Anda di halaman 1dari 13

Electr Eng (2008) 90:479491

DOI 10.1007/s00202-008-0097-3
ORIGINAL PAPER
Self tuning control of wind turbine using neural network identier
M. Sedighizadeh A. Rezazadeh
Received: 26 September 2007 / Accepted: 15 December 2007 / Published online: 13 February 2008
Springer-Verlag 2008
Abstract The nonlinear characteristics of the wind turbines
and electric generators necessitate that grid connected wind
energy conversion systems (WECS) use nonlinear controls.
The present paper proposes an adaptive self tuning control
strategy with neural network Morlet wavelet for WECS con-
trol. The proposed strategy is based on single layer feedfor-
ward neural networks with hidden nodes of adaptive Morlet
wavelet functions controller and an innite impulse response
recurrent structure. The neuro controller is based on a cer-
tain model structure to approximately identify the system
dynamics of WECS, and control its response. The proposed
controller is studied in three situations: without noise, with
measurement input noise and with disturbance output noise.
Finally, the results of the performance of the new controller
were comparedwitha multilayer perceptronnetworkproving
a more precise modeling and control of WECS.
Keywords Adaptive control Morlet wavelets
Wind energy conversion system
1 Introduction
The increasing interest in environmental concerns has forced
the industrial and academic communities to look for clean
sources of energy. The wind energy is one of the viable can-
didates to replace the conventional energies. The newenergy
sources require new power devices technologies, new circuit
topologies and novel control strategies for their efciency.
M. Sedighizadeh (B) A. Rezazadeh
Faculty of Electrical and Computer Engineering,
Shahid Beheshti University, Tehran, Iran
e-mail: mos_sedighizadeh@yahoo.com
Therefore, the environmental and safety energy demands
have led to worldwide research efforts in achieving control
strategies compatible with the renewable energy sources.
Currently, the wind energy conversion systems (WECS)
are constructed with standalone topology or hybrid topol-
ogy or grid topology. The turbines are traditionally linked
with induction generators (squirrel cage or wound rotor) to
achieve a robust, lowmaintenance and cost-effective system.
However, there is a drawback in this structure. Being a highly
nonlinear system, it requires a nonlinear control strategy to
set the system in its optimal operation point. So in the previ-
ous works, different methods based on classic methods and
intelligent methods for controlling of WECS are introduced.
But regarding to nonlinear structure of WECS, stochastic
parameters of wind and other uncertainties in system, the
intelligent methods are more attractive than classic meth-
ods. Many authors [14] survey fuzzy logic control, neural
network control, expert system control and synthesis intelli-
gent control methods that used in the stability, speed control
system and maximum-power transfer of WECS and showed
that the intelligent control approaches are robust and exhibit
a superior performance to classic control methods.
Traditional self tuning adaptive control approaches are
interesting alternatives for identifying and control of WECS
nonlinear dynamics systems. But they cannot deal with com-
plex nonlinear systems. The problem is exacerbated when
the complex functions describing the systems are unknown
and change with time. Developments in the self tuning adap-
tive neuro controller design have proved to be useful for a
wide class of practical situations [5]. Mayosky and Cancelo
[6] used this idea for controlling the WECS. They proposed
a neural-network-based structure that consists of two com-
bined control actions, a supervisory control and an RBF
(radial basis function) network-based self tuning adaptive
controller.
123
480 Electr Eng (2008) 90:479491
Lekutai and VanLandingham [7] presented an innovative
combination of the wavelet transform theory with the basic
concept of neural networks, proposing a new mapping net-
work. The resulting network called neural network adaptive
wavelets or wavenets is presented as an alternative to feed-
forward neural network to approximate arbitrary nonlinear
functions. It is found that this new controller is very useful
for identication and control of systems with unknown and
highly nonlinear dynamics [7].
Sedighizadeh et al. [8, 9] used idea of self-tuning control
of nonlinear systems using neural network adaptive frame
wavelets and WECS model presented by Mayosky and
Cancelo, for identication and control of WECS. They sug-
gested an adaptive PI and PID controller using rational func-
tion with second-order poles (RASP1) wavenets for Wind
turbine control. Sedighizadeh et al. [10] also suggested an
adaptive controller using Morlet wavelets frames neural net-
work for identication and control of WECS. Their wavenet
consists of a single layer feed forward neural network with
hidden nodes of adaptive wavelet functions followed with
an innite impulse response (IIR) recurrent structure. The
IIR cascaded to the neural network to provide local structure
network, to improve the speed of learning. A neural network
estimator approximates the unknown dynamics of the plant,
and then the parameters of the proposed controller are set
within a feedback loop based on algebraic computations fol-
lowing the sampled input-output data.
In the present paper, the suggested controller in [10] was
simulated in three situations: without noise, with measure-
ment input noise and disturbance output noise to control
WECS. The results have also been compared with the MLP
based performance.
This paper is organized as follows. Section 2 illustrates
the WECS system dynamics and model. Section 3 presents a
wavenet control strategy and discusses the adaptive network
algorithmic implementation, providing the neuro controller
design architecture. Section 4 identies WECS dynamics
using the proposed neural network with various number of
Morlet wavelets in a hidden layer. After identication, the
section also discusses controlling the system in three differ-
ent situations regarding to the noise of the system. Finally,
Sect. 5 presents the conclusions.
2 Wind energy conversion systems
2.1 Wind turbine characteristics
The dynamic model of horizontal-axis type wind turbine
which is the most commonly used wind turbine is discussed
in this section. The output mechanical power gained from a
wind turbine is calculated as [6].
P = 0.5C
p
(V

)
3
A, (1)
Fig. 1 Power coefcient C
p
versus turbine speed
where is the air density, A is the area swept by the blades,
and V

is the wind speed. C


p
, is power coefcient and is
approximated as follows,
C
p
= +
2
+
3
, (2)
where,
= R/V

(3)
R is the radius of the turbine and is the rotational speed.
, and are constructive parameters for a given turbine.
Figure 1 illustrates typical C
p
versus turbine speed curves,
with the wind speed or V

as a parameter. It can be noticed


that maximum value for C
p
which is represented as C
p
max
,
is a constant for any given turbine. Replacing this value in
(1), yields the maximum output power for the given wind
speed. This maximum output power is achieved in a certain
rotational speed,
opt
, according to a certain wind speed, V

.
The resulting torque by the wind turbine is calculated as
[6]:
T
l
= 0.5
_
C
p

_
(V

)
2
R
2
(4)
The torque/speed curves of a typical wind turbine, with V

as
a parameter is shown in Fig. 2. It can be seen that maximum
generated power do not coincide with maximum developed
torque points.
Superimposed to those curves is the curve of P
max
. As it
can be seen (Fig. 2), the maximum generated power and the
maximumtorque are not achieved at the same speed. Optimal
performance of the turbine is achieved when it operates at the
P
max
condition. Setting the turbine at the P
max
in specied
wind speed is the control objective of the present paper.
123
Electr Eng (2008) 90:479491 481
Fig. 2 Torque/speedcurves (solid) of a typical windturbine. The curve
of C
p max
is also plotted (dotted)
Fig. 3 Slip recovery using a static Kramer drive
2.2 Induction generators and slip power recovery
WECSs mainly use induction generators to produce elec-
tricity and are available in two basic congurations namely,
variable speed and constant speed. Due to their higher per-
formance, variable speed conguration WECSs are more
popular. The induction generators used for variable speed
constant frequency (VSCF) applications are of two types:
cage rotor machines and wound rotor machines.
Figure 3 illustrates a typical wound machine congura-
tion. The power generation and generators torque can be
controlled by varying the ring angle , of inverter in Static
Kramer drive [6].
Static Kramer drive which is a combination of a rectier
and an inverter is used to inject slip power to the ACline as in
Fig. 3. As the wind speed changes, the C
p
will also change.
The ring angle of the inverter must be controlled to achieve
the maximum power with changing the Torque- speed curve
of the generator.
The torque developed by the generator/Kramer drive com-
bination is [6]:
T
g
=
3V
2
s R
eq

s
_
(s R
s
+ R
eq
)
2
+(s
s
L
ls
+ s
s
L
lr
)
2
_ , (5)
where
R
eq
=
s
_
n
2
2
s R
b
+(n
1
|cos()|)
2
R
s
1
n
1
|cos()|

_
((n
2
s) (n
1
|cos()|)
2
)
R
b
= R
r
+ 0.55R
f
= 2n
2
2
R
b
s R
s
+(n
2
s R
s
)
2
+ n
2
2
(s
s
L
ls
+ s
s
L
lr
)
2
+(n
2
R
b
)
2

_
n
1
|cos()|(
s
L
ls
+
s
L
lr
)
2
_
(6)
and
n
1
: transformation ratio between rotor and stator
wounds;
n
2
: turn ratio of the transformer between the Kra-
mer drive output and the AC line;
R
r
, R
s
, R
f
: Rotor, stator, and dc link resistances;
L
ls
: stator dispersion inductance;
L
lr
: rotor dispersion inductance;
: ring angle;

s
: synchronous pulsation;

s
: synchronous machine rotational speed;
V: stator voltage
s: slip
: ring angle of inverter
(all values referred to the rotor side).
2.3 Turbine/generator model
The dominant dynamics of the whole system (turbine plus
generator) are those related to the total moment of inertia.
Therefore, ignoring torsion in the shaft, generators electric
dynamics, and other higher order effects, the approximate
systems dynamic model is
J

= T
l
(, V

) T
g
(, ), (7)
where J is the total moment of inertia. Regarding (4) and (5),
123
482 Electr Eng (2008) 90:479491
Fig. 4 Control strategy proposed. The ring angle is adjusted so that
the turbines operation point settles to the C
p max
condition
systems model becomes

=
1
J
_
0.5
_
C
p

_
(V

)
2
R
2

3V
2
s R
eq

s
_
(s R
s
+ R
eq
)
2
+(s
s
L
ls
+ s
s
L
lr
)
2
_
_
, (8)
where R
eq
depends nonlinearly on the control action cos()
according to (6). C
p
, and are nonlinear functions of V

(2). Variation of generator parameters due to aging and tem-


perature, leads to using a nonlinear adaptive control strategy.
This control strategy system aims at placing the turbine in
its maximum power generation point, despite the variations
in the wind speed and generators parameters. The turbine
torque,T
l
, for a given V

, and the generated torque,T


g
, for
a given cos(), are sketched in Fig. 4. It should be mentioned
that for a given wind speed, the turbines operational curve
and optimum generation point are xed. According to (7),
the intersection of T
l
and T
g
curves represents the equi-
librium point (

= 0) of the turbine-generator pair. The


control strategy converges the rotational speed, , and tur-
bine torque,T
l
, to their optimal values by changing the ring
angle of the inverter, as the wind speed changes [6].
The general form of

, is a nonlinear function of and


depending on the turbine and generator characteristics as
in (8). The designing of system is so that the maximum tur-
bine torque occurs 0.5 to 0.7 of the generator torque peak.
Regarding to the generator torque curves in this region T
g
is
considered as a linear expression [6]. The generated torque
curve in optimal point is shown in Fig. 4. The expression for
T
g
in (5), can be rewritten as:
T
g
= K
1
+ K
2
cos(). (9)
Thus the whole system will have the following expression.

=
1
J
_
0.5
_
C
p

_
(V

)
2
R
2
+ K
1
+ K
2
cos()
_
.
(10)
The standard normal form for (10) is

= f () + bu, (11)
where f is a nonlinear function of rotational speed, , b is a
constant and u is the system input which is the cosine of the
ring angle, .
3 Control strategy
3.1 Structure and algorithms
In order to deal with the tracking operation using a neural net-
work based controller, the unknown nonlinear WECS should
be identied according to a particular model. In this particu-
lar identication process, the model consists of a neural net-
work topology with the wavelet transformused in the hidden
units. This network is named wavenet. The concept of wave-
net introduces a super-wavelet, which is a linear combina-
tion of daughter wavelets that is also a wavelet. The daughter
wavelets are simply a dilated and shifted version of the ori-
ginal wavelet or mother wavelet. The super-wavelet allows
the shape of the wavelet to adapt to a particular problem,
a concept which goes beyond adapting the parameters of a
xed shape wavelet. This network has shown good results in
nonlinear system and signal identication and control [11].
A local innite impulse response (IIR) block structure is
cascaded with the network (Fig. 5). The IIR synopsis net-
work is used to create double local network architecture. This
architecture provides a computationally efcient method for
training the system, resulting in quick learning and fast
convergence [7]. The algorithm of proposed neural network
adaptive wavelets is similar to those in [7] where any desi-
red signal y(t ) can be modeled by a linear combination of
Morlet daughter wavelets h
a,b
(t ). Here h
a,b
(t ) are gener-
ated by dilation, a, and translation, b, from a Morlet mother
wavelet:
h
a,b
(t ) = h
_
t b
a
_
= cos
_

o
_
t b
a
__
exp
_
0.5
_
t b
a
_
2
_
(12)
With the dilation factor a > 0. The
o
is the wavelet fre-
quency which is chosen
o
= 4 which meets approximately
the admissibility condition [7].
123
Electr Eng (2008) 90:479491 483
Fig. 5 IIR adaptive wavelet
network structure: a neural
network architecture, b IIR
model
+
+
+
The approximated signal of the network y(t ) can be mod-
eled by [7]:
y(t ) =
M

i =0
c
i
z(t i )u(t ) +
N

j =1
d
j
y(t j )v(t ), (13)
where
z(t ) =
K

k=1
W
k
h
a
k
,b
k
(t ) (14)
K is the number of wavelets, w
k
is the kth weight coef-
cient. M is the number of feedforward delays and c
j
is the
feedforward coefcient of the IIR lter. N is the number of
feedback delays and d
j
is the recursive lter coefcients. The
signals u(t ) and v(t ) are the input (cosine of ring angle) and
co-input to the system at time t , respectively. Input v(t ) is
usually kept small for feedback stability purposes [7].
The neural network parameters a
k
, b
k
, c
i
, w
k
and d
j
can
be ned by optimizing the following objective function by
means of least mean square (LMS) minimization
E =
1
2
T

t =1
e
2
(t ), (15)
where e(t ) is time varying error function and y(t ) is the desi-
red response (rotational speed of wind turbine).
e(t ) = y(t ) y(t ). (16)
To minimize the cost function, we may use the method of
steepest descent which requires the gradients
E
w
k
,
E
b
k
,
E
a
k
,
E
c
i
and
E
d
j
for updating the incremental changes of each
particular parameter w
k
, b
k
, a
k
, c
i
and d
j
respectively. For
Morlet mother wavelet, gradients of E are
E
w
k
=
T

t =1
u(t )e(t )
M

i =0
c
i
h( i ), (17)
E
b
k
=
T

t =1
u(t )e(t )
M

i =0
c
i
w
k
h( i )
b
k
, (18)
E
a
k
=
T

t =1
u(t )e(t )
M

i =0
c
i
w
k
h( i )
b
k
=
E
b
k
, (19)
E
c
i
=
T

t =1
u(t )e(t )z(t i ), (20)
E
d
j
=
T

t =1
v(t )e(t ) y(t i ), (21)
where = (t b
k
)/a
k
and we have
h
a,b
(t )
b
=
1
a
_

o
sin
_

o
_
t b
a
__
exp
_
0.5
_
t b
a
_
2
_
+
_
t b
a
_
h
a,b
_
t b
a
_
_
.
123
484 Electr Eng (2008) 90:479491
The incremental changes of each parameter are simply the
negative of their gradients,
w =
E
w
, b =
E
b
, a =
E
a
,
c =
E
c
, and d =
E
d
, (22)
Thus each coefcient vector w, b, a, c and d of the network
is updated in accordance with the following rules
w(n + 1) = w(n) +
w
w, (23)
b(n + 1) = b(n) +
b
b, (24)
a(n + 1) = a(n) +
a
a, (25)
c(n + 1) = c(n) +
c
c, (26)
d(n + 1) = d(n) +
d
d, (27)
where the coefcient values are xed learning rate param-
eters.
3.2 System model and controller design
Consider a general single input single output (SISO) dynami-
cal system, similar to (11) represented by the state equations:
x

= f (x(t ), u(t ), t ) (28)


y(t ) = g(x(t ), t ) (29)
The Eqs. (28) and (29) can be written in discrete time space
as:
x(k + 1) = f (x(k), u(k), k),
(30)
y(k) = g(x(k), k),
where x(k) R
n
and u(k), y(k) R. The only accessi-
ble data are the input u and output y. if the linear system
around the equilibrium state is observable, an input-output
representation exists which has the form:
y(k + 1) = (y(k), y(k 1), . . . , y(k n + 1),
u(k), u(k 1), . . . , u(k n + 1)) (31)
i.e. a function (.) exists that maps y(k) and u(k), and their
n 1 past values, onto y(k + 1). In this light, a neural net-
work model can be trained to approximate over the inter-
est domain. Practically if an exact model of the plant were
available, approximate models would be adapted to update
the control parameters. The alternative model of an unknown
plant that can simplify the computation of the control input
is described by the following equation
y(k + 1) = (y(k), y(k 1), . . . , y(k n + 1),
u(k)u(k 1), . . . , u(k n + 1))
+(y(k), y(k 1), . . . , y(k n + 1),
u(k), u(k 1), . . . , u(k n + 1))u(k). (32)
Fig. 6 Closed loop block diagram
Because the system in (11) is rst order, we can express the
above equation as follows:
y(k + 1) = (y(k)) +(y(k)).u(k), (33)
where y(k) and u(k) denote the input and the output at the
kth instance of time.
If the nonlinearity terms (.) and (.) are known exactly,
the requiredcontrol u(k) for trackinga desiredoutput r(k+1)
can be computed at every time instance using the formula
u(k) =
r(k + 1) (y(k))
(y(k))
. (34)
However, if (.) and (.) are unknown, the idea is to use
the neural network adaptive wavelets model to approximate
the system dynamics i.e.,
y(k + 1) =

(y(k),

) +

(y(k),

)u(k). (35)
Comparing the model of (35) with the (13) we can conclude
that

(y(k),

) =
N

j =1
d
j
y(k j )v(k), (36)

(y(k),

) =
M

i =0
c
i
z(k i ). (37)
After approximation of (.) and (.) nonlinearities as

(.)
and

(.), by two distinct neural network functions with
adjustable parameters (including weights w
k
, dilations a
k
,
translations b
k
, IIRfeedforward coefcients c
k
, IIRfeedback
coefcients d
k
), represented by

and

respectively, the
control u(k) for tracking a desired output r(k + 1) can be
obtained from
u(k) =
r(k + 1)

(y(k), )

(y(k),

)
. (38)
The neuro controller for self-tuning control WECS is pro-
vided in Fig. 6.
The optimum shaft rotational speed
opt
is obtained for
each wind speed V
w
, and used as a reference for the close
loop control of WECS. Note that wind speed also acts as a
perturbation on the turbines model. Actually, the turbine is
coupled with the generators shaft using a gearbox, which
imposes an additional unknown dynamic to the model.
123
Electr Eng (2008) 90:479491 485
The characteristics of the turbine/generator pair used for
the simulations in this paper are summarized in [6], but they
are considered unknown for the controller. For this reason,
the number of wavelets was obtained on a trial-and-error
basis.
4 Simulation results
4.1 Identication of WECS
Using the wind turbine data extracted from [12], the wave-
net network with different size of Morlet mother wavelets is
employed to identify the wind turbine model. IIRblock struc-
ture with 3 feedforward delay blocks and 3 feedback blocks
is also implemented. Wavelets are local basis functions pro-
viding less interference than global ones. This leads to a non-
complex dependency in the neural network parameters [7].
We will now conrm the aforementioned idea by presenting
several observations derived from the results of the MAT-
LAB simulations. Assuming the training data are stationary
and sufciently rich, optimal performance can usually be
achieved with a small learning rate. Therefore, all learning
rate parameters for weights, dilations, translations, IIR feed-
forward coefcients, and feedback coefcients are xed at
0.005, 0.025, 0.025, 0.01, and 0.01, respectively. All initial
weights w
k
and dilations a
k
are set to 0 and 10, respectively.
The learning epoch will terminate when the desired normal-
ized error of 0.032 is reached. The following simulations
will describe the results of the wavenet network performance
employing Morlet super-mother wavelet. Figures 7, 8, 9, 10
capture the learning performance of the wavenet network
using 13 and 24 Morlet wavelets, respectively. We can con-
clude that the wavenet network composed of more wavelets
can reach initial convergence with reference to the number
of iterations very rapidly. However, to reach the desired error
goal 0.032, networks with a large number of wavelets cannot
converge easily and the error performance starts to oscillate.
Large choosing of the step size of learning rate, will cause
the iteration process to bounce between two opposite sides
of a valley rather than following the natural gradient contour
(as shown in Fig. 10). As we can see, when the number of
wavelets K is small, for example, K = 4, it takes 36 itera-
tions to reach error of 0.85 while it takes 28 iterations to reach
error of 0.35 for K = 40, but when the error of 0.032 is the
target, K = 13 takes 75 iterations while K = 40 takes 212
iterations. Large K is also undesirable because of more coef-
cients to be updated. Small K can also take a large amount
of time to compute; as for K = 4, it takes more than 1,109
iterations to reach error of 0.05. Table 1 provides numerical
values of the simulation for different number of wavelets. In
conclusion, the number of Morlet wavelets between K = 13
to K = 21 is sufcient to identify the unknown WECS model.
Fig. 7 Wavenet simulations with 13 Morlet wavelets Solid: plant out-
put and dotted: NN output
4.2 Control
After the identicationmodel is completed, the trackingoper-
ation takes command of the neuro process control to track
the desired setpoint
opt
. The co-input v(t ) is set to 0.95. In
123
486 Electr Eng (2008) 90:479491
Fig. 8 Wavenet parameter updates with 13 Morlet wavelets
the following gures, the results of the wind turbine control
using the proposed self-tuning neuro wavenet controller with
16 Morlet is compared with the results of the wind turbine
control using the MLP networks. In these gures, a sequence
of step-shaped wind gusts is applied to the system.
4.3 Controller with prior NN training
Figure 11(up) illustrates the results of the setpoint control
using the proposed self-tuning neuro wavenet controller with
16 Morlet wavelets. The same control u(k) is fed to both
the actual plant and the neural network identier. The mean
squared error between the setpoint reference,
opt
and actual
plant output, (k) is obtained as 0.022 and the one bet-
ween the setpoint reference,
opt
and the NN output res-
ponse, (k) is obtained as 0.007. Figure 11 (down) shows
the control input u(k). In input, the uctuations period is the
Fig. 9 Wavenet simulations with 24 Morlet wavelets (Solid: plant
output and Dotted: NN output)
result of the training effort in controller which would take
longer without Prior training [12]. Figure 12a provides the
network parameter updates. This demonstrates that after the
123
Electr Eng (2008) 90:479491 487
Fig. 10 Wavenet parameter updates with 24 Morlet wavelets
identication process of the wavenet network, the network
parameters become very stable around the equilibrium state.
Figure 12b shows the update of the nonlinearity terms

(.)
and

(.).
A base line comparison is demonstrated in this paper by
providing a traditional feed forward MLP neural network
structure based on the back propagation (BPP) algorithms.
50 100 150 200 250 300 350 400 450 500
350
400
450
500
550
S
p
e
e
d
(
r
p
m
)
50 100 150 200 250 300 350 400 450 500
0
0.2
0.4
0.6
0.8
Time(sec)
c
o
s
(
a
l
p
h
a
)
1
2
3
Fig. 11 Up: Self-tuning neuro wavenet controller responses to a
sequence of wind gusts: 1 setpoint reference 2 NN output response,
3 plant output response, Down: wavenet control input
Two layer perceptron networks (1 hidden layer, 1 output
layer) are simulated with 16 hidden nodes.
The bias for each node is set at 0.5. Initial weights of
both layers are set at random. Adaptive learning rate and
its momentum coefcient are initially set at 0.005 and 0.95,
respectively. Hyperbolic tangents are employed in the hid-
den nodes as an activation function. Figure 13 illustrates the
results of the set point control using the BPP algorithms. The
tracking response of the plant needs a longer time to reach the
desiredtarget. The neural networkemulator inFigure 13indi-
cates that its response could not capture the dynamic changes
on real time and thus, could never provide the good one step
predictive control performance to the tracking operation.
4.4 Controller with the input noise problem
The block diagram of an input noise model can be equiv-
alently interpreted the same as the block diagram of Fig. 6
with the measurement noise input is added to the input port
or output port of the NN. In Fig. 14 the noise is added to
the output port of the NN. The wavenet output response with
noisy input w(k) added to the input port of the network is
represented by
Table 1 Number of iterations
vs. number of Morlet wavelets
employed
Number of iterations Number of wavelets
4 10 13 21 24 33 40 53
Error of 0.85 36 8 7 5 5 4 3 2
Error of 0.35 303 22 25 18 16 25 28 33
Error of 0.05 1109 55 50 39 62 77 90 234
Error of 0.032 5003 80 75 74 150 170 212 892
123
488 Electr Eng (2008) 90:479491
Fig. 12 Self-tuning wavenet parameters tracking to set-point refer-
ence
Fig. 13 MLP controller responses to a sequence of wind gusts. 1 Set-
point reference, 2 plant output response, 3 NN output response
opt

V

opt

) ( k e
) ( k
) ( k u ) ( k
Wavenet Controller WECS
NN Aproximator
+

noise
Fig. 14 Equivalent identication to input noise immunity system
y(k + 1) =

(.) +

(.)(u(k) +w(k))
=

(.) +

(.)u(k) +

(.)w(k)
=

(.) +

(.)u(k) + w(k). (39)
The wavenet response to the noise added to the network out-
put is represented by
y(k + 1) =

(.) +

(.)u(k) +w(k). (40)
It is seen that (39) and (40) provide the same results to the
identication performance.
The measurement noise w(k) of random distribution with
a variance of 0.01 is inserted at the input port of the wavenet
network. Simulations of wavenet and MLP control methods
are shown in Figs. 15 and 17. All initial values of parameters
are assumed the same as previous section. The simulation of
the plant and NN output responses with application of neuro
wavenet controller, is shown in Fig. 15(up). Figure 15(down)
shows the corresponding control effort to the actual plant and
the noisy control to the wavenet network simulator. Figure 16
provides the necessary parameter updates of the network and
the control. It can be seen that when noises are mixed in the
control process, the tracking operation of the actual plant to
123
Electr Eng (2008) 90:479491 489
50 100 150 200 250 300 350 400 450 500
35
400
450
500
550
S
p
e
e
d
(
r
p
m
)
50 100 150 200 250 300 350 400 450 500
0.1
0.2
0.3
0.4
0.5
0.6
Time(sec)
c
o
s
(
a
l
p
h
a
)
1
2
3
1
2
Fig. 15 Up: Self-tuning neuro wavenet controller responses to a
sequence of wind gusts (1 setpoint reference, 2 NN output response, 3
plant output response), down: wavenet control input (1 control Action
to NN Emulator, 2 control action to actual plant)
the desired reference is inferior with the noise level. Note
that the measurement noise is only added to the input to the
NN emulator, but not to the actual plant. However, the plant
responses to the control action are noisy. As in Fig. 15 is
shown, the wavenet network setpoint tracking response is
better than the plant response. It can be concluded that with
this type of control scheme, noises are the main distraction
for the actual plant system of the tracking operation.
With the same initialization, the simulation result of the
input noise problem study for the two-layer feed forward
NN with back propagation algorithms is shown in Fig. 17.
The neural network output hardly responds to the tracking
operation, and the plant output is noisy. Unlike wavenets, the
back propagation algorithms do not signify the interpolation
characteristic, thus noise interferes the NN responses.
4.5 Controller with the output noise problem
The block diagram of an output noise model can be equiv-
alently interpreted the same as the block diagram of Fig. 6
with the disturbance noise output is added to the input port
or output port of the of the plant. In Fig. 18 the plant with
the input noise is shown. In (41) and (42) it is shown that we
can use the input noise to the plant rather than modeling the
noise in the plant output. When the model that noise added
to output of the plant is used, the error becomes
e(k) = (y(k) +w(k)) y(k),
(41)
e(k) = [(.)

(.)]+[(.)u(k)

(.)u(k)]+w(k).
And the error of the model with input added noise to the
plant is
Fig. 16 Self-tuning wavenet parameters tracking to set-point refer-
ence
e(k)=[(.)

(.)]+[(.)u(k)

(.)u(k)]+(.)w(k),
e(k)=[(.)

(.)]+[(.)u(k)

(.)u(k)]+ w(k).
(42)
123
490 Electr Eng (2008) 90:479491
Fig. 17 MLP controller responses to a sequence of wind gusts. 1 Set-
point reference, 2 plant output response, 3 NN output response
Wavenet Controller
opt
WECS
NN Aproximator

V opt

) (k e
(k)
) (k u (k)
noise
-
+
Fig. 18 Equivalent identication to output noise immunity system
50 100 150 200 250 300 350 400 450 500
-500
0
500
1000
1500
2000
S
p
e
e
d
(
r
p
m
)
50 100 150 200 250 300 350 400 450 500
-0.5
0
0.5
1
1.5
2
Time(sec)
c
o
s
(
a
l
p
h
a
)
1
2
3
Fig. 19 Up: Self-tuning neuro wavenet controller responses to a
sequence of wind gusts: 1 setpoint reference, 2 NN output response,
3 plant output response, Down: wavenet control input
Thus, the two identication models provide similar algorith-
mic approximation.
After the identication process for the unknown WECS
system has been completed, the control action is activated to
track the same desired set-point reference. The disturbance
Fig. 20 Self-tuning wavenet parameters tracking to set-point refer-
ence
123
Electr Eng (2008) 90:479491 491
50 100 150 200 250 300 350 400 450 500
200
300
400
500
600
S
p
e
e
d
(
r
p
m
)
50 100 150 200 250 300 350 400 450 500
0.1
0.2
0.3
0.4
0.5
0.6
Time(sec)
c
o
s
(
a
l
p
h
a
)
1
2
3
Fig. 21 MLP controller responses to a sequence of wind gusts. 1 Set-
point reference, 2 plant output response, 3 NN output response, Down:
wavenet control input
noise w(k), with a variance of 0.01 is inserted at the input port
of the nonlinear plant. Simulations of two control methods
(wavenet and MLP) are shown in Figs. 19 and 21. The simu-
lation result of the plant and NNoutput responses with neuro
wavenet controller is shown in Fig. 19(up). Figure 19(down)
demonstrates the corresponding control action to both the
actual plant and the wavenet network identier. Figure 20
provides the necessaryparameter updates of the network. The
results showthat there is an offset between the plant response
and the set-point due to the disturbance at the output port of
the plant.
With this control scheme, the NN performance is noisy,
with fast response to the set point, while the plant output
performance is affected by the disturbance of the noise.
Figure 21(up) shows the setpoint tracking response of the
output noise using the same two-layer perceptron network
with BPP algorithms. The control effort to the plant is illus-
trated in Fig. 21(down). With this conventional algorithm,
the plant response is noisy with a small offset and the NN
response never adjusts to the changes. Finally, the output
noise immunity studies show the worst scenario in the adap-
tive self-tuning control.
5 Conclusion
This paper emphasizes on self-tuning control applications
of an efcient neural network architecture based on a wave-
let theory called wavenets. The wavenet based controllers
improve the performance of the trained network for fast
convergence, robustness to noise interference, and high com-
plex ability to learn and track of WECSs. Two control sche-
mes were shown. One is based on an assumed plant model
and a wavenet -identier that is used to construct adaptive
controllers. The other is a traditional neuro-control scheme
based on the feedforward neural network structure with back-
propagation (BPP) algorithms. This scheme is used for a base
line comparison to wavenets. The rst scheme tends to pro-
vide faster tracking adjustment to control changes. Finally,
the conventional scheme with BPP shows that it requires a
longer time for adapting to changes and performs poorly to
added noises. From the results studies, the worst scenario to
all of the control schemes, in terms of MSE, occurs when
noise is contaminated to the output port of the plant systems.
References
1. Junhua Y, Jie W, Jinming Y, Ping Y (2004) Apply intelligent con-
trol strategy in wind energy conversion system. In: Fifth world
congress on intelligent control and automation, WCICA2004. vol
6, pp 51205124
2. Chedid R, Mrad F, Basman M (1999) Inteligent control of a class
of wind energy conversion system. IEEE T-EC 14(4):15971604
3. Kanellos FD, Hatziargyriou ND (2002) A new control scheme for
variable speed wind turbine using neural networks. IEEE Power
Eng Soc Winter Meeting 1(1):360365
4. Kyoungsoo Ro, Han-ho Choi (2005) Application of neural network
controller for maximumpower extraction of a grid-connected wind
turbine system. Electr Eng (Archiv Elektrotech) 88(1):4553
5. Narendra KS, Parthasarathy K (1990) Identication and control of
dynmical systems using neural networks. IEEE Trans Neural Netw
1(1):427
6. Mayosky MA, Cancelo GIE (1999) Direct adaptive control of wind
energy conversion systems using gaussian networks. IEEE Trans
Neural Netw 10(4):898906
7. Lekutai G, VanLandinghamHF (1997) Self-tuning control of non-
linear systems using neural network adaptive frame wavelets. IEEE
Int Conf Systems Man Cybern 2:10171022
8. Sedighizadeh M, Kalantar M (2004) Adaptive PID control of wind
energy conversion systems using RASP1 mother wavelet basis
function networks. IEEE TENCON 2004, Chiang Mai
9. Sedighizadeh M, et al. (2005) Nonlinear model identication and
control of wind turbine using wavenets. In: Proceedings of the 2005
IEEE conference on control applications, Toronto, pp 10571062
10. Kalantar M, Sedighizadeh M (2004) Adaptive self tuning control
of wind energy conversion systems using Morlet mother wavelet
basis functions networks. In: 12th Mediterranean IEEE conference
on control and automation MED04, Kusadasi
11. Szu HH, Telfer BA, Kadambe S (1992) Neural network adap-
tive wavelets for signal representation and classication. Opt Eng
31(9):19071916
12. Sedighizadeh M (2005) Modeling and adaptive-neural control of
windenergyconversionsystems usingwavelet functions. PhDThe-
sis, Iran University of Science and Technology
123

Anda mungkin juga menyukai