Sie sind auf Seite 1von 28

Journal of Hydrology 253 (2001) 4168

www.elsevier.com/locate/jhydrol

Fuzzy conceptual rainfallrunoff models


zelkan a,*, Lucien Duckstein b
Ertunga C. O
b

a
i2 Technologies, 909 E. Las Colinas Blvd, Irving, TX 75039, USA
Ecole Nationale du Genie Rural, des Eaux et des Forets 19, avenue du Maine, 75732 Paris Cedex 15, France

Received 19 February 1999; revised 26 April 2001; accepted 27 April 2001

Abstract
A fuzzy conceptual rainfallrunoff (CRR) framework is proposed herein to deal with those parameter uncertainties of
conceptual rainfallrunoff models, that are related to data and/or model structure: with every element of the rainfallrunoff
model assumed to be possibly uncertain, taken here as being fuzzy. First, the conceptual rainfallrunoff system is fuzzied and
then different operational modes are formulated using fuzzy rules; second, the parameter identication aspect is examined using
fuzzy regression techniques. In particular, bi-objective and tri-objective fuzzy regression models are applied in the case of
linear conceptual rainfallrunoff models so that the decision maker may be able to trade off prediction vagueness (uncertainty)
and the embedding outliers. For the non-linear models, a fuzzy least squares regression framework is applied to derive the
model parameters. The methodology is illustrated using: (1) a linear conceptual rainfallrunoff model; (2) an experimental twoparameter model; and (3) a simplied version of the Sacramento soil moisture accounting model of the US National Weather
Services river forecast system (SAC-SMA) known as the six-parameter model. It is shown that the fuzzy logic framework
enables the decision maker to gain insight about the model sensitivity and the uncertainty stemming from the elements of the
CRR model. q 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Rainfallrunoff; Conceptual rainfallrunoff models; Uncertainty analysis; Fuzzy logic; Parameter estimation; Fuzzy regression

1. Introduction
The purposes of this paper are (1) to develop a
fuzzy conceptual rainfallrunoff framework to
analyze those parameter uncertainties of conceptual
rainfallrunoff (CRR) models that are related to
data and/or model structure and (2) to apply the methodology to CRR parameter identication under those
uncertainties.
The rainfallrunoff process is usually modied
through (1) CRR models and (2) system theoretic
approaches (Klemes, 1982; O'Connell and Clarke,
1981; Sorooshian, 1983; Young and Wallis, 1985;
* Corresponding author.
zelkan).
E-mail address: ertunga_ozelkan@i2.com (E.C. O

Singh, 1988). CRR models are considered as physically based (although not in the strict scientic sense
of rigid adherence to known physical laws) because
they are designed to reect, albeit in a simplied
manner, the mechanisms that govern the hydrologic
cycle. On the other hand, system theoretic approaches
establish an inputoutput mapping without physical
considerations. For this reason, these models are
referred to as black-box approaches.
In general, a CRR model is composed of submodules that reect mechanisms governing elements
of the hydrologic cycle, where each sub-module is
coupled to others. In addition, because of the threshold parameters representing the capacity of the
conceptual storage tanks, CRR models are `modal'
in structure. As noted by Gupta and Sorooshian

0022-1694/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0022-169 4(01)00430-9

42

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 1. (a) An algorithmic ow diagram resulting from `IFTHEN' rules. Each `IF' represents a branching decision process (e.g. IF X . Y) and
Ci represents the model subcomponent computation on each branch; (b) modal decomposition of the hypothetical ow diagram given in (a).

(1983), this modality behavior is equivalent to the


branching structure in an algorithmic ow diagram
resulting from `IFTHEN' rules. This concept is illustrated in Fig. 1 and represented in a statespace
framework as described in Section 2.
Some of the factors contributing to the uncertainty
of model parameters are the imprecision in data
collection, inadequacy of the data to explain the
physical process, errors in model structure, choice of
objective function used in the parameter identication
process, and global convergence properties of the
optimization algorithm used to calibrate the CRR
model. Sorooshian and Dracup (1980) and Sorooshian
et al. (1983) have pointed out that the choice of objective function based on the stochastic properties (e.g.
based on maximum likelihood theory) of the rainfall
runoff data might improve parameter estimation.
Gupta and Sorooshian (1983) and Sorooshian and
Gupta (1985) have proposed reparameterization to
improve uniqueness and observability properties of
the CRR parameters. Duan et al. (1992) have developed a global optimization algorithm named Shufed
Complex Evolution of the University of Arizona
(SCE-UA), that consistently locates the global optimum, and generally proves to be an efcient and
effective technique; but, as stated by Gan and Biftu
(1996), because of the identiability problem in CRR
models, even sophisticated optimization techniques
might still fail to solve real life problems. Yapo et
al. (1996) have analyzed the sensitivity of the CRR
models to calibration data and concluded that para-

meter uncertainty is greatest when the wettest data


periods on record are used. They also pointed out
that strategies to account for model error during calibration are required. A slightly different perspective
was then followed by Yapo et al. (1998), who have
formulated the parameter estimation problem in a
multiobjective optimization framework. In that
research, the authors proposed a new multiobjective
global optimization algorithm, namely, Multiobjective Complex Evolution, and presented an application
to a bi-objective CRR model using daily root mean
square and heteroscedastic maximum likelihood estimators to analyze the trade-off between tting high
and low ow values (the former tting better the high
ows, and the latter, the low ows).
Based on the above cited research endeavors, it is
usually desirable (1) to establish the sensitivity of the
CRR model to parameter uncertainty, and (2) to
obtain interval estimates for parameters so as to
account for data uncertainties. Traditional uncertainty
analysis utilizes probability theory, that requires
distributional assumptions concerning the random
variables and often exhibits computational difculties
arising from multiple convolutions in the usual case of
dealing with several non-normal dependent random
variables. In this paper, alternative fuzzy logicalbased approaches are investigated for the purpose of
CRR modeling. Fuzzy logic provides a framework for
dealing with uncertainty of model elements. The
proposed fuzzy CRR methodology is not meant to
be a substitute for probabilistic or stochastic

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

approaches, but, as will be shown later, can be useful


when distributional assumptions cannot be justied
and/or computational problems cannot be overcome.
In recent years, fuzzy logic has also been applied in
hydroclimatological systems modeling (Bardossy et
al., 1990; Bardossy and Disse, 1993; Bardossy and
Duckstein, 1995; Hsu et al., 1993; Pesti et al., 1996;
zelkan, 1997; O
zelkan et al., 1996, 1998a, 1999a; ;
O
Galambosi et al., 1998, 1999). Here, we will utilize
concepts such as fuzzy arithmetic (Kaufmann and
Gupta, 1991), fuzzy rule-based modeling (Bardossy
zelkan,
and Duckstein, 1995) and fuzzy regression (O
1997).
In the next section, the fuzzy logic methodology is
illustrated using a linear CRR model (Hsu et al.,
1993), the experimental two-parameter (TWOPAR)
linear reservoir model, and a simplied version of
the Sacramento soil moisture accounting model of
the US National Weather Service river forecast
system (SAC-SMA) known as the six-parameter
(SIXPAR) model. The TWOPAR and the SIXPAR
models, that provide insight into some of the important behavior of more complex CRR models, have
been extensively studied in the literature under
controlled experimental settings in order to demonstrate the difculties faced in the calibration of CRR
models (e.g. Sorooshian and Dracup, 1980;
Sorooshian and Gupta, 1983; Gupta and Sorooshian,
1985a,b; Sorooshian, 1985; Duan et al., 1988, 1992).
In this paper, using these experimental models, it is
shown that the fuzzy logic framework enables the
decision maker to gain insight into the model sensitivity to any element of the CRR model exhibiting
uncertainty, especially sensitivity to the input data.
2. Fuzzy CRR models
A CRR model can be represented in a statespace
framework as follows (Gupta and Sorooshian, 1983):
)
Xtm fm Xt21 ; Ut ; Q
if mode m of the overall
Z tm gm Xt21 ; Ut ; Q system is activated at time t
1
where m 1, , M denotes the CRR system modes
due to the threshold parameters, t 1, , T is the
time index and Xtm and Ztm are the vector of system
states (e.g. amount of ground water storage) and

43

outputs (e.g. runoff) at time t for operational mode


m, Ut is the vector of system inputs (e.g. rainfall)
and Q is the vector of system parameters. As a foundation for the fuzzy CRR models, we will reformulate
Eq. (1) to represent the statespace for each submodule of the CRR system as
)
s
X ts;m fs;m Xt21
; Uts ; Q s
if mode m of the overall
s
Zts;m gs;m Xt21
; Uts ; Q s sub 2 module s is
activated at time t 2
where m 1, , M denotes the system modes due to
the threshold parameters and t 1, , T is the time
index as before, s 1, , S denotes the sub-modules
(e.g. the percolation process), and Xts,m and Zts,m are the
vectors of sub-module states (e.g. amount of ground
water storage) and outputs (e.g. runoff), respectively,
at time t for operational model m, Uts is the vector of
sub-module inputs (e.g. rainfall) and Q s is the vector
of system parameters for sub-module s. Using Eq. (2),
the overall sub-module state and output vectors can be
expressed as
M
X

X ts

s
Is;m Xt21
; Uts ; Q s Xts;m

s
Is;m Xt21
; Uts ; Q s Zts;m

m1

and
Zts

M
X
m1

where
s
Is;m Xt21
; Uts ; Q s

if mode m of submodule s is activated at time t

otherwise

PM

s
s
s
m1 Is;m Xt21 ; Ut ; Q

1. Note that the actiwith


vation of a mode is equivalent to saying that all the
`IF' switches along the path of the mode have a
`TRUE' value. The formulation given in Eqs. (2)
(5) is slightly different from that in Eq. (2) of Gupta
and Sorooshian (1983), who have decomposed the
entire CRR system into modes in spite of the modal
decomposition of each sub-module. In fact, for crisp
(i.e. non-fuzzy) CRR models, both representations
are equivalent, but for the development of the fuzzy
CRR model, the consequences (outputs) from each

44

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

system can be imprecise and therefore modeled as


fuzzy. The relationships for the fuzzy CRR model
can be then represented as follows:
9
X^ s;m
f^s;m X^ st21 ; U^ st ; Q^ s =
t
if mode m of sub6
Z^ s;m
g^s;m X^ st21 ; U^ st ; Q^ s ;module s is activated
t
at time t

Fig. 2. Representation of a linear CRR model.

sub-module need to be evaluated and fed into (or


coupled with) the subsequent process(es) sub-module
zelkan, 1997).
(O
The rst step is to fuzzify the parameters and variables of the CRR model using fuzzy numbers (FNs)
and fuzzy functional operations (as described for
example in Zimmermann, 1985, or Kaufmann and
Gupta, 1991, and briey presented in Appendix A).
Here, it is assumed that every element of the CRR

where the `^' is used to denote a fuzzy number. The socalled fuzzy extensions of Eqs. (3) and (4) can be
written as:
!
M
X
s
^
Xt
I^s;m X^ st21 ; U^ st ; Q^ s X^ s;m
7
t
m1

and
Z^ st

M
X
m1

!
I^s;m X^ st21 ; U^ st ; Q^ s Z^ s;m
t

where I^s;m X^ st21 ; U^ st ; Q^ s [ 0; 1 measures the extent


to which operational mode m 1, , M is activated
within a sub-module s 1, , S with

Fig. 3. Location of the Lucky Hills and the Walnut Gulch experimental watersheds in Arizona.

PM

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

^
^s
^s ^ s
m1 I s;m X t21 ; U t ; Q

1, and is referred to as the


degree of fulllment of operational mode m. The
parentheses are used to emphasize the fuzzy arithmetic
P operations. For example, in Eqs. (7) and (8),
M
m1 represents the summation of fuzzy numbers.
Thus, the crisp CRR model given in Eqs. (2) and (3) is
a special case of the fuzzy CRR model given in Eqs.
(6) and (7), where I^s;m X^ st21 ; U^ st ; Q^ s [ {0; 1}. Notice
that, because I^s;m X^ st21 ; U^ st ; Q^ s takes on values in [0,
1], unlike the crisp CRR model, each mode of the
fuzzy CRR system might be applied with a certain
degree of fulllment. Note also that operations on
fuzzy numbers will be performed at a selected fulllment level, also referred to as membership or a -level
(Kaufmann and Gupta, 1991).
Next, the details of the concept will be illustrated
using a linear CRR (Hsu et al., 1993).
2.1. Fuzzy linear CRR model
Fig. 2 gives a representation of the linear CRR
model analyzed here. The model is based on the
Lucky Hills sub-watershed of the Walnut Gulch
experimental watershed located southeast of Tucson
near the San Pedro River in Arizona (Fig. 3). The
Walnut Gulch experimental watershed (150 km 2),
that is operated by the US Department of Agriculture,
offers unparalleled data collection instrumentation
and a historical database for semi-arid rainfall and
runoff analysis as well as erosion analysis. The
watershed is especially equipped with xed crosssection, critical depth, runoff measuring umes.
These umes are especially designed to accurately
measure sediment laden runoff typical of ephemeral
semi-arid channels. The watershed contains about 100
gauges, about 25 umes and weirs to monitor 43 subwatersheds, ranging in size from 0.3 to 112 km 2, that
are operated for a variety of studies including the
effects of land use management, erosion monitoring
and small scale rainfall variability effects on rainfall
runoff modeling. Soil texture is mostly gravelly sandy
loam and the vegetation is mostly desert shrub (whitethorn, creosotebush, and tarbush) and grassland (black
grama and curly mesquite). A more detailed description of the Walnut Gulch watershed can be found in
Renard et al. (1993) as well as in the world wide web
pages of the US Department of Agriculture (http://
www.tucson.ars.ag.gov/gis/wg.html). Lucky Hill,

45

that is one of the relatively smaller sub-watersheds


of Walnut Gulch, has a catchment area of
44,000 m 2. In Fig. 2, the upstream tank represents
the combined upstream runoff from sub-watersheds
`102' and `106' of Lucky Hills. Similarly, the downstream runoff from the sub-watershed `104' is
modeled by the downstream tank. This linear CRR
model has only one mode due to the innite capacity
reservoir level assumption, i.e. 0 # Xt # 1, and is
characterized by the following continuity relation:
Zj;t 2 Zj11;t 1 Ut

dXt
dt

where Xt f Zj;t ; Zj11;t ; Ut (units of length) is the


reservoir level (state of the model) of the downstream
tank, Ut is the rainfall, Zj,t and Zj11;t (units of length/
time) are the channel ows from the upper and lower
storage tanks, respectively. Assuming a linear functional relationship for the reservoir level Xt, and replacing the derivative term in Eq. (9) with a forward
difference relation, yields the following equation:
Zj11;t C0 1 C1 Zj11;t21 1 C2 Zj;t 1 C3 Zj;t21 1 C4 Ut
1 C5 Ut21
10
where C0 is constant and Ci, i 1, 5 are parameters
that can be expressed as:
C1

K2
;
Dt 1 K1

C3

K1
Dt 1 K1

C5

K3
:
Dt 1 K1

C2
C4

Dt 2 K2
;
Dt 1 K1
Dt 2 K3
;
Dt 1 K1

The fuzzication of the linear CRR model (Eq. (10))


consists of replacing each variable and parameter by a
fuzzy number, and ordinary arithmetic operations by
fuzzy operations written as:
Z^ j11;t C^ 0 1C^ 1 Z^ j11;t21 1C^ 2 Z^ j;t 1C^ 3
Z^ j;t21 1C^ 4 U^ t 1C^ 5 U^ t21 :

(11)

Note that Eq. (11) is essentially a linear model in all of


its parameters, so that the parameters can be obtained

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

46

the fuzzy reservoir level (state of the model) at the


beginning of the time interval t and U^ t be the fuzzy
rainfall (units of length) during the time interval t.
Then, performing the same fuzzication procedure
as for Eq. (10), the fuzzy intermediate state can be
written as follows:
X^ t X^ t21 1U^ t :

Fig. 4. Representation of the TWOPAR model.

using a so-called fuzzy linear regression model as


developed later in this paper.
2.2. Fuzzy TWOPAR model
Fig. 4 gives a sketch of the TWOPAR model. There
are two parameters to be estimated: K (1/units of
time), the recession coefcient and C (units of length),
the threshold parameter that reects the capacity of
the reservoir. Here, St (units of length/time), and Rt
(units of length/time) are the interow and overland
ow from the reservoir, respectively. The combination of the inter- and overland ows gives the channel
ow (Zt).
The fuzzication of the TWOPAR model is
performed as follows: let X^ t21 (units of length) be

12

Let the two modes of the TWOPAR system be reformulated as fuzzy rules.
Rule 1.

if X^ t #C^ then

^
X^ t;1 12K
X^ t21 1U^ t ; R^ t;1 0; S^ t;1
^
K
X^ t21 1U^ t :

Rule 2.

if X^ t .C^ then

^ R^ t;2 X^ t21 1U^ t 2C;


^ S^ t;2
^
X^ t;2 12K
C;
^
^ C:
K

13

Let I^t;1 , I^t;2 be, respectively, the degrees of fulllment

Fig. 5. Representation of the SIXPAR model.

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

of rules 1 and 2, that can be computed as described in


Appendix A in Eqs. (40) and (41). The consequence
state and output from the fuzzy rules 1 and 2 can then
be evaluated as
X^ t I^1;t X^ t;1 1I^2;t X^ t;2

14

R^ t I^1;t R^ t;1 1I^2;t R^ t;2

15

S^ t I^1;t S^ t;1 1I^2;t S^t;2 :

16

47

ciency ratio. Then, the percolation computation is


carried out as follows:
Rule P1.

^
if D#
A^ then

^U
^L
^U
^
^L
P^ t;1 X^ U
t21 1U t :C C K 1C 2
^ W^ ; E^ U
^ A
^U
^
C^ L K^ L D:
t;1 X t21 1U t
18

Finally, the TWOPAR runoff is obtained as


Z^ t R^ t 1S^t :

17

A numerical application of this equation is given in


the `Application' Section 3.
2.3. Fuzzy SIXPAR model
Fig. 5 gives a schematic representation of the
SIXPAR model which, as mentioned earlier, is a
simplied version of the SAC-SMA model. Although
some of the hydrologic components, such as evaporation and tension water reservoirs are not included in
SIXPAR, the major components of the SAC-SMA
model, that are the two-layer structure and the percolation feature, are still preserved. There are six parameters to be studied: K U and K L (1/units of time) are
the recession coefcients of the upper and lower zone
storages, respectively, and C U and C L (units of length)
are the threshold parameters that show the capacity of
the storages. The other two parameters, A and W, that
are dimensionless, are part of the percolation Pt relation. Here, Rt, St and Bt (units of length/time) are the
overland ow from the upper reservoir and interows
from the upper and lower storages, respectively. The
combination of the inter- and overland ows gives the
channel ow Zt. XtU and XtL (units of length) represent
the storage levels (states of the model) and Ut (units of
length) the rainfall (input) during the time interval t.
The SIXPAR model computes percolation rst, then
lower zone relations, and nally the upper zone state
and outputs. The SIXPAR model ow diagram and
the modal decomposition are shown in Fig. 5. Next,
the relations associated with each of these processes
are presented in terms of fuzzy operations.
2.3.1. Percolation
Let D^ C^ L 2X^ Lt21 :C^ L be the lower zone de-

^ ^U
^
Rule P2. if D.
A^ then P^ t;2 X^ U
t21 1U t ; E t;2 0;
U
^
where E denotes the remaining water in the upper
storage after the percolation process has taken place.
The consequences from the percolation sub-module
are then obtained as:
P^ t I^P1;t P^ t;1 1I^P2;t P^ t;2

19

^
^U
^
^U
E^ U
t I P1;t Et;1 1I P2;t E t;2

20

where I^P1;t and I^P2;t are rule fulllments computed


similar to Eqs. (40) and (41).
2.3.2. Lower zone
Let X^ Li X^ Lt21 1P^ t be the fuzzy lower zone intermediate state. Then the lower zone computations
follow as:
Rule L1.

L
if X^ t #C^ L then

B^ t;1 X^ t K^ L ; X^ Lt;1 2B^ t ; E^ Lt;1 0

Rule L2.

L
^
if X^ t .C^ L AND T#
C^ L then

^ E^ Lt;2 0
B^ t;2 C^ L K^ L ; X^ Lt;2 T;

Rule L3.

21

L
^
if X^ t .C^ L AND T.
C^ L then

^
B^ t;3 C^ L K^ L ; X^ Lt;3 C^ L ; E^ Lt;3 T2
C^ L
where E^ L denotes the
excess water from the lower
L
storage and T^ X^ t 2CL K^ L : Then, the lower
zone consequences from the lower zone rules can be

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

48

computed as:
B^ t I^L1;t B^ t;1 1I^L2;t Bt;2 1I^L3;t B^ t;3

22

X^ Lt I^L1;t X^ Lt;1 1I^L2;t X^ Lt;2 1I^L3;t X^ Lt;3

23

E^ Lt I^L1;t E^ Lt;1 1I^L2;t E^ Lt;2 1I^L3;t E^ Lt;3 :

24

2.3.3. Upper zone


U
^L
Let X^ t E^ U
t 1E t be the upper zone intermediate
state. Then similar to the percolation and lower zone
processes, the upper zone computations follow as:
Rule U1.

U
if X^ t #C^ U then

U
if X^ t .C^ U then

^U ^
^U ^
^U
^U
X^ U
t;2 12K C ; St;2 K C ; Rt;2
U
X^ t 2C^ U

^
^U
^
^U
X^ U
t I U1;t X t;1 1I U2;t X t;2

25

S^ t I^U1;t S^ t;1 1I^U2;t S^ t;2

26

R^ t I^U1;t R^ t;1 1I^U2;t R^ t;2 :

27

Finally, the SIXPAR runoff can be obtained as


Z^ t R^ t 1S^ t 1B^ t :

Step 1: Obtain the CRR model parameters using


classical optimization techniques.
Step 2: Select the input variable(s) and/or parameter(s) to be analyzed.
Step 3; Introduce fuzziness into the selected CRR
element(s).
Step 4. Run the fuzzy CRR using the dened fuzziness to generate fuzzy runoff.
Step 5. Measure the impact of fuzziness using the
sum of square vagueness (SSV) dened as
X
SSVa
z2 a 2 z1 a2
t

U
^ U ^
^ ^
^U
^U
X^ U
t;1 12K X t ; St;1 K X t ; Rt;1 0

Rule U2.

excluded from this analysis because linearity in the


parameters renders a sensitivity analysis trivial.
The sensitivity analysis methodology can be
summarized as follows:

28

3. Application of the fuzzy CRR method for


sensitivity analysis purposes
In this section, a methodology is presented to use
the fuzzy CRR technique to enable the decision maker
to gain insight about the model sensitivity with respect
to the elements constituting the CRR model. The
concept is illustrated using the TWOPAR and the
SIXPAR models. The fuzzy linear CRR model is

where z1(a ) and z2(a ) correspond to the lower and


upper condence levels of the fuzzy runoff Z^ t a at
a selected a -level and t, , T is the time. Thus, the
fuzzy runoff is represented as a fuzzy interval, at a
selected a -level, as Z^ t a z1 a; z2 a:
Suppose that the optimal parameter values have
been determined as crisp (i.e. non-fuzzy) numbers
using classical methods, but that the rainfall data
and thus the parameters are considered to be fuzzy.
As stated in the Introduction, there can be a combination of factors contributing to the uncertainty of the
model parameters, namely, imprecision and inadequacy of the data, not just rainfall, but also runoff
data, errors in the chosen model structure, the sensitivity and/or bias of the chosen objective function
used in the parameter identication process and the
fundamental weaknesses and inefciencies associated
with and, indeed, inherent in the standard optimization methods. In the case under consideration here, the
selected true parameter values are as follows: for the
TWOPAR model: K 0.5 l/h and c 50 mm, and for
the SIXPAR model: C U 10 mm, K U 0.2 l/h,
C L 20 mm, K L 0.5 l/h, A 0.31, and W 3.
The parameter values for the SIXPAR model are
taken as the experimental values given by Duan et
al. (1992). The sensitivity scenarios include: (1) the
rainfall Ut being fuzzy; (2) the individual parameters
being fuzzy, i.e. two cases for the TWOPAR model,
and six cases for the SIXPAR model; and (3) all the

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

49

Fig. 6. Sensitivity analysis of the TWOPAR model: sum of square vagueness (SSV) of simulated fuzzy channel ow for different sensitivity
scenarios.

parameters and rainfall being fuzzy. All together, four


experimental scenarios for the TWOPAR model and
eight for the SIXPAR model are analyzed in order to
identify the individual and combined effects of the
parameters and rainfall data vagueness on the CRR
system output. Furthermore, each sensitivity scenario
is computed for 5, 10, 15, 20, 25, and 30% fuzziness,
where the percent fuzziness denes the spread of the
parameters, e.g. a 10% fuzzy recession coefcient
with the most likely value K 0.5 l/h would mean
that the parameter uncertainty ranges between 0.45
and 0.55 l/h, denoted as K^ [0.45, 0.55]. All computations on fuzzy numbers are carried out at the a 0
level (see Appendix A), that corresponds to an interval analysis, but does not restrict the generality of the
approach (Kaufmann and Gupta, 1991). Again, the

reason for introducing the fuzziness independently


into each parameter is to measure the sensitivity of
the model with respect to each individual parameter
uncertainty.
Figs. 6 and 7 give the SSV for the channel ow the
for the different sensitivity cases. In the case of the
TWOPAR model, the output vagueness is seen to be
more sensitive to the parameter C than to the rainfall
or to K (Fig. 6). Similarly, for the SIXPAR model, it is
found that for the vagueness #15%, the parameter C L
accounts for most vagueness, then comes the vagueness effect of rainfall, C U, K L, and K U. On the other
hand, for vagueness $20%, the effect of rainfall, C U,
and K L is followed by C L in terms of runoff vagueness.
In general, the SIXPAR model is least affected by A
and W. For both the TWOPAR and the SIXPAR

Fig. 7. Sensitivity analysis of the SIXPAR model: sum of square vagueness (SSV) of simulated fuzzy channel ow for different sensitivity
scenarios.

50

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 8. Sensitivity analysis of the TWOPAR model by introducing 5% vagueness to rainfall data, C (mm) and K (l/h): fuzzy rainfall series,
storage level, and the channel ow ( upper fuzzy condence level; lower fuzzy condence level).

models, increasing magnitude of vagueness results in


an increasing runoff vagueness except for C L of the
SIXPAR model, for which a decrease from 10 to 20%
is observed (Fig. 7). Figs. 8 and 9 give the rainfall
series, storage levels, percolation (for the SIXPAR
model) and channel ow for sensitivity case 1, for
the 5% vagueness scenario. As can be seen from the
results, the effect of uncertainty on the outcome is
well observable through the application of the
proposed fuzzy methodology.
4. Parameter identication
In order to explore more fully the fuzzy CRR model
just developed, the associated parameter identication
issues are now investigated.
Calibration of a fuzzy CRR model results in a fuzzy
optimization problem that can be analyzed under the
general fuzzy regression framework developed by
zelkan (1997); O
zelkan et al. (1998b), and O
zelkan
O
and Duckstein (2000). Note that regression is a general

term used to denote a relation between variables and,


without loss of generality, many optimization problems
can be treated within a regression analysis framework.
In this context, the calibration issues of fuzzy CRR
models using fuzzy regression techniques are now
addressed. As shown in the literature (e.g. Tanaka et
al., 1982; Bardossy, 1990) in general, fuzzy regression
gives rise to substantial mathematical programming
problems. Since CRR models are usually non-linear
and non-convex in structure, one would expect to
face difculties in applying fuzzy regression. For this
reason, we have selected the three CRR models
explained earlier in Section 2, which possesses different degrees of complexity. The rst model is the linear
CRR model that would yield a fuzzy linear regression
problem that can be solved using linear programming
techniques (Tanaka et al., 1982). Additionally, here, a
newly developed technique called multi-objective
fuzzy regression (MOFR) is used, that has been
shown to overcome the shortcomings of regular fuzzy
zelkan and Duckstein, 2000). The
linear regression (O
other two models are also those used earlier in this

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

51

Fig. 9. Sensitivity analysis of the SIXPAR model by introducing 5% vagueness into rainfall data, C U (mm), K U (l/h), C L (mm), K L (l/h), A and
W: rainfall series, percolation, lower storage level, upper storage level, and the channel ow ( upper fuzzy condence level; lower
fuzzy condence level).

paper, namely the TWOPAR model and the SIXPAR


model. The non-linear TWOPAR and SIXPAR models
are calibrated in a fuzzy least squares regression framework where the resulting optimization problem is
solved using the so-called SCE-UA technique of
Duan et al. (1992).
5. Fuzzy regression
5.1. Principles
In general, regression analysis can be considered as
the search for relations among variables (Bardossy,
1990), having the general form
Ot PXt ; Q;t;

29

where Xt is the vector of input data, that includes


the `modal' states (e.g. the amount of ground

water storage, and rainfall), Q is the parameter


vector (e.g. C and K for TWOPAR) and Ot is
the output (runoff), that we are trying to predict.
Generally, there is no exact solution to Eq. (29),
because errors associated with the model estimates
constitute conicting objectives, thus, resulting in
zela multi-objective decision making problem (O
kan and Duckstein, 2000). Therefore, some form
of relaxation method is needed. In classical regression analysis, this is done by introducing a
stochastic error term e , with Ee t 0, and, if
possible, homoscedasticity of the errors such that
Varet s 2 , giving
Ot PXt ; Q 1 e t :

30

Another form of relaxation used in the fuzzy


regression methodology is to take the parameter
vector, Q as a fuzzy quantity, with
^ X^ t ; Q^ :
O^ t P

31

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

52

For the TWOPAR model, the fuzzy parameter


^ K},
^ that can be expressed at a
vector is Q^ {C;
selected a -level as

^ a c 1 a; c2 a; K
^ a
Q^ a {C

k1 a; k2 a}:
Thus, for fuzzy regression at a selected a -level,
each fuzzy parameter requires that two parameters
(lower and upper condence limits) be estimated.
We thus have 2 2 4 parameters to be estimated for the TWOPAR model, and 6 2 12
parameters for the SIXPAR model. As shown in
the literature (e.g. Tanaka et al., 1982; Bardossy,
1990), there is an innite number of solutions to
the fuzzy regression problem; the aim is, then, to
seek the optimal fuzzy parameter vector Q^ p , such
that a measure of vagueness (or fuzziness) or the
prediction is minimized.
There are mainly two directions in fuzzy regression model building: fuzzy linear regression as
proposed by Tanaka et al. (1982) and its variations, and fuzzy least squares regression as developed by Diamond (1988) and Celmins (1987a,b).
Before summarizing the research done in each of
these two directions, the rationale for the why and
when to use fuzzy regression is discussed.
5.2. Why and when to use fuzzy regression?
Some of the major reasons listed in the literature for using fuzzy regression analysis instead of
statistical regression analysis can be summarized
as follows:
when the data set is insufcient to support

statistical regression analysis (Bardossy et al.,


1990, 1992; Bardossy, 1990)
when statistical distributional assumptions
cannot be justied (Tanaka et al., 1982)
if the aptness of the regression model is poor
(Kim et al., 1996)
when human judgements are involved, i.e.
inputs and/or outputs are fuzzy numbers
(Tanaka et al., 1982)
if the errors are associated with the indeniteness of the model structure and with the vagueness of human perception of the model, in
contrast with the non-Bayesian statistical case
where the errors are associated with observations (Tanaka et al., 1982).

5.3. Direction 1: fuzzy linear regression


Since the development of fuzzy linear regression by
Tanaka et al. (1982), there have been many modications and alternatives to the original model. In general,
a fuzzy regression model can be formulated as
min V such that
^ X^ t a; Q^ a;t
O^ t a # P

(32)

where V denotes the vagueness measure to be


minimized, and O^ t a and P^ t X^ t a; Q^ a are the
a -level sets for observations and predictions, respectively. In the remainder of this section we will use P^ t
to denote P^ t X^ t ; Q^ for the sake of simplicity. The
constraints in Eq. (32) basically state that the
predicted values must include the observed ones
(Fig. 10). Note that P^ t is a function of the input
vector X^ t a x^1;t a; ; x^N;t aT and parameters

Fig. 10. Illustration of the conditions for fuzzy linear regression: at a selected a -level; predicted runoff Z^ pt includes the observed Z^ t runoff value.

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Q^ a u^ 1 a; ; u^ N a . For fuzzy linear


regression, we have P^ t a X^ t aT Qa.
Furthermore, recognizing that the observations and
predictions can be represented as intervals at chosen
a -levels, such that O^ t a OL;t a; OR;t a and
P^ t a PL;t a; PR;t a, each constraint in formulation (32) can be replaced by two equivalent
constraints as follows:
T

min V such that


OL;t a $ PL;t a; OR;t a # PR;t a;t:
The original fuzzy linear regression model treated
uncertainties associated with system structure, and,
thus, considered the linear regression parameters to
be fuzzy. Although the model allowed output data to
be fuzzy, it considered only crisp (non-fuzzy) input
(`independent variables') data. The objective function V was taken as the spread of the parameters, that
were represented by symmetrical triangular fuzzy
numbers (TFNs). In a subsequent paper, Tanaka
(1987) proposed instead the minimization of the
spread of the predictions.
A slightly different point of view on fuzzy regression was presented by Sakawa and Yano (1992), who
constructed three bi-objective fuzzy regression
formulations using vagueness as one objective, and
maximization of the a -level as another.
5.4. Direction 2: fuzzy least squares regression
This other direction in fuzzy regression model
building was developed by Diamond (1988). His
method solves an unconstrained least squares problem
and yields solutions analogous to ordinary least
squares. The fuzzy least squares regression formulation is
X
min D2 O^ t ; P^ t
33
t

where P^ t a X^ t aT Qa, as in fuzzy linear


regression, and D 2() is a distance measure between
observations and predictions. The distance measure
dened by Diamond (1988) has the form
X
D 2 O^ t ; P^ t
OL;t 0 2 PL;t 02 1 OL;t 1
t

2 PL;t 12 1 OR;t 0 2 PR;t 02 : 34

53

Along the same lines, Bardossy et al. (1992) used a


somewhat more general distance measure, (that is a
generalization of the Euclidian distance), namely the
so-called Hagaman distance:
X Z1
{OL;t a 2 PL;t a2 1 OR;t a
D2 O^ t ; P^ t
t

2 PR;t a2 }f ada

(35)

on [0, 1] such
where f(a ) is a continuous function
R
that f(a ) . 0 if a . 0 and 10 f ada 1=2. As
shown in Bardossy et al. (1992), Eq. (35) yields the
ordinary least squares estimators when the model
input and output are assumed to be crisp (i.e. nonfuzzy).
5.5. Drawbacks of fuzzy regression
Besides the advantages of fuzzy regression that
were listed in Section 5.2, there have been many criticisms of the original fuzzy linear regression formulation. In contrast, the fuzzy least squares regression has
received very little criticism because of its similarity
with the traditional least squares regression formulation, whose problems have long been recognized. The
major problems with fuzzy linear regression can be
summarized as follows: (1) it does not allow all data
points to inuence the estimated parameters (Savic
and Pedrycz, 1991); (2) it is sensitive to data outliers;
(3) prediction intervals become wider as more data
are collected (Redden and Woodall, 1994, 1996);
(4) it may yield some crisp parameter estimates
(Celmins, 1987a,b); and (5) it is scale dependent
(Jozsef, 1992).
To remedy these drawbacks, one may cite the methods developed by Celmins (1987a,b), Savic and
zelkan
Pedrycz (1991), and Peters (1994). Recently, O
(1997) has proposed a large set of multi-objective
fuzzy regression models based on the trade-off
between data outliers and prediction vagueness:
these models are applied below to the parameter estimation problem.
5.6. Multi-objective fuzzy regression (MOFR) models
The rst MOFR formulation, that results in a biobjective fuzzy regression (BOFR), addresses the
data outlier problem in Tanaka et al.'s (1982) fuzzy

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

54

linear regression. The BOFR formulation is as


follows:
BOFR.

min-dominate {V; E } such that

PL;t a 2 OL;t a # eL;t ; OR;t a 2 PR;t a # eR;t ;t

eL;t ; eR;t $ 0;t:

(36)

PL;t a 2 OL;t 1 2 a # eL;t ; OR;t 1 2 a 2 PR;t a


# eR;t ;t; eL;t ; eR;t $ 0;t

TOFR3.

max-dominate {V; a; 2Ep } such that

OL;t a 2 PL;t 1 2 a # eL;t ; PR;t 1 2 a 2 OR;t a


# eR;t ;t; eL;t ; eR;t $ 0;t

The solution of Eq. (36) is the set of non-dominated solutions known as the Pareto solution set.
The nal compromise solution is chosen according
to the decision maker's relative preference
between the two objectives. In Eq. (36), from
which `min-dominate' means nding the nondominated Pareto optimum set for minimizing a
multi-objective problem, eL;t and eR;t can be
considered
as relaxation variables and Ep
P p
p
t eL;t 1 eR;t is the Lp norm, where 1 # p # 1
indicates the compensation level between the
relaxation variables. The most commonly used
compensation levels are p 1, meaning compensation from which we obtain the sum of relaxation
variables, p 2, for the sum of squares of the
relaxation variables, and p 1, yielding the
max{eL;t ; eR;t } and indicating that no compensation
is allowed. Generally, the vagueness V is also
dened as an aggregated expression of the individual prediction vagueness. In Eq. (36), all the
constraints are active and, therefore, all the data
points contribute to the parameter estimation.
Next, using similar concepts to those embodied in
Eq. (36), the models of Sakawa and Yano (1992) are
extended to a tri-objective fuzzy regression (TOFR)
formulation, as follows:
TOFR1.

min-dominate {V; 2a; Ep } such that

PL;t a 2 OR;t a # eL;t ; OL;t a 2 PR;t a


# eR;t ;t; eL;t ; eR;t $ 0;t

TOFR2.

min-dominate {V; 2a; Ep } such that

(37)

(38)

(39)

An investigation of the properties of the BOFR and


the TOFR formulations has shown that both the existing fuzzy regression and the classical regression techniques are specic cases of the proposed methodology
zelkan, 1997; O
zelkan and Duckstein, 2000).
(O
6. Application and results
In this section, the linear CRR model (Hsu et al.,
1993), the experimental TWOPAR model, and the
SIXPAR model are used to illustrate the methodology
proposed above. The parameters of the linear CRR
model are obtained using the linear BOFR and
TOFR techniques described in the previous section.
The non-linear TWOPAR and SIXPAR models are
then calibrated using fuzzy least squares regression
that is suitable for non-linear parameter identication.
6.1. Use of MOFR to calibrate the linear CRR model
In this application, the MOFR solution is obtained:
(1) using compromise programming that seeks to
minimize an Lp-norm of the distance between an
ideal point (the non-feasible point whose coordinates
are the optimum of each objective function taken
separately) and the Pareto optimum; and (2) using
the e -constraint technique, that gives the decision
under the ability to specify bounds on the objective
function in a sequential manner. Further details are
given in texts such as those of Goichoechea et al.
(1982); Szidarovszky et al. (1986). More specically,
L1 and L1 are used for compromise programming in
order to obtain linear programming formulations. For
the e -constraint solution, the vagueness objective

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

55

Fig. 11. Obtaining Pareto optimal solutions for BOFR for the rainfallrunoff example using the e -constraint technique (output 10% fuzzy).

56

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 12. Rainfallrunoff estimation using BOFR. Calibration event Lucky Hills, 1 August 1974, and comparison of the results with Tanaka et
al.'s (1982, 1987) model (output 10% fuzzy).

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

function V has been constrained in all cases. The


relaxation variables are assumed to be symmetric
(i.e. eL;t eR;t et ) and fuzzy quantities are represented by symmetrical triangular fuzzy numbers,
where each fuzzy parameter C^ i , i 0, , 5 is represented by the most likely value xip and the fuzzy spread
b i.
The data are obtained for the sub-watershed `Lucky
Hills' of the experimental watershed `Walnut Gulch'
in Arizona, as described in Section 2.1. Three
rainfallrunoff events occurring on 27 July 1973, 1
August 1974, and 17 July 1975, were selected among
multiple events to illustrate the approach. The idea for
selecting these three events was to get a representation
of a small (1 August 1974), medium (27 July 1973),
and large (17 July 1975) rainfallrunoff event for this
experimental watershed. Computations were done by
using one event as a model calibration event and the
other two for model validation. Here, we will present
the case where the 1 August 1974, event is used for
model calibration and the other two events for model
validation purposes.
Results are provided for the case when runoff
observations are assumed to be 10% fuzzy. Fig.
11 shows the Pareto optimal solutions for the
BOFR problem, obtained using the e -constraint
technique, for a range of values of the vagueness
objective function V. Comparison of the BOFR
solutions with that of Tanaka et al. (1982) is
provided in Fig. 12 for the calibration event. Similarly, Fig. 13 shows the corresponding results for
TOFR1, TOFR2, and TOFR3. The validation
results for BOFR are shown in Figs. 14 and 15.
The TOFR validation outputs, that are not shown,
yield similar results. In general, the results show
that BOFR and TOFR are capable of modeling the
rainfallrunoff events. As seen from the results,
BOFR is more robust to data outliers, whereas it
results in a wider prediction interval than that of
Tanaka et al.'s model for which the constraints are
crisp. In the case of TOFR, Sakawa and Yano's
third model does not yield feasible solutions,
whereas TOFR3 does. For purposes of comparison, the parameter estimates for BOFR are
provided in Table 1. For the TOFR results, the
a -level objective is optimized among the a 0,
0.5, and 0.75 levels. The results presented in Fig.
13 correspond to the a 0.5 level.

57

6.2. Use of fuzzy least squares regression to calibrate


the TWOPAR and SIXPAR models
In this part of the analysis, we have tried to identify
the parameters of the fuzzy CRR models after perturbing synthetically produced rainfall and runoff data.
For the TWOPAR model, the parameters are C^ and
^ and for the SIXPAR model the parameters C^ U ; K^ U ,
K,
C^ L , K^ L ; A^ and W^ need to be estimated. Since we are
making our analysis at the a 0 level, each parameter is replaced by an interval, resulting in two parameters for each fuzzy parameter. Thus, four
parameters need to be estimated for the TWOPAR
model and 12 for the SIXPAR model. For experimental design purposes, both heteroscedastic (error
variance is proportional to the magnitude of measurement) and homoscedastic (constant error variance)
error effects for the 0% (error free), 10 and 20%
cases are investigated. The 25 possible different
experimental scenarios are simulated, and the parameter estimates are obtained using the fuzzy least
squares regression technique, as explained in Section
5.3. The resulting non-linear optimization problem is
solved using the SCE-UA method of Duan et al.
(1992). For fuzzy least squares regression, the fuzziness of the rainfall and runoff are assumed to be equal
to the contaminating error magnitude, i.e. if the rainfall is 10% error contaminated (heteroscedastic or
homoscedastic), 10% fuzziness is assumed, e.g. if
rainfall U1 10 mm then fuzzy rainfall is U^ 1 [9,
11] mm. As noted in Section 5.3, if both the rainfall
and runoff are assumed to be crisp (non-fuzzy), then
ordinary least squares estimates are obtained for the
parameters. Here, we have computed the ordinary
least squares estimates for comparison purposes.
The results of the experimental study, for both parameters of the TWOPAR model and for two of the six
parameters of the SIXPAR model are given in Figs.
1619. The x-axis in these gures shows the percent
of error contamination and the type of error. For
example, `10 he' indicates 10% heteroscedastic
error. Thus, in Figs. 16 and 17, for the `0-10 he'
case, one nds C^ [8.95, 10.13] and K^ [0.480,
0.515], meaning that the storage capacity and recession parameters can take on values in these intervals at
the a 0 level.
One can verify that, in the no-error and no-fuzziness case, the fuzzy least squares regression scheme

58

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 13. Rainfallrunoff estimation using TOFR. Calibration event Lucky Hills, 1 August 1974, and comparison of the results with Sakawa and
Yano's (1992) model, a 0.5 (output 10% fuzzy).

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

59

Fig. 14. Rainfallrunoff estimation using BOFR. Validation event Lucky Hills, 27 July 1973, and comparison of the results with Tanaka et al.'s
(1982, 1987) model (output 10% fuzzy).

60

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 15. Rainfallrunoff estimation using BOFR. Validation event Lucky Hills, 17 July 1975, and comparison of the results with Tanaka et al.'s
(1982, 1987) model (output 10% fuzzy).

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

61

Table 1
Estimated parameters for the rainfallrunoff example: runoff 10% fuzzy
Parameter

x0p
b0
x1p
b1
x2p
b2
x3p
b3
x4p
b4
x5p
b5
V

Tanaka et al.

BOFR: e -const. V # S ie i

BOFR: L1

BOFR: L1

0.5

0.75

0.5

0.75

0.5

0.75

0.5

0.75

0.045
0.024
0
13.132
0
8.205
0.844
0
1.013
0.355
0.254
0
25.9

0.043
0.044
0
30.978
0
16.516
1.053
0
0.877
0.372
0.225
0
47.2

0.041
0.084
0
66.670
0
33.136
1.158
0
0.809
0.408
0.210
0
89.8

0.062
0.202
0
0
0
0
0
0
2.210
0
0.215
0
7.7

0.062
0.216
0
0
0
0
0
0
2.210
0
0.215
0
8.2

0.062
0.227
0
0
0
0
0
0
2.210
0
0.215
0
8.6

0.060
0.010
0
0
0
0
0
0
2.210
0
0.215
0
0.4

0.062
0.021
0
0
0
0
0
0
2.210
0
0.215
0
0.1

0.062
0
0
0
0
0
0
0
2.210
0
0.215
0
0.1

0.064
0.012
0
0.630
0
1.489
0
0
2.210
0.167
0.212
0.026
7.8

0.063
0.020
0
0.364
0
3.111
0
0
2.210
0.136
0.215
0.024
8.3

0.060
0.016
0
0
0
5.944
0
0.141
2.211
0
0.215
0.001
8.7

should recover the true model parameters. Our studies


have shown that, although the true parameters were
easily recovered for the fuzzy TWOPAR model, the
true parameters for the fuzzy SIXPAR model could
not be recovered when all six parameters were estimated. More specically, when the percolation parameters A and P were included in the optimization
process, multiple optimization runs using the SCEUA algorithm resulted in premature convergence of
the model parameters to different local optima. This
result is consistent with the ndings of other researchers, who have similarly experienced difculties in the
identiability of CRR models (e.g. Gupta and

Sorooshian, 1983; Sorooshian and Gupta, 1985; Gan


and Biftu, 1996). This point has, indeed, been emphasized in the literature, especially the problems associated with the percolation process (Gupta and
Sorooshian, 1983). For this reason, in the present
study, the percolation process parameters were
excluded from the optimization process and xed at
their true values. After this adjustment, the remaining
four parameters of the SIXPAR model were successfully recovered. The results provided for the SIXPAR
model, in Figs. 18 and 19, include the optimization
results of the upper zone fuzzy parameters.
For the TWOPAR model, it has been observed that

Fig. 16. Fuzzy regression estimates for the TWOPAR parameter C and their comparison with the crisp estimates under different experimental
error scenarios (he: heteroscedastic; ho: homoscedastic).

62

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

Fig. 17. Fuzzy regression estimates for the TWOPAR parameter K and their comparison with the crisp estimates under different experimental
error scenarios (he: heteroscedastic; ho: homoscedastic).

Fig. 18. Fuzzy regression estimates for the SIXPAR parameter C U and their comparison with the crisp estimates under different experimental
error scenarios (he: heteroscedastic; ho: homoscedastic).

Fig. 19. Fuzzy regression estimates for the SIXPAR parameter K U and their comparison with the crisp estimates under different experimental
error scenarios (he: heteroscedastic; ho: homoscedastic).

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

even in cases where crisp optimization resulted in


unstable parameter estimates, the fuzzy case gave
more reasonable estimates (see cases 20 he-0, 20 he20 he, 20 he-20 ho, 10 ho-20 he, and 20 ho-20 he).
Similarly, for the SIXPAR model, more reasonable
estimates were obtained for the 0-20 he, 20 he-20
he, 10 ho-10 he, 10 ho-20 he, and 20 ho-20 he
cases, when both rainfall and runoff data are assumed
to be fuzzy.
For an objective comparison of the crisp and fuzzy
least squares parameter estimation, the relative sum of
squares of error (RSSE) between the estimated parameter and the true parameter value is computed as
!
X up;j 2 uT;j 2
RSSE
uT;j
j
where up;j and uT;j represent the predicted and true
parameter values, respectively. The comparison of
these results is provided in Table 2. The RSSE values
conrm that, for the TWOPAR and the SIXPAR
models, the fuzzy estimates are more stable than the
crisp ones.
7. Discussion and conclusions
After developing fuzzy conceptual rainfallrunoff
models, the calibration of these models has been
investigated using fuzzy regression techniques. First,
multi-objective fuzzy regression (MOFR), more
specically bi-objective fuzzy regression (BOFR)
and tri-objective fuzzy regression (TOFR) techniques
were applied to identify the parameters of a linear
conceptual rainfallrunoff model. These models
were applied to the sub-watershed `Lucky Hills' of
the experimental watershed `Walnut Gulch' in
Arizona. Three widely different events were selected
for the study; using one event for calibration, validation gave good results for the other two. Although it is
rather difcult to make a wide generalization from the
few data sets used in this study, the results show that
MOFR has potential for reliable rainfallrunoff
modeling and for reducing possible model error in
the case of uncertain data. While the aim of this
paper is not to compare the MOFR results with the
traditional regression approaches, the reader should
zelkan (1997), standard regresnote as shown by O

63

sion, that includes traditional least squares regression,


is a special case of the proposed MOFR methodology.
Second, two non-linear fuzzy conceptual rainfall
runoff models, namely the TWOPAR and the
SIXPAR have been analyzed using fuzzy least squares
regression, leading to unconstrained optimization
problems. This analysis has been performed using
synthetic rainfallrunoff data under controlled heterogeneous and homogeneous error scenarios. For the
case of fuzzy least squares regression, the assumption
of no fuzziness is equivalent to a crisp (ordinary) least
squares optimization: the corresponding results
obtained are summarized in Table 2. These results
show that, in general, the fuzzy parameter estimates
tend to be more robust than their corresponding crisp
counterparts.
The main conclusions of this analysis can be
summarized as follows.
Calibration of fuzzy conceptual rainfallrunoff
models using fuzzy least squares regression technique, with fuzziness introduced into both rainfall
and runoff, yielded usually more stable parameter
estimates than those obtained assuming non-fuzziness or crispness of the rainfall and runoff.
The fuzzy conceptual rainfallrunoff model
requires more parameters to be estimated than the
standard conceptual rainfallrunoff model since,
for each parameter, upper and lower condence
limits need to be evaluated. Essentially, the total
number of parameters is doubled, but an estimate
of output error is also obtained. Does the increase
in the number of parameters outweigh the advantages of fuzzy conceptual rainfallrunoff models?
Does this increase introduce more uncertainty in
the parameter identication process? Due to the
emergence of such powerful computers in the
past decade, it would probably not be signicantly
disadvantageous from a computational perspective
when set against the benets of being able to represent the uncertainty. Clearly, there is, indeed, some
trade-off involved, but further research is required
to answer this question fully. One should also keep
in mind that all uncertainty analysis techniques
require additional parameters to represent the
variability (e.g. in the case of condence interval
analysis, the variance would be computed). In the
case of the fuzzy regression, the estimation of the

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

64

Table 2
Relative sum of squares error of the crisp (i.e. non-fuzzy) and fuzzy
estimated parameter values from the true parameter value for the
TWOPAR and the SIXPAR models
Parameter
TWOPAR
C
K
SIXPAR
CU
KU
CL
KL

Crisp

Fuzzy

91.4
2.2

3.5
0.5

4.1
5.1
4.7
8.7

1.7
2.7
1.8
5.6

parameter variability is embedded in the optimization process.


As in the case of the SIXPAR model, model
components, such as the percolation process,
involve identiability problems and do not allow
even a sophisticated global optimization algorithm,
such as SCE-UA, to converge to the true parameter
values. This conclusion holds true also for the nonoise scenario.

Acknowledgements
The research reported here has been partially
supported by the US National Science Foundation
and the US Army Corps of Engineers, especially
during the rst author's doctoral studies in the Department of Systems and Industrial Engineering at The
University of Arizona, Tucson. Both authors would
like to thank Hoshin V. Gupta, Agnes Galambosi
and the referees for their constructive suggestions.

Appendix A. Elements of fuzzy logic


This section summarizes the fuzzy framework used
in this paper. For more details on fuzzy logic, the
reader is referred to Bardossy and Duckstein (1995),
Dubois and Prade (1980, 1988), Kaufmann and Gupta
(1991), and Zimmermann (1985).
A.1. Fuzzy sets and fuzzy numbers
Generally speaking, a fuzzy set is a quantity dened
vaguely, such as `large runoff event'. Formally, A^ is a
fuzzy subset of X if A^ {x; mA^ x; x [ X}, where X
is a set (universe). For hydrologic modeling, X is the
set of real numbers, usually the positive ones. Here,
mA^ xx [[0, 1] is the membership of x in a fuzzy set
^ The more x belongs to A,
^ the closer m ^ x is to 1
A.
A
^
and the less x belongs to A, the closer mA^ x is to 0. For
example, if we dene a `medium runoff' as a fuzzy set
in the interval 2040 mm, with a most likely value of
30 mm, and a `large runoff' as a fuzzy set in the interval 3050 mm, with a most likely value of 40 mm, as
shown in Fig. 20, then a runoff value of 38 mm would
have a membership of 0.2 to the `medium runoff' and
0.8 to the `large runoff' fuzzy sets. If the membership
interval [0, 1] is replaced by the binary set {0, 1}, then
A can be regarded as a regular subset of X.
The a -level set of the fuzzy subset A^ is the set of
those elements that have at least a membership:
^ a {x : m ^ x $ a}. For example, in Fig. 20,
A
A
a 0.5-level set for the `large runoff' is dened as
the runoff values between 35 and 45 mm.
A Fuzzy Number (FN) is a fuzzy subset that is
normal, convex and dened on the set of real
numbers. Normality is satised if the membership
function value reaches 1, that is 'z s:t: mA^ z 1.

Fig. 20. Illustration of a fuzzy number, membership function, with the a -level set for observed Z^ t runoff.

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

65

Convexity is satised if the membership function


branches have no `dips', namely,
;a; b; c [ X; s:t: a # b # c; mA^ b
$ minmA^ a; mA^ c:
A FN may be considered as a generalization of the
interval of condence (Kaufmann and Gupta, 1991). It
represents an innite number of intervals whose size
is non-increasing with increasing membership
function value and may be considered as the
fuzzy counterpart of a statistical condence interval.
Here, a 0 would mean the 100% condence interval, also called the support of the FN. FNs can take
any non-linear form satisfying the above stated properties, but most commonly used FNs have piecewise
linear membership functions known as triangular
fuzzy numbers (TFN). Fig. 20 gives an example
for TFNs and an a -level set for an observed runoff
value.
A.2. Functions of FNs and fuzzy arithmetic
Functions of FNs are calculated by means of the socalled extension principle. Let A^ 1 ; ; A^ K be fuzzy
subsets dened on X1 ; ; XK ; y f x1 ; ; xK ; x^k [
Xk ;k; and B^ f A^ 1 ; ; A^ K ; then
(

mB^ y

^
^
Fig. 21. Computation of the fulllment of a fuzzy statement A.
B.

where
SM {a1 a b1 a; a1 a b2 a; a2 a
b1 a; a2 a b2 a}
^ a:B
^ a minSD ; maxSD
A
and where
SD {a1 a : b1 a; a1 a : b2 a; a2 a : b1 a; a2 a
: b2 a};
^ a does not
dened only for the condition that B

supminmA^ 1 x1 ; ; mA^ K xK ; y f x1 ; ; xK ; x^k [ Xk ;k


0 elsewhere

For operational purposes on FNs, it is usually more


practical to use the a -level set notation as
^ a a1 a; a2 a, where a1 a and a2 a are the
A
^ a (Fig. 20).
lower and upper bounds of the interval A
Through application of the extension principle, it can
be shown that the following arithmetic operations
hold for FNs (Kaufmann and Gupta, 1991):

include zero.
^ a
^ aB
A
minSP ; maxSP ;

where
SP {a1 ab1 a ; a1 ab2 a ; a2 ab1 a ; a2 ab2 a }. If
A^ and B^ are dened in R1 , then the multiplication
and division operations further simplify as follows:

^ a1B
^ a a1 a 1 b1 a; a2 a 1 b2 a
A

^ aB
^ a a1 a b1 a; a2 a b2 a
A

^ a2B
^ a a1 a 2 b2 a; a2 a 2 b1 a
A

^ a:B
^ a a1 a : b2 a; a2 a : b1 a:
A

^ aB
^ a minSM ; maxSM
A

If A^ does not include zero, the inverse of a FN is dened


^ a21 1 : a2 a; 1 : a1 a.
as A^ 21 such that A

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

66

Using this notation, division can alternatively be


^ a:B
^ aB
^ a A
^ a21 : Furtherdenoted as A
more, because of piecewise linearity, trapezoidal FNs
(TrFN) and triangular FNs can be represented as A^
a1 0; a1 1; a2 1; a2 0TrFN and A^ a1 0; a1 1;
a2 0TFN , respectively, that allow fuzzy arithmetic
operations to be performed more conveniently.
In this paper, all fuzzy operations are performed at a
selected a -level. Note that fuzzy arithmetic operations are completely different from stochastic variable
operations, that basically rely on the convolution of
random variables. As noted earlier, in the case of nonnormal dependent random variables it might be rather
difcult, if not impossible, to obtain closed formulations for uncertainty modeling that includes convolution operations (e.g. Resnick, 1992). On the other
hand, compared to stochastic variable operations,
fuzzy algebra or arithmetic is typically relatively
easy, and usually yields closed formulation of uncertainty.

^
B^ is
Similarly, the truth value at level a of A.
computed as the complement of I^L a, i.e. as

A.3. Comparison of FNs: fuzzy ranking

where ( is a logical operator specied according to


the application. Usually, rules are formulated using
`AND' or `OR' operators. The choice of the logical
operator in Eq. (42) depends on the physics of the
problem. In general, for the CRR models, `AND'
operators follow naturally from the algorithmic ow
of the models, `AND' operators can be computed
using the product norm:

Comparison of FNs is required in many decisionmaking problems (e.g. Shrestha and Duckstein, 1997,
zelkan and Duckstein, 1999b). As shown in this
and O
paper in Sections 2.2 and 2.3, with reference to the CRR
model structure, the degree of dominance of one FN
over another needs to be evaluated. For example, for the
TWOPAR model, the truth value of the statement
X^ t #C^ should be evaluated to determine the degree
of fulllment of the model state to rules 1 and 2.
For building fuzzy CRR models, we will use a
similar idea to that in Shrestha and Duckstein
(1997) to compute the degree of dominance of a FN.
More specically, a degree of dominance relation will
be used to compute the fulllment of fuzzy rules in
fuzzy CRR models. The computation is performed at
a selected a -level, as follows: let A^ and B^ be two FNs;
^ B^ let
to evaluate the truth value of a statement A
^
^
^
^
D A2B and Da d1 a; d2 a, let the truth
value at level a be (Fig. 21)

I^L a

8
>
>
>
<

2d1 a
>

a
d
>
2 2 d1 a
>
:
1

if d1 a $ 0
if d1 a , 0 and d2 a . 0
if d1 a # 0

40

I^g a 1 2 I^i a

41

For example, for the TWOPAR model, if the truth


value of the statement X^ i #C^ has been computed
using Eq. (40) as 0.7, this would imply that the degree
of fulllment or the membership of the model state to
rule 1 is 0.7 and to rule 2 is 0.3.
A.4. Fuzzy rule-based modeling
In general, a fuzzy rule consists of a set of explanatory variables called premises A^ i;n given in the
form of FNs with membership functions mA^ i;k ak ,
and a consequence B^ i also in the form of a FN,
such that:
if a1 [ A^ i;1 (a2 [ Ai;2 ((aK [ A^ i;K ; then b [ B^ i
42

vi t

K
Y
k1

mA^ i;k ak t:

Rules are usually constructed from a data set or


dened by experts based on the physics of the problem
zelkan et al., 1996; Bardossy and Duckstein,
(e.g. O
1995). In the case of CRR models, each operational
mode of the CRR system can be considered as an
expert dened rule. For the CRR model, the membership function mA^ i;k ak measures the truth value of the
`IF' switches along the algorithmic ow of each mode
(Fig. 1). Each mode shows a possible route on the ow
diagram. In a traditional crisp (non-fuzzy) model,
only one of these modes (routes) would apply,
depending on the branch directed by the `IFTHEN'
switches. In a fuzzy model, the three modes may
occur simultaneously to different extents, so as to
reect the uncertainty in the process. Here, Eqs. (40)

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

and (41) will be used to compute the degree of applicability of any given CRR rule, also called the degree
of fulllment.
Finally, the response of the fuzzy rules to a given
input vector a1 t; ; aN t can be computed as
!
X
^
vi tB^ i ;
Bt
i

where the parentheses are used to denote a fuzzy arith^ is the estimated fuzzy
metic operation and the Bt
value of b(t).
Appendix B. Table of acronyms

BOFR Bi-objective fuzzy regression


FN
Fuzzy number
he
Heteroscedastic error
ho
Homoscedastic error
MOFR Multi-objective fuzzy regression
RSSE Relative sum of squares of error
SAC-SMA Sacramento Soil Moisture Accounting
model of the US National Weather Service
river forecast system
SCE-UA Shufed Complex Evolution method of
The University of Arizona
SIXPAR The SIX PARameter model
SSV
Sum of square vagueness
TFN
Triangular fuzzy number
TOFR Tri-objective fuzzy regression
TrFN Trapezoidal fuzzy number
TWOPAR The TWO PARameter model

References
Bardossy, A., 1990. Notes on fuzzy regression. Fuzzy Sets and
Systems 37, 6575.
Bardossy, A., Bogardi, I., Duckstein, L., 1990. Fuzzy regression in
hydrology. Water Resources Research 26 (7), 14971508.
Bardossy, A., Hagaman, R., Duckstein, L., Bogardi, I., 1992.
Fuzzy least squares regression: theory and application. In:
Kacprzyk, J., Fedrizi, M. (Eds.). Fuzzy Regression Analysis.
Omnitech Press/Physica-Verlag, Warsaw/Heidelberg, pp. 21
44.
Bardossy, A., Disse, M., 1993. Fuzzy rule-based models for inltration. Water Resources Research 29 (2), 373382.
Bardossy, A., Duckstein, L., 1995. Fuzzy Rule-Based Modeling

67

with Applications to Geophysical, Biological and Engineering


Systems. CRC Press, Boca Raton.
Celmins, A., 1987a. Multidimensional least-squares tting of fuzzy
moels. Mathematical Modelling 9 (9), 669690.
Celmins, A., 1987b. Least-squares model tting to fuzzy vector
data. Fuzzy Sets and Systems 22, 245269.
Diamond, P., 1988. Fuzzy least squares. Information Sciences 46,
141157.
Duan, Q., Sorooshian, S., Ibbit, R.P., 1988. A maximum likelihood
criterion for use with data collected at unequal time intervals.
Water Resources Research 24 (7), 11631173.
Duan, Q., Sorooshian, S., Gupta, V.J., 1992. Effective and efcient
global optimization for conceptual rainfallrunoff models.
Water Resources Research 28 (4), 10151031.
Dubois, D., Prade, H., 1980. Fuzzy Sets and Systems: Theory and
Applications. Academic Press, New York.
Dubois, D., Prade, H., 1988. Possibility Theory: an Approach to
Computerized Processing of Uncertainty. Plenum Press, New
York.
zelkan, E.C., Bogardi, I., 1998. A
Galambosi, A., Duckstein, L., O
fuzzy rule-based model to link circulation patterns, ENSO, and
extreme precipitation. In: Haimes, Y.Y., Moser, D., Stakhiv,
E.Z. (Eds.). Risk-Based Decision Making in Water Resources,
vol. VIII. ASCE Press, Reston VA, pp. 83103.
zelkan, E.C., Duckstein, L., Bogardi, I., 1999. A
Galambosi, A., O
fuzzy rule-based model for precipitation analysis under climate
change in the US Southwest. Proceedings of American Meteorological Society '9914 Hydrology Conference, 1015 January, Dallas, TX, pp. 2930.
Gan, T.Y., Biftu, G.F., 1996. Automatic calibration of conceptual
rainfallrunoff model: optimization algorithms, catchment
conditions and model structure. Water Resources Research 32
(12), 35133524.
Goichoechea, A., Hansen, D.R., Duckstein, L., 1982. Multiobjective Decision Analysis with Engineering and Business Applications. John Wiley, New York.
Gupta, V.J., Sorooshian, S., 1983. Uniqueness and observability of
conceptual rainfallrunoff model parameters: the percolation
process examined. Water Resources Research 19 (1), 269276.
Gupta, V.J., Sorooshian, S., 1985a. Automatic calibration of
conceptual rainfallrunoff models using derivative-based
optimization algorithms. Water Resources Research 21 (4),
473485.
Gupta, V.J., Sorooshian, S., 1985b. The relationship between data
and the precision of parameter estimates of hydrologic models.
Journal of Hydrology 81, 5777.
zelkan, E.C., Duckstein, L., 1993. Fuzzy regressionHsu, K.-L., O
based rainfallrunoff estimation. Proceedings of Sixth Annual
Symposium of the Arizona Hydrological Society, Casa Grande,
AZ.
Jozsef, S., 1992. On the effect of linear data transformations in
possibilistic fuzzy linear regresson. Fuzzy Sets and Systems
45, 185188.
Kaufmann, A., Gupta, M.M., 1991. Introduction to Fuzzy Arithmetic: Theory and Applications. Van Nostrand Reinhold, New
York.
Kim, J.K., Moskowitz, H., Koksalan, M., 1996. Fuzzy versus

68

zelkan, L. Duckstein / Journal of Hydrology 253 (2001) 4168


E.C. O

statistical linear regression. European Journal of Operational


Research 92, 417434.
Klemes, V., 1982. Empirical and causal models in hydrology.
Scientic Basis of Water Resources Management. National
Academy Press, Washington, DC, pp. 95106.
O'Connel, P.E., Clarke, R.T., 1981. Adaptive hydrological forecastinga review. Hydrological Sciences Bulletin 26 (2), 179205.
zelkan, E.C., Ni, F., Duckstein, L., 1996. Relationship between
O
monthly atmospheric circulation patterns and precipitation:
fuzzy logic and regression approaches. Water Resources
Research 32 (7), 20972103.
zelkan, E.C., 1997. Multi-objective fuzzy regression applied to the
O
calibration of conceptual rainfallrunoff models. Unpublished
PhD dissertation, The University of Arizona, Systems and
Industrial Engineering Department, Tucson, Arizona.
zelkan, E.C., Galambosi, A., Duckstein, L., Bardossy, A., 1998a.
O
A multi-objective fuzzy classication of large scale atmospheric
circulation patterns for precipitation modeling. Applied Mathematics and Computation 91 (2), 127142.
zelkan, E.C., Duckstein, L., Galambosi, A., 1998. Analysis of
O
trade off between data outliers and prediction vagueness in
fuzzy regression using a bi-objective framework. In: Proceedings of EUFIT '98Sixth European Congress on Intelligent
Techniques and Soft Computing, 710 September, Aachen,
Germany.
zelkan, E.C., Duckstein, L., 1999. Fuzzy logic to analyze the
O
uncertainty in conceptual rainfallrunoff models. Proceedings
of American Meteorological Society '9914th Hydrology
Conference, 1015 January, Dallas, TX, pp. 376379.
zelkan, E.C., Duckstein, L., 1999b. Optimal fuzzy counterparts of
O
scheduling rules. European Journal of Operational Research 113
(3), 593599.
zelkan, E.C., Duckstein, L., 2000. Multi-objective fuzzy regresO
sion: a general framework. Computers and Operations Research,
Special Issue on Articial Intelligence and Decision Support
with Multiple Criteria 27(7 8, 635652.
Pesti, G., Shrestha, B.P., Duckstein, L., 1996. A fuzzy rule-based
approach to drought assessment. Water Resources Research 32
(6), 17411747.
Peters, G., 1994. Fuzzy linear regression with fuzzy intervals. Fuzzy
Sets and Systems 63, 4555.
Redden, D.T., Woodall, W.H., 1994. Properties of certain fuzzy
linear regression methods. Fuzzy Sets and Systems 64, 361
375.
Redden, D.T., Woodall, W.H., 1996. Further examination of fuzzy
linear regression. Fuzzy Sets and Systems 79, 203211.
Renard, K.G., Lane, L.J., Simanton, J.R., Emmerich, W.E., Stone,
J.J., Weltz, M.A., Goodrich, D.C., Yakowitz, D.S., 1993. Agricultural impacts in an arid environment: Walnut Gulch case
study. Hydrological Science and Technology 9(1 4, 145190.
Resnick, S.I., 1992. Adventures in Stochastic Processes. Birkhaeuser, Boston.

Sakawa, M., Yano, H., 1992. Multiobjective fuzzy linear regression


analysis for fuzzy inputoutput data. Fuzzy Sets and Systems
47, 173181.
Savic, D.A., Pedrycz, W., 1991. Evaluation of fuzzy linear regression models. Fuzzy Sets and Systems 39, 5163.
Shrestha, B.P., Duckstein, L., 1997. A fuzzy reliability measure for
engineering applications. In: Bilal, A. (Ed.). Uncertainty Modeling and Analysis in Civil Engineering. CRC Press, Boca Raton,
pp. 121136.
Singh, V.P., 1988. Hydrologic Systems: Rainfallrunoff Modeling.
Prentice Hall, Englewood Cliffs, NJ.
Sorooshian, S., Dracup, J.A., 1980. Stochastic parameter estimation
procedures for hydrologic rainfallrunoff models: correlated
and heteroscedastic error cases. Water Resources Research 16
(2), 430442.
Sorooshian, S., 1983. Surface water hydrology: on-line estimation.
Reviews of Geophysics and Space Physics 21 (32), 706721.
Sorooshian, S., Gupta, V.J., 1983. Automatic calibration of conceptual rainfallrunoff models: the question of parameter observability and uniqueness. Water Resources Research 19 (1), 260
268.
Sorooshian, S., Gupta, V.J., Fulton, J.L., 1983. Evaluation of maximum likelihood parameter estimation techniques for conceptual
rainfallrunoff models: inuence of calibration data variability
and length on model credibility. Water Resources Research 19
(1), 251259.
Sorooshian, S., 1985. Synthesis of hydrologic and system sciences
in the development of rainfallrunoff models. Applied Mathematics and Computations 17, 279298.
Sorooshian, S., Gupta, V.J., 1985. The analysis of structural identiability: theory and application to conceptual rainfallrunoff
models. Water Resources Research 21 (4), 487495.
Szidarovszky, F., Gershon, M.E., Duckstein, L., 1986. Techniques
for Multiobjective Decision Making in Systems Management,
Advances in Industrial Engineering. Elsevier, New York.
Tanaka, H., Uejima, S., Asai, K., 1982. Linear regression analysis
with fuzzy model. IEEE Transactions on Systems, Man, and
Cybernetics 12 (6), 903907.
Tanaka, H., 1987. Fuzzy data analysis by possibilistic linear models.
Fuzzy Sets and Systems 24, 363375.
Yapo, P.O., Gupta, H.V., Sorooshian, S., 1996. Automatic calibration of conceptual rainfallrunoff models: sensitivity to calibration data. Journal of Hydrology 181 (14), 2348.
Yapo, P.O., Gupta, H.V., Sorooshian, S., 1998. Multi-objective
global optimization for hydrologic models. Journal of Hydrology 204 (14), 8397.
Young, P., Wallis, S., 1985. Recursive estimation: a unied
approach to the identication, estimation, and forecasting of
hydrological systems. Applied Mathematics and Computations
17, 299334.
Zimmermann, H.J., 1985. Fuzzy Sets Theory and Its Application.
Kluwer Academic.

Das könnte Ihnen auch gefallen