runoff under climate change conditions by
ing problems. Numerical and mathematical or logical operators means of an ANN model that has two hidden layers, each of which
are considered as decision variables in GP whose goal is finding has two neurons, and is programmed in MATLAB software. The
a mathematical relation between inputs and output(s). Fig. 1 rep- ANN’s structure is achieved by sensitivity analysis.
resents a mathematical relational function produced by GP.
Genetic programming first defines the set of variables on oper-
ators employed in the search algorithm. These sets are named ter- Support Vector Machine
minal sets (T) that contain variables, and function sets (F) that
contain operators. Next, GP produces trees by selecting a set of Support vector machine is a data-mining method by Vapnik (1995)
basic solutions of terminal and function sets. Trees in GP are the that is based on statistical learning theory (Kashif Gill et al. 2007).
same as chromosomes in the GA. In each iteration of GP, the ob- Support vector machine performs classification and regression
jective function is calculated for each tree. Penalties are assigned to (Aboutalebi et al. 2015). Guyon et al. (2002) applied SVM as a
the objective function to avoid answers that would violate con- classification technique. The regression form of SVM is known
straints. Crossover and mutation are implemented in GP to improve
the quality of intermediate solutions iteratively until fulfilling a ter-
minal criterion.
The GP model that is used in this paper is programmed in the
MATLAB software.

Artificial Neural Network


An ANN is a data-mining algorithm that identifies relations be-
tween input and output variables employing a learning process.
These relations can be complex, yet the ANN processes input
and their corresponding outputs to identify mathematical relations
Fig. 2. Schematic of ANN’s operations
between them. An ANN includes three steps: training, validation,

© ASCE 04017026-2 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


tool and includes data analysis, learning machine, and statistical
training techniques. The kernel function that used in this study
is the radial basis function (RBF), which has a parameter named
γ. The amounts of γ and ζ are achieved by trial and error in order
to find the optimal fitting.

Runoff Projection
This paper’s objective is to estimate runoff under climate change
conditions. Monthly precipitation and temperature inputs for runoff
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

simulation in the period 2026–2039 were generated by statistical


downscaling of the precipitation and temperature projection by
the Hadley Centre Coupled Atmosphere-Ocean General Circula-
tion Model version 3 (HadCM3), driven by the A2 greenhouse
gas (GHG) emissions scenario.
The next step of model development is the creation of the data-
base. Precipitation and temperature are used for estimating runoff,
Fig. 3. Geometric illustration of the first function, which determines and these data are scaled as follows:
the error [variables are presented in Eqs. (4) and (5)]
x − xmin
xscaled ¼ ð6Þ
xmax − xmin
as support vector regression (SVR). Support vector machine is a
where xscaled = scaled data; x = actual data value before scaling; and
tool to similar ANN, but a major difference with ANN is that it
xmin and xmax = minimum and maximum data values before scaling,
does not converge to local optima. SVR is defined by two func-
respectively.
tions. The first function is the error function [Eq. (1)] and the sec-
Selection of suitable inputs to create the model structure is im-
ond function calculates the model output from weight values, input,
portant (Kashif Gill et al. 2007). The correlation criterion is applied
and bias [Eq. (2)]
as a statistical measure to choose the input parameters. The corre-
 lation between runoff at time t and precipitation and temperature
0 ifjy − fðxÞj ≤ ε
jy − fðxÞj ¼ ð1Þ with various lag times were analyzed and the parameters related
jy − fðxÞj − ε ¼ ζ otherwise to the lag times that have meaningful correlation were detected.
No statistically meaningful correlation was detected unless be-
fðxÞ ¼ wT · x þ b ð2Þ
tween the precipitation and temperature at time t and t − 1 and
where y = actual output value; fðxÞ = output value calculated by the the observed runoff. Thus the projected runoff at time t is a function
model; ε = sensitivity function; ξ = penalty; w = vector of weights of the precipitation and temperature at times t and t − 1
applied to input x; b = amount of bias of wT · x from the actual Qt ¼ fðPt ; T t ; Pt−1 ; T t−1 Þ ð7Þ
values; and T = transpose operator.
Fig. 3 shows that the first function does not apply a penalty to where Qt = predicted runoff at time t; Pt = precipitation at time t;
the predicted output values if the difference between actual and cal- T t = temperature at time t; Pt−1 = precipitation at time t − 1; T t−1 =
culated values are in the range ðþε; −εÞ. Otherwise a penalty ξ is temperature at time t − 1; f = function that produces output from
applied. input; and t = index counter in month.
SVR models have as objective function the minimization of pre- Then GP, ANN, and SVM are trained with the input data. The
diction errors to increase accuracy and reduce the weight vector w runoff is projected by the IHACRES model. A total of 75% of the
to simplify the model. Therefore, SVR is posed as an optimization data are randomly selected and used for training of the three data-
model according to Eq. (3) mining models. The key components of the predictive approach
Xm implemented in this work are presented in Fig. 4.
1 þ
min kwk2 þ C ðξ −
i − ζi Þ ð3Þ The remaining 25% of the data are used to test the GP, ANN,
2 i¼1 and SVM. The coefficient of determination ðR2 Þ, root-mean-square
subject to error (RMSE), mean absolute error (MAE), and overall index of
model performance (OI) are used as the standard indicators of mod-
ðwT · x þ bÞ − yi < ε þ ζ þ
i i ¼ 1; 2; 3; : : : ; m ð4Þ els’ performance (Bayram and Al-Jibouri 2016). The closer the

yi − ðwT · x þ bÞ ≤ ε þ ζ −
i i ¼ 1; 2; 3; : : : ; m ð5Þ

where C = penalty coefficient; m = number of input data to the


model in the training step; ξ þ −
i and ξ i = penalties for the upper
and lower bounds (þε; −ε); yi = real value of the ith data; and
k k = vectoral magnitude. The optimization problem Eqs. (3)–(5)
is solved with the Lagrange method, the values of w and b are cal-
culated and substituted in Eq. (2) to calculate the SVR model’s out-
put. SVR can also simulate nonlinear data. In such cases, it applies
a transfer (kernel) function that linearizes the nonlinear data.
Runoff projection with the SVM data-mining model was
Fig. 4. Schematic of simulation method by GP, ANN, and SVM
performed with the Tanagra software, which is a data-mining

© ASCE 04017026-3 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


50 values of R2 and OI are to 1 and the closer the values of RMSE and
45 MAE are to 0, the better the model performance, which are pre-

Precipitation (mm)
40 sented in Eqs. (8)–(10)
35 rP ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T ðQ 2
t¼1 obs;t − Qsim;t Þ
30 RMSE ¼ ð8Þ
T
25
20
1X T
15 MAE ¼ jQ − Qsim;t j ð9Þ
10
n t¼1 obs;t

2031
2032
2033
2026
2027
2028
2029
2030

2034
2035
2036
2037
2038
2039
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

 PT 2

Time (year) 1 RMSE t¼1 ðQobs;t − Qsim;t Þ
OI ¼ 2 − − PT 2
ð10Þ
2 Qobs;max − Qobs;min t¼1 ðQobs;t − Q̄obs Þ
Fig. 5. Average annual precipitation under climate change in the
Aydoghmoush River Basin (data from Ashofteh et al. 2013) where Qobs;t = scaled observed runoff in month t; Qsim;t = scaled
simulated runoff in month t; T = number of total months; Qobs;max
and Qobs;min = maximum and minimum of scaled observed runoff,
17 respectively; and Q̄obs = average of scaled observed runoff.

16
Temperature (˚C)

15 Case Study
14 The Aidoghmoush River Basin is located in eastern Azarbaijan
13 province in Iran. It covers an area of approximately 1,802 km2
and its height is between 1,060 and 2,500 m. Most of the area
12
is semiarid. Average annual precipitation and temperature are
11 336.2 mm and 11.6°C, respectively. The maximum temperature
10 is 31.9°C in July and the minimum temperature is −16.8°C in
2027

2030

2034

2038
2026

2028
2029

2031
2032
2033

2035
2036
2037

2039

February. Ten meteorological stations provided historical precipi-


tation and temperature variables in the basin for the baseline period
Time (year) 1986–2000. The HadCM3 general circulation model driven by the
A2 GHG emissions scenario, using the change factor downscaling
Fig. 6. Average annual temperature under climate change in
method, produced the monthly precipitation and temperature in the
Aydoghmoush River Basin (data from Ashofteh et al. 2013)
14-year period of 2026–2039 in the study region. The average

1.0 1
Scaled simulated runoff

y = 0.7733 x + 0.0354 0.8 y = 0.8382 x + 0.0293


Scaled simulated runoff

0.8
R² = 0.7864 R² = 0.8040
0.6 0.6

0.4 0.4

0.2 0.2

0.0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Scaled observed runoff (b) Scaled observed runoff

1
Scaled simulated runoff

0.8 y = 0.8147 x + 0.0216


R² = 0.8487
0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
(c) Scaled observed runoff

Fig. 7. Runoff prediction diagram for (a) GP; (b) ANN; (c) SVM

© ASCE 04017026-4 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


annual precipitation and temperature under climate change condi- Results and Discussion
tions are shown in Figs. 5 and 6 (Ashofteh et al. 2013). As shown in
Fig. 6, there is an increasing trend in average annual temperature Recall that 75% of data were used for training and 25% were used
under climate change conditions. for testing the runoff-prediction model. Variables Pt , T t , Pt−1 , and
Average annual runoff in the Aidoghmoush River Basin is T t−1 were inputs and the runoff Qt was the projected variable
approximately 170 × 106 m3 and the minimum and maximum [Eq. (7)].
baseline period is 23 and 342 × 106 m3 , respectively. Maximum It is seen in Fig. 7 that the R2 values for runoff projections in the
monthly average runoff is approximately 55 × 106 m3, which testing step equaled 0.7864, 0.8040, and 0.8487 with the GP, ANN,
occurs in April and May. and SVM data-mining models, respectively. As was previously
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

Table 1. Performance Results for Runoff Prediction with Data-Mining Algorithms


Standard indicators of models’ performance
R2 RMSE MAE OI
Model Training Testing Training Testing Training Testing Training Testing
GP 0.8025 0.7864 0.1033 0.0979 0.0640 0.0712 0.8468 0.8260
ANN 0.8907 0.8040 0.0759 0.0935 0.0470 0.0631 0.9073 0.8447
SVM 0.9006 0.8487 0.0887 0.0838 0.0398 0.0580 0.9118 0.8643

70
Observed
60
Simulated
50
Runoff (m 3 /s)

40

30

20

10

0
0 20 40 60 80 100 120 140 160
(a) Time (month)

70
Observed
60
Simulated
50
Runoff (m 3 /s)

40

30

20

10

0
0 20 40 60 80 100 120 140 160
(b) Time (month)

70
Observed
60
Simulated
50
Runoff (m 3 /s)

40

30

20

10

0
0 20 40 60 80 100 120 140 160
(c) Time (month)

Fig. 8. Observed and simulated runoff with (a) GP; (b) ANN; (c) SVM

© ASCE 04017026-5 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


mentioned, the closer the R2 value is to 1, the more accurate is the Ahmad, S., and Simonovic, S. (2001). “Developing runoff hydrograph us-
model in simulation. ing artificial neural networks.” Proc., World Water and Environmental
Table 1 lists the runoff- projections results from ANN, GP, and Resources Congress, ASCE, Reston, VA.
SVM in the training and testing steps. The R2 , RMSE, MAE, and Anderson, M. L., Chen, Z., Kavvas, M., and Feldman, A. (2002).
“Coupling HEC-HMS with atmospheric models for prediction of water-
OI were used to compare the models’ performances. According to
shed runoff.” J. Hydrol. Eng., 10.1061/(ASCE)1084-0699(2002)7:
Table 1, all of the standard indicators of models’ performance in- 4(312), 312–318.
dicate that SVM’s performance has better fitness to observed runoff Andrade, M., Choi, C., Mondaca, M., Lansey, K., and Kang, D. (2013).
in general. On the other hand, ANN can simulate observed runoff “Enhancing artificial neural networks applied to the optimal design
better than GP. of water distribution systems.” World Environmental and Water
The outputs from the data-mining models in the training and Resources Congress, ASCE, Reston, VA.
testing steps were descaled to their actual range of values for Asefa, T., Kemblowski, M. W., Urroz, G., McKee, M., and Khalil, A.
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

2026–2039. (2004). “Support vectors-based groundwater head observation networks


It is clear from Fig. 8(a), that the GP provides poor runoff pro- design.” J. Water Resour. Res., 40(11), W11509.
jections in several months. The ANN model produced better results Ashofteh, P. S., Bozorg-Haddad, O., Akbari-Alashti, H., and Mariño, M. A.
than the GP, especially for peak runoff values [Fig. 8(b)]. However, (2014). “Determination of irrigation allocation policy under climate
change by genetic programming.” J. Irrig. Drain. Eng., 10.1061
SVM exhibited better performance than GP and ANN in most
/(ASCE)IR.1943-4774.0000807, 04014059.
months. The graphs in Fig. 8 clearly indicate that SVM has the best
Ashofteh, P. S., Bozorg-Haddad, O., and Mariño, M. A. (2013). “Scenario
performance in estimating runoff among the three data-mining assessment of streamflow simulation and its transition probability in
models. The values of R2 , RMSE, MAE, and OI listed in Table 1 future periods under climate change.” J. Water Res. Manage., 27(1),
prove this statement. 255–274.
According to Fig. 8, in the period of time that starts in the 140th Bayram, S., and Al-Jibouri, S. (2016). “Efficacy of estimation methods in
month to the end, it is clear that the simulated runoff produced by forecasting building projects’ costs.” J. Constr. Eng. Manage., 10.1061
the ANN and SVM models is very accurate and there is small error /(ASCE)CO.1943-7862.0001183, 05016012.
between simulated and observed runoff. Accurate prediction also Behzad, M., Asghari, K., Eazi, M., and Palhang, M. (2009). “Generaliza-
occurs under low-flow conditions. It seems these results indicate tion performance of support vector machines and neural networks in
that data-mining methods have high acceptable performance in pre- runoff modeling.” Expert Syst. Appl., 36(4), 7624–7629.
diction of low flows. Citakoglu, H. (2015). “Comparison of artificial intelligence techniques
via empirical equations for prediction of solar radiation.” J. Comput.
The three data-mining models indicate that annual average run-
Electron. Agric., 118, 28–37.
off in the Aidoughmoush River Basin is declining. In other words, Cramer, N. L. (1985). “A representation for the adaptive generation of sim-
it seems there is a decreasing trend in runoff in the future period of ple sequential programs.” Proc., Int. Conf. on Genetic Algorithms and
time 2026–2039 under climate change conditions. The data mining the Applications, John J. Grefenstette, ed., Carneige-Mellon Univ.,
models applied in this paper have identified a runoff trend. Pittsburgh, 183–187.
Dye, P. J., and Croke, B. F. W. (2003). “Evaluation of streamflow predic-
tions by the IHACRES rainfall-runoff model in two South African
Concluding Remarks catchments.” J. Modell. Hydrol. Syst., 18(8–9), 705–712.
Fallah-Mehdipour, E., Bozorg-Haddad, O., and Mariño, M. A. (2013). “Ex-
Genetic programming, ANN, and SVM were employed to project traction of optimal operation rules in an aquifer-dam system: Genetic
runoff from precipitation and temperature under climate change programming approach.” J. Irrig. Drain. Eng., 10.1061/(ASCE)IR
.1943-4774.0000628, 872–879.
conditions. This paper’s results established that the SVM had
Fallah-Mehdipour, E., Bozorg-Haddad, O., and Mariño, M. A. (2014).
approximately 5 and 7% better performance than ANN and GP,
“Genetic programming in groundwater modeling.” J. Hydrol. Eng., 10
respectively, in projection runoff in the Aidoghmoush River Basin, .1061/(ASCE)HE.1943-5584.0000987, 04014031.
Iran. The ANN achieved approximately 2% better performance Fernandes, L., and Haie, N. (2001). “Neural networks in water resources
than GP. Support vector machine produced better results than management.” Proc., World Water and Environmental Resources
ANN and GP in the training step, and ANN had better performance Congress, ASCE, Reston, VA.
than GP in the training step as well. Moreover, this work’s results Guyon, L., Weston, J., Barnhill, S., and Vapnik, V. (2002). “Gene selection
indicate that the SVM achieved R2 , RMSE, MAE, and OI values for cancer classification using support vector machine.” J. Mach.
equal to approximately 0.84, 0.08, 0.06, and 0.86, respectively, Learn., 46(1), 389–422.
demonstrating its accuracy in predicting runoff under climate Hancock, K., Chung, C., and Mills, W. (2004). “Climate change and its
change conditions that outperforms those of other leading data- effects on California water resources.” Proc., Water and Environmental
mining algorithms, especially in a basin for which there are not Resources Management, ASCE, Reston, VA.
Hong, S., Jung, S., Kim, B., and Kim, H. (2013). “The impact assessment
enough hydrological and climate data. Sensitivity analysis was
of climate change on the long-term runoff on the Han river in south
shown to be helpful in identifying the best structure for data-mining Korea based on RCP climate change scenarios.” Proc., World Environ-
methods for runoff prediction. mental and Water Resources Congress, ASCE, Reston, VA.
Huang, W. (2001). “Neural network method in real-time forecasting
of Apalachicola river flow.” Proc., World Water and Environmental
References Resources Congress, ASCE, Reston, VA.
Ingol-Blanco, E., and McKinney, D. (2009). “Hydrologic modeling
Aboutalebi, M., Bozorg-Haddad, O., and Loáiciga, H. A. (2015). “Optimal for assessing climate change impacts on the water resources of the
monthly reservoir operation rules for hydropower generation derived Rio Conchos basin.” Proc., World Environmental and Water Resources
with SVR-NSGA II.” J. Water Resour. Plann. Manage., 10.1061 Congress, ASCE, Reston, VA.
/(ASCE)WR.1943-5452.0000553, 04015029. Kalin, L., and Isik, S. (2010). “Prediction of water quality parameters using
Aboutalebi, M., Bozorg-Haddad, O., and Loáiciga, H. A. (2016). “Simu- an artificial neural network model.” Proc., World Environmental and
lation of methyl tertiary butyl ether concentration in river-reservoir sys- Water Resources Congress, ASCE, Reston, VA.
tems using support vector regression.” J. Irrig. Drain. Eng., 10.1061 Kashif Gill, M., Asefa, T., Kaheil, Y., and Mckee, M. (2007). “Effects of
/(ASCE)IR.1943-4774.0001007, 04016015. missing data on performance of learning algorithms for hydrologic

© ASCE 04017026-6 J. Irrig. Drain Eng.

J. Irrig. Drain Eng., 2017, 143(8): 04017026


predictions: Implications to an imputation technique.” J. Water Resour. She, N., and Basketfield, D. (2005). “Long range forecast of streamflow
Res., 43(7), W07416. using support vector machine.” Proc., World Environmental and Water
Khalil, A., McKee, M., Kemblowski, M., and Asefa, T. (2005). “Basin Resources Congress, ASCE, Reston, VA.
scale water management and forecasting using artificial neural net- Sidhu, R. S., Vatta, K., and Lall, U. (2011). “Climate change impact and
works.” J. Am. Water Res. Assoc., 41(1), 195–208. management strategies for sustainable water-energy-agriculture out-
Khan, M., and Coulibaly, P. (2006). “Application of support vector machine comes in Punjab.” Indian J. Agric. Econ., 66(3), 328–339.
in lake water level prediction.” J. Hydrol. Eng., 10.1061/(ASCE)1084- Su, J., Wang, X., Liang, Y., and Chen, B. (2014). “GA-based support vector
0699(2006)11:3(199), 199–205. machine model for prediction of monthly reservoir storage.” J. Hydrol.
Maity, R., and Kumar, D. (2006). “Artificial neural network approach for Eng., 10.1061/(ASCE)HE.1943-5584.0000915, 1430–1437.
streamflow forecasting in India using ENSO and EQUINOO.” Proc., Tanagra 1.4 [Computer software]. Ricco Rakotomalala, France.
World Environmental and Water Resources Congress, ASCE, Reston, Tripathi, S., Srinivas, V. V., and Nanjundiah, R. S. (2006). “Downscaling of
VA. precipitation for climate change scenarios: A support vector machine
Downloaded from ascelibrary.org by "University of California, Santa Barbara" on 05/03/18. Copyright ASCE. For personal use only; all rights reserved.

Makkeasorn, A., Chang, N. B., and Zhou, X. (2008). “Short-term stream-


approach.” J. Hydrol., 330(3–4), 621–640.
flow forecasting with global climate change implications—A compar-
Vapnik, V. (1995). The nature of statistical learning theory, Springer,
ative study between genetic programming and neural network models.”
New York.
J. Hydrol., 352(3–4), 336–354.
Wagesho, N., Jain, M., and Goel, N. (2013). “Effect of climate change
Marquardt, D. W. (1963). “An algorithm for least-squares estimation of
nonlinear parameters.” J. Soc. Ind. Appl. Mathe., 11(2), 431–441. on runoff generation: Application to Rift Valley Lakes basin of
MATLAB [Computer software]. MathWorks, Natick, MA. Ethiopia.” J. Hydrol. Eng., 10.1061/(ASCE)HE.1943-5584.0000647,
Mimikou, M., Baltas, E., and Varanou, E. (2001). “Climate change impacts 1048–1063.
on water resources: Quantity and quality aspects.” Proc., World Envi- Whigham, P. A., and Crapper, P. E. (2001). “Modelling rainfall-runoff us-
ronmental and Water Resources Congress, ASCE, Reston, VA. ing genetic programming.” Math. Comput. Modell., 33(6–7), 707–721.
Nazif, S., Karamouz, M., and Zahmatkesh, Z. (2012). “Climate change im- Xie, H., and Eheart, J. (2003). “Assessing vulnerability of water resources
pacts on runoff evaluation: A case study.” Proc., World Environmental to climate change in Midwest.” Proc., World Water and Environmental
and Water Resources Congress, ASCE, Reston, VA. Resources Congress, ASCE, Reston, VA.
Romero-Gomez, P., Austin, R. G., and Choi, C. Y. (2007). “Prediction of Zhang, J., et al. (2013). “Using hydrologic simulation to explore the
contaminants in water distribution systems using artificial neural net- impacts of climate change on runoff in the Huaihe River Basin of
works.” Proc., World Environmental and Water Resources Congress, China.” J. Hydrol. Eng., 10.1061/(ASCE)HE.1943-5584.0000581,
ASCE, Reston, VA. 1393–1399.

© ASCE 04017026-7 J. Irrig. Drain Eng.

View publication stats J. Irrig. Drain Eng., 2017, 143(8): 04017026

You might also like