Design and implementation of a fuzzy controller for wind generators

performance optimisation

Vito Calderaro, Vincenzo Galdi, Antonio Piccolo, Pierluigi Siano

DIIIE, UNIVERSITY OF SALERNO
Fisciano, (SA) Italy
Tel.: +39/089964294
Fax: +39/089964284
E-Mail: vcalderaro@unisa.it;vgaldi@unisa.it, apiccolo@unisa.it; psiano@unisa.it
URL: http://www.unisa.it/Dipartimenti/DIIIE

Keywords
«Fuzzy control», «Wind energy», «Adaptive control», «Power supply».

Abstract
Actual wind power costs together with incentives and financing options for developing renewable energy
facilities make wind energy source competitive with conventional generation sources and it is believed
that wind energy will be the most cost effective source of electrical power in the next future. However, the
wind power production diffusion involves the development of efficient control systems able to improve
wind systems effectiveness. Therefore, a design methodology, able to generate an adaptive fuzzy model
for maximum energy extraction from variable speed wind turbines is proposed in this paper. The fuzzy
model is designed by using fuzzy clustering combined with genetic algorithms (GA) and recursive least-
squares (LS) optimisation methods. Some simulation results on a doubly-fed induction generator
confirmed that the proposed design methodology is able to identify a Takagi-Sugeno-Kang (TSK) fuzzy
model exhibiting adaptivity, learning capability, high speed of computation and low memory occupancy.

Introduction
Wind energy represents a renewable natural energy resource which can provide a considerable effort to
electricity production, allowing a reduction of the environmental impact because of its feature of
generating carbon-emission-free electricity. Therefore, nowadays many European countries energy
policies have the purpose to enhance wind energy utilization, also by means of incentives and financing
options [1]. In order to establish wind energy expansion, the development of efficient control systems, that
can improve the wind systems effectiveness, is essential. Wind turbines can operate at fixed speed or
variable speed, however variable speed operation is preferable as it allows a reduction of the mechanical
structure stresses and the acoustic noise and the control of active and reactive power. Moreover, variable
speed operation increases the energetic efficiency and reduces the drive train torque and generated power
fluctuations [2].
Because the effectiveness of control systems for variable speed wind turbines is fundamental and,
considering that fuzzy control can offer many advantages over traditional controls, such as fast
convergence, adaptivity and parameter insensitiveness [3], a sensorless peak power tracking adaptive
fuzzy control for variable speed wind turbines is proposed in this paper.
The proposed adaptive fuzzy control is able to counterbalance the non-linearities and time variances of the
system under control by means of its adaptability and learning capability.
In fact, even if a linear controller is designed using realistic and accurate simulation models described in
terms of transfer functions, simulated systems don’t describe a wind turbine to its full extent [4] and it is
probable that the controller would need further adjustments before it can be applied in a real wind turbine.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 A real wind turbine, in fact, cannot be precisely described by a simulation model as it contains several
dynamics that are not fully known. Conversely, the proposed adaptive fuzzy controller, without the need
of an accurate model of the system to be controlled and thanks to its adaptive characteristic based on real
data, can continuously optimise its internal parameters, thus adapting them to the system dynamic
behavior and time variances. As a consequence of its adaptive and learning ability on the site, the
proposed fuzzy approach can represent an alternative to the lookup table based approach (power or torque
vs. rotor speed) used in most of modern wind turbines today. In fact, it is likely that the wind turbine
manufacturers have assumed a lookup table which, in some real occasion does not exactly correspond to
how the real turbine blades perform. This can happen because of inaccurate aerodynamic calculations,
tolerance in mounting the turbine, dirt on blades and other unpredictable parameter variations. As the
proposed fuzzy controller can learn this on the site and tune itself accordingly to exploit more energy from
the wind system, it could be relevant for wind turbine manufacturers.
The proposed controller is based on a Takagi-Sugeno-Kang (TSK) that is identified by using a data driven
designing methodology. The methodology, on the basis of given input-output numerical data, generates
the ‘best’ TSK fuzzy model, able to estimate with high accuracy the maximum extractable power from a
variable speed wind turbine.
In order to obtain the model, fuzzy clustering combined with genetic algorithms (GA) and recursive least-
squares (LS) optimisation methods are used.

Wind generation system

A power system consisting of a wind turbine with a doubly-fed induction generator connected to a three-
phase grid has been considered [5], as shown in Fig. 1.
The stator of the induction generator is directly connected to the grid, while the rotor winding is connected
via slip rings to a converter.

Fig. 1: Wind Turbine and Doubly-Fed Induction Generator System

The induction generator converts the power captured by the wind turbine into electrical power and
transmits it to the grid.
The AC/DC/AC converter consists of the rotor-side converter (Crotor) and the grid-side converter (Cgrid).
Both Crotor and Cgrid converters are voltage-sourced converters using forced commutated power electronic
devices to synthesize an AC voltage from a DC voltage source. A capacitor connected on the DC side acts
as the DC voltage source and a coupling inductor L is used to connect the grid-side converter to the grid.
The three-phase rotor winding is connected to Crotor by slip rings and brushes and the three-phase stator
winding is directly connected to the grid.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 The pitch angle command and the voltage command signals Vr and Vgc for Crotor and Cgrid converters,
respectively, are generated by the control system controlling the power of the wind turbine, the DC bus
voltage and the voltage at the grid terminals.
The rotor-side converter is used to control the wind turbine output power and the voltage measured at the
grid terminals.
The converter Cgrid is used to regulate the voltage of the DC bus capacitor. In addition, this model allows
using Cgrid converter to generate or absorb reactive power. A detailed description of the control systems
for both the converters and the pitch angle command can be found in [5].
Wind turbine characteristics and proposed fuzzy control approach
The power extracted from a wind turbine results from the wind power available, the machine power curve
and the machine reactiveness to wind variations. The power and torque extracted from the wind can he
expressed as [6]:
1
Pω = ρC P (λ , β ) AVω3 (1)
2
P 1
Tω = ω = ρCT (λ , β )rm AVω2 (2)
ωr 2
where Pω is the rotor mechanical power (W); Tω is the turbine torque (Nm); Vω the wind speed at the
λV
centre of the rotor (m/s); A = πrm2 is the wind rotor swept area (m2); ρ the air density (kg/m3); ωr = ω is
rm
the rotor angular velocity (rad/sec); rm is the turbine radius (m); C P is the rotor power coefficient, CT is
ωr
the torque coefficient; β is the pitch angle of rotor blades [deg]; λ = r m is the tip speed ratio.
Vω
The maximum power and the corresponding rotor speed of the wind turbine can be derived from (1):
Pmax = K1Vω3 , K1 = 0.5 ρAC P _ max
λ (3)
ωropt = K 2Vω , K 2 = opt
rm
where Pmax is the turbine maximum power at a certain wind velocity and is proportional to the wind
velocity with the scaling constant K1 , ωropt is the desired rotor speed when the turbine operates at
maximum power point.

Maximum power line

Turbine mechanical
power P (kW)

PB

PA

s)
m/
t y(
el oci
ωΒ ω n dv
Rotor speed
(rad/s) A Wi
Fig. 2: Turbine power curves

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 The proposed maximum power point tracking TSK fuzzy controller has two inputs and one output: the
measured rotor speed and power generated by the generator are the inputs, while the output is the
estimated maximum power, that is the maximum power that can be generated. The TSK fuzzy system, by
acquiring and processing at each sample time the inputs, is able to calculate the maximum power that may
be generated by the wind generator.
The approach can be better explained by considering the turbine power curves shown in Fig. 2. Assuming
that the wind turbine operates initially at point A, the controller, by the knowledge of measured rotor
speed ω A and turbine power PA , can derive the corresponding optimum operating point B, giving the
desired rotor speed reference ω B . Therefore, the generator speed will be controlled in order to reach the
speed ω B allowing the extraction of the maximum power PB from the turbine.

TSK fuzzy model identification

This section describes the data-driven identification procedure able to identify the structure and
parameters of the best TSK fuzzy model that minimizes, on a typical wind pattern, an objective function
based on the Mean Squared Error (MSE) between the maximum power estimated by the fuzzy controller
and the maximum power that the wind generator can supply.

Start

Generate initial population

Cluster the input-output measured data

with each individual number of clusters
Create new
Identify the TSK fuzzy model generation by
with each individual number of clusters -reproduction,
(rules) and spread by using recursive -crossover
least-square procedure -mutation

Evaluate objective function for each

individual by using modified Akaike
information criterion

Stop criterion NO
reached?

YES

Print solution values, and TSK

model output

Fig. 3: Flow chart of the TSK identification procedure

The data-driven algorithm for the generation of a TSK fuzzy model is based on GA, fuzzy clustering and
recursive least-squares procedure. It is based on a GA having a chromosome of two elements: (N r , r ) ,
where N r is the number of clusters and r is the spread of the membership functions.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 The flow chart of the algorithm is reported in Fig. 3. For each possible chromosome, the corresponding
TSK model is identified in two steps:
1. a fuzzy clustering technique is applied, with a number of clusters equal to N r ;
2. assuming the centers furnished by the previous step, the number of rules equal to N r and the
spreads of the memberships functions equal to r, the model’s parameters are identified by a
recursive least-squares procedure.
Once the TSK model is identified, its fitness function is evaluated and the GA stops when the prefixed
stop criterion is reached.
Input-output space creation and partitioning technique
The identification of the TSK model implies the existence of a knowledge base consisting of a set of
input/output measured data samples.
In particular, the proposed method requires, for each couple of rotor speed and generated power, the
knowledge of the maximum power that can be extracted by controlling the rotor speed. The generation of
power curves, similar to the ones shown in Fig. 4 is consequently required. In fact, by using these curves
the corresponding rotor speed command can be decided by moving from point A to point B, following the
curve shape as shown in Fig. 4 and without the knowledge of wind velocity.
The power curves are generated considering a range of wind velocities by the following procedure [6]:

1. for each wind velocity value Vω(j) in the considered range the rotor speed is regulated to a
constant value ω r (i) ;
2. the corresponding turbine power Pm(i, j) is measured;
3. rotor speed is updated to the next constant ω r (i + 1) ;
4. the previous steps are repeated until data of most operation points have been collected.

Once generated, the .power curves allow to obtain, for an assigned time varying wind velocity profile, a
data set of samples (one for each time step) consisting of two inputs (measured rotor speed and generated
power) and corresponding output (maximum extractable power).
In order to perform partitioning of the input-output space, various approaches can be used. Amongst them,
pattern-recognition methods of fuzzy clustering, such as fuzzy c-means (FCM) [7-10], are suitable tools
for the partitioning process.
Using such a method, data samples are organized in clusters, each of which is associated to a centre by
using the FCM algorithm and the TSK model is based on a set of fuzzy IF-THEN rules, extracted by using
the FCM clustering technique.
Turbine mechanical
Turbine mechanical

power P (kW)
power P (kW)

Ro
tor
sp
e ed
(ra
d/s (m/s)
Rotor speed (rad/s) ) Wind velocity
Fig. 4: Example of power curves

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 Identification of model structure and parameters
For multi-input single-output system, the typical TSK model consists of a set of IF-THEN rules having the
following form [11,12]:
Rh : IF x1 is Ah1 and … and x p is Ahp THEN y is f h ( x ),
(4)
h = 1,…, R
where
f h ( x ) = a0h + a1h x1 + a2 h x2 + + a ph x p (5)
in which x1,…, p are the input variables, y is the output variable, A1h,…, p are the fuzzy sets and f h ( x ) is a
linear function. The h-th fuzzy rule of the collection is able to describe the local behaviour associated to
the fuzzy input region characterized by the antecedent of the fuzzy rule.
For each input, ~ x , the inferred value of the TSK model, is calculated as
~ ~ ~
~y = ∑ h =1 Ah ( x ) ∗ f h ( x ) = ∑ h =1τ h ∗ f h (x )
n n
(6)
(~ )
∑h =1 h ∑h =1 h
n n
A x τ
where the degree of firing of each rule, τ , for the current input ~ h x is determined by the Gaussian law,
which ensures the greatest possible generalization
2
x − x*h
−α ~
τh = e , h = 1,2,..R; (7)
where xh* is the centre of a rule and α = 4 / r 2 and r is a positive constant defining the zone of influence
of the rule. As the consequent of each rule is linear, its parameters, which minimize the overall error
between the TSK fuzzy model and the system being modeled, can be determined recursively by least-
squares procedure. Substantially, the TSK model has to be identified by the following two subtasks:
1. identification of the antecedent part of the model, by determining the centres ( xh* , h = 1,2,..R; )
and spreads ( r ) of the membership functions;
2. identification of the parameters ( aih , i = 0,1,...p; h = 1,2,...R ; ) of the consequent part by least
squares technique.
Genetic algorithm
The trade-off between fitting the training data samples and reducing the model complexity, in terms of
rule-base structure represents an important issue in the model identification, consequently the modified
Akaike information criterion (AIC) is used for model construction. It considers a penalty for the over
fitting so that the complexity of the fuzzy model is determined by the number of fuzzy rules in the model
and not only by the number of antecedent and consequent parameters of the rules.
The modified AIC, which tends to minimize both the output error and the order of the model, can be
defined as [13]:
AIC = N clog (MSE ) + 2m p (8)
where N c is the number of data samples, m p is the number of parameters in the model and MSE is the
Mean Squared Error that the identified model exhibits, when used to estimate a input-output association
for a given data set. Based on the considerations stated in [13], the single parameter m p in the above
statistical information criteria is replaced and a ‘complexity function’ is used, which is defined as:
s (ma , mc , mr ) = ma + mc + cmr (9)
where ma is the number of antecedent parameters, mc is the number of consequent parameters, mr is the
number of fuzzy rules constituting the model, and c is a constant allowing the user to incorporate
heuristics regarding the relative importance for reducing the number of fuzzy rules.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 The data-driven approach for the identification of the TSK model is based on GA considering a
chromosome of two elements: (N r , r ) , where N r represents the number of clusters to be used for
clustering, corresponding also to the number of rules, and r represents the spreads of the membership
functions.
The fitness function to be minimized in the implemented GA is based on the modified AIC and defined as
follows:
2(ma + mc + cmr )
f ( N , r ) = MSE + (10)
Nc
where the MSE is written as:

∑ (Pkestimated )
NC
2
− Pkmax
k =1
MSE = (11)
Nc
Pkmax and Pkestimated are the maximum extractable powers obtained, respectively, from the power curves
and the TSK fuzzy model, for each time step k.
Normalized geometric ranking scheme is used as selection mechanism and simple crossover and binary
mutation are used as genetic operators. An elitism mechanism is also adopted. The iteration process stops
when improvement of the best objective function value is below a threshold for a given number of
generations or if the total number of generations is higher than a maximum number.

Simulation results
The generator power curves were generated by means of the previously described procedure and by using
the Matlab/SimPowerSystems model of the doubly-fed induction generator connected to a 25-kV
distribution system with resistive and motor loads and connected to a 120-kV grid through a 30-km, 25-
kV feeder [5]. The TSK fuzzy model was identified following the formerly described identification
procedure and by using a training set, obtained assuming the wind profile shown in Fig. 5.

14

1.4
13

12 1.2
Wind velocity [m/s]

11
1
Power [MW]

10
0.8
9
0.6
8

0.4 Estimated maximum power

7
Maximum power
6 0.2
0 50 100 150 200 250 300 0 50 100 150 200 250 300
time [s]
time [s]

a) b)
Fig. 5: Training set
a) Wind profile
b) Maximum power estimated by the fuzzy model and effective maximum extractable power

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 The wind pattern is relative to a 300 seconds observation period with N C = 300 data samples. For
clustering the Euclidean norm was used as matrix norm, the fuzziness parameter, m, was set to 2.0, the
value of ε was set to 10-5 and the maximum number of iterations was set to 100.
For tuning the GA, a population size of 10 and a maximum number of 100 generation was set. The
simulation experience tells that these values, for the problem under study, guarantee the convergence of
the algorithm to a satisfactory solution. For the mapping of each term set on the domain of the
corresponding linguistic variable, the Gaussian membership functions were used. In particular, the
learning procedure has identified an optimal TSK model characterized by 2 rules and 2 membership
functions with spreads equal to 0.7286, which guaranties a performance, expressed in terms of the MSE,
equal to 7.8 × 10 −4 . The comparison between the maximum power estimated by the TSK fuzzy model and
the effective maximum extractable power, for the considered training set, shown in Fig. 5, evidences that
the TSK fuzzy model can estimate with a great accuracy the maximum power, with a percentage
Maximum Absolute Error (MAE) of 8.4%.
Two wind speed profiles, different from the first one and shown in Fig. 6 and Fig. 7, were then selected to
evaluate the generalization capability and accuracy of the TSK model.
Two validation sets of samples, specifically validation set 1 and validation set 2, deriving from the wind
profiles, were thus generated.

14
1.4
13

1.2
12
Wind velocity [m/s]

11 1
Power [MW]

10
0.8
9
0.6
8

0.4
7 Estimated maximum power
Maximum power
6 0.2
0 50 100 150 200 250 300 0 50 100 150 200 250 300
time [s] time [s]
a) b)
Fig. 6: Validation set 1
a) Wind profile
b) Maximum power estimated by the fuzzy model and effective maximum extractable power

A great accuracy of the TSK fuzzy model can be evidenced by comparing the maximum power estimated
by the TSK fuzzy model and the effective maximum extractable power for the two validation sets as
shown in Figs. 6 and Fig. 7.
The analysis of the MSE and the percentage MAE, shown in table I, for the training set and the two
validation sets, reveals that the minimum MSE value is obtained for validation set 1 but with the worse
percentage MAE, corresponding to 10.4%. Instead the TSK fuzzy model is able to obtain, for the
validation set 2, the minimum MAE, but with the worst MSE value. The values of the MSE and of the
percentage MAE for the training set are included between the corresponding values of the two validation
sets: this is mainly due to the choice of the GA objective function which is based on the modified AIC,
allowing for a balancing between reducing the fitting error and generalization capability of the identified
model.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 14

1.4
13

12 1.2
Wind velocity [m/s]

11
1

Power [MW]
10
0.8
9
0.6
8
Estimated maximum power
7 0.4 Maximum power

6 0.2
0 50 100 150 200 250 300 0 50 100 150 200 250 300
time [s]
time [s]
a) b)
Fig. 7: Validation set 2
a) Wind profile
b) Maximum power estimated by the fuzzy model and effective maximum extractable power

Table I: Simulation results

MAE % MSE [ MW 2 ]

Training set 8.4 7.8 × 10 −4

Validation set 1 10.4 5.7 × 10 −4

Validation set 2 7.6 9.4 × 10 −4

The different shape between the wind profile and the maximum extractable power that can be evidenced
in the range of high wind velocities is mainly due to the pitch control action limiting rotor speed below the
value of 1.21 p.u..
The implemented TSK fuzzy model, therefore, exhibits, generalization capability in addition to a great
accuracy. Furthermore, measuring the CPU time consumption with reference to a personal computer with
a Pentium IV processor, 3 GHz and 2 Gb RAM, it can be evidenced an high speed of computation:
specifically, TSK fuzzy system takes 100 µs to generate the output and is also characterized by memory
occupancy of less then 1 Kbytes.
Moreover, if compared to the lookup table based approach, the TSK fuzzy model is able to compensate for
unpredictable system’s parameters variation producing a lookup table not accurately corresponding to real
turbine blades performances.
Therefore, because of its adaptivity and learning capability on the site, that allows compensating for the
inherent non-linearities in a wind turbine and the unpredictability of the whole system, such as parameters
variation with time, the proposed fuzzy system can relevant for wind turbine manufacturers.
In addition, the proposed fuzzy system overcomes some disvantages of classical sensorless peak power
tracking control methods and well adapts to be implemented on microprocessor to control a real variable
speed wind turbine.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.
 On these basis, the research activity is focusing on the implementation of a TSK fuzzy controller on a
microprocessor and an experimental study with a real variable speed wind turbine and synchronous
generator system of 20kW is in progress.

Conclusion
Sensorless peak power tracking control methods allow to overcome problems linked to the inaccuracy of
measuring wind speed and to the effect of system parameters variation with time and varied environment.
A methodology for designing a sensorless maximum power point tracking TSK fuzzy system for variable
speed wind generators has been proposed and tested in the present paper.
The TSK fuzzy system is characterized by a low memory occupancy, fault tolerance and learning
capability allowing to cope with some disvantages of classical sensorless peak power tracking control
methods and well adapts to be implemented on microprocessor.
Moreover, because of its adaptivity and learning ability on the site, the proposed TSK fuzzy system can be
significant for wind turbine manufacturers and can represent an alternative to the lookup table based
approach, which, in some real situations is not based on the real turbine blades performances.

References
[1] Boukhezzara, L. Lupua, H. Siguerdidjanea, M. Handb: Multivariable control strategy for variable speed,
variable pitch wind turbines, Renewable Energy, Vol. 32, Issue 8, pp, 1273-1287, 2007.
[2] H. Camblonga, I. Martinez de Alegriab, M. Rodriguezc and G. Abadc: Experimental evaluation of wind turbines
maximum power point tracking controllers, Energy Conversion and Management, Vol. 47, issues 18-19 , pp. 2846-
2858, 2006.
[3] Marcelo Godoy Simões, Bimal K. Bose, and Ronald J. Spiegel: Design and Performance Evaluation of a Fuzzy-
Logic-Based Variable-Speed Wind Generation System, IEEE Transactions on Industry Applications, Vol. 33, no. 4,
pp. 956-965, 1997.
[4] Clemens Jauch, Tom Cronin and Poul Sorensen: A Fuzzy Logic Pitch Angle Controller for Power System
Stabilization Wind Energy 2007; Wiley Interscience, Vol.10 pp. 19-30, 2007.
[5] The MathWorks: SimPowerSystems For Use with Simulink, User’s Guide Version 4.
[6] Hui Li; Shi, K.L.; McLaren, P.G.: Neural-network-based sensorless maximum wind energy capture with
compensated power coefficient, IEEE Transactions on Industry Applications, Vol. 41, issue 6, pp.1548–1556, 2005.
[7] Bezdek, J.C. :Pattern Recognition with Fuzzy Objective Function Algorithms, New York, Plenum Press, 1981.
[8] Nascimento, S., Mirkin, B., Moura-Pires, F.: A Fuzzy Clustering Model of Data and Fuzzy c-Means, Proc. of
Ninth IEEE International Conference on Fuzzy Systems, Vol. 1 , pp. 302-307, 2000.
[9] Delgado, M., Gomez-Skarmeta, A.F., Martin, F.: A fuzzy clustering-based rapid prototyping for fuzzy rule-
based modelling, IEEE Transactions on Fuzzy Systems, Vol. 5, Issue 2, pp. 223–233, 1997.
[10] Chiu, S.L: A cluster estimation method with extension to fuzzy model identification Fuzzy Systems,
Proceedings of the Third IEEE World Congress on Computational Intelligence, Vol.2, pp. 1240 –1245, 1994.
[11] Hatanaka, T., Uosaki, K., Manabe, N.: Structure identification in Takagi-Sugeno fuzzy modelling, Proc. of the
2002 IEEE International Conference on Fuzzy Systems, Vol. 1, pp. 69-74, 2002.
[12] Johansen, T.A., Shorten, R., Murray-Smith, R.: On the interpretation and identification of dynamic Takagi-
Sugeno fuzzy models, IEEE Transactions on Fuzzy Systems, Vol. 8, Issue 3, pp. 297-313, 2000.
[13] Yen, J., Wang, L.: Application of statistical information criteria for optimal fuzzy model construction, IEEE
Transactions on Fuzzy Systems, Vol. 6, Issue 3, pp. 362 -372, 1998.

Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 22, 2009 at 08:59 from IEEE Xplore. Restrictions apply.