Professional Documents
Culture Documents
Abstract: Scour at a downstream vertical gate of a dam is investigated using a fuzzy logic inference system. The application is presented
for experimental data sets. A comparison is made between a regression formulation 共RM兲 and a fuzzy logic approach. Some restrictive
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.
assumptions for RM are discussed. Here the Takagi–Sugeno 共TS兲 approach is used with a constant function in the consequent part of the
fuzzy rules. It is demonstrated that the TS model gives a lower error than the RM. Furthermore, scatter diagrams indicate that the TS
approach has superiority over the RM.
DOI: 10.1061/共ASCE兲0733-9429共2006兲132:10共1069兲
CE Database subject headings: Scour; Fuzzy sets; Hydraulic engineering; Regression analysis; Gates; Dams.
Takagi–Sugeno 共TS兲. For the Mamdani 共1974兲 approach, the out- Fig. 4. Experimental setup
come of each IF-THEN rule is a fuzzy set for the output variable,
and hence the step of defuzzification is necessary to obtain a crisp The general algorithm of a TS fuzzy inference system can
value of the output variable. The Takagi–Sugeno approach be expressed as follows. It is assumed that there are
共Takagi and Sugeno 1985兲, however, does not have a classical Rr共r = 1 , 2 , ¯ , n兲 rules in the above mentioned form:
defuzzification procedure and the conclusion of each IF-THEN 1. For each implication Ri, y i is calculated by the function f i in
inference rule is a scalar rather than a fuzzy set for the output the consequence
variable. In this paper the TS method is used due to its higher
performance in the training phase when compared with the Mam- y r = f r共x1,x1, . . . ,x p兲 = br0 + br1x1 + ¯ + brpx p 共1兲
dani approach. The main problem with TS fuzzy logic modeling
2. The weights are calculated as
is related to the choice of the parameters 共Chen et al. 1998; Chin
and Qi 1998; Stoeva and Nikov 2000兲. There is a need for effec- rr = 共mr1⌳mr2⌳ . . . mrk兲 * Rr 共2兲
tive methods for tuning the membership functions so as to mini-
mize the error measures. For this reason, the adaptive network where mr1, mr2 . . . mrk denote ␣ cuts of membership functions
based fuzzy inference system 共ANFIS兲 methodology first pro- according to input values for the rth rule. The occurrence
posed by Jang 共1993兲 is applied to optimize the parameters of the probability is shown by Rr and ⌳ represents the minimum
membership functions and the consequent part by using a hybrid operator. For the sake of simplicity Rr is equal to 1, and
learning algorithm. Various approaches have been proposed to finally;
determine the fuzzy model parameters. Many of these approaches 3. The final output y inferred from n implications is given as
use clustering techniques, genetic algorithms, gradient descent al- the weighted average of all y r with the weights rr as
gorithms, and numerical data. However in this study the simple n
the experiment are presented in Fig. 5. After conducting experi- application is possible under the light of a set of assumptions
ments with two types of materials, the relationship among grain which includes linearity, normality, variance constancy, ergodic-
size, depth of scour, height of water at the outlet, total discharge, ity, and independence of the residuals. The TS fuzzy logic model
and head has been investigated. The result is that scour is directly replaces Eq. 共5兲 without any of these restrictions and expresses a
proportional to head, h, and to discharge passing through a unit natural relationship between qo / qu and w. On the other hand, the
width, q, and inversely proportional to grain size d. regression model consists of two parts, namely parabolic and lin-
The change of w values for Material I, Material II, and Mate- ear. Thus it is necessary to consider the transition region which
rial I + Material II with respect to the ratio 共qo / qu兲 is considered depends on the ratio 共qo / qu兲. However, the fuzzy logic model
for the implementation of the TS fuzzy logic methodology so as does not contain such a transition region.
to depict the common behavior of these two variables. The fol- The system with four membership functions per input and four
lowing equation gives the result of the experimental work as: rules is determined as the most suitable one after trial and error.
h0.5q0.6 The input membership functions shown in Fig. 9 consist of “low,”
d0 + h1 = w 共4兲 “medium,” “high,” and “very high” fuzzy subsets. The Gaussian
d0.4
90 membership function provides a smaller relative error compared
where d0⫽scour depth 共m兲; hl⫽downstream depth 共m兲; h⫽head to the other types and therefore is selected for the modeling. Table
共m兲; q = qo⫹qu⫽discharge per unit length 共m3 / s / m兲; d90⫽granule 3 indicates the fuzzy rule base of the system. The last column in
size of bed material 共mm兲, 共diameter in which 90% of the mate- Table 3 shows the parameters of the output functions which has
rial by weight goes through the screen兲; and w⫽coefficient. The the following form:
冉冊
coefficient w is calculated separately for each experiment. The
change of w values for Material I, Material II, and Material I ⫹ qo 1
wr = br0 + b 共6兲
Material II with respect to the ratio 共qo / qu兲 is shown, respectively, qu r
in Figs. 6–8. The given curves are obtained using the least where br⫽consequent parameters of the TS fuzzy model. The
squares method. The input variable is the ratio qo / qu and the comparison of the TS fuzzy model and RM can be achieved on
output variable is w. Uyumaz 共1988兲 expressed such an equation the basis of different error definitions and graphical representa-
as tions. Average relative error percentage 共AREP兲 is employed in
w=A 冉冊 冉冊
qo
qu
2
−B
qo
qu
+C
qo
qu
⬍2
this paper for numerical error comparisons which is defined as
follows:
冉冊
n
1
p
兩w pi − wmi兩
w=E
qo
+F
qo
ⱖ2 共5兲
共AREP兲i = 100 兺
n p i=I wmi
共7兲
qu qu
in which A, B, C, D, and E⫽coefficients. The application of such where w pi and wmi⫽predicted and measured values at point i. In
a model requires, prior to any prediction procedure, the parameter order to validate the TS fuzzy model, the data was divided into
estimations from the available data 共Table 2兲. Furthermore, its training and testing parts. For Material I, Material II, and Material
I + Material II 4, 8, and 15 points of the entire data were left for
validation, respectively, whereas 20, 21, and 41 data points that
are the rest of the data are employed for training.
Material number A B C E F
I 3.17 −6.62 10.63 0.35 8.8
I 1.63 −4.68 11.65 0.35 8.8
Fig. 7. Change of coefficient w by qo / qu 共Material II兲 I + II 2.10 −5.65 11.56 0.35 8.8
Input W
Rule number qo / qu br0 br1
Material I
1 Low −26.03 1.71
2 Medium −3.933 31.35
3 High −12.5 38.31
4 Very high −9.44 44.2
Material II
1 Low −25.03 10.93
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.
Fuzzy Fuzzy
Date qo / qu w Model Regression model Regression
Training 0.115 10.092 9.944 9.911 1.468 1.797
part 0.151 9.917 9.598 9.703 3.220 2.161
Fig. 9. Input membership functions for: 共a兲 Material I; 共b兲 Material 0.127 9.565 9.824 9.840 2.634 2.799
II; and 共c兲 Materials I + II 0.218 8.197 8.970 9.337 8.616 12.214
0.238 9.018 8.839 9.234 1.986 2.339
0.323 9.135 8.510 8.822 6.839 3.421
The overall prediction results for validation and training for 0.384 8.920 8.417 8.555 5.642 4.088
Material I are given in Table 4. As can seen from the table for the 0.428 7.571 8.390 8.377 9.756 9.625
observed w value of 8.803, which produces the highest relative 0.463 8.060 8.380 8.244 3.819 2.238
error for regression, the predictions by fuzzy and regression 0.483 8.588 8.373 8.172 2.507 4.843
model are 8.064 and 7.618, respectively. The maximum relative 0.670 8.803 8.064 7.618 8.398 13.466
errors for the fuzzy and regression model are 11.502 and 0.709 7.024 7.937 7.530 11.502 6.719
13.466, respectively. On the other hand, the R2 coefficient for 0.783 8.158 7.666 7.390 6.030 9.414
the performance relative to a naive baseline is defined as 0.931 7.024 7.203 7.214 2.490 2.639
R2 = 1 − 共mean-square-error兲 / 共variance of observations兲. R2 and 1.240 7.590 7.566 7.295 0.320 3.882
AREP values for Material I, Material II, and Material I + Material 1.530 7.767 7.760 7.922 0.089 1.957
II are presented in Table 5. On the basis of these error criteria, the 1.720 8.607 8.637 8.622 0.351 0.171
TS fuzzy model yields smaller prediction error and is, therefore, 1.840 9.780 9.749 9.182 0.317 6.119
recommended.
2.520 9.193 10.261 9.444 10.408 2.658
Fig. 10 indicates the observed and predicted w values for train-
2.690 10.249 10.010 9.742 2.332 4.952
ing. It is obvious that they follow each other very closely and on
Average 4.436 4.875
average the observed and predicted w values have almost the
same statistical parameters. Predictions are successful especially Validation 0.294 8.470 8.600 8.958 1.515 5.445
part 0.604 8.080 8.232 7.788 1.849 3.614
at low or high values and have a very high correlation with ob-
servations. A comparison is also made for validation in Fig. 11. 1.120 8.177 7.221 7.192 11.694 12.045
As can be seen from Table 4, in the majority of the predictions, 3.620 10.014 10.306a 10.067a 2.833 0.526
the TS fuzzy model relative error is lower than the RM prediction. Average 4.473 5.408
a
In Table 4, the AREP is also given as 4.436 and 4.875 for the Validation points that are extrapolated.
FM RM FM RM
Material
2 2 2 2
number R AREP R AREP R AREP R AREP
I 0.73 4.44 0.70 4.99 0.68 4.47 0.60 5.41
II 0.66 5.61 0.59 6.21 0.48 5.15 0.26 6.11
I + II 0.61 6.45 0.53 6.86 0.50 5.99 0.49 6.03
Fig. 11. Comparison of models for testing data for: 共a兲 Material I; 共b兲
Material II; and 共c兲 Materials I + II
Conclusions
Acknowledgments
The writers wish to sincerely thank Professor Dr. Zekai Şen for
Fig. 10. Comparison of models for: 共a兲 Material I; 共b兲 Material II; his helpful suggestions. They are also grateful to Ömer Beğendik
and 共c兲 Materials I + II who contributed to the improvement of this paper.
108共10兲, 1147–1162.
h ⫽ head of flow;
Jang, J. S. R. 共1993兲. “ANFIS: Adaptive network based fuzzy inference
h1 ⫽ water depth at downstream;
system.” IEEE Trans. Syst. Man Cybern., 23共3兲, 665–684.
i ⫽ data point; Johnson, P. A., and Ayyub, B. M. 共1996兲. “Modeling uncertainty in pre-
m ⫽ coefficient; diction of pier scour.” J. Hydraul. Eng., 122共2兲, 66–72.
mrk ⫽ ␣ cuts of membership functions; Kindler, J. 共1992兲. “Rationalizing water requirements with aid of fuzzy
n ⫽ number of fuzzy rules; allocation model.” J. Water Resour. Plan. Manage., 118共3兲, 308–323.
p ⫽ number of input variables; Laursen, E. M. 共1952兲. “Observations on the nature of scour.” Proc., 5th
q ⫽ discharge per unit length; Hydraulics Conf., State Univ. of Iowa, Iowa City, Iowa, Bulletin
qo ⫽ discharge per unit length over gate; Number No. 34, 179–197.
qo / qu ⫽ ratio of discharge per unit length over and Mamdani, E. H. 共1974兲. “Application of fuzzy algorithms for simple
under gate; dynamic plant.” Proc. IEEE, 121, 1585–1588.
qu ⫽ discharge per unit length under gate; Mohamed, M. S., and McCorquodale, J. A. 共1992兲. “Short-term local
Rr ⫽ rth rule; scour.” J. Hydraul. Res., 30共5兲, 685–699.
Rr ⫽ occurrences probability; Pappis, C. P., and Mamdani, E. H. 共1977兲. “A fuzzy controller for a traffic
junction.” IEEE Trans. Syst. Man Cybern., 7共10兲, 707–717.
R2 ⫽ correlation coefficient;
Pongracz, R., Bogardi, I., and Duckstein, L. 共1999兲. “Application of
r ⫽ rule number; fuzzy rule-based modeling technique to regional drought.” J. Hydrol.,
rr ⫽ weight of rth inference rule; 224共3–4兲, 100–114.
t ⫽ time; Ross, J. T. 共1995兲. Fuzzy logic with engineering applications, McGraw-
V ⫽ volume of scour hole; Hill, New York.
w ⫽ coefficient; Sarma, K. V. N. 共1971兲. “Discussion of similarity laws for localized
wmi ⫽ measured values in point i; scour.” J. Hydraul. Div., Am. Soc. Civ. Eng., 93共2兲, 67–71.
w pi ⫽ predicted values in point i; Şen, Z. 共2004兲. Fuzzy logic and system models in water sciences, Turkish
wr ⫽ output value of coefficient for rth rule; Water Foundation, Istanbul, Turkey.
x1, x2, x p ⫽ input variables; Şen, Z., and Altunkaynak, A. 共2004兲. “Fuzzy awakening in rainfall-runoff
y r ⫽ output of rth inference rule; modeling.” Nord. Hydrol., 35共1兲, 31–43.
␣ ⫽ membership function truncation level; Şen, Z., Altunkaynak, A., and Özger, M. 共2003兲. “Autorun persistence of
⌳ ⫽ minimum operator; and hydrologic design.” J. Hydrologic Eng., 8共6兲, 329–338.
2 ⫽ chi-square value. Şen, Z., Altunkaynak, A., and Özger, M. 共2004兲. “Sediment concentration
and its prediction by perceptron Kalman filtering procedure.” J. Hy-
draul. Eng., 130共8兲, 816–826.
Shrestha, B. P., Duckstein, L., and Stakhiv, E. Z. 共1996兲. “Fuzzy rule-
References based modeling of reservoir operation.” J. Water Resour. Plan. Man-
age., 122共4兲, 262–269.
Altunkaynak, A. Özger, M., and Çakmakcl, M. 共2004a兲. “Water con- Stoeva, S., and Nikov, A. 共2000兲. “A fuzzy backpropagation algorithm.”
sumption prediction of Istanbul city by using fuzzy logic approach.” Fuzzy Sets Syst., 112, 27–39.
Water Resour. Manage., 19共5兲, 641–654. Takagi, T., and Sugeno, M. 共1985兲. “Fuzzy identification of systems and
Altunkaynak, A., Özger, M., and Çakmakcl, M. 共2004b兲. “Fuzzy logic its application to modeling and control.” IEEE Trans. Syst. Man Cy-
modeling of the dissolved oxygen fluctuations in golden horn.” Ecol. bern., 15, 116–132.
Modell., 189共3–4兲, 436–446. Tusuchiya, Y., and Iwagaki, Y. 共1967兲. “On the mechanism of the local
Bardossy, A., and Disse, M. 共1993兲. “Fuzzy rule based models for infil- scour from flows downstream of an outlet.” Proc., 12th Int. Associa-
tration.” Water Resour. Res., 29共2兲, 373–382. tion of Hydraulics Research, 12th Congress, Fort Collins, Colo.
Breuser, H. N. C. 共1966兲. “Conformity and time scale in two-dimensional Uyumaz, A. 共1988兲. “Scour downstream of vertical gate.” J. Hydraul.
local scour.” Proc., Symp. on Model and Prototype Conformity, Hy- Eng., 114, 811–816.
draulics Research Laboratory, Poona, India, 1–8. Xiong, L., Shamseldin, A. Y., and O’Connor, M. K. 共2001兲. “A nonlinear
Breuser, H. N. C. 共1967兲. “Time scale of two-dimensional local scour.” combination of the forecasts of rainfall-runoff models by the first
Proc., 12th Congress, International Association for Hydraulic Re- order Takagi–Sugeno fuzzy system.” J. Hydrol., 245, 196–217.
search, Fort Collins, Colo., C32-1–C32-8. Zadeh, L. A. 共1965兲. “Fuzzy sets.” Information and Control, 12, 94–102.