SUGENO FUZZY INFERENCE SYSTEM FOR RIVER FLOW ESTIMATION

Mahmut FIRAT
Research Asist. Pamukkale University, Civil Engineering Department, Denizli, Trkiye, mfirat@pamukkale.edu.tr

Mahmud GNGR
Assist. Prof. Dr. Pamukkale University, Civil Engineering Department, Denizli, Trkiye, mgungor@pamukkale.edu.tr

ABSTRACT

River flow estimation is quite important for reservoir operation studies, flood planning and control, modeling and management of water resources. In this study, applicability and capability of Sugeno Fuzzy Inference System (ANFIS) is investigated for river flow forecasting. To illustrate the applicability and capability of the ANFIS, the Seyhan River, located the south of Turkey and the most important water resource of Seyhan Catchments, is chosen as a case study area. Totally 2192 daily data sets collected in 1986-1992 years are used for river flow forecasting. The models having various input structures are constructed and the best structure is investigated. In addition four various training/testing data sets are built by cross validation methods and the best data set is obtained. The performance of the ANFIS models in training and testing sets are compared with the observations and also evaluated. In order to get the accurate and reliable comparison, the best fit model structure is also trained and tested by Artificial Neural Networks (ANN). The results of ANFIS and ANN models are compared and evaluated. The results indicated that the ANFIS can be applied successfully and provide high accuracy and reliability for time series forecasting Keywords: River Flow Estimation, Seyhan River, ANN, Fuzzy Logic
1. INTRODUCTION

In last decades, the forecasting of river flow in hydrological processes is quite important to provide the accurate and sustainable use, planning and management of

584

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

the water resources. In order to estimate the hydrological processes such as precipitation, runoff and change of water level by using existing methods, some parameters such as the physical properties of the watershed region and river network and observed detail data are necessary. In the literature, there have conventional methods of the river flow forecasting and modeling. In the last year, the new approaches, Artificial Neural Networks (ANN) and Fuzzy Logic, are used modeling of hydrological process. ANN have been recently accepted as an efficient alternative tool for modeling of complex hydrologic systems and widely used for prediction. Some specific applications of ANN to hydrology include modeling rainfall-runoff process [Sajikumar et. al., 1999],, rainfall forecasting [Bodri ve di., 2000], river flow forecasting [Hsu ve di., 1998; Dibike ve di., 2001], hydrologic time series modeling [Jain ve di., 2006], sediment transport prediction [Firat et.al., 2004], sediment concentration estimation [Nagy et.al., 2002], and reservoir operation [Jain et.al., 1999]. Moreover, The ASCE Task Committee reports (2000) did a comprehensive review of the applications of ANN in the hydrological forecasting context. Fuzzy logic method was first developed to explain the human thinking and decision system by Zadeh (1965). Several studies have been carried out using fuzzy logic in hydrology and water resources planning [Chang et. al., 2001; Ertunga et. al., 2001; Liong et. al., 2000; Mahabir et. al., 2000; Mitra et. al., 1998; Nayak et. al., 2004a]. Recently, Adaptive Neuro-fuzzy inference system (ANFIS), which consists of the ANN and fuzzy logic methods, have been used for many application such as, database management, system design and planning/forecasting of the water resources [Chen et. al., 2006; Chang et. al., 2006; Chang et. al., 2001; Da Silva et. al., 1999; Nayak et. al., 2004b; Firat et. al., 2006]. The main purpose of this study is to present a novel ANFIS to the river flow forecasting. To verify the application of this approach, the Seyhan River catchments location in the south part of Turkey is chosen as the case study area. Seyhan Catchments is one of the most important water resources in Turkey. In this region Seyhan River has a quite significant effect on drinking water, irrigation, hydroelectric energy and recreation. In the face of these impacts, forecasts of future River flow can be of help in making efficient operating decisions of water demand, for a wise and sustainable use of the river. To exemplify its applicability and demonstrate that the adaptive network fuzzy inference system has the ability to deal with human activities, we developed ANFIS models for time series forecasting with having various input structures. In the following, Firstly, the main network structure of the ANFIS model and the parameters estimating algorithms are defined and a description of the study area, available data and the model construction are described. Finally, the results of all ANFIS models are discussed

RIVER BASIN FLOOD MANAGEMENT

585

2. Adaptive Neural Fuzzy Inference System (ANFIS) The fuzzy logic approach is based on the linguistic uncertain expression rather than numerical uncertainty. It is an artificial intelligence technique that has been widely used in hydrological processes. Since Zadeh (1965) proposed the fuzzy logic approach to describe complicated systems, it has become popular and has been successfully used in various engineering problems, especially on control processes [Chen et. al., 2006; Chang et. al., 2006; Chang et. al., 2001; Liong et. al., 2000; Mahabir et. al., 2000; Nayak et. al., 2004a; Da Silva et. al., 1999; Firat et. al., 2006; Nayak et. al., 2004b; en, 2001]. Nonetheless, the main problem with this approach is that there is no systematic procedure for a design of fuzzy controller. However, a neural network system has the ability to learn its structure from the input-output sets and adapt itself in an interactive manner. ANFIS, consisted of the combination of the ANN and the fuzzy logic, has been used by many researchers to organize network structure itself and to adapt the parameters of fuzzy system for many engineering problems such as the estimating a controlled reservoir water level and time series forecasting. Fuzzy inference system is a rule based system consists of three conceptual components. These are: (1) a rule-base, contains fuzzy if-then rules, (2) a data-base, defines the membership function and (3) an inference system, combines the fuzzy rules and produces the system results [en, 2001]. First phase of fuzzy logic modeling is the determination of membership functions of input output variables, second is the construction of fuzzy rules and the last is the determination of output characteristics, output membership function and system results [Frat et. al., 2006]. Two methods, called as back propagation algorithm and hybrid-learning algorithm, provide learning of the ANFIS and construction of the rules, are used To determine the membership function of the input-output variables. The ANFIS is a multilayer feed-forward network uses ANN learning algorithms and fuzzy reasoning to characterize an input space to an output space. A general structure of fuzzy system is demonstrated in Figure 1.
Knowledge Base INPUT Fuzzification OUTPUT Database Rulebase
Defuzzification

Decision System
Figure 1: The General Structure of the fuzzy Inference System

586

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

ANFIS has been shown to be powerful in modeling numerous processes such as wind speed time series and real-time reservoir operation [Chen et. al., 2006; Chang et. al., 2006; Frat et. al., 2006]. ANFIS possesses properties such as capability of learning, constructing, expensing and classifying. It has the advantage of allowing the extraction of fuzzy rules from numerical data or expert knowledge and adaptively constructs a rule base. Moreover, it can adapt the complicated conversion of human intelligence to fuzzy systems. The main difficulty of the ANFIS predicting model is the time required for training structure and determining parameters. In this study, the ANFIS method consisting of the combination of the artificial neural networks and fuzzy logic approach has been used to estimate the River flow. ANFIS uses the learning ability of the ANN to define the input-output relationship and construct the fuzzy rules by determining the input structure. The system results were obtained by thinking and reasoning capability of the fuzzy logic. The hybrid-learning algorithm and subtractive function were used to determine the input structure. The detailed algorithm and mathematical background of the hybrid-learning algorithm can be obtained from the research of Jang et.al., [Jang et. al., 1997]. The consequence parameter in Sugeno inference system (Takagi T. and Sugeno M.1985) is a linear equation or constant coefficient. The linear equation is called first-order Sugeno inference system and the constant type is called zero-order Sugeno inference system. It was assumed that the fuzzy inference system had two inputs, x and y, and one output, z. Definitions of the symbols are given in appendix. For the first-order Sugeno inference system, typical two rules can be expressed as; Rule 1: Rule 2: Where,

IF x is A1 and y is B1 THEN IF x is A2 and y is B2 THEN

f1 = p1 * x + q1 * y + r1 f 2 = p 2 * x + q 2 * y + r2
and

and

y are the crisp inputs to the node i , Ai

Bi

are the linguis-

tic labels (low, medium, high, etc.) characterized by convenient membership functions and

pi , qi

and ri are the consequence parameters (i

= 1 or 2) . The struc-

ture of this fuzzy inference system is shown in Figure 2. The scheme of ANFIS consists of five layers (Figure 2). The model is briefly presented step by step in the following way; Input notes (Layer 1): Each node in this layer generates membership grades of the crisp inputs which belong to each of convenient fuzzy sets by using the membership functions. Each nodes output

Oi1

is calculated by:

RIVER BASIN FLOOD MANAGEMENT

587

Oi1 = A i ( x )

for i = 1, 2

Oi1 = B i 2 ( y )
w1 w1 + w2

for i = 3, 4

(1)

A1

x
A2

w1 ( x, y)

w1 =
N

f ( x, y )

w1 f1

f ( x, y )

B1

y
B2 Layer 1

w2 ( x, y)
Layer 2 Layer 3

w2 w2 = w1 + w2

f ( x, y )

w2 f2

Layer 4

Layer 5

Figure 2: The Scheme of Adaptive Neural Fuzzy Inference System

Where

Ai and Bi

are the appropriate membership functions for

Ai

and

Bi fuzzy sets, respectively. Many various membership functions such as trapezoidal,

triangular, Gaussian function, etc. can be applied to determine the membership grades. In this study, the gauss membership function is used, as;
( x c)2 2 2

Oi1 = A i ( x) = e
Where,

(2)

{ai , bi , ci } is the membership functions parameter set that changes the

shape of membership function from 1 to 0. These parameters are referred to as the premise parameters. Rule nodes (Layer 2): In this layer, the AND/OR operator is applied to get one output that represents the results of the antecedent for a fuzzy rule, that is, firing strength. It means the degrees by which the antecedent part of the rule is satisfied and it indicates the shape of the output function for that rule. The outputs of the second layer, called as firing strengths Oi , are the products of the corresponding degrees obtaining from layer 1, named as
2

in below. (3)

Oi2 = wi = Ai ( x) Bi ( y ),

i = 1,2

588

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

Average nodes (Layer 3): Main target is to compute the ratio of firing strength of each ith rule to the sum of all rules firing strength. Thus the firing strength in this layer is normalized as;

Oi3 = wi =

wi wi
i

i = 1,2

(4)

Consequent nodes (Layer 4): The contribution of ith rules towards the total output or the model output and/or the function defined is calculated in Equation (5);

Oi4 = wi f i = wi ( pi x + qi y + ri )
Where,

i = 1,2

(5)

wi

is the ith nodes output from the previous layer (i.e., demonstrated in

the third layer).

{pi , qi , ri } is the parameter set in the consequence function and

also the coefficients of linear combination in Sugeno inference system. Output nodes (Layer 5): This layer is called as the output notes in which the single note computes the overall output by summing all the incoming signals and is the last step of the ANFIS. Hence, each rules fuzzy results are transformed into a crisp output in this layer by defuzzification process, as;

f ( x, y ) =

w1 ( x, y ) f1 ( x, y ) + w2 ( x, y ) f 2 ( x, y ) w1 f1 + w2 f 2 = w1 ( x, y ) + w2 ( x, y ) w1 + w2

(6)

Qi5

wi f i f ( x, y ) = wi . f i = wi f1 + wi f 2 = i wi i
i

(7)

Neural network is trained based on supervised learning. The objective is to train adaptive networks to be able to have convenient unknown functions given by training data and be able to find the proper value of the input and output parameters. For this aim, ANFIS applies the hybrid learning algorithm, consists of the combination of the gradient descent method and the least-squares method. The gradient descent method is used to assign the nonlinear input parameters (ai , bi , ci ) , as the least-squares method is employed to identify the linear output parameters ( pi , qi , ri ) . The antecedent parameter, i.e., membership function given in layer 2 is applied to construct the rules of the ANFIS model. Since the input variables

RIVER BASIN FLOOD MANAGEMENT

589

within a range might be clustered into several classes, the structure of input layer needs to be determined accurately. The subtractive fuzzy clustering function offering the effective result by less rules, is applied to solve the problem in ANFIS modeling.
3. STUDY AREA AND DATA

The applicability of ANFIS as a river flow forecasting model is investigated. To illustrate the validity and capability of ANFIS method for time series forecasting and modeling, Seyhan River, located the south part of Turkey is chosen. It has been operated for irrigation, hydropower generation, domestic use and recreation facilities. The Seyhan River and its drainage basin are shown in Figure 3. There is one River flow gauging stations, (1818), equipped with automatic daily flow recorders, on Seyhan River as shown in Figure 3.

Figure 4. The Seyhan River and its Drainage Area

4. RIVER FLOW ESTIMATION BY SUGENO FUZZY INFERENCE SYSTEM

The total observed 2192 daily data sets were obtained from River flow gauging stations of the 1818. To discriminate the process of segregation, the statistical parameters (i.e., minimum value;

x min ,

maximum value;

x max ,

mean; x , standard

590 deviation; s x , variation coefficient

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

cvx , skewness coefficient; c sx ) for the total ob-

served daily data sets are given in Table 1.

Table 1: The Statistical Parameters for data sets

Variables

x min
3

x max
1154.00

x
144.46

sx
115.69

c sx
2.64

Q(t )1818 (m /s)

11.00

One of the most important steps in developing a satisfactory forecasting model is the selection of the input variables. Because, these variables determine the structure of the ANFIS model and affect the weighted coefficient and the results of the model. Hence, Pearson and Spearman correlation coefficients between input and output parameters are calculated in order to apply ANFIS model. Different combinations of the antecedent flows of the river flow gauges stations were used to construct the appropriate input structure in the time series forecasting model. The general structures of the time series forecasting models are given in Equation (8). The structures of the forecasting models are shown in Table 2.

Q(t )1818 = Q(t 1)1818 + Q (t 2)1818 + ...........Q(t n)1818

(8)

Tablo 2. Seyhan Nehri Ak Tahmini iin Kurulan Modeller

Model R-I M1 R-I M2 R-I M3 R-I M4 R-I M5 R-I M6

Input structure

Numbers of the variable 1 2 3 4 5

Output
Q(t )1818 Q(t )1818 Q(t )1818 Q(t )1818 Q(t )1818 Q(t )1818

Q(t 1)1818 Q(t 1)1818 Q(t 2)1818 Q(t 1)1818 Q(t 2)1818 Q(t 3)1818 Q(t 1)1818 Q(t 2)1818 Q(t 3)1818 Q(t 4)1818
Q(t 1)1818 Q(t 2)1818 Q(t 3)1818 Q(t 4)1818 Q(t 5)1818
Q(t 1)1818 Q(t 2)1818 Q (t 3)1818 Q(t 4)1818 Q(t 5)1818 Q(t 6)1818

Where;

Qt

represents the River flow at time (t),

Q(t 1) , Q(t n) are the

river flow respectively at times (t1) (tn). It is evident that the training data sets should cover all the characters of the problem in order to get effective estimation. For this aim, data set were divided three training/testing subsets, which are m1, m2 and m3 by cross validation method as a systematic process to get effective and sensitive modeling. The structures of data set are given in Table 3, Table 4 and Figure 4.

RIVER BASIN FLOOD MANAGEMENT

591

Table 3 The Cross Validation Data Sets for River Flow Estimation

Cross Validation m1 m2 m3

Dates of the data 01.10.1986 30.09.1988 01.10.1988 30.09.1990 01.10.1990 30.09.1992 Numbers of training Data 1461 1462 1461
Training m2 Testing m2 Training m2

(2 yl) (2 yl) (2 yl)

Numbers of Data 731 730 731 Numbers of testing Data 731 730 731

Table 4 The Structure of the Training and Testing Data Sets

MY1 MY2 MY3

Dates of the training set 1988 1992 1986 1988 ve 1990 1992 19861990
MY1 MY2 MY3 Testing m1 Training m1 Training m1

Dates of the testing set 1986 1988 1988 1990 1990 1992

Training m3 Training m3 Testing m3

Figure 4. The Training and Testing Data Sets for River Flow Estimation

In each model every input variables might be clustered into several class values in layer 1 to build up fuzzy rules. Each fuzzy rule is constructed through several parameters of membership function in layer 2. If the number of parameters that needs to be determined increases with the fuzzy rule increment, the model structure will becomes more complicated. In this study, the subtractive fuzzy clustering function is used to establish the fuzzy rule based on the relationship between the inputoutput variables. In order to determine the nonlinear input and linear output parameters, the hybrid algorithm is used. The learning procedure and the construction of the rules are provided by this algorithm. The performances of the ANFIS models both training and testing data are evaluated and the best training/testing data set is selected according to Relative Error (AARE), Correlation Coefficient (CORR), Efficiency (E) and Root Mean Square Error (RMSE). The performances of river flow forecasting models are given in Table 7.

Correlation Coefficient (R); R =

(Q D Q D ).(QY QY )
i =1

(9)

(Q D Q D ) 2 .(QY QY ) 2
i =1
N

Efficiency (E);

E=

E1 E 2 E1

E1 =

(Q D Q D ) 2 ,
t =1

E2 =

(QY Q D ) 2
t =1

(10)

592

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

Where, QY is the estimated river flow by ANFIS, Q D is the field observation river flow, Q y is the average of the estimated flows, Q D is the average of the observation flows.
Table 5. The Performances of the ANFIS Models for River Flow Estimation

Criteria RMSE E

AARE

CORR

Testing Data Set 1986 1988 1988 1990 1990 1992 1986 1988 1988 1990 1990 1992 1986 1988 1988 1990 1990 1992 1986 1988 1988 1990 1990 1992

R-I M1 65.13 45.54 30.64 0.815 0.763 0.885 0.089 0.073 0.063 0.911 0.874 0.942

R-I M2 64.96 36.67 29.88 0.816 0.762 0.891 0.088 0.078 0.067 0.916 0.876 0.945

R-I M3 59.44 33.51 28.94 0.846 0.801 0.898 0.082 0.081 0.064 0.920 0.896 0.948

R-I M4 59.88 33.90 30.85 0.844 0.796 0.884 0.083 0.081 0.066 0.919 0.893 0.940

R-I M5 61.29 37.04 29.28 0.836 0.757 0.895 0.088 0.084 0.071 0.915 0.874 0.946

R-I M6 61.21 36.68 29.10 0.837 0.761 0.896 0.089 0.084 0.071 0.915 0.873 0.946

As can be seen in Table 5, the ANFIS models are evaluated based on their performance in training and testing sets. The models have shown significant variations in the criteria of the performance evaluation given in Table 5. It appears that the ANFIS models are accurate and consistent all the values of RMSE and AARE are smaller, and all correlation coefficients and efficiencies are also very close to unity. It also shows that the lowest value of the RMSE is 28.94 (in R-I M3) and the highest value of the RMSE is 65.13 (in R-I M1). The values of the AARE (% 6.40) and RMSE (28.94) in forecasting R-I M3 are much lower than other models. In addition, the values of the efficiency (0.898) and correlation coefficients (0.948) are higher than other models. R-I M3, which consists of three antecedent flows in input, has shown the highest efficiency, correlation and the minimum RMSE and AARE. As a result, RI M3 was selected as the best-fit model to estimate the River flow in the Great Menderes catchments. The performance of the best model (R-I M3) was compared for the all cross validation data sets and the best cross validation training/testing data set was determined. As seen in Table 5, in the testing of R-I M3, the lowest value of the RMSE is 28.94 (for MY3), the highest value of the RMSE is 59.44 (for MY1). On the other hand, it appears that all the values of AARE, efficiency (E) and CORR of the MY3 data set are higher than other data sets. As a result, the ANFIS model consisting of three antecedent flows was chosen as the best fit model structure. This ANFIS model has shown the highest performance in training of the MY3 cross validation data set. The general structure of R-I M3 estimation model is given in equation (11). In addition, the performances of the all ANFIS models for MY3 data set are shown in Figure 5.

RIVER BASIN FLOOD MANAGEMENT

593

Q(t )1818 = Q(t 1)1818 + Q (t 2)1818 + Q(t 3)1818

The Peformances of ANFIS M odels (for M Y3)
0.95 0.945 0.94 CORR E 0.88 E 0.9
AARE 0.1 0.08 0.06

(11)

CORR

30 28

0.935

0.93 0.87 R-I M1 R-I M2 R-I M3 R-I M4 R-I M5 R-I M6 Models

26 0.04 R-I M1 R-I M2 R-I M3 R-I M4 R-I M5 R-I M6 Models

Figure 5. The Performances of the all ANFIS Models for MY3 data set

The training parameters of the best-fit river flow estimation model, R-I M3 are given Table 6 and the structure of this model is shown in Figure 6.

Q(t1)1818

Q(t)1818

Q(t2)1818

Q(t3)1818
Input Membership Function Fuzzy Rules

Output Output Membership Function

Tablo 6. The Training Parameters of R-I M3 ANFIS Model Numbers of Rules Epoch Numbers of Membership functions AND method Defuzzification Method 8 25 Gauss- 8 prod wtaver

Figure 6. The Structure of the best fit River Estimation Model

RMSE

0.89

T he Peformances of ANFIS Models (for MY3) 34 AARE RMSE 32

594

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

RESULTS AND DISCUSSION

The results of the ANFIS models and observation are compared with the observed flows in order to evaluate the performance of the training/testing of the runoff models. Figure 6 shows the scatter diagrams of the estimated values of the training/testing of the ANFIS models and observation values.

700 600 Daily Flow (m3 /s) 500 400 300 200 100 0 1

T esting Data Set

R-I M3 ANFIS Observation

800 Estimated Flow (m 3 /s) 600 400 200 0 0

T est ing Data Set- R-I M3 ANFIS River Flow

366 Data Set (1990-1992)

731

200 400 600 Observation Flow (m 3 /s)

800

1400 1200 Daily Flow (m3 /s) 1000 800 600 400 200 0 1

T raining Data Set

R-I M3 ANFIS Observation

800 Estimated Flow (m 3 /s) 600 400 200 0

T raining Data Set - R-I M3 ANFIS River Flow

366 731 Data Set (1986-1990)

1096

200 400 600 Observation Flow (m 3 /s)

800

Figure 7. Comparison of ANFIS model and Observation

As can be seen in Figure 7, the ANFIS model, in general, has good performance for river flow estimation. Although the ANFIS model was trained with limited training data set, the results of the ANFIS model has shown that ANFIS can be applied for river flow estimation. The best fit model was trained and test by ANN in order to get more sensitive and accurate comparison. The ANN models having various the numbers of hidden neuron in hidden layer were tried and the best fit structure of ANN was determined for river flow estimation. The performances of the ANN models are shown in Figure 8.

RIVER BASIN FLOOD MANAGEMENT

595
The Performances of ANN M odels
1 0.6
CORR E

The Performances of ANN M odels 0.95 CORR 0.9 0.85 0.8 0.75 0.7 4 5 6 7 8 9 10 11 12 13 The Numbers of Hidden Neurons

60 RMSE 55 50 45 40 4

0.8 0.4 0.2 0 E

RMSE

5 6 7 8 9 10 11 12 The Numbers of Hidden Neurons

13

Figure 8. The Performances of ANN Models

The back propagation algorithm and tangent activation function were used for training and testing of ANN models. In addition, the learning rate (0.01), momentum coefficient (0.5), epoch (1000) and the numbers of hidden neuron in hidden layer were selected by trial and error method during training process. The comparison of the ANFIS and ANN models are given in Table 6.

Tablo 6. Comparison of ANFIS and ANN models

Model R-I M3 ANFIS R-I M3 ANN

Testing Data Set RMSE 28.94 41.50 E 0.898 0.789 CORR 0.948 .894

Training Data Set RMSE 50.09 64.96 E 0.840 0.731 CORR 0.916 0.856

Comparing the two forecasting models, it can be seen that the values of the AARE and RMSE of the ANFIS model is much lower than ANN model. The RMSE value of the ANFIS model is also lower (28.94) than ANN model (41.50). In addition, the values of the efficiency (E) and correlation coefficients (CORR) of the ANFIS model is higher than the ANN model. The efficiency of the ANFIS model is much higher (0.898) than ANN model (0.789). The CORR of the ANFIS model is also higher (0.948) than ANN (0.894). It may be noted that a trial and error procedure has to be performed for ANN model to develop the best network structure, while such a procedure is not required in developing an ANFIS model. Moreover, in the current study, the ANFIS model is trained by using just 25 epochs, while the ANN model took 1000 epochs. The results suggest that the ANFIS method is superior to the ANN method in the modeling and forecasting of the River flow. The comparison of the results of the ANFIS and ANN models are given in Figure 9.

596
800 Estimated Flow (m 3 /s) 600 400 200 0 0 200 400 600 Observation Flow (m 3 /s) T esting Data Set- R-I M3 ANFIS River Flow

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

800 Estimated Flow (m3 /s) 600 400 200 0

T esting Data Set- R-I M3 ANN River Flow

800

200 400 600 Observation Flow (m 3/s)

800

800 Estimated Flow (m3 /s) 600 400 200 0 0

T raining Data Set- R-I M3 ANFIS River Flow

Estimated Flow (m 3 /s) 1400 1200 1000 800 600 400 200 0 0 200

T raining Data Set- R-I M3 ANN River Flow

200 400 600 Observation Flow (m 3 /s)

800

400 600 800 1000 1200 1400 Observation Flow (m 3 /s)

Figure 9. Comparison of the R- II M1 ANFIS and ANN Models

5. CONCLUSIONS

In this study, the applicability and capability of ANFIS method was investigated for river flow forecasting. To verify the application of this approach, the Seyhan River catchments is chosen as the case study area. The data set include 2192 daily runoff data for the period at 1986-1992 years and data set was divided three subsets by cross validation method. The models having various input structure were trained and tested with all cross validation data sets. The performances of ANFIS models and observation are compared and evaluated. It appears that the ANFIS models are accurate and reliable all the values of RMSE and AARE are smaller, and all correlation coefficients and Efficiency are also very close to unity. According to the criteria of performance evaluation, R-I M3 ANFIS model having three input variables, was selected as the best fit model. The best fit model was also trained and tested by ANN and the performances of ANFIS and ANN models were compared. Comparing the two estimation models, it can be seen that the values of the AARE and RMSE of the ANFIS

RIVER BASIN FLOOD MANAGEMENT

597

model is much lower than ANN model. The values of the E and CORR of the ANFIS model are higher than the ANN model. The results suggest that the ANFIS method is superior to the ANN method and MR in the modeling and forecasting the of the time series. The results of the ANFIS model showed that the ANFIS can be successfully applied to establish for river flow forecasting and modeling

6. REFERENCES

Bodri L, Cermak V. 2000. Prediction of extreme precipitation using a neural network: application to summer flood in Moravia. Advances in Engineering. 31: 311-321. Chang F.J, Hu H.F, Chen Y.C. 2001. Counter propagation fuzzyneural network for River flow reconstruction. Hydrological Processes, 15: 219232. Chang F.J, Chang Y.T. 2006. Adaptive neuro-fuzzy inference system for prediction of water level in reservoir. Advances in Water Resources, 29: 1-10. Chang, Y.T., Chang, L.C., ve Chang, F.J. 2005. Intelligent control for modeling of realtime reservoir operation, part II: artificial neural network with operating rule curves. Hydrological Process, 19: 14311444 Chen, S.H., Lin, Y.H., Chang, L.C., Chang, F.J., The strategy of building a flood forecast model by neuro fuzzy network, 2005 Dibike, Y. B., Solomatine, D. P., 2001. River Flow forecasting Using Artificial Neural Networks, Phys. Chem. Earth (B), 26, 1-7. Ertunga C., Duckstein L. 2001. Fuzzy conceptual rainfall-runoff models. Journal of Hydro., 253: 41-68. Frat, M., Gngr, M., 2004. Ask Maddesi Konsantrasyonu ve Miktarnn Yapay Sinir Alar ile Belirlenmesi, MO Teknik Dergi, Cilt 15, Say 3, 3267- 3282. Frat M, Gngr M. 2006a. River flow estimation using adaptive neuro-fuzzy inference system. Mathematics and Computers in Simulation, (baskda) Hsu K, Gupta H.V, Sorooshian S. 1998. Riverflow Forecasting Using Artificial Neural Networks, ASCE, Water Resources Engin.Conference98, 967-972. Jain, A., Kumar, A.M. 2006. Hybrid neural network models for hydrologic time series forecasting. Applied Soft Computing, (baskda) Jain S.K, Das A, Srivastava D.K. 1999. Aplication of ANN for Reservoir Inflow Prediction and Operation. ASCE J. of Water Res. Planning and Mang., 25 (5) Mahabir C, Hicks F.E, Fayek A.R. 2000. Application of fuzzy logic to the seasonal runoff, Hydrological Process, 17: 37493762. Mitra B, Scott H.D, Dixon J.C, Mckimmey J.M. 1998. Application of fuzzy logic to the prediction of soil erosion in a large watershed. Geoderma, 86: 183209.

598

INTERNATIONAL CONGRESS ON RIVER BASIN MANAGEMENT

Liong, S.Y., Lim, W.H., Kojiri, T., Hori, T. (2000). Advance Flood forecasting for Flood stricken Bangladesh with a fuzzy reasoning method, Hydrological Process,14: 431448. Nagy H.M, Watanabe K, Hirano M. 2002. Prediction of sediment load concentration in rivers using artificial neural network model, Journal of Hydr. Eng, 128: 588-595. Nayak, P.C., Sudheer, K.P., Ramasastri, K.S. (2004a). Fuzzy computing based rainfallrunoff model for real time flood forecasting. Hydrological Process, 17: 37493762. Nayak, P.C., Sudheer, K.P., Ragan, D.M., Ramasastri, K.S. 2004b. A Neuro Fuzzy computing technique for modeling hydrological time series. Journal of Hydrology, 29: 5266 Sajikumar N, Thandaveswara B.S. 1999. A non-linear rainfallrunoff model using an artificial neural network. Journal of Hydrology, 216: 3255. en, Z. (2001). Bulank Mantk ve Modelleme lkeleri, ISBN: 9758509233, 172 sayfa, Bilge Kltr Sanat Yaynevi, stanbul. en, Z., 2004. Yapay Sinir Alar lkeleri, Su Vakf Yaynlar, stanbul.