Sunteți pe pagina 1din 9

ISA Transactions 63 (2016) 233241

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research Article

Adaptive robust control of a class of non-afne variable-speed


variable-pitch wind turbines with unmodeled dynamics
Pedram Bagheri n, Qiao Sun
Department of Mechanical and Manufacturing Engineering, University of Calgary, 2500 University Dr. NW, Calgary, AB, Canada, T2N 1N4

art ic l e i nf o a b s t r a c t

Article history: In this paper, a novel synthesis of Nussbaum-type functions, and an adaptive radial-basis function neural
Received 11 January 2016 network is proposed to design controllers for variable-speed, variable-pitch wind turbines. Dynamic
Received in revised form equations of the wind turbine are highly nonlinear, uncertain, and affected by unknown disturbance
20 March 2016
sources. Furthermore, the dynamic equations are non-afne with respect to the pitch angle, which is a
Accepted 8 April 2016
Available online 2 May 2016
control input. To address these problems, a Nussbaum-type function, along with a dynamic control law
This paper was recommended for publica- are adopted to resolve the non-afne nature of the equations. Moreover, an adaptive radial-basis function
tion by Dr. Jeff Pieper. neural network is designed to approximate non-parametric uncertainties. Further, the closed-loop sys-
tem is made robust to unknown disturbance sources, where no prior knowledge of disturbance bound is
Keywords: assumed in advance. Finally, the Lyapunov stability analysis is conducted to show the stability of the
Wind turbines
entire closed-loop system. In order to verify analytical results, a simulation is presented and the results
Adaptive control
are compared to both a PI and an existing adaptive controllers.
Robust control
Adaptive neural network & 2016 ISA. Published by Elsevier Ltd. All rights reserved.
Nussbaum-type functions

1. Introduction control perspective, xed speed, variable-speed, and variable-


pitch turbines are of the most importance. Variable-speed wind
Due to limitations and pollution concerns of fossil fuels, the turbines offer many benets as opposed to xed-speed ones. It has
search for alternative ways of energy production has attracted been shown that their annual energy production can reach up to
much attention recently. Among a few available alternatives, the 10% above that of xed-speed wind turbines [2]. It has also been
wind energy is known to be virtually one of the most tantalizing recognized that controllers in such turbines play the important
energy resources. Wind turbines are mechanical machines, role of the attenuation of the unfavorable effects of loads on the
exploited to capture the kinetic energy of the wind. The captured turbine structure [3,4]. However, owing to the simplicity of
energy can be later converted to useful electrical or mechanical schemes, the majority of wind turbine control systems have been
energy. Regarding the advantages of the wind energy, cleanness designed based on linear models [5,6]. Considering the nonlinear
and global accessibility are the most prominent features. However, and uncertain nature of the wind turbine dynamics, designing
variations in wind intensity, cause wind turbines to function with nonlinear controllers, developed to accommodate the complexity
low efciency for a long time-span throughout the year. To over- of the turbine dynamics can improve efciency dramatically.
come this challenge, well-designed control systems with a cap- The presence of control systems enables wind turbines to
ability of adjusting and controlling wind turbines components can function in a wide range of wind speeds, which eventuates in
be successfully utilized. Such controllers can increase the power higher power capture. The control operating region in wind tur-
capture, as well as the turbine efciency. Alongside an increase in bines can be divided into two main parts, which are: 1. Region
the energy capture, such controllers are employed to reduce loads with wind speeds below the rated speed; the control objective in
on the turbine structure as well. Consequently, higher- this region is to enable the turbine to reach desired rotor speeds,
performance wind turbines with longer longevity are achieved [1]. resulting in the maximization of the power capture. 2. Region with
The objective of the control design for wind turbines is to wind speeds above the rated speed; the control goal in this region
optimize the captured wind energy. For this purpose, several is to decrease electrical power and structural loads on different
designs for wind turbines have been introduced, where each one parts of the turbine. It is worth noting that the turbine does not
of them has its own advantages and disadvantages. From the operate at the highest efciency in the second region [7]. To
achieve these control objectives, there are two control inputs
n
Corresponding author. available: generator torque and blades pitch angle. One way to
E-mail address: pbagheri@ucalgary.ca (P. Bagheri). manipulate these inputs is to employ the generator torque as the

http://dx.doi.org/10.1016/j.isatra.2016.04.008
0019-0578/& 2016 ISA. Published by Elsevier Ltd. All rights reserved.
234 P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241

input for low wind speeds (i.e., the rst region). In this case, the capture. As previous works in the literature suggest, there exits an
pitch angle is maintained at a constant value. On the other hand, optimal rotor speed in the torque control phase, which results in
for high wind speeds, in the second region, the pitch angle is the maximum power extraction [7]. The optimal rotor speed is
adopted as the control input, while the generator torque is kept time-varying and proportional to the wind speed. Thus, the
constant as it reaches the vicinity of its rated value [8,9]. achievement of the asymptotic stability in the torque control
Various linear control methods have been developed to address phase is of prime importance. However, the establishment of such
the aforementioned control objectives in the literature. Examples stability for the nonlinear dynamics of wind turbines in the pre-
of such methods are, LQ and LGQ controllers [10,11], PI and PID sence of uncertainties and disturbance sources is not achieved in
controllers [12], and liner robust methods [13]. Although linear existing methods. Contrary to existing methods in the literature
controllers have been widely utilized for wind turbines, highly [19], the proposed scheme can guarantee the asymptotic stability
nonlinear nature of wind turbines compromises the performance of the closed-loop system in the torque control phase. It is worth
of such controllers [14]. mentioning that such stability is obtained in the presence of
As mentioned earlier, the existence of highly nonlinear terms, parametric and non-parametric uncertainties, and unknown dis-
uncertainties, and unknown disturbance sources in the dynamic turbance sources. In addition, in contrast to the existing methods,
equations of wind turbines affect the performance of linear con- the direction of the non-afne input (i.e., pitch angle), and the sign
trollers adversely. Hence, nonlinear approaches suit the nonlinear of its derivatives are assumed to be unknown. Thus, no prior
dynamics of wind turbines, which contribute to better perfor- knowledge of the non-afne function is presumed in the extrac-
mance. Several research works have been conducted in attempts tion of control laws for the non-afne input. Consequently, the
to design nonlinear controllers for wind turbines. A combination proposed controller enables the turbine to reach desired rotor
of a sliding mode control and a PI controller is adopted in [15]. speeds, which leads to the maximum power capture. On the
Authors in [15] propose a sliding mode design for the torque
contrary, the existing nonlinear methods are incapable of estab-
control, and a PI controller for the pitch angle. The proposed
lishing the asymptotic stability in the presence of unknown dis-
controller accommodates nonlinearities in the torque control
turbance and uncertainties, which compromises power capture
phase. However, in order to avoid the complexity of the non-afne
efciency. In the proposed scheme, a Nussbaum-type function is
nature of the pitch angle, a PI control is designed for the pitch
utilized to address the non-afne nature of the dynamic equations.
angle. A nonlinear model predictive controller is designed in [16].
Moreover, an adaptive neural network is designed to approximate
It is assumed that the pitch angle is xed, hence, the controller is
uncertain and unmodeled dynamics. Furthermore, the entire
applied to the torque control phase. In addition, uncertainties and
closed-loop system is made robust to unknown disturbance
disturbance sources are not considered in the design. Neural net-
sources, without any knowledge of disturbance bound. At the end,
works are proven to be effective ways of coping with non-
the stability of the closed-loop system is demonstrated by the use
linearities and uncertainties [17]. One of the benets of using
of the Lyapunov stability analysis.
neural networks is their adaptation capabilities. Consequently,
they can approximate parametric and non-parametric uncertain-
ties real-time, without any previous knowledge of those uncer-
tainties [18]. An example of such nonlinearities in the dynamics of 2. Problem statement
the turbine is corresponding to the pitch angle, which is also a
control input. The dynamic equations is non-afne with respect to The objective of this paper is to develop a nonlinear controller
the pitch angle, which creates an uncertain non-afne system. To for variable-speed, variable-pitch wind turbines in the presence of
cope with this challenge, there exist a few studies that suggest the unknown disturbance sources and uncertainties. The proposed
use of neural network-based approximators. Authors in [19] controller accounts for turbines non-afne and uncertain
developed adaptive controllers for variable-speed variable-pitch dynamics, in order to maximize the power capture.
wind turbines, where neural networks are used to approximate The wind turbine, considered in this study has two manipu-
uncertainties and disturbance. Nevertheless, the asymptotic sta- lated inputs: 1. Generator torque. 2. Pitch angle. To exploit these
bility of the closed-loop system is not achieved, and restrictive control inputs, the following two-phase control strategy is
assumptions are made to tackle the non-afne pitch angle. A adopted:
control scheme to design RBF-PID controllers for wind turbines is
proposed in [20]. However, the method does not guarantee the 1. Torque Control: For wind speeds below the rated speed, the
asymptotic stability in the torque control phase, which adversely generator torque is used as the manipulated input. The goal in
affects the power capture efciency. The existing nonlinear and this phase is to achieve desired rotor speeds. Depending on the
neural network-based methods in the literature on addressing wind speed, there exists an optimal rotor speed, which leads to
both non-afness and uncertainties of the turbine equations, the maximum power capture. Therefore, the control objective is
introduce two major problems: 1. Such methods can merely to produce the right amount of torque, such that the rotor
establish stability in the sense of Lyapunov, as opposed to the rotates at the optimal velocity. Moreover, in the torque control
asymptotic stability. It means that the turbine does not operate at phase, the pitch angle is maintained at a constant value. The
desired speeds, required to achieve the maximum power capture optimal rotor velocity can be calculated as follows [7]:
efciency. 2. In order to establish stability, these methods require
opt v
prior knowledge of the directions of the non-afne input and its _ d 1
Rb
derivatives. Given the complex nature of the turbine equations,
such information is not available in practical applications. Thus, where _ d is the desired rotor speed; v is the wind speed; opt is
the objective of this study is to propose a method to eliminate the optimal tip-speed relation which depends on the blade
such restrictive assumptions, with the goal of the achievement of design and aerodynamics; Rb stands for the blades length.
the maximum efciency. 2. Pitch Control: As the wind speed approaches its rated value, the
In this paper, nonlinear control of variable-speed variable-pitch generator torque and rotor speed reach their rated values as
wind turbines is considered. The contribution of the paper is to well. Thus, the generator torque can be no longer used as a
develop a novel synthesis of Nussbaum-type functions and adap- manipulated input. Hence, the strategy in this phase is to keep
tive radial-basis function neural networks to maximize the power the generator torque at its rated value, and employ the pitch
P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241 235

angle as a manipulated input. Thus, the pitch angle is used to established as follows:
hold the rotor speed at its rated value.
f ; _ ; v RT ; _ ; vQ h 9

where R; _ ; v is the vector containing radial-basis functions of


x  cij
3. Wind turbine modeling
the form r ij x e ij , and x; _ ; v ; cij is the center of the


receptive eld; ij is the width of the basis functions; h is the


In this section, the nonlinear equations governing the dynamics
approximation error, where hr , and is assumed to be
of the rotor of a variable-speed, variable-pitch wind turbine are
unknown. Q is also the optimal weight parameters vector, which
presented. For control purposes, the rotor is the main dynamical
minimizes :
component, whose dynamics is written as follows:
Substituting eqs. (9) into (2), the dynamic equation of the rotor
J r J g C s _ f ; _ ; v T g  T a dt 2 is obtained as follows:

where J r and J g are the equivalent moments of inertia of the rotor J r J g C s _ RT Q h T a T g dt 10


and generator, respectively. A denotes the angular displace-
ment of the rotor, and _ A is the rotor speed. T g stands for the eq. (10) is used in the next section for control design. Note that
torque, produced by the generator;T a represents the aerodynamic Cs, Q, h, and d(t) are unknown.
torque imposed on the blades. C s is the damping coefcient of the
shaft, which is not available. f ; _ ; v is an unknown continuous
nonlinear function which accounts for the unmodeled dynamics, 4. Control design
affecting the rotor. d(t) represents an unknown bounded dis-
turbance source. That is, dtr where is not available, and In this section, a two-phase control strategy is adopted to
3 denotes the Euclidean norm of a vector. design a controller for a variable-speed, variable-pitch wind tur-
The aerodynamic torque, acting on the blades can be calculated bine, in the presence of disturbance and uncertainties. A combi-
as follows [14]: nation of an adaptive neural network, adaptive controller, and
Nussbaum-type function is proposed to develop a controller for
1
Ta Rb 3 C q ; v2 3 the considered wind turbine. The adaptive neural network and
2
adaptive controller address unmodeled dynamics, and parametric
where C q is the torque coefcient. is the air density; is the uncertainties, respectively. In addition, a Nussbaum-type function
blades pitch angle, Rb stands for the blades length, and denotes is employed to cope with the non-afne nature of the pitch angle.
the tip-speed ratio, calculated by For this purpose, the proposed control scheme is divided into two
_ Rb phases, which are: 1. Torque control. 2. Pitch control. The gen-
4 erator torque and pitch angle are exploited as the inputs in the
v
rst and second phases of control, respectively. The proposed
The captured power is also achieved from control scheme is explained in detail in the subsequent sections.
1
P a Rb 2 C p ; v3 5
2 4.1. Torque control
where Cp is the power coefcient, dened as
In this control phase, the generator torque is adopted as an
C p ;
C q ; 6 input to the system. Simultaneously, the pitch angle is kept at a
constant value. It is worth noting that, in this phase, wind and
The power coefcient depends on both the tip-speed ratio and rotor speeds are below their rated values. The control objective is
pitch angle, described by the following equation [21]: to achieve the desired rotor speed which leads to the achievement
  of the maximum power capture. In order to proceed with control
c2  c5
C p c1   c3  c4 e  c6 7 design, the following variables are dened.

Tracking error is dened as e _  _ d , and _ d represents the
where desired rotor speed given by eq. (1). C~ is the estimation error of the
1 1 0:035 damping coefcient, dened as C~ C^  C s :; Q~ stands for the esti-
 8
 0:08 3 1 mation error of the neural network weight parameters, expressed
as Q~ Q^  Q ; ~ ^  represents the estimation error of the
and c1 0:5176; c2 116; c3 0:4; c4 c5 5; c6 0:0068: As disturbance bound, ~ ^  exhibits the estimation error of
seen from eqs. (7) and (8), the function containing the pitch angle, which is the error, stemming from the neural network approx-
; is non-afne. Therefore, an appropriate strategy yet to be imation. p1; p2 and p3 are positive constants; a J r J g ; which is a
adopted, in order to utilize the pitch angle as a manipulated input. measure of the moments of inertia, and its estimation error is
dened as a~ a^  a: Moreover ^: stands for the estimation of a
3.1. Neural network design parameter.

As stated earlier, nonlinear function f in eq. (2) represents Theorem 1. For a system whose dynamics is governed by eq. (10),
unmodeled dynamics. Hence, a radial-basis function neural net- the following control and adaptive laws can establish the asymptotic
work is designed to approximate f. Based on the universal stability, and eliminate the tracking error asymptotically (i.e.,
approximation theorem, continuous functions on a compact sub- eA L2 \ L1 ). Furthermore, such stability is achieved in the presence of
set DC can be approximated by a nite number of neurons [22]. uncertainties and unknown disturbance sources. In addition, all
That is, for a function f x A L2 DC ; an approximator of the form closed-loop signals remain uniformly bounded.
f^ RT xQ can approximate f with a bounded approximation

R The proposed control laws are expressed as follows:
error. Moreover, Q arg min Dc f x  f^ x : Q dx : Given the
explanation above, the structure of the proposed neural network is u1  1 e a^ d RT Q^ C^ _ T a 11
236 P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241

^ 2 e Using the adaptive laws given by eqs. (16)(18), (21) becomes


u2  12
^ e 1 t V_ r  e2 eu2 eu3 e^ e^ 23
2 Since the control law was split into three parts, the second and
^ e
u3  13 third parts of the control law, dened in eqs. (12) and (13) are
^e t
2 introduced into eq. (23) as follows:
where T g u1 u2 u3 ; i i 1; 2 is an arbitrary function which 2
belongs to L1, and 1 is a positive constant. The proposed adaptive ^ 2 e ^ e
V_ r  e2  e e^  e e^
laws are dened as follows: ^ e 1 t ^e t
2
2
_
C~  e_
^ 2 e2 ^ e2
14 r  e2  e^  e^
^ e 1 t ^ e 2 t
_
Q^ R; _ ; ve
2 2
15 ^ 2 e2 e2 ^ 2 1 te^ ^ e2 e2 ^ 2 te^
r  e2  
^ e 1 t ^ e 2 t
_^ p2 e 16
1 te^ 2 te^
_ r  e2 24
^ p1 e 17 ^ e 1 t ^ e 2 t
i t is arbitrary, thus, it can be chosen to take on merely non-
a_^ p3 e d 18 negative values. Note that for b d Z 0; bbd r b; Therefore
d
where eqs. (14) and (15) update the estimations of the parametric V_ r  e2 1 t 2 t 25
uncertainties and neural network weight parameters, respectively.
Eqs. (16) and (17) estimate the bounds of the neural network error Note that eqs. (16) and (17) imply that non-negative initial
and disturbance, respectively, and eq. (18) estimates the moments conditions for ^ and ^ always lead to nonnegative values for ^ and
of inertia. ^ . Nevertheless, ~ ; ~ and a~ are the arguments of the Lyapunov
function. Therefore, the stability results achieved from the Lya-
Proof. The following Lyapunov function is proposed to conduct punov analysis guarantee their boundedness.
the stability analysis Taking integral from both side of eq. (25) leads to
1 1 2 1 T 1 ~2 1 1 Z 1 Z 1 Z 1
V ae2 C~ Q~ Q~ ~ 2 a~ 2 19 V1  V0 r  e2 dt 1 tdt 2 tdt 26
2 2 2 2p1 2p2 2p3
0 0 0
Taking time derivative of eq. (19), and using eq. (10) lead to eq. (26) can be written as follows:
  Z Z 1 Z
V_ e  C s _ T g  T a  RT Q  h dt  a d 1 1
e2 dt r V0  V1 1 tdt 2 tdt 27
_ T _ 1 1 _ 1 0 0 0
C~ C~ Q~ Q~ ~ ~_ ~ ~ a~ a~_
p2 p1 p3 i t is arbitrary and can be chosen to be non-negative
R1
at all
 
_ ^ d times. Further, the fact that i t A L1 implies that 0 i tdt o 1.
e  C^  C~ T g  T a R Q^  Q~  h dt a~  a
T
Moreover, by denition, V(0) R 1is positive, and the denition of the
_ _ T _ 1 _ 1 1
C~ C~ K~ K~ Q~ Q~ ~ ~ ~ ~_ a~ a~_ 20 vector norm implies that 0 e2 dt is non-negative. Therefore, it is
p1 p2 p3 concluded that V0  V1 is positive (i.e., V 1 r V0 ). Hence,
since V(0) is bounded, it is learnt that V1 is bounded as well.
In order to achieve the control objectives, the controller is
Consequently, all the termsR1 on the right side of eq. (27) are
required to fulll tracking and robustness purposes. In addition,
bounded. As a result, 0 e2 dt is also bounded (i.e., e A L2 ). Fur-
the controller needs to account for the error of the neural network
thermore, it is straightforward to check that e_ A L1 . Thus, it is
approximation as well. Hence, the control input is divided into
concluded from Barbalats lemma that lim et 0; which implies
three parts as T g u1 u2 u3 : Substituent for T g leads to t-1
  the asymptotic stability of the closed-loop system. Therefore, it
_ T _
V_ e  C^ _ u1 u2 u3  RT Q^  T a  a~ C~ C~ Q~ Q~
d was demonstrated that the use of the control laws given by eqs.
1 1 _ 1 (11)(13), and adaptive laws in eqs. (14)(18) eventuates in the
~ ~_ ~ ~ a~ a~_ eC~ _ Q~ Re
T
p2 p1 p3 asymptotic stability of the closed-loop system. Consequently, the
turbine can operate precisely at the desired rotor speed, max-
 eh edt ea~ d 21 imizing the power capture. As noted earlier, it is shown that there
where, as dened earlier, T g u1 u2 u3 . Using the control law exist rotor speeds proportional to the wind speeds, which even-
^ Q^ given by eqs.
for u1 dened in eq. (11) and adaptive laws for C; tuate in the maximum power extraction [7]. The relationship
(14) and (15), (21) becomes between the optimal rotor speed and wind speed is given by eq.
(1). The achievement of the asymptotic stability leads to the
1 1 _ 1 maintaining of the optimal rotor speed, which results in the
V_  e2 eu2 eu3  eh ~ ~_ ~ ~ edt ea~ d a~ a~_
p2 p1 p3 maximum efciency. It is worth noting that despite unknown
1 1 _ disturbance, parametric and non-parametric uncertainties, the
r  e2 eu2 eu3 eh ~ ~_ edt ~ ~ ea~ d
p2 p1 proposed controller can establish robustness and the asymptotic
1 1   stability of the closed-loop system.
a~ a~_ r  e2 eu2 eu3 e^  ~ ~ ~_ e ^  ~
p3 p2
1 ~ ~_ 1 _ 4.2. Pitch control

ea~ d a~ a~ r  e eu2 eu3  e~  e~ e^
2
p1 p3
In the second phase of the control strategy, the pitch angle is
1 _ 1 ~ ~_ 1
e^ ~ ~ ea~ d a~ a~_ 22 employed as an input. In addition, in this phase, the generator
p2 p1 p3 torque is maintained at its rated value. As it follows from eqs.
P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241 237

(6) to (8), the dynamic equation is non-afne with respect to the Rk


can be established to show lim inf 1k 0 Ns wds  1: Thus,
k-1
pitch angle. To address the non-afne nature of the pitch angle,
N(s) w is a Nussbaum-type function as well.
utilizing the mean value theorem [23], the following can be
written: Lemma. [25] For a smooth function s(t) on interval [0, t], and a
smooth Nussbaum-type function N(s), if a positive denite function V
T a 
T a T a  28 (t) on the same interval satises the following inequality:
 Z t Z t
Vt r c0 e  c1 t GxNss_ ec1 d e  c1 t s_ ec1 d 32
0 0
where the mean value theorem is adopted to expand function T a
with respect to the pitch angle. Note that although eq. (28) where G(x) is a non-zero time-varying function, then V(t), S(t) and
Rt
0 Gx Nssd remain bounded on the same interval.
resembles the linearized presentation of a function, it is indeed an _
exact presentation of T a . The difference between the linearization
of T a at ; and the use of the mean value theorem is that the Theorem 2. Given an uncertain system subjected to unknown dis-
  turbance and unknown control direction, whose dynamics is repre-
point is not available. In addition, has a time-varying nature
sented by eq. (29), there exist adaptive Nussbaum function-based
as well, which adds to the complexity of the problem. The time-
 control and estimation laws, which establish the stability of the
varying nature of stems h fromithe fact that for each value of ;
 closed-loop system. Furthermore, the proposed control and estima-
the length of interval changes. Therefore, point tion laws guarantee that all signals remain UUB.

depends on at any time,  which makes time-varying, and
a
The proposed adaptive control and estimations laws are as
unavailable. As a result, T
d  
becomes an unknown nonlinear
follows
function. As seen from eq. (28), using the mean value theorem, the  
non-afne nature of the pitch angle is resolved.  1 Ns  C^ _ 2 e T g  T a 0  RT Q^  u1 u2 33
 However, the
a
existence of the unknown nonlinear function T d  
makes the

input direction unknown, which is yet to be addressed. Note that s_  eC^ _ 2 e2 eT g  eT a 0  eRT Q^ eu1  eu2 34
T a 
d  
being zero implies that the pitch angle cannot be used as
where eq. (33) denotes the dynamic control law for the pitch
an input, thus, it is assumed that T
d
a
is non-zero. angle; N(s) is a Nussbaum-type function; s is the argument of the
Using eqs. (28) in (10), the dynamic equation of the rotor is Nussbaum-type function, whose dynamics is given by eq. (34). 1
obtained as follows: and 2 are two positive constants. As seen from eqs. (33) and (34),
the use of Nussbaum-type functions in the control law introduces
a C s _ RT Q h T a T g  G dt

29 a new state into the equationsi:e:; s: Therefore, the designer
needs to choose appropriate dynamics for the new state, in order
where a J r J g ; T a denotes the aerodynamic torque at ; and to establish the stability of the closed-loop system. The choice of

a
 suitable dynamics for the new state depends on the structure of
G T
d  
is an unknown nonlinear function. It is worth
the system, and presence of other phenomena such as uncer-
noting that can be chosen based on the initial condition of the tainties and disturbance. Thus, it varies from system to system.
pitch angle, used in the torque control phase. As mentioned earlier, Moreover, u1 and u2 are two introduced control variables, pro-

G creates a control problem with an unknown input direction. posed to account for uncertainties and unknown disturbance,
The problem of unknown input direction can be addressed which are dened as
with the use of the properties of Nussbaum-type functions. A
function N(s) is Nussbaum-type if it satises the following condi- ^ 2 e
u1  35
tions [24]:
^ e 1 t

Z 2
1 k ^ e
lim sup Nsds 1 30 u2  36
k-1 k 0
^e t
2

Z The proposed adaptive laws are also presented as follows:


k
1
lim inf Nsds  1 31  
_
k-1 k 0 C^ p6  e_  C^ 37
2  2 
An example of such functions is es cos s ;which is also  
_
used in this paper. Q^ p7  eR  Q^ 38

Property. If N(s) is a Nussbaum-type function, N(s) w is also a  


Nussbaum-type function, where w A  ; and is a positive _^ p8 e  ^ 39


constant.
_  
 R  ^ p9 e  ^ 40
Rk 1 k
Rk
Proof. lim sup1k 0 Ns wds lim sup k 0
Nsds 1k 0 wds :
k-1 k-1
where pi i 6; :::9 are positive constants. Eqs. (37) and (38)
Since N(s) is a Nussbaum-type function and w is bounded, it fol-
 R Rk  estimate the damping coefcient and neural network weight
k
lows directly that lim sup 1k 0 Nsds 1k 0 wds 1: The same parameters, respectively. Eqs. (39) and (40) update the estimations
k-1
of the bounds of the neural network error and disturbance,
respectively.
238 P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241

2
Proof. To analyze the stability of the closed-loop system using the ^ 2 e ^ e
aforementioned adaptive and control laws, the following Lyapu- e e^  e e^
^ e 1 t ^ e t
2
nov function is proposed
r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^  ~ ^  ~ ^
 T
1 1 ~2 1 ~T ~ 1 ~2 1
V ae2 C Q Q ~ 2 41
 ^ e2 e2 ^ e
2 2
2 2p6 2p7 2p8 2p9 ^ 1 t

Taking time derivative of eq. (41) gives ^ e 1 t
  2 2
 ^ e2 e2 ^ 2 te^
V e  C s _ T g  T a 0  G a  RT Q  h  a d dt
_ n

^ e 2 t
1 _ 1 T 1 1 _
C~ C~ Q~ Q~ ~ ~_ ~ ~ r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^  ~ ^  ~ ^ 1 t 2 t
 T
p6 p7 p8 p9
 
46
e C^  C~ _ T g  RT Q^  Q~  h  a d  T a 0  G a
n

The completion of square suggests that for two real numbers m


1 ~ ~_ 1 ~ T ~ 1 1 _
C C Q Q ~ ~_ edt ~ ~ and n, m  n2 m2 n2  2mn: Exploiting this relation between
p6 p7 p8 p9
    the estimate of Cs and its estimation error, C~ ; leads to
1 ~_
e C^_ T g  RT Q^  d  T a 0  G C~ C e_
n
2 1 1 2 1 2
C~  C^ 2 C~ C^  2C~ C^  C~ C^ C~  C^ 2  C~  C^
2
p6
  2 2 2
~ T 1 ~_ 1 _ 1 ~ ~_
Q Q eR ~ ~ edt  eh 42 1 1 2 1 2
p7 p8 p9 C s 2  C~  C^ 47
2 2 2
Introducing eqs. (33) and (34) into eq. (42) and considering that The same equation can be extracted for ; Q and as follows:
d 0 in the pitch control phase, eq. (42) is expressed as  ~ ^ 12  12~  12^ ;  Q~ Q^ 12Q T Q  12Q~ Q~  12Q^ Q^ , and  ~ ^
2 2 T T T
 
V_ e  C^_ T g  RT Q^ T a 0 2  2  2 : Introducing these equations into eq. (46) gives
1 2 1~2 1^2

V_ r  2 e2 1 G Ns 1s_ 1 t 2 t

1 G Nss_ eC^ _  2 e2  eT g eT a 0 eRT Q^ s_
n

    1 2 1 2 1 2 1 T 1 T 1 T 1 1 2 1 2
1 ~_ 1 ~_ 1 1 _  ~  ^ Q Q  Q~ Q~  Q^ Q^ C s 2  C~  C^
eu1 eu2 C~ C e_ Q~ Q eR ~ ~_ ~ ~
T
2 2 2 2 2 2 2 2 2
p6 p7 p8 p9
  1 2 1 2 1 2
edt  eh  2 e2 1 G Ns 1s_ C~
n 1 ~_
C e_  ~  ^ 48
p6 2 2 2
 
1 _ 1 1 _ Considering the structure of the Lyapunov function given by
Q~ Q~ eR eu1 eu2 ~ ~_  eh edt ~ ~
T
43 eqs. (33), (48) can be rewritten as follows:
p7 p8 p9
V_ r  c1 V 1 G Ns 1s_ 1 t 2 t

To reach the desired form of V, _ the proposed adaptive laws
1 1 1 1 2 1 T 1 2 1 2 1 2
given by eqs. (37) and (38) are injected into eq. (43) as follows: 2 Q T Q C s 2  Q^ Q^  C^  ^  ^
2 2 2 2 2 2 2 2
1 
V_  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^ eu1 eu2 ~ ~_ r  c1 V 1 G Ns 1s_ 1 t 2 t
 T
p8
1 T 1 2 1 2 1 2
1 _  Q^ Q^  C^  ^  ^ 49
 eh edt ~ ~ r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^
 T
2 2 2 2
p9
where c1 is a positive constant; 12 2 12Q T Q 12C s 2 12 is
2
1 1 _
eu1 eu2 eh ed ~ ~_ ~ ~ r  2 e2 bounded. The boundedness of stems from the fact that it merely
p8 p9
contains
 the actual parameters
 of the system. Note that
1 G Ns 1s_  C~ C^  Q~ Q^ eu1 eu2 e e
 T
2
1^
T
1^ ^ 1~2
 Q Q C  
2
1^2
2 2 is always non-positive. Therefore, eq.
2
1 1 _
~ ~_ ~ ~ r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^ eu1
 T
p8 p9 (49) can be rewritten as follows:
V_ r  c1 V 1 G Ns 1s_ 1 t 2 t

  1 50
~ ~ e ^  ~ ~ ~_ r  2 e2
1 _
eu2 e^  ~
p8 p9 Solution to eq. (50) can be written as follows:
Z t
1 G Ns 1s_  C~ C^  Q~ Q^ eu1 eu2  e~  e~
 T

e  c1 t  1 G Ns 1s_ d

Vt r e  c1 t V 0
1 1 _
e^ e^ ~ ~_ ~ ~ 44 0
Z t
p8 p9
e  c1 t  1 2 d 51
0
Substituting eqs. (39) and (40), which were dened to estimate
the bounds of the neural network error and disturbance source, It is concluded from the discussed property of Nussbaum-type
into eq. (44) leads to functions that if N(s) is a Nussbaum-type function, N(s) 1 is a
Nussbaum-type function as well. Using this property, eq. (51) is
V_ r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^ eu1 eu2
 T
rewritten as
 e~  e~ e^ e^  ~ ^  ~ ^ 45 Z t

Vt r e  c1 t V0 e  c1 t ec1 1 G N 1 ss_ d
As noticed, the control law has been divided into three parts. 0
Z t
The rst part, introduced in eqs. (35) and (36) is chiey respon-
e  c1 t ec1 1 2 d
sible for tracking purposes. In addition, the remaining parts of the 0
control law, given by eqs. (35) and (36) are responsible for Z t Z t
ec1 1 G N 1 ss_ d e  c1 t ec1 d

unknown bounds of the neural network error and disturbance r e  c1 t V 0 e  c1 t
0 0
source, respectively. Adopting eqs. (35) and (36), (45) becomes Z t

r e  c1 t V 0 e  c1 t ec1 1 G N 1 ss_ d 52
V_ r  2 e2 1 G Ns 1s_  C~ C^  Q~ Q^  ~ ^  ~ ^
 T
0 c1
P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241 239

where Z max f 1 t 2 t g . Thus, employing the Lemma, are compared. The considered wind turbine is an NREL-offshore-
given by eq. (32), implies that the closed-loop system is UUB, and all baseline 5 MW wind turbine. In order to make the dynamics
uncertain, _ was added to the equation as an unknown non-
2
the control signals remain bounded. As a result, the proposed
controller is able to adopt the pitch angle as a control input, despite linear function. In addition, unknown disturbance was added to
its non-afne nature. Furthermore, the designed controller also the wind prole, which can be seen in Fig. 1. The characteristics of
accounts for both parametric and non-parametric uncertainties, the considered wind turbine can be found in Table 1.
along with unknown disturbance sources. It is worth noting that no The control strategy is to, rst, increase the rotor speed up to
prior knowledge of the direction of the non-afne input (i.e., pitch the cut-in speed. Then, proportional to wind speeds, the desired
angle) was used in the derivation of the proposed method. Hence, rotor speed, given by eq. (1), is achieved so that the turbine
given the restrictive assumptions imposed on the non-afne por- operates at the maximum efciency. It is worth noting that the
tion of the dynamics and its derivatives in the literature [19], the generator torque is utilized as the input in this phase. However, as
proposed method is at advantage in practical applications. the rotor speed increases and approaches its rated value, the
Remark 1. As mentioned earlier, there exist methods in the lit- generator torque reaches its rated value as well. Thus, it can be no
erature, which can address the non-afne nature of the pitch longer employed as an input. Hence, the pitch angle is adopted as
angle, along with uncertainties [19]. However, such methods an input to keep the rotor speed at its rated value. The selected
require to impose restrictive assumptions on the non-afne control gain for the torque control phase is 1 5; and for the
function and its derivatives. However, the fulllment of such pitch control phase are 1 1 and 2 5: The wind speed prole,
assumptions for systems subjected to uncertainties cannot be used for this case study can be seen in Fig. 1. As seen from Fig. 1,
guaranteed. Be that as it may, no such restrictive assumptions are the considered wind prole has both increasing and uctuating
made in the proposed method. behavior, which makes it a suitable case to investigate the
adaptability of the proposed controller. Fig. 2 exhibits the perfor-
Remark 2. In the second control phase (i.e., pitch control), the mance of the proposed controller, compared to those of the PI
control objective is not the achievement of the maximum power
capture efciency. It is because in the pitch control phase, the rotor
should not exceed its rated value. In addition, the rotor speed rated
value is different from the desired speed for the maximum power
capture efcacy, given by eq. (1). Thus, the establishment of the
asymptotic stability is not of prime control importance during
this phase.

5. Simulation results

In this section, the simulation results of the proposed con-


troller, alongside a PI controller and the controller designed in [19]

Fig. 2. Rotor speed versus desired rotor speed.

Fig. 1. Wind prole.

Table 1
NREL-offshore-baseline 5 MW wind turbine.

Hub height 90 m Rotor mass 110,000 kg


Rotor, hub diameter 126 m, 3 m Optimal tip-speed- 7.55
ratio
Cut-in, rated rotor 6.9 rpm, Rated generator 43,100 N m
speed 12.1 rpm torque
Fig. 3. Zoomed-in version of the rotor speed.
240 P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241

Fig. 4. Generator torque required by the proposed controller versus PI controller.

Fig. 7. Estimation of neural network parameters.


Fig. 5. Pitch angle activities of the proposed controller versus PI controller.

However, the PI controller and the one in [19] fail to follow the
desired rotor speeds in the second region. Nevertheless, the per-
formance of the controller proposed in [19] is superior to that of
the PI controller. However, the controller in [19] is not able to
establish the asymptotic stability, which eventuates in lower
power capture efciency. On the contrary, the proposed controller
can precisely track the desired rotor speed, which guarantees the
maximum power capture. Fig. 4 displays the required generator
torque for each controller. As seen from the gure, the PI controller
and the one in [19] demand a larger amount of torque, in contrast
to the proposed method. The main reason for this is due to the fact
that the tracking error of the proposed controller is smaller, which
results in lower generator torque. Since the tracking error of the PI
controller is larger than the one in [19], the required generator
torque of the one in [19] is less than the PI controller. It is worth
mentioning that the method proposed in [19] does not account for
the error caused by unknown disturbance source. Hence, it
experiences larger tracking error, which leads to greater deman-
ded torque. There also exits a larger amount of overshooting in the
beginning of the simulation in case of the PI controller, which
Fig. 6. Estimation of the damping coefcient. leads to higher generator torque. The pitch angle behavior can be
observed in Fig. 5. As it can be noticed, the proposed controller
controller and the one designed in [19]. As it follows from Fig. 2, results in smoother behavior, contrary to that of the PI controller.
and its zoomed-in version Fig. 3, all the three controllers function The pitch action of the method presented in [19] becomes steady
successfully to enable the rotor to reach the cut-in speed. at a higher degree, compared to the proposed method. It is
P. Bagheri, Q. Sun / ISA Transactions 63 (2016) 233241 241

because the control law for the pitch angle in [19] is not directly a References
function of the tracking error. Therefore, it does not vanish as the
tracking error becomes small. Moreover, pitch angle activities of [1] Camblong H, Nourdine S, Vechiu I, Tapia G. Control of wind turbines for fatigue
loads reduction and contribution to the grid primary frequency regulation.
the PI controller are relatively more intense, particularly at the Energy 2012;48:28491.
time when the pitch angle is engaged as the input (around Second [2] Vihril H. Control of variable speed wind turbines; 2002.
[3] Burton T, Sharpe D, Jenkins N, Bossanyi E. Wind energy handbook. John Wiley
310). As it follows from Fig. 5, the proposed controller exhibits
& Sons; 2001.
smoother and atter behavior, compared to the PI controller. [4] Munteanu I, Bratcu AI, Cutululis N-A, Ceanga E. Optimal control of wind
However, as Fig. 5 suggests, the required pitch angle to maintain energy systems: towards a global approach. Springer Science & Business
Media; 2008.
the rotor at its rated speed is less for the PI controller. Finally, [5] Bianchi F, Mantz R, Christiansen C. Gain scheduling control of variable-speed
Figs. 6 and 7 display the estimations of the damping coefcient wind energy conversion systems using quasi-LPV models. Control Eng Pract
2005;13:24755.
and neural network parameters, respectively. As Fig. 7 suggests, [6] Munteanu I, Cutululis NA, Bratcu AI, Ceang E. Optimization of variable speed
three basis functions of the type given in eq. (9) are used to model wind power systems based on a LQG approach. Control Eng Pract
2005;13:90312.
non-parametric uncertainties; Q1, Q2 and Q3 are the coefcients of
[7] Boukhezzar B, Siguerdidjane H. Comparison between linear and nonlinear
these basis functions, which are updated using adaptive laws eqs. control strategies for variable speed wind turbines. Control Eng Pract
(15) and (38) in the torque and pitch control phases, respectively. 2010;18:135768.
[8] Van der Hooft E, Van Engelen T. Feed forward control of estimated wind
It is worth noting that the reason these estimations converge to speed. Energy research Centre of the Netherlands ECN; 2004.
zero after the pitch angle is used as the input, stems from the fact [9] Van der Hooft E, Van Engelen T. Estimated wind speed feed forward control for
wind turbine operation optimisation. In: Proceedings of European wind
that their adaptive laws change from eqs. (14) and (15) to (37) and energy conference in London, UK; 2004.
(38), which dictate converging exponential behavior. [10] Novak P, Ekelund T, Jovik I, Schmidtbauer B. Modeling and control of variable-
speed wind-turbine drive-system dynamics. Control Syst IEEE 1995;15:2838.
[11] Stol K, Fingersh L. Wind turbine eld testing of state-space control designs.
National Renewable Energy Laboratory, NREL/SR-500-35061, Golden, Co.;
2004.
6. Conclusion [12] Hand MM. Variable-speed wind turbine controller systematic design metho-
dology: a comparison of non-linear and linear model-based designs. Golden,
CO (US): National Renewable Energy Laboratory; 1999.
Due to the existence of uncertainties and unknown disturbance [13] Bongers PMM. Modeling and identication of exible wind turbines and a
sources, the design of controllers to increase power capture ef- factorizational approach to robust control design. TU Delft, Delft University of
Technology; 1994.
ciency in wind turbines is a challenging problem. The objective of [14] Boukhezzar B, Lupu L, Siguerdidjane H, Hand M. Multivariable control strategy
this paper was to develop a novel controller for variable-speed, for variable speed, variable pitch wind turbines. Renew Energy 2007;32:
127387.
variable-pitch wind turbines. In addition, the controller was [15] Saravanakumar R, Jena D. Validation of an integral sliding mode control for
required to achieve the maximum power capture, in the presence optimal control of a three blade variable speed variable pitch wind turbine. Int
of unknown disturbance and uncertainties. Alongside unknown J Electr Power Energy Syst 2015;69:4219.
[16] Dang D, Wang Y, Cai W. Nonlinear model predictive control (NMPC) of xed
disturbance and uncertainties, the non-afne nature of the pitch pitch variable speed wind turbine. In: 2008 ICSET 2008 IEEE international
angle (i.e., a control input) was among the main complications of conference on sustainable energy technologies, IEEE; 2008. p. 2933.
[17] Bagheri P, Shahrokhi M. Neural network-based synchronization of uncertain
the control design. To address these problems, a Nussbaum-type chaotic systems with unknown states. Neural Comput Appl 2015:18.
function was employed in the control law, and a neural network [18] Sharma N, Gregory CM, Johnson M, Dixon WE. Closed-loop neural network-
based NMES control for human limb tracking. IEEE Trans Control Syst Technol
was designed to account for unmodeled dynamics. Further, a 2012;20:71225.
multi-part control law was proposed to guarantee the stability of [19] Jafarnejadsani H, Pieper J, Ehlers J. Adaptive control of a variable-speed vari-
the entire closed-loop system, in the presence of unknown dis- able-pitch wind turbine using radial-basis function neural network. IEEE Trans
Control Syst Technol 2013;21:226472.
turbance and neural network approximation error. It was [20] Wang Z, Shen Z, Cai C, Jia K. Adaptive control of wind turbine generator sys-
demonstrated that the proposed controller is able to achieve the tem based on RBF-PID neural network. In: Proceedings of the 2014 interna-
tional joint conference on neural networks (IJCNN), IEEE; 2014. p. 53843.
asymptotic stability in the torque control phase, in which the [21] Hui J, Bakhshai A. A new adaptive control algorithm for maximum power point
turbine is required to operate at the maximum efciency. More- tracking for wind energy conversion systems. In: Proceedings of the power
electronics specialists conference, 2008 PESC 2008, IEEE; 2008. p. 40037.
over, it was proven that the proposed controller is able to provide
[22] Farrell JA, Polycarpou MM. Adaptive approximation based control: unifying
arbitrary small convergence error in the pitch control phase. At the neural, fuzzy and traditional adaptive approximation approaches. John Wiley
end, the performance of the proposed controller was compared to & Sons; 2006.
[23] Matkowski J. A mean-value theorem and its applications. J Math Anal Appl
those of a PI controller, and a recent adaptive method via simu- 2011;373:22734.
lations. The simulation results conrmed the analytical results [24] Ryan E. A universal adaptive stabilizer for a class of nonlinear systems. Syst
Control Lett 1991;16:20918.
regarding the stability of the closed-loop system. Further, the [25] Zhang T, Wen H, Zhu Q. Adaptive control of time delay nonlinear systems with
proposed controller exhibited a more successful performance in unknown control direction.In: Proceedings of the 2010 Chinese control and
decision conference (CCDC), IEEE; 2010. p. 205762.
comparison to the PI Controller and the existing adaptive method.