Energy 89 (2015) 324e333

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Adaptive neuro-fuzzy estimation of diffuser effects on wind turbine

performance
Vlastimir Nikoli

c a, Dalibor Petkovi

c a, *, Shahaboddin Shamshirband b, Zarko

Cojbasi

ca
a
University of Nis, Faculty of Mechanical Engineering, Deparment for Mechatronics and Control, Aleksandra Medvedeva 14, 18000 Nis, Serbia
b
Department of Computer System and Technology, Faculty of Computer Science and Information Technology, University of Malaya, Kuala Lumpur, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: Wind power is generating interest amongst many countries to produce sustainable electrical power. It is
Received 14 December 2014 well known that the main drawback of wind power is the inherent variable behavior of wind speed.
Received in revised form Signicant research has been carried out to improve the performance of the wind turbines and establish
18 May 2015
the power system stability. As power output is proportional to the cubic power of the incident airspeed,
Accepted 20 May 2015
Available online 4 July 2015
any small increase in the incident wind yields a large increase in the energy output. One of the more
promising advanced concepts for overcoming the inherent variable behavior of wind speed is the DAWT
(diffuser-augmented wind turbine). The diffuser or anged diffuser generates separation regions behind
Keywords:
Ducted wind turbine
it, where low-pressure regions appear to draw more wind through the rotors compared to a bare wind
Shrouded wind turbine turbine. Thus, the output power of the DAWT is much larger than for a unshrouded turbine. To estimate
Diffuser augmented wind turbine rotor performance of the diffuser-augmented wind turbine, this paper constructed a process which
Neuro-fuzzy simulates the power output, torque output and rotational speed of the rotor in regard to diffuser effect
ANFIS and wind input speed with ANFIS (adaptive neuro-fuzzy) method. This intelligent estimator is imple-
mented using Matlab/Simulink and the performances are investigated.
2015 Elsevier Ltd. All rights reserved.

1. Introduction wind turbine blades and increases in cross-sectional area in the

The usage of renewable, clean energy has become a very
important issue in recent years. Among all renewable energy of
different styles, wind energy possesses many advantages such as
extensive distribution, high efciency, low cost, low maintenance
and environmental friendliness. The power in the wind is well
known to be proportional to the cubic power of the wind velocity
approaching a wind turbine. This means that even small amount of
its acceleration gives large increase of the energy generation. Peo-
ple have long been seeking designs of enhancements to the con-
ventional wind turbines, and one of the most assuring proposals is
the DAWT (diffuser augmented wind turbine) [1e3]. Some attrac-
tive studies regarding augmented wind turbines were reported by
Refs. [4e7]. A DAWT [8e10] has a duct or shroud that surrounds the
* Corresponding author. 381 643283048.
E-mail address: dalibortc@gmail.com (D. Petkovi

c).
stream wise direction. The resulting sub-atmospheric pressure
within the diffuser draws more air through the blade plane, and
hence more power can be generated compared to a conventional
turbine of the same rotor blade diameter [11e13]. Several re-
searchers have examined the benets and economics of placing a
diffuser around a wind turbine [14,15]. It has been reported that the
speed of wind passing through the shroud is dramatically increased
[16,17]. Shrouding (diffuser augmented) horizontal axis wind tur-
bine has been shown to be an effective way to potentially improve
the performance of wind turbine for applications in built or urban
environments [9,10]. The ducted wind turbine overcomes many of
the problems associated with the use of conventional wind tur-
bines in an urban environment [18e20], which are hampered by
high levels of turbulence in the air stream, and are also constrained
by concerns over visual impact, noise and public safety [21].
Generating energy from the wind in an urban environment is an
attractive idea [22,23]. However, there are major problems over its
practical implementation at a signicant scale [24,25]. In such
location, the wind is normally weak, turbulent and unstable in

http://dx.doi.org/10.1016/j.energy.2015.05.126
0360-5442/ 2015 Elsevier Ltd. All rights reserved.
 V. Nikoli
c et al. / Energy 89 (2015) 324e333 325

Nomenclature n number of blades

c chord length
ANFIS adaptive neuro-fuzzy inference system CD Drag coefcient
DAWT diffuser-augmented wind turbine CL lift coefcients
CFD computational uid dynamics r0 radius of blade
Cp power coefcient rh radius of hub
Ct torque coefcient Ch rotation coefcient
FIS fuzzy inference system m membership function
P power of wind {p, q, r, s} premise parameters
r air density {a,b,c} variable parameters
r rotor radius {pi, qi,ri,si}
u wind velocity consequent parameters
h efciency of the wind turbine RMSE root-mean-square
Ux stream wise velocity R2 coefcient of determination
Ctrq torque coefcient r Pearson coefcient

terms of direction and speed [26,27], because of the presence of
buildings and other adjacent obstructions. To yield a reasonable
power output from a wind turbine located in this turbulent envi-
ronment the turbines have to improve their energy capture [28,29].
This means that turbines need to be specically designed to work
effectively in low and turbulent wind resource areas [30,31].
In this study is an analyzed diffuser effect on wind turbine rotor
performance. Since the using of CFD (computational uid dy-
namics) for the ducted wind turbine performance analyzing could
be very challenging and time consuming, soft computing tech-
niques are preferred. It is attempted to estimate the rotor perfor-
mance with diffuser for different number of blades and for different
wind input speed by soft computing methodology i.e. ANFIS
(adaptive neuro-fuzzy inference system). As the rotor performance,
power output, torque output and rotor rotational speed of the rotor
are tracked. Cp (power coefcient) and Ct (torque coefcient) are
used as the measure for power output and torque output
respectively.
ANFIS is one of the most powerful types of neural network
system [32]. ANFIS shows very good learning and prediction ca-
pabilities, which makes it an efcient tool to deal with encountered
uncertainties in any system. ANFIS, as a hybrid intelligent system
that enhances the ability to automatically learn and adapt, was used
by researchers in various engineering systems [33e35]. So far, there
are many studies of the application of ANFIS for estimation and
real-time identication of many different systems [36e38]. FIS
(fuzzy inference system) is the main core of ANFIS. FIS is based on
expertise expressed in terms of IFeTHEN rules and can thus be
employed to predict the behavior of many uncertain systems. FIS
advantage is that it does not require knowledge of the underlying
physical process as a precondition for its application. Thus ANFIS
integrates the FIS with a back-propagation learning algorithm of
neural network.
The key goal of this investigation is to establish an ANFIS for
estimation of the wind turbine rotor performances, power output,
torque output and rotor rotational speed in regard to diffuser effect
and wind input speed. The basic idea behind the soft computing
methodology is to collect input/output data pairs and to train the
proposed network from these data. This technique gives fuzzy logic
the capability to adapt the membership function parameters that
best allow the associated FIS to track the given input/output data
[39e41]. A CFD simulation is carried out to extract the training and
checking data for the ANFIS network.
2. Materials and methods
2.1. Diffuser augmented wind turbines
The power in wind is proportional to the cubic power of the
wind velocity approaching the wind turbine:
1 2 3
P rr pu h (1)
2
where the air density is represented by Refs. r, rotor radius is r, u is
the wind velocity and h is the efciency of the wind turbine. The
ratio of captured power to available power is referred to as the
power coefcient Cp. Equation (1) shows that a small amount of
acceleration leads to a large increase in the energy output. Fig. 1
illustrates an overview of the present DAWT (diffuser
augmented-wind turbine system). A ange generates a large sep-
aration behind it, where a very low-pressure region appears to
draw more wind compared to a diffuser with no ange. Owing to
this effect, the ow coming into the diffuser can be effectively
concentrated and accelerated. In this system, the maximum ve-
locity is obtained near the inlet of diffuser and thus a wind turbine
is located there as shown in Fig. 1.
The goal of this research is to determine the diffuser effect on
wind turbine performances (power output, torque output and
rotational speed) by ANFIS methodology.

2.2. Numerical method and computational conditions

The present ow eld is generally expressed by the continuity

Fig. 1. Schematic view and a ow mechanism around a diffuser-shrouded wind and the incompressible Reynolds-averaged NaviereStokes equa-
turbine. tions as follows [2]:
326 V. Nikoli
c et al. / Energy 89 (2015) 324e333

Zr0 Zr0
vUi 1 n 2 o
0; T dT r Ux ru2 CL sin b  CD cos bncrdr
vxi 2
rh rh
Zr0 n o
1 2
rUx CL sin b  CD cos b 1 ru=Ux 2 ncrdr
( ! ) 2
vUi vP v vUi vUj rh
pUj  pv  rui uj Fi ; (2)
vxj vxi vxj vxj vxi Zr0
1
r Ux2 Ch 2pr 2 dr (5)
2
where denotes a Reynolds-averaged value. In the equations, r, P, rh
Ui, ui and v respectively denote the density, mean static pressure,
mean velocity, turbulent uctuation and kinematic viscosity. In where r0 and rh denote the radius of blade and hub, respectively. Ch
Equation (2) Fi is the body-force term imposed for the represen- is rotation coefcient since it merges drag and lift coefcient and it
tation of a load. is dened as [2]:
The computational conditions are shown in Fig. 2. In Fig. 2, L, D, n o
f and h respectively denote the diffuser length, diffuser diameter at ncCL sin b  CD cos b 1 ru=Ux 2
the throat, diffuser opening angle and height of ange used in the Ch (6)
2pr
present experiment. The subscripts 0, 1, 2 and b in the Fig. 2 denote
values at the inlet (free stream), in front of the load (approaching), Therefore, the total torque coefcient generated by the blades
behind the load and at the exit of the diffuser, respectively. can be expressed as [2]:
In this study, the load inside the diffuser was represented by the
Zr0 Zr0
following general expression [2]: T 1 1
Ctrq 1 1 2 dT Ch K 2 2pr 2 dr; (7)
2 rU 2 AD
0 2 rU 0 AD
AD
rh rh

Ctrq 1 where Ap(D/2)2 and K U2/U0 is the acceleration factor. The po-
Fx rUx jUx j; Fr 0; (3)
D 2 wer coefcient is estimated as follows [2]:

where Ux is the stream wise velocity. In Equation (3) Ctrq and D are Zr0 Zr0  
the loading coefcient and its stream wise width imposed, Tu u 1 ru
Cp 1 1 2 dT Ch K 2 2pr 2 dr: (8)
respectively. In this study, Ctrq was determined by use of a disk- 2 rU 2A
0 2 rU0A
A Ux
rh rh
loading method in the two-dimensional cylindrical (xr) coordi-
nate as follows [2]:

n o 2.3. Adaptive neuro-fuzzy inference system

ncCL cos b CD sin b 1 ru=Ux 2
Ctrq ; (4) ANFIS model will be established in this study to estimate
2pr
diffuser effect on wind turbine performances. Since there are three
where n is the number of blades, u is the angular velocity of the outputs to determine, power coefcient Cp, torque coefcient Ctrq
wind turbine, c is the chord length and btan1(Ux/ru). In Equation and rotational speed of rotor, three ANFIS networks are modeled.
(4) CD and CL are, respectively, the drag and lift coefcients for the Each ANFIS network has three inputs: wind input speed, number of
relative angle of attack, abg, where g is the angle of blade blades and presence of the diffuser. Training and checking data for
setting. the ANFIS networks is extracted from the above presented nu-
On the other hand the torque generated by the blades can be merical analyze and by CFD simulations. With a proper training
estimated as follows [2]: scheme and ne ltered data-sets, ANFIS is capable to estimate the
diffuser effect on the wind turbine performances quite accurately
since it learns from training data.
FIS (fuzzy inference system) is the main core of ANFIS. FIS is
based on expertise expressed in terms of IFeTHEN rules and can
thus be employed to predict the behavior of many uncertain sys-
tems. FIS advantage is that it does not require knowledge of the
underlying physical process as a precondition for its application.
Thus ANFIS integrates the fuzzy inference system with a back-
propagation learning algorithm of neural network. The basic
structure of a FIS consists of three conceptual components: a rule
base, which contains a selection of fuzzy rules; a database, which
denes the MFs (membership functions) used in the fuzzy rules;
and a reasoning mechanism, which performs the inference proce-
dure upon the rules and the given facts to derive a reasonable
output or conclusion. These intelligent systems combine knowl-
edge, technique and methodologies from various sources. They
possess human-like expertise within a specic domain e adapt
themselves and learn to do better in changing environments. In
Fig. 2. Computational conditions of the diffuser system. ANFIS, neural networks recognize patterns, and help adaptation to
 V. Nikoli
c et al. / Energy 89 (2015) 324e333 327

environments. ANFIS is tuned with a back propagation algorithm

based on the collection of inputeoutput data.
There are three ANFIS networks with three inputs used in the
study and two membership functions on each input. In this study
bell-shaped membership functions are chosen with maximum
equal to 1 and minimum equal to 0. Fuzzy logic toolbox in MATLAB
was used for the entire process of training and evaluation of fuzzy
inference system. Fig. 3 shows an ANFIS structure with three inputs.
Nodes of the same layer have similar functions. The output of the
ith node in layer l is denoted as Ol,i.
In this work, the rst-order Sugeno model with three inputs and
fuzzy IF-THEN rules of Takagi and Sugeno's type is used. Since there
are 3 inputs and 2 membership functions on each input there are 8
Fig. 4. Bell-shaped membership function (a 2, b 4, c 6).
IF-THEN rules as follows (23 8):

Rule 1 : if x is A and y is C and z is E then f1 p1 x q1 y r1 z s1 is used here as it has the best abilities for the generalization of
Rule 2 : if x is A and y is C and z is F then f2 p2 x q2 y r2 z s2 nonlinear parameters:
Rule 3 : if x is A and y is D and z is E then f3 p3 x q3 y r3 z s3
Rule 4 : if x is A and y is D and z is F then f4 p4 x q4 y r4 z s4
Rule 5 : if x is B and y is C and z is E then f5 p5 x q5 y r5 z s5 1
m  2b (11)
Rule 6 : if x is B and y is C and z is F then f6 p6 x q6 y r6 z s6 xc
1
Rule 7 : if x is B and y is D and z is E then f7 p7 x q7 y r7 z s7 a
Rule 8 : if x is B and y is D and z is F then f8 p8 x q8 y r8 z s8
(9) where {a,b,c} is the variable set. The bell-shaped function varies
accordingly as the values of the variables change, therefore mani-
The parameters {p, q, r, s} are premise paremters in the ANFIS festing different types of membership functions for fuzzy sets
model. These parameters should be optimized during trianing A, B, C, D, E and F . The bell-shaped function depend on three pa-
procedure of the ANFIS network. rameters a, b, and c. The parameter b is usually positive. The
The rst layer consists of input variables membership functions parameter c located the center of the curve as it is shown in Fig. 4.
(MFs). This layer just supplies the input values to the next layer. In The second layer (membership layer) checks for the weights of
the rst layer every node is an adaptive node with a node functions each MFs. It receives the input values from the 1st layer and acts as
O1,i (i1,2,3,4,5,6): MFs to represent the fuzzy sets of the respective input variables.
Every node in the second layer is non-adaptive and this layer
O1;1 mA x; multiplies the incoming signals and sends the product out like
O1;2 mB x;
O1;3 mC y;
(10) O2;1 w1 mA xmC y;
O1;4 mD y;
O1;5 mE z; O2;2 w2 mB xmD y;
(12)
O1;6 mF z O2;3 w3 mC ymE z;
O2;4 w4 mD ymF z
where x or y or z is the input to node i (i1,2,3,4,5,6) and A or B or C
Each node output represents the ring strength of a rule or
or D or E or F is an associated linguistic label (such as small or
weight.
large). In other words, O1,i is the membership grade of a fuzzy sets
The third layer is called the rule layer. Each node (each neuron)
A, B, C, D, E and F . It stipulates the extent to which the specied
in this layer performs the pre-condition matching of the fuzzy
input x or y or z satises the quantiers A or B or C or D or E or F. In
rules, i.e. they compute the activation level of each rule, the number
this instance, the membership function for can be any suitable
of layers being equal to the number of fuzzy rules. Each node of
parameterized membership function. The generalized bell function
these layers calculates the weights which are normalized. The third
layer is also non-adaptive and every node calculates the ratio of the
rule's ring strength to the sum of all rules' ring strengths like

w1
O3;1 w1 ;
w1 w2
w2
O3;2 w2 ;
w1 w2
(13)
w3
O3;3 w3 ;
w3 w4
w4
O3;4 w4
w3 w4
The outputs of this layer are called normalized ring strenghts
or normalized weights.
The fourth layer is called the defuzzication layer and it pro-
vides the output values resulting from the inference of rules. Every
Fig. 3. ANFIS structure with three inputs. node in the fourth layer is an adaptive node with node function
328 V. Nikoli
c et al. / Energy 89 (2015) 324e333

where Pi and Oi are known as the experimental and forecast values

O4;1 w1 fi w1 pi x qi y ri z si ; of Cpmax, Ctmax amd RPMmax, respectively, and n is the total
O4;2 w2 fi w2 pi x qi y ri z si ; number of test data.
(14)
O4;3 w3 fi w3 pi x qi y ri z si ;
O4;4 w4 fi w4 pi x qi y ri z si
3.2. ANFIS results
where {pi, qi,ri,si} is the parameter set and in this layer is referred to
as consequent parameters. At the beginning the ANFIS networks is trained with the
The fth layer is called the output layer which sums up all the extracted data from the computational uid dynamics simulations
inputs coming from the fourth layer and transforms the fuzzy and numerical analysis. Table 1 shows training input and output
classication results into a crisp (binary). The output represents data parameters. The output parameters represent optimal
estimated modulation transfer function of the optical system. The (maximal) values for the wind turbine parameters. These optimal
single node in the fth layer is not adaptive and this node computes output parameters varied according to input values. One can note
the overall output as the summation of all incoming signals that the maximal Cp is 0.773 and the maximal Ct 0.409 is during the
CFD simulations. These values are higher than usually power co-
X
4 P4 efcient since diffuser was implemented during the simulations.
wi f
O5 wi fi Pi1
4
(15) However the obtained maximal values of Cp and Ct are not practical
i1 i1 wi veried. Many factors will inuence to make the coefcient smaller
The hybrid learning algorithms were applied to identify the in real applications. In other words the ANFIS networks were
parameters in the ANFIS architectures. In the forward pass of the trained with the optimal data and the ANFIS output should be
hybrid learning algorithm, functional signals go forward until Layer optimal values in relation to different inputs.
4 and the consequent parameters are identied by the least squares There are three ANFIS networks: ANFIS-1 for power coefcient
estimate. In the backward pass, the error rates propagate back- Cp estimation, ANFIS-2 for torque coefcient Ct estimation and
wards and the premise parameters are updated by the gradient ANFIS-3 for rotational speed estimation of turbine. Two bell-
descent. shaped membership functions are used for each input during the
training procedure. It is not appropriate to further increase the
number of the membership functions since there are too many
3. Results parameters to estimate.
The nal decision surfaces after training procedure of the ANFIS
3.1. Evaluating accuracy of proposed model networks are shown in Fig. 5, Fig. 6 and Fig. 7 respectively for power
coefcient prediction, torque coefcient prediction and rotational
Predictive performances of proposed model were presented as speed prediction of the turbine. These ANFIS surfaces represent
RMSE (root means square error), Coefcient of determination (R2) optimal values for the output parameters. The ANFIS provided
and Pearson coefcient (r). These statistics are dened as follows: optimal values since the ANFIS networks were trained with opti-
mized data in initial stage. In other words the optimization was
1) RMSE (root-mean-square error) performed during CFD simulations and training data acquiring. One
can note that the maximal Cp value 0.773 was not predicted by
s ANFIS model. It means the CFD training data is not real data and
Pn 2 cannot be achieved in real applications. It should be noted as the
i1 Pi  Oi
RMSE ; (16) number of blades input it was used solidity percentage of each
n
blade. Each blade represents 10% solidity. For example 40% solidity
represents 4 blades. The method of estimation diffuser effect and
wind input speed inuence on wind turbine performances, power
coefcient Cp, torque coefcient Ct and rotational rotor speed RPM
2) Pearson correlation coefcient (r)
(rotation per metres), is impemented in MATLAB Simulink block
diagram as it shown in Appendix.
! ! !
Pn Pn Pn
n i1 Oi $Pi  i1 Oi $ i1 Pi 3.3. Wind tribune performances
r v
u0 !2 1 0 !2 1 Figs. 8e10 shows power coefcient, torque coefcient and
u
u@ Pn 2 Pn
A @
Pn 2 Pn
A rotational speed percentage changes of rotor for different number
t n i1 Oi  i1 Oi $ n i1 P i  i1 Pi

(17) Table 1
Training data for the ANFIS network: (a) input parameters, (b) output parameters.

(a)

Inputs Min Max

3) coefcient of determination (R2)
1. Solidity 30% 60%
2. Wind speed 10 m/s 20 m/s
3. Flanged diffuser 0 1
" #2
Pn     (b)
Oi  Oi $ Pi  Pi Outputs Min Max
i1
1. Cpmax 0.228 0.773
R2 P   P   (18)
n 2. Ctmax 0.069 0.409
i1 Oi  Oi $ ni1 Pi  Pi 3. RPMmax 1077 3251
 V. Nikoli
c et al. / Energy 89 (2015) 324e333
Fig. 5. ANFIS prediction of maximal power coefcient Cpmax.
Fig. 6. ANFIS prediction of maximal torque coefcient Ctmax.
of blades and different input speed after adding diffuser. As can be
seen, the maximal changing of power and torque coefcient was
achieved for 3 blades. In other words the Cp was increased in 160%
after adding diffuser for three blades. Fig. 11 shows that the
Fig. 7. ANFIS prediction maximal rotational speed of turbine RPMmax.
329
Fig. 8. Increasing of the maximal power coefcient Cp in percentage after adding
diffuser for different number of blades and different wind input speed.
Fig. 9. Increasing of the maximal torque coefcient Ct in percentage for different
number of blades and different wind input speed.
maximal changing of rotational speed of rotor for three blades was
achieved for wind input speed of 20 m/s.
3.4. Performance evaluation of proposed ANFIS model
In this section performance results of ANFIS predictive model
are reported. Fig. 11(a) presents the accuracy of developed ANFIS
predictive model for Cpmax, while Fig. 11(b) presents the accuracy
of developed ANFIS predictive model for Ctmax. Finally Fig. 11(c)
shows ANFIS prediction of RPMmax. It can be seen that the most of
Fig. 10. Increasing of the maximal rotational speed of rotor RPM in percentage for
different number of blades and different wind input speed.
330 V. Nikoli
c et al. / Energy 89 (2015) 324e333

Fig. 11. Scatter plots of actual and ANFIS predicted values of (a) Cpmax, (b) Ctmax and (c) RPMmax.

Table 2 Table 4
Comparative performance statistics of the ANFIS Cpmax predictive model. Comparative performance statistics of the ANFIS RPMmax predictive model.

RMSE R2 r RMSE R2 r

ANFIS 0.011739 0.9954 0.997695 ANFIS 39.49314 0.9932 0.996583

ANN 0.030613 0.9695 0.98464 ANN 103.7439 0.9592 0.979378
GP 0.0313 0.9673 0.9835 GP 103.8132 0.9592 0.979375
SVM 0.029755 0.9708 0.985314 SVM 86.64905 0.9773 0.988571

the points fall along the diagonal line for the ANFIS prediction (genetic programming) [45e47] and SVM (support vector machine)
models. Consequently, it follows that prediction results are in very [48e50]. Conventional error statistical indicators i.e. RMSE, r and R2
good agreement with the actual values of the wind turbine char- were used for comparison. Tables 2e4 summarize the prediction
acteristics. This observation can be conrmed with very high value accuracy results for two analyzed models.
of coefcient of determination. The number of either overestimated Fig. 12 presents ANFIS prediction results in comparison with
or underestimated values is limited. It is obvious that the predicted experimental results for Cpmax, Ctmax and RPMmax.
values enjoy high level precision.
In order to demonstrate the merits of the proposed ANFIS
approach on a more denite and tangible basis, ANFIS models 4. Conclusion
prediction accuracy was compared with other soft computing
methodologies like which were used as a benchmark. These Diffuser-augmented wind turbines are capable of concentrating
benchmark models are ANN (articial neural network) [42e44], GP the energy in the wind and permits more power to be extracted
from the wind. Since wind power is well known to be directly
proportional to the cube of wind speed approaching a wind turbine
Table 3
it leads that with a slight acceleration in wind speed, it will cause a
Comparative performance statistics of the ANFIS Ctmax predictive model.
large increase in the wind power output. It is important issue that a
RMSE R2 r wind power generation is improved to raise the efciency as one of
ANFIS 0.018587 0.9537 0.976557 natural energy sources in order to promote the usage of sustainable
ANN 0.020435 0.9444 0.971799 energy. According to the background, the wind turbine with a
GP 0.020628 0.9436 0.971414 diffuser shroud is developed as one of high performance wind
SVM 0.019195 0.9555 0.975075
turbines. In particular, the diffuser augmented-wind turbines can
 V. Nikoli
c et al. / Energy 89 (2015) 324e333
Fig. 12. ANFIS prediction results in comparison with experimental results for (a) Cpmax, (b) Ctmax and (c) RPMmax.
generate electric power even in low velocity wind since the diffuser
shroud increases the wind velocity at rotor.
In this study was analyzed diffuser effect on wind turbine rotor
performances. The impact of the diffuser and wind input speed on
the rotor performance was investigated. As the parameters of the
rotor performance, power output (power coefcient Cp), torque
output (torque coefcient Ct) and rotational speed of the rotor were
used. A Simulink model was developed in MATLAB with the ANFIS
network for the diffuser effect estimation. Simulations were run in
MATLAB and the results were observed on the corresponding
output blocks. The main advantages of the ANFIS scheme are:
computationally efcient, well-adaptable with optimization and
adaptive techniques. ANFIS can also be used with systems handling
more complex parameters. Another advantage of ANFIS is its speed
of operation, which is much faster than in other control strategies;
the tedious task of training membership functions is done in ANFIS.
The main contribution of the investigation is numerical results.
Further analyzing should include experimental verication of the
numerical results. Also the ANFIS network is data-driven method
which indicates that the method is dependable on training data
selection. In this investigation was examined different training data
combination to conrm the results. However for further investi-
gation other soft computing method should be examined also for
this topic.
331
Acknowledgment
This paper is supported by Project Grant 35005 00 Research and
development of new generation wind turbines of high-energy ef-
ciency (2011e2014) nanced by Ministry of Education, Science
and Technological Development, Republic of Serbia.
Appendix
There are 6 ANFIS networks in the block diagram since three
ANFIS networks estimate cases when the diffuser is active as it
shown in Fig. 13. The other three ANFIS networks estimate cases
when the diffuser is inactive. As can be seen wind input speed range
is from 10 m/s to 20 m/s. This approach is very useful for fast
estimation of the diffuser effect for the DAWT project since
computational uid dynamics simulations could be very chal-
lenging and time consuming. As can be seen in the SIMULINK block
diagram outputs give percentage of power coefcient, torque co-
efcient amd rotational speed increasing after implementation of
diffuser. Fig. 8 shows a case for 3 blades and wind input speed
15 m/s. For the case power coefcient is increased 132.1%, torque
coefcient is increased 108% and rotational speed of rotor is in-
crease 12.8%.
332 V. Nikoli
c et al. / Energy 89 (2015) 324e333

Fig. 13. Simulink block diagram for estimation of diffuser effect on power coefcient Cp, torque coefcient Ct and rotational rotor speed RPM.

References [12] Matsushima T, Takagi S, Muroyama S. Characteristics of a highly efcient

propeller type small wind turbine with a diffuser. Renew Energy 2006;31:
1343e54.
[1] Igra O. ANFIS: cost-effectiveness of the vortex-augmented wind turbine. En-
[13] Chong WT, Poh SC, Fazlizan A, Yip SY, Chang CK, Hew WP. Early development
ergy 1979;4:119e30.
of an energy recovery wind turbine generator for exhaust air system. Appl
[2] Abe K, Nishida M, Sakurai A, Ohya Y, Kihara H. Experimental and numerical
Energy 2013;112:568e75.
investigations of ow elds behind a small wind turbine with a anged
[14] Chen TY, Liao YT, Cheng CC. Development of small wind turbines for moving
diffuser. J Wind Eng Industrial Aerodynamics 2005;93:951e70.
vehicles: Effects of anged diffusers on rotor performance. Exp Therm Fluid
[3] Ohya Y, Karasudani T, Sakurai A, Abe K-i, Inoue Masahiro. Development of a
Sci 2012;42:136e42.
shrouded wind turbine with a anged diffuser. J Wind Eng Indus Aero-
[15] Grassmann H, Bet F, Cabras G, Ceschia M, Cobai D, DelPapa C. A partially static
dynamics 2008;96:524e39.
turbinedrst experimental results. Renew Energy 2003;28:1779e85.
[4] Rachman A, Muhtar LA, Levi S, Salea N. An experiment on horizontal and
[16] Rhee D-H, Cho HH. Effect of vane/blade relative position on heat transfer
vertical wind turbines with incorporation of rounded shroud device using
characteristics in a stationary turbine blade: Part 1. Tip and shroud. Int J
wind simulation in a vehicle. Int J Innovation Appl Stud June 2013;3(2). ISSN:
Therm Sci 2008;47:1528e43.
2028-9324:365e74.
[17] Igra O. Research and development for shrouded wind turbines. Energy
[5] Isensee GM, Abdul-Razzak H. Modeling and analysis of diffuser augmented
Convers Manag 1981;21:13e48.
wind turbine. Int J Energy Sci IJES 2012;2(3):84e8.
[18] Hu S-Y, Cheng J-H. Innovatory designs for ducted wind turbines. Renew En-
[6] Ohya Y, Karasudani T. A shrouded wind turbine generating high output power
ergy 2008;33:1491e8.
with wind-lens technology. Energies 2010;3:634e49. http://dx.doi.org/
[19] Li QS, Chen FB, Li YG, Lee YY. Implementing wind turbines in a tall building for
10.3390/en3040634.
power generation: a study of wind loads and wind speed amplications.
[7] Khunthongjan P, Janyalertadun A. A study of diffuser angle effect on ducted
J Wind Eng Ind Aerodyn 2013;116:70e82.
water current turbine performance using CFD, Songklanakarin. J Sci Technol
[20] Chong WT, Fazlizan A, Poh SC, Pan KC, Ping HW. Early development of an
Jan. - Feb. 2012;34(1):61e7.
innovative building integrated wind, solar and rain water harvester for urban
[8] Kosasih B, Tondelli A. Experimental study of shrouded micro-wind turbine.
high rise application. Energy Build 2012;47:201e7.
Procedia Eng 2012;49:92e8.
[21] Kirke BK. Tests on ducted and bare helical and straight blade Darrieus hy-
[9] Toshimitsu K, Kikugawa H, Sato K, Sato T. Experimental investigation of
drokinetic turbines. Renew Energy 2011;36:3013e22.
performance of the wind turbine with the anged-diffuser shroud in sinu-
[22] Wang F, Bai L, Fletcher J, Whiteford J, Cullen D. The methodology for aero-
soidally oscillating and uctuating velocity ows. Open J Fluid Dyn 2012;2:
dynamic study on a small domestic wind turbine with scoop. J Wind Eng
215e21.
Industrial Aerodynamics 2008;96:1e24.
[10] Igra O. Compact shrouds for wind turbines. Energy Convers 1977;16:149e57.
[23] Alam F, Golde S. An aerodynamic study of a micro scale vertical axis wind
[11] Foreman KM, Gilbert B, Oman RA. Diffuser augmentation of wind turbines. Sol
turbine. Procedia Eng 2013;56:568e72.
Energy 1978;20:305e11.
 V. Nikoli
c et al. / Energy 89 (2015) 324e333
[24] Mithraratne N. Roof-top wind turbines for microgeneration in urban houses
in New Zealand. Energy Build 2009;41:1013e8.
[25] Millward-Hopkins JT, Tomlin AS, Ma L, Ingham DB, Pourkashanian M.
Assessing the potential of urban wind energy in a major UK city using an
analytical model. Renew Energy 2013;60:701e10.
[26] Walker SL. Building mounted wind turbines and their suitability for the urban
scaledA review of methods of estimating urban wind resource. Energy Build
2011;43:1852e62.
[27] Abohela I, Hamza N, Dudek S. Effect of roof shape, wind direction, building
height and urban conguration on the energy yield and positioning of roof
mounted wind turbines. Renew Energy 2013;50:1106e18.
[28] Ayhana D, Saglam Safak. A technical review of building-mounted wind power
systems and a sample simulation model. Renew Sustain Energy Rev 2012;16:
1040e9.
[29] Booker JD, Mellor PH, Wrobel R, Drury D. A compact, high efciency contra-
rotating generator suitable for wind turbines in the urban environment.
Renew Energy 2010;35:2027e33.
[30] Booker JD, Mellor PH, Wrobel R, Drury D. A numerical investigation into the
effect of diffusers on the performance of hydro kinetic turbines using a
validated momentum source turbine model. Renew Energy 2010;35:
1152e8.
[31] Grant A, Johnstone C, Kelly N. Urban wind energy conversion: the potential of
ducted turbines. Renew Energy 2008;33:1157e63.
[32] Jang J-SR. ANFIS: adaptive-network-based fuzzy inference systems. IEEE
Trans. Syst Man, Cybern 1993;23:665e85.
[33] Petkovi
c D, Issa M, Pavlovi
c ND, Pavlovi

c NT, Zentner L. Adaptive neuro-fuzzy
estimation of conductive silicone rubber mechanical properties. Expert Syst
Appl 2012;39. ISSN: 0957-4174:9477e82.
[34] Petkovi


c D, Cojba si
 Adaptive neuro-fuzzy estimation of automatic nervous
c Z.
system parameters effect on heart rate variability. Neural Comput Appl
2012;21(8):2065e70.
[35] Petkovi
c D, Issa M, Pavlovi


c ND, Zentner L, Cojbasi
 Adaptive neuro fuzzy
c Z.
controller for adaptive compliant robotic gripper. Expert Syst Appl 2012;39.
ISSN: 0957-4174:13295e304.
333
[36] Ekici BB, Aksoy UT. Prediction of building energy needs in early stage of design
by using ANFIS. Expert Syst Appl 2011:5338e52.
[37] Khajeh A, Modarress H, Rezaee B. Application of adaptive neuro-fuzzy infer-
ence system for solubility prediction of carbon dioxide in polymers. Expert
Syst Appl 2009;36:5728.
[38] Inal M. Determination of dielectric properties of insulator materials by means
of ANFIS: a comparative study. Expert Syst Appl 2008;195:34.
[39] Wahida Banu RSD, Shakila Banu A, Manoj D. Identication and control of
nonlinear systems using soft computing techniques. Int J Model Optim
2011;1(1):24e8.
[40] Grigorie TL, Botez RM. Adaptive neuro-fuzzy inference system-based con-
trollers for smart material actuator modelling. J Aerosp Eng 2009:655e68.
[41] Akcayol MA. Application of adaptive neuro-fuzzy controller for SRM. Adv Eng
Softw 2004;35:129e37.
[42] Karunanithi N, Grenney WJ, Whitley D, Bovee K. Neural networks for river
ow prediction. J Comput CivilEngineering 1994;8(2):201e20.
[43] ASCE Task Committee. Articial neural networks in hydrologyd1: pre-
liminary concepts. J Hydrologic Eng 2000a;5(2):115e23.
[44] ASCE Task Committee. Articial neural networks in hydrologyd2: hydrologic
applications. J HydrologicEngineering 2000b;5(2):124e37.
[45] Koza JR. Genetic programming: on the programming of computers by natural
selection. Cambridge, MA: MIT Press; 1992.
[46] Babovic V, Keijzer M. Rainfall runoff modeling based on genetic programming.
Nord Hydrol 2000;33:331e46.
[47] Khu ST, Liong S-Y, Babovic V, Madsen H, Muttil N. Genetic programming and
its application in real-time runoff forecasting. J Am Water Resour Assoc
2001;37:439e51.
[48] Asefa T, Kemblowski M, McKee M, Khalil A. Multi-time scale stream ow
predictions: the support vector machines approach. J Hydrology 2006;318(1):
7e16.
[49] Ji Y, Sun S. Multitask multiclass support vector machines: model and exper-
iments. Pattern Recognit 2013;46(3):914e24.
[50] Sun S. A survey of multi-view machine learning. Neural Comput Appl
2013;23(7e8):2031e8.