Sunteți pe pagina 1din 12

Renewable Energy 105 (2017) 483e494

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Adaptive pitch control of wind turbine for load mitigation under


structural uncertainties
Yuan Yuan, J. Tang*
Department of Mechanical Engineering, University of Connecticut, 191Auditorium Road, Unit 3139, Storrs, CT 06269, USA

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

Article history: In this research, a new adaptive control strategy is formulated for the pitch control of wind turbine that
Received 17 June 2016 may suffer from reduced life owing to extreme loads and fatigue when operated under high wind speed
Received in revised form and internal structural uncertainties. Specically, we aim at making a trade-off between the maximum
14 November 2016
energy captured and the load induced. The adaptive controller is designed to both regulate generator
Accepted 26 December 2016
speed and mitigate component loads under turbulent wind eld when blade stiffness uncertainties exist.
Available online 27 December 2016
The proposed algorithm is tested on the NREL offshore 5-MW benchmark wind turbine. The control
performance is compared with those of the gain scheduled proportional integral (GSPI) control and the
Keywords:
Wind turbine
disturbance accommodating control (DAC) that are used as baselines. The results show that with the
Pitch control proposed adaptive control the blade root apwise load can be reduced at a slight expense of optimal
Load mitigation power output. Moreover, the blade load mitigation performance under uncertain blade stiffness
Adaptive control reduction is improved over the baseline controllers. The control approach developed in this research is
Turbulent wind eld general, and can be extended to mitigating loads on other components.
Uncertainties 2016 Elsevier Ltd. All rights reserved.

1. Introduction changes, icing, assemblage error of the turbine, and rotor asym-
metries caused by damaged in blades, etc. Intuitively, the long-term
Wind energy is commonly recognized as a major environmen- productivity of a wind turbine may be higher if it is operated
tally friendly energy source that is renewable. Modern wind tur- somewhat conservatively at the presence of these uncertainties.
bines are large, exible structures operating in highly turbulent and Blade pitch control adjusts the aerodynamic angle of attack.
sometimes unpredictable wind eld [1]. There are three regions in Currently, collective pitch angle control is widely utilized to adjust
wind turbine operation (Fig. 1). In Region 1 where wind speed is the generator speed in Region 3 (above rated-speed). Typically,
very low, the brake is equipped to stop the wind turbine operation, linear system control methods can be adopted when accurate
yielding no power generation. In Region 2, from the cut-in wind model can be acquired. Jonkman et al. developed the PID approach
speed to the cut-out speed, the objective is to achieve the to regulate the generator speed, and a gain scheduling part was
maximum power output, i.e., to maximize the aerodynamic power added to deal with the aerodynamic sensitivity of the Cp curve to
coefcient Cp . The generator torque control is utilized to track the blade pitch angle, which yields a gain scheduled proportional in-
optimal tip speed ratio in specic Cp curve which is usually ob- tegral (GSPI) control [2]. Wright investigated the state-space
tained from benchmark test data. In Region 3 where wind speed is feedback control after linearizing the nonlinear model under
high, the objective is to track the rated generator speed. There selected operating points. As the model depends on the operating
generally exists a trade-off between tracking the generator speed point selection and wind disturbance may not be directly
(i.e., power output) and maintaining the mechanical loads to tur- measured, a Disturbance Accommodating Control (DAC) was
bine components (e.g., blades), due to the signicant uncertainties developed to achieve disturbance rejection and to attenuate loads
in operating conditions and system properties. Such external and by accounting for wind disturbances with additional state estima-
internal uncertainties include varying wind speeds, weather tors [3]. Nevertheless, the actual performance may still be affected
by the wind estimator accuracy, and the wind speed disturbances
cannot be precisely cancelled out in high-order controllers [4]. Also
the wind disturbance gain should be redened to track the set-
* Corresponding author.
point. Frost et al. employed an adaptive control for speed
E-mail address: jtang@engr.uconn.edu (J. Tang).

http://dx.doi.org/10.1016/j.renene.2016.12.068
0960-1481/ 2016 Elsevier Ltd. All rights reserved.
484 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

adaptive Disturbance Tracking Control (DTC) technique to track the


optimal tip speed ratio based on a wind speed estimator with the
linearized CART model [25]. Magar and Balas employed a simplied
model in adaptive control design, and demonstrated performance
robustness in tracking optimal tip speed ratio under various oper-
ating conditions [26].
In this research, we take into consideration the varying wind
speeds and model uncertainties, and aim at striking a balance be-
tween tracking the generator speed (i.e., near-term performance)
and mitigating mechanical loads (i.e., long-term reliability). We
investigate the collective pitch control that can be applied to many
existing wind turbines. We specically focus on above-rated wind
speeds, and assume the turbine is subjected to blade stiffness re-
ductions (possibly caused by minor damage). Specically, a model
reference adaptive controller is formulated towards this goal.
Mathematically, our control objective is to balance the generator
speed-tracking and the structural loads under blade stiffness un-
Fig. 1. Three regions of a typical wind turbine.
certainties. A closed-loop system based on the DAC strategy that is
originally suggested by Lee and Balas [34] is selected as the refer-
ence model in the adaptive control. The reference signal is essen-
regulation, which does not require knowing precisely the wind
tially the outcome of a trade-off between the generator speed and
turbine model parameters [5]. Connor et al. formulated an H
structural loads, acquired based on an optimal control of apwise
controller to mitigate fatigue damage of wind turbine with known
deection modes and generator speed modes through selecting
linear model [6]. As the wind speed varies, the linear model is no
proper weighting matrices. The reduction in apwise moment at
longer time-invariant. Stol and Balas thus extended the periodic
the slight expense of uctuation of power and generator speed can
DAC based on time-varying turbine model, aiming at improving
lead to the aforementioned trade-off. Meanwhile, adaptive control
blade load attenuation level [7]. To better reject wind disturbance,
strategy has the capability of dealing with various uncertainties by
Hand and Balas further incorporated properties of coherent tur-
online updating the parameters at every sampling step according to
bulent inow structures to DAC to mitigate the transient blade load
the difference of the nominal output and reference output while
response caused by the Rankine vortex in the ow [8]. Since it is
maintaining the stability robustness. While our formulation is
difcult to optimize all performance metrics simultaneously, Model
mainly based on blade load consideration, the approach can be
Predictive Control (MPC) was proposed to realize trade-off between
extended to loads on other components.
multiple objectives [9e11]. In order to realize better balance of the
The rest of the paper is organized as follows. In Section 2 we
conicting objectives such as power capture and tower fore-aft
briey introduce the 5-MW benchmark wind turbine model
fatigue load in the entire operating region, multi-objective MPC
employed. The two existing controllers, i.e., the GSPI control and
tuned by Pareto curves was formulated [12,13]. In addition, the
the DAC approach, are outlined in Section 3, followed by the new
dynamic inow can be included in the prediction model of MPC
adaptive control proposed. Case studies are presented in Section 4.
design to facilitate the robustness of the controller [14].
In this research both the model development and the case simu-
Along with the hardware advancement in newer wind turbines,
lations are conducted by using the Fatigue, Aerodynamics, Struc-
there have also been recent studies focusing on individual blade
tures, and Turbulence (FAST) code [29]. The performance resulted
pitch control to attenuate blade root loads to counteract wind shear
from the new adaptive control are compared with those from the
and asymmetric disturbance across the rotor disc [15e17]. Hassan
baseline GSPI controller and DAC controller. Section 5 summarizes
investigated a mixed H2 /H with pole placement to regulate
the concluding remarks.
generator speed and reduce once-per-revolution (1P) load [18].
More recently, light detection and ranging (LIDAR) system has been
developed to remotely measure the incoming wind speed [19]. 2. Model development
Using LIDAR measurement, Dunne and Pao [20] designed a com-
bined feedback/feedforward blade pitch controller to reduce the We employ the NREL offshore 5-MW benchmark wind turbine
structural loads in blade root, tower base and top. In another study, in this research, which is a popular choice in related studies
FX-RLS-based feedforward control was suggested to mitigate both [2,18,20,26]. This is a three-bladed, upwind, variable-speed variable
blade bending and tower moment while regulating the rotor speed blade-pitch-to-feather-controlled horizontal axis turbine [2]. The
with slightly sacriced energy, which takes the advantage of the rotor diameter is 126 m, and the hub height is 90 m. The cut-in
robustness characteristics and good disturbance rejection of the wind speed is 3 m/s, the rated wind speed is 11.4 m/s, and the
adaptive control [21]. cut-out wind speed is 25 m/s. The rated generator speed is
In wind turbine blade pitch control, the fundamental challenge 1173.7 rpm. We additionally assume a rigid tower, a rigid hub and
stems from the nonlinearity of the model, unmodeled dynamics, as locked yaw position.
well as the internal and external uncertainties [22,23]. Conceptu- A variable-speed wind turbine generally consists of blades, a low
ally, the strategy of adaptive control is appealing in tackling such speed shaft, a gearbox, a high speed shaft, and a generator. The
challenge, as it has the potential of dealing with highly nonlinear aerodynamic power captures by the rotor is given as [1]
model with unknown parameters and complicated operating con-
ditions. Indeed, this strategy has been explored on wind turbines to 1
Pwind rpR2 Cp l; bv3 (1)
achieve different goals. In order to realize the maximum power at 2
below-rated wind speeds (Region 2), Johnson et al. presented an
where R is the rotor radius and r is the air density. The power co-
adaptive pitch controller to obtain high aerodynamic efciency by
efcient Cp is a nonlinear function of the tip speed ratio l and the
using a discrete hill-climbing method [24]. Balas et al. developed an
pitch angle b, and can be obtained from look-up table generated by
Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494 485

led test data. Here elastic equation of motion can be expressed as [29],

ur R Mq; u; tq fq; q;
_ u; ud ; t 0 (8)
l (2)
v
where M is the mass matrix, f is the nonlinear forcing function
where v is the wind speed and ur is the rotor speed. From Eqs. (1) vector that includes the stiffness and damping effects, q is the
and (2) we can see that the aerodynamic coefcient depends on response vector, u is the vector of control inputs, ud is the vector of
wind speed, rotor speed and blade pitch angle. Pitch angle control wind disturbance, and t is time. In our analysis, f is calculated by
can be used to limit the aerodynamic power captured by the rotor using AeroDyn through the Blade Element Momentum (BEM)
to realize load mitigation. The aerodynamic torque applied to the Theory [28]. In this approach, aerodynamic forces and moments are
rotor can be expressed as calculated in each segment of the blade, and the distributed forces
are obtained by integration from blade root to tip. FAST then
1 pCp l; brR2 v3 numerically linearizes Eq. (8) by perturbing each variable about
Ta (3)
2 ur their respective operating points. After the Taylor series expansion,
we obtain
When we assume that the low speed shaft is rigid, we have the
equations of motion for the rotor and the generator as, respectively, Mq Cq_ Kq Fu Fd ud (9)
Jr u_ r Ta  Cr ur  Tls ; (4a)
where M, C, and K are, respectively, the linearized mass, damping,
and stiffness matrices, F is the control input vector and Fd is the
Jg u_ g Ths  Cg ug  Tem (4b) wind disturbance vector.
Here we employ the linearized model that is simplied from the
where Jr and Jg are the moment of inertia of the rotor and that of the nonlinear model of the actual wind turbine, which is obtained
generator, ug is the generator speed, Tls and Ths are the low speed when the rst apwise blade mode DOF (degree-of-freedom), the
shaft torque and high speed shaft torque, Cr and Cg are the external generator DOF, and the drivetrain torsional exibility DOF are
damping coefcient of the rotor and that of the generator, and Tem switched on. One can cast the linearized equation into the state-
is the generator electromagnetic torque. space representation,
We assume an ideal gearbox (i.e., 100% efciency) with the
transmission ratio ng given as x_ Ax Bu Bd ud (10a)

Tls ug y Cx Du Dd ud
ng (5) (10b)
Ths ur
T
Substituting Eq. (5) into Eq. (4) yields a simplied relation where x qT ; q_ T is the state vector. The rst ve state variables
are, respectively, the generator deection, the drive-train torsional
Jt u_ r Ta  Kt ur  Tg (6) deection perturbation, and the ap displacement perturbations of
three blades. The remaining ve state variables are the corre-
where sponding velocities. A, B, C, and D are the state matrix, the control
input matrix, the output matrix, and the control input transmission
Jt Jr n2g Jg ; (7a) matrix. Bd and Dd are the wind disturbance input matrix and the
wind disturbance input matrix. u is the control input (i.e., per-
turbed blade collective pitch angle), ud is the disturbance input (i.e.,
Kt Kr n2g Kg ; (7b) perturbed wind speed), and y is the output. The design point for
linearization is chosen to be in Region 3. An operating point with
Tg ng Tem (7c) wind speed v 15 m=s, pitch angle q 10:45 deg, and rotor speed
ur 12:1 rpm is used in the subsequent control investigations.
The governing equation (i.e., Eq. (6)) can represent dynamics in
all operating regions. It indicates that the generator speed can be
regulated by the aerodynamic torque Ta and the electromagnetic 3. Control syntheses
torque Tem . Hence, two separate SISO control loops, i.e., torque
controller and blade pitch controller, can be used. Usually in Region In this section we rst outline a GSPI controller that is one of the
2, torque controller is utilized to regulate the generator speed while baselines to illustrate the mathematical background and also the
the blade pitch angle is held constant to maintain the maximum challenge, and then present the formulation of the proposed model
aerodynamic coefcient. In above-rated region (i.e., Region 3), reference adaptive control with disturbance accommodation. For
however, pitch controller is often implemented to limit the aero- these controllers, we assume the ltered generator speed mea-
dynamic torque Ta to avoid extreme loads. Here we adopt the tor- surement is the only input and the controllers provide the collec-
que controller suggested by Jonkman et al. [2] that was designed for tive pitch command.
the same NREL 5-MW benchmark wind turbine, and focus on the
blade pitch control. 3.1. Baseline gain scheduled proportional integral control and
Observing Eq. (3), one can see that there exists complex generator torque control
nonlinear relation between the pitch angle b and the aerodynamic
torque Ta . In addition, the apwise vibration of a blade caused by One baseline controller used for comparison in this research is
aerodynamic load needs to be suppressed to avoid damage. the gain scheduled proportional integral control (GSPI) originally
Therefore a more accurate dynamic model including the ap mode developed by Jonkman et al. [2]. The conventional approach to
is needed. In this research we employ the FAST code developed by controlling power output involves two independent control loops,
National Renewable Energy Laboratory (NREL) [29] to establish the i.e., the generator torque control and the full-span rotor collective
mathematical model of the wind turbine. The nonlinear aero- blade pitch control. The purpose of blade pitch control is to regulate
486 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

the generator speed to avoid excessive structural loads at above- systems. In wind turbines, nevertheless, the system parameters are
rated wind speed. The generator torque control aims at achieving usually not completely known, due to, for example, the inherent
the maximum power capture in various regions. nonlinearities of the model, unmodeled modes, manufacturing and
The form of GSPI can be written as assemblage tolerances, and external operating uncertainties. One
0 1 potential solution is to use an adaptive control approach which can
Zt deal with various uncertainties by online updating parameters to
Dq GSq@KP Du KI DuA (11) force the error between reference and plant output to approach
0 zero. Here in this research we adopt the direct model reference
adaptive control approach with a disturbance accommodation
where Dq is the small perturbation of blade pitch angle around the controller (DAC) (Fig. 2) [33]. There are two parts in this augmented
operating point, and Du is the error between measured generator controller. One is the reference model which is chosen as the DAC
speed and the rated set point value. KP and KI are rst calculated scheme. The other is the adaptive part, where the algorithms can
under the initial operating point. The gain correction factor GSq is adapt itself to the internal and external uncertainties. These two
a function of the blade pitch angle to regulate the gains under parts will be discussed as follows.
different operating conditions [2],

1
GSq (12)
1 q=qk
3.2.1. Reference model formulation with disturbance
where q is the blade pitch angle, and qk is the blade pitch angle at
accommodation
which the pitch sensitivity value is doubled from its value at the
The concept of Disturbance Accommodating Control (DAC) was
rated operating point. The pitch sensitivity refers to the sensitivity
originally proposed by Johnson [31], and Balas et al. [32] further
of aerodynamic power to the rotor collective blade pitch angle.
extended the DAC concept to large-scale horizontal axis wind tur-
As mentioned, another component of wind turbine control is
bines. In the adaptive control to be developed in this research, we
the generator torque control. In Region 2, the control objective is to
employ the DAC strategy to formulate the reference model. As the
reach the maximum power by tracking the optimum tip speed ratio
error between the reference and plant output approaches zero, the
(TSR) whereas the pitch angle is kept constant zero. Region 2 12 is a
wind disturbance effect is reduced and the component loads are
linear transition to capture the maximum power when the gener-
mitigated. Furthermore, an optimal control inner loop is embedded
ator speed approaches the rated-speed. According to [2], the
in the reference model in order to realize the trade-off between the
baseline torque control can be formulated as
load mitigation and generator speed tracking.
8 Recall the linearized state-space equation of the system, Eq. (10).
> 0 Region 1
>
> In DAC, we assume that wind disturbance is the variance from the
>
>
>
< ku2 Region 2 wind speed at the operating point and has a known waveform but
t trated  t1 (13) unknown amplitude. Specically, we can model it as step distur-
>
> t1 u  u1 Region 2 1
>
> urated  u1 bance [3,4,7,8], where the amplitude of wind speed changes from
>
> 2
: trated one to another within a relatively short sampling interval. The wind
Region 3
disturbance is denoted as ud , and is characterized by the following
where u is the generator speed, k is an optimal constant, i.e., disturbance wave generator,
0:0255764 Nm=rpm2 for this benchmark wind turbine. The tran-
sitional generator speed u1 is 1161.963 rpm (i.e., 99% of the rated-
speed), the generator rated-speed urated is 1173.7 rpm, the rated Wind Reference Model
generator torque trated is 43,093.55 Nm, and t1 is 43,092.38 Nm to
yield the generator-slip percentage constant of 10%.
Linearized Model
It is worth noting that GSPI is designed for generator speed
tracking in the above-rated region, which may indeed cause large
uctuation in component loads. In addition, the proportional and
integral gains are tuned without considering the possible un- +
certainties in the plant. Consequently, it may not work well when LQR Plant Es mator
inherent and external uncertainties exist. +

Disturabnce Disturbance
3.2. Model reference adaptive control with disturbance Cancella on Es mator
accommodation
-
The life span of wind turbines is expected to be 20 years or Nonlinear Wind +
longer. In reality, wind turbines are often subjected to harsh envi- Pitch Actuator
Turbine
ronment with highly turbulent wind eld. There is a large uctu-
ation in aerodynamic forces and moments to the blade. Such
situation, if not handled properly, may lead to unexpected failure of
turbine components. In high wind speeds, we also want to regulate Adap ve Upda ng
the generator speed at the rated value and reduce the oscillation.
Thus an intelligent control system is desired to ensure both the
near-term performance (constant speed) and long-term reliability
(less failure). Fig. 2. Direct model reference adaptive control approach with a disturbance accom-
The model delity of the plant is essential for all control modation controller (DAC).
Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494 487

between speed regulation and apwise displacement response.


z_ d Fzd ; (14a) An illustration will be presented in the case studies section.
The second term at the right hand side of Eq. (18) concerns the
ud Qzd (14b) disturbance. To mitigate the effect of disturbance, we can generally
choose Gd in such a manner that the coefcient matrix of distur-
For the step disturbance, F and Q are assumed to be known as
bance BGd Bd Q in Eq. (18) is equal to zero, i.e., to realize distur-
bance cancellation. In most cases, this matrix does not have full
F 0; Q 1 (15)
rank and is ill-conditioned. In those cases, we can opt to minimizing
In Eq. (10b), D and Dd are both equal to zero according to the its L2 norm kBGd Bd Qk and obtain
FAST linearization. When we give a bounded control input um (i.e.,
 1
blade pitch angle signal) to the linearized model and maintain the Gd  BTm Bm BTm Bdm (20)
stability of the closed-loop system (i.e., the eigenvalues of the
closed-loop system all having negative real parts), the closed-loop As the only measurement considered to be available is the
DAC system, which is used as the reference model in the proposed generator speed, here state estimator is used to predict the un-
adaptive control strategy, can be expressed as measured states. The plant state and output estimator are given as

x_ m Am xm Bm um Bdm udm (16a)  


^_ m Am x
x b Bdm u
b m Bm um K1 y  y b dm (21a)
y m Cm x m (16b)
bm Cx
y bm (21b)
where Am , Bm , Bdm , and Cm are the state, input, disturbance input,
and output matrices of the linearized model. In DAC, the control b dm , and y
bm, u b are the estimates of xm , udm and y. The
where x
law is expressed as combination of full state feedback (which is
disturbance state estimator is used to predict the unknown wind
intended to take load mitigation into consideration) and distur-
disturbance,
bance state feedback to reject disturbances,
b d qb
u zd (22a)
um Gxm Gd zdm (17)
 
z_ d Fb
where G is the state gain and Gd is the disturbance state gain. We b b
z d K2 y  y (22b)
then have

x_ m Am Bm Gm xm Bm Gd Bdm Qzdm (18) where ub d and b


z d are the estimates of ud and zd . Recall Eq. (10).
These two estimators can be joined as an augmented state esti-
As mentioned, the rst term at the right hand side of Eq. (18) is
mator form. K1 and K2 are adjusted by pole placement such that the
related to the regulation of generator speed with load mitigation
estimator has acceptable transient behavior and accuracy.
being considered. As such, G can be obtained by Linear Quadratic
Regulator (LQR) approach, i.e., minimizing the quadratic perfor-
mance index [38]:
3.2.2. Adaptive algorithm development
Z  The formulation of the DAC-type reference model, as explained
1
J xm tT Q xm t um tT Rum t dt (19) in the preceding subsection, enables us to reach a trade-off by
2 balancing load mitigation and speed regulation while reducing the
0
disturbance effect. DAC controller, on the other hand, is not
where Q is a positive-semidenite, real, symmetric matrix and R is designed to be robust and it may be sensitive to errors in the tur-
a positive-denite, real, symmetric matrix. For the NREL 5-MW bine model. As shown in Fig. 2, the second part of the proposed
benchmark wind turbine, the linearized system described by Eq. adaptive control scheme is the adaptation which allows us to take
(10) is actually not state controllable because the rank of its uncertainties into consideration. The uncertainties include 1) those
controllability matrix is less than the order of the system. There- caused by the nonlinear aerodynamic loads on the turbine, 2)
fore, further reduction of the system is necessary in order to allow varying wind speed, and 3) unmodeled dynamics.
arbitrary selection of system poles to enhance damping and For adaptive algorithm formulation, we assume the actual wind
improve system responses (with G) [3]. Here we reduce the original turbine is a linear, time-invariant, and nite dimensional plant
six states (displacements and velocities) for apwise mode of three shown as follows,
blades to one symmetric displacement state and one symmetric
velocity state. Meanwhile, as the only available measurement is x_ p Axp Bup Bd ud (23a)
generator speed, the corresponding system is not observable (i.e.,
the rank of the observability matrix is less than the order of the yp Cxp (23b)
system). We eliminate the state for generator deection mode and
only reserve generator speed state to enable the placement of the where xp is the plant state vector, up is the plant input vector, and
poles of the estimators which will be discussed later. yp is the plant output vector. In this research, the input is the col-
After the reduction of the linearized model, we can use Eqs. lective pitch angle, and the output is the generator speed. It is
(17)e(19) to facilitate the design of reference model with desired worth noting that, in simulation, we will still employ the nonlinear
response. The diagonal elements in Q and R need to be adjusted NREL 5-MW benchmark as our actual plant. This proposed linear-
properly to move the poles further to the left in the complex plane ized model is only adopted for theoretical analysis of the adaptive
[23]. In some scenarios, the improvements of apwise mode per- control.
formance may deteriorate the generator mode response. Thus we The error between the reference output ym and the plant output
should carefully choose Q and R to obtain an optimal trade-off y is dened as ey , i.e.,
488 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

Table 1
ey y  ym (24) Wind le parameters.

Case 1 Case 2 Case 3


where y is the actual plant output, and ym is the reference output
Mean wind speed (m/s) 15 12 17
obtained by the closed-loop DAC applied in Eq. (16). Here y and ym
Turbulence model IEC Kaimal spectral model
are both actually scalars, because the only available measurement is Turbulence intensity IEC category B (14%)
the generator speed. The reference signal is essentially the result of Height of reference wind speed (m) 90
a trade-off between the generator speed and blade apwise loads.
The purpose of the adaptive algorithm is to force the actual plant
output y to follow the reference output ym , i.e. lim ey 0, by the Matlab/Simulink environment. The simulation duration is 600 s
t/
online updating the parameters at every sampling step with un- with an integration step size of 0.0125 s. As mentioned, the model
known or varied system dynamics or other specic uncertainties. involves ve DOFs including the rst apwise blade mode DOF of 3
The general theory of model reference adaptive control and the blades, the generator DOF, and the drivetrain torsional exibility
associated stability analysis can be found in Ref. [34]. Here we have DOF. Aerodynamic forces and moments are calculated by using
AeroDyn. Turbulent wind elds are generated by using TurbSim
b m Gu um Gz b
u Gy ey Gx x zd (25) [35]. Table 1 lists all the simulation parameters used to obtain the
Correspondingly, the adaptive gain laws are wind elds. The mean wind speeds are set to be below, above and
equal to the operating point value (15 m/s), respectively, corre-
G_ y Ky ey eTy ; (26a) sponding to three cases studied in this section. The Simulink model
of the adaptive pitch controller is shown in Fig. 3.

G_ x Kx ey x
b Tm (26b)

4.1. Selection of parameters in closed-loop DAC used as reference


G_ u Ku ey uTm ; (26c) model

G_ z Kz ey b
T
zd (26d) The design of closed-loop DAC is important, because it will be
used as the reference model in the proposed adaptive control, and
where Ky , Kx , Ku , Kz can be arbitrary, positive-denite matrices. In DAC will also be used for performance comparison. As indicated in
order to guarantee the stability, we must have the following con- Section 3.2, we should carefully choose the weighting matrices Q
ditions satised [5,26,34]: and R in Eq. (19) to facilitate an optimal trade-off between speed
regulation and apwise displacement response. We may choose
1) All inputs to the reference model are bounded. In this research, different Q and R combinations and examine the closed-loop poles.
the pitch angle input bound is guaranteed by the pitch angle Here the open-loop poles are 1:901013:892i, 2:46423:9884i,
limits (0e90 ). The disturbance bound is also guaranteed since and 0:4524. The rst pair of poles corresponds to the drive-train
the wind turbine is stopped to avoid excessive loads when the torsional mode, the second pair corresponds to the symmetric
wind speed exceeds the cut-out value (25 m/s). rst apwise blade mode, and the last pole corresponds to the
2) The reference model (closed-loop DAC) described by Eq. (18) is generator mode. When we select, for example,
stable. In this research, all the eigenvalues of closed-loop system 2 3
are placed in the left-half s-plane. 105 0 0 0 0
3) The system described by Eq. (23) is Almost Strict Positive Real 6 0 0:1 0 0 0 7
6 7
(ASPR). This is satised when the transfer function R 1; Q 6
6 0 0 0:06 0 0 7 7
Cp sI  Ap 1 Bp is Strictly Positive Real [5]. The linearized model 4 0 0 0 100 0 5
has been examined at different operating points to validate this 0 0 0 0 103
condition.
we can obtain the closed-loop poles as2:557813:6131i,
The selection of parameters in Eqs. (25) and (26) is based on the 13:55633:1939i, 0:4180. The damping in the ap mode (i.e.,
control objective which is to regulate the generator speed accord- the second pair) is increased as the corresponding poles are moved
ing to the closed-loop DAC output and achieve simultaneously load from 2:46423:9884i to 13:55633:1939i. However, at the
mitigation. Since we have considered the disturbance rejection in same time, the damping in the generator mode is reduced as the
reference model by DAC approach, the gain Gz in Eq. (25) is chosen corresponding pole is moved from 0:4524 to 0:4180. Clearly, a
to be 0. Our main task is to track the generator speed and we do not trade-off exists between the apwise mode and generator mode.
need to track the control input. Thus we can also set Gu to be 0. Alternatively, we may select
Similarly, all the elements in matrix Gx are chosen to be 0 except the 2 3
one that corresponds to the generator speed state. Therefore, in Eq. 106 0 0 0 0
6 0 102 0 0 0 7
(26), the adaptive gains Ku and Kz are 0. We can adjust Ky and Kx to 6 7
achieve the desired generator speed regulation and load mitigation R 1; Q 6
6 0 0 0:8 0 0 7 7
4 0 0 0 0:01 0 5
effect.
0 0 0 0 104
4. Results and discussions The closed loop poles now become 14:062620:3352i,
6:07252:2044i, and 2:2319, indicating both the apwise mode
In this section, case analysis results based on the GSPI, DAC, and and generator mode can be enhanced after careful selection of Q
the proposed adaptive control algorithm are compared. The simu- and R. The corresponding gain matrix G is
lations on the NREL offshore 5-MW benchmark wind turbine are 802:7443; 0:4439; 21:7998; 19:0430; 0:0735. This result is
carried out by connecting FAST with the respective controllers in adopted in the reference model for the following three case studies.
Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494 489

Fig. 3. Controller block diagram.

Table 2 shown in Fig. 4(c). It is observed that the GSPI yields higher ap-
Parameters of controllers. wise root moment from 360 s to 480 s, compared with DAC and
GSq Gain correction factor 1 adaptive control, as shown in Fig. 4(d). The overall performances
1q=6:302336
KP Proportional gain 0.01,882,681 are compared in Fig. 5, where all performance indices are
KI Integral gain 0.008,068,634 normalized with respect to those obtained based on GSPI. It can be
R Weighting matrix 1 seen that the adaptive controller reduces the apwise moment DEL
Q Weighting matrix diag106 ; 102 ; 0:8; 0:01; 104  by 9% which is similar to DAC controller. The normalized generator
G State gain 802:7443; 0:4439; 21:7998; 19:0430; 0:0735
speed errors of GSPI, DAC and adaptive control are 0.1693, 0.2960,
Gd Disturbance state gain 0.0264
Ky Adaptive gain 107
and 0.1727, respectively. The DAC controller exhibits a much larger
Kx Adaptive gain 1010 error than the adaptive controller. The average power output of the
adaptive controller is slightly less than those of GSPI and DAC. The
extreme apwise moment is decreased about 11% in DAC and
4.2. Results of GSPI, DAC and adaptive control under different adaptive control. Therefore, at operating point, the apwise fatigue
turbulent wind elds can be reduced by adaptive controller, although the generator
speed regulation is not as good as GSPI since GSPI is only designed
The DAC and the adaptive controller are simulated in all three for speed regulation without consideration of load mitigation.
abovementioned (below-rated, rated, above-rated) wind elds, and
compared to the baseline GSPI controller. The parameters of the
controllers involved in the case studies are summarized in Table 2.
In order to compare these three different controllers, four perfor- 4.2.2. Case 2 results, under turbulent 12 m/s wind led
mance indices are analyzed: Root Mean Square (RMS) error of For the situation where the wind speed is below the operating
generator speed, mean power average, blade root apwise moment point, the turbulent wind eld with mean wind speed 12 m/s is
DEL, and maximum apwise moment. Here DEL refers to the fa- chosen as Case 2. This wind le covers the switch region between
tigue damage equivalent load (DEL), which serves as an important Region 2 and Region 3 (since the rated wind speed is 11.4 m/s for
metric for comparing fatigue loads across the entire spectrum of this 5-MW turbine). The parameters of DAC controller and adaptive
turbulent wind les. The equivalent fatigue damage is represented controller are kept the same as those used in Case 1. The wind le,
by a constant load and calculated by MLife [36]. MLife is a post- generator speed, pitch angle and blade root apwise moment are
processing code developed by NREL to assess fatigue according to shown in Fig. 6(a)e6(d), respectively. GSPI yields the highest ac-
rainow counting algorithm [37]. In high-cycle fatigue situations, tivity of the pitch actuator, as shown in Fig. 6(c). The overall value of
materials performance is usually characterized by an S-N curve (i.e., apwise root moment is higher under GSPI than those under DAC
Wo hler curve). Here the S-N slope of 10 is used which is typical for and adaptive controller, as shown in Fig. 6(d). Therefore, when
composite materials (blade). pitch angle is adjusted rapidly to maintain constant generator
speed, it will cause large oscillation in blade apwise deection.
The normalized performances are compared in Fig. 7. Similar to
4.2.1. Case 1 results, under turbulent 15 m/s wind eld DAC, the adaptive controller reduces the apwise moment DEL by
Case 1 concerns the results under turbulent 15 m/s wind eld. 38% as compared to the baseline GSPI. However, as a trade-off, the
15 m/s is the mean speed of wind distribution from 0 to 600 s. The power output exhibits a decrease of 36% (under DAC) and 25%
wind le, generator speed, pitch angle and blade root apwise (under adaptive control), respectively, as compared to that of GSPI.
moment are shown in Fig. 4(a)e4(d), respectively. From Fig. 4(b) we The normalized generator speed errors of three controllers are
can see that there is less uctuation of generator speed under 0.1822, 0.5740, and 0.3316, respectively. The maximum apwise
adaptive control than that under DAC. The pitch angle change rate moments are also reduced by 10% under DAC and 9% under adap-
is smaller under adaptive control than that under DAC or GSPI, as tive control.
490 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

Fig. 4. Time-domain performance comparison of GSPI, DAC, and adaptive control under 15 m/s turbulent eld. (a) Wind speed; (b) Generator speed; (c) Pitch angle; and (d)
Flapwise root moment.

values of three controllers are nearly the same while the extreme
moment is decreased by 5% under DAC and adaptive control. This
can be ascribed to the fact that the mean wind speed in Case 3 is far
away from the operating point. GSPI still shows better ability to
regulate the generator speed, and the RMS errors are 0.1964,
0.6513, and 0.3217, respectively. Because of the larger wind speed,
the average power output increases by 15% (under DAC) and 6%
(under adaptive), respectively, compared to that of GSPI. Therefore,
in the above-rated region, the load mitigation performance of both
DAC and adaptive control is still effective, and the adaptive control
leads to better speed regulation than DAC as well as increased
average power than both GSPI and DAC.

4.3. Uncertainties analysis

The performance improvement and robustness of the adaptive


controller are demonstrated in the preceding section under
different turbulent wind elds. Here we will further investigate the
Fig. 5. Performance comparison of GSPI, DAC and adaptive control under 15 m/s tur-
bulent eld.
effect of inherent structural uncertainty of wind turbine to the
controllers. Wind turbine operates in a very complicated environ-
ment with many sources of uncertainties such as potential blade
4.2.3. Case 3 results, under turbulent 17 m/s wind eld damage and mass imbalance caused by rain and ice. Uncertainties
In Case 3 we consider turbulent 17 m/s wind eld. The wind le, may also be caused by manufacturing tolerance and in-service
generator speed, pitch angle and blade root apwise moment are degradation. The benchmark 5-MW wind turbine blade has 49 el-
shown in Fig. 8(a)e8(d), respectively. The curve of generator speed ements with individual structural and aerodynamic properties
under adaptive control is generally lower than the curve under DAC dened by NREL [29]. Each element has different blade properties,
and higher than the curve under GSPI, as shown in Fig. 8(b). The i.e. aerodynamic center, structural twist angle, mass density, ap-
pitch angle curve under adaptive controller is very close to that wise and edgewise stiffness. In this research Latin Hypercube
under DAC, as shown in Fig. 8(c). The apwise root moment trends Sampling is adopted to generate blade uncertainty which is used to
are shown in Fig. 8(d), where three curves generally overlap. The assess the controller performances. For n set of input stiffness
normalized performance comparison is given in Fig. 9. When the values of blade elements DK DK1 ; DK2 ; DKn , the distribution
wind speed is above the operating point, the apwise moment DEL of the corresponding inputs is assumed to be a multivariate
Gaussian distribution. The mean function mDKi and the
Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494 491

Fig. 6. Time-domain performance comparison of GSPI, DAC, and adaptive control under 12 m/s turbulent eld. (a) Wind speed; (b) Generator speed; (c) Pitch angle; and (d)
Flapwise root moment.

In GSPI control, the gain scheduled part is only designed to deal


with the varying operating point and cannot deal with model un-
certainty. The performance of GSPI will then be sensitive to state
matrix change. In DAC approach, the wind disturbance Gd is not
affected by the change of state matrix A while the response vector
will be affected. Theoretically, the initial G (without the blade
stiffness uncertainty) is not able to maintain the exact same per-
formance of speed regulation and load mitigation. However, the
gain G also possesses robustness. In practice, we still adopt the
initial G to obtain the reference signal and examine the robustness
of the DAC and the adaptive controller under structural uncertainty.
The results obtained under GSPI, DAC and adaptive control with
the blade stiffness uncertainty/reduction are shown in Fig. 10,
where box whisker plots for 20 different blade stiffness les are
provided. The apwise moment DELs of both DAC and adaptive
control exhibit a signicant decrease compared to those of GSPI.
The ap DEL is decreased by 20%e28% at 15 m/s. Similar trend of
DEL decrease of 21%e23% and 21%e31% are observed at 12 m/s and
Fig. 7. Performance comparison of GSPI, DAC and adaptive control under 12 m/s tur-
17 m/s respectively under DAC and adaptive control. We can then
bulent eld.
conclude that GSPI is more sensitive to model uncertainty than DAC
and adaptive control. In nominal condition without uncertainty, a
P reduction of 9% in ap DEL is achieved by both DAC and adaptive
covariance function DKi ; DKj are pre-specied. In addition, we control as compared to GSPI. With uncertainty, the ap DEL re-
assume DKi and DKj are independent of each other. Here we let the ductions by DAC and adaptive control have been increased to over
mean value of the stiffness uncertainty be 0:05 (i.e., 5% mean 20% than GSPI. For generator regulation, we can see from the 20
reduction, possibly caused by minor damage), and let the variance sets of results that DAC shows the largest variation and the
be 105 . 20 samples of stiffness uncertainty are generated and generator speed error of adaptive control is smaller than that of
injected into the analysis. For each sample, FAST linearization is GSPI. GSPI still exhibits the best speed regulation performance,
~ which is different from
utilized to obtain the new stiffness matrix K, primarily because it only tracks the reference value and does not
the original one. One example is shown in Appendix. The mass contribute to load mitigation. When uncertainty is present, DAC
matrix M,~ the damping matrix C, ~ the wind input matrix F ~ and the and adaptive control lead to much better performance as compared
~
wind input disturbance matrix Fd all remain to be very close to the to GSPI. Also, we can observe that all the absolute values of ap DEL
nominal values. The state matrix is changed to A, ~ while other with blade stiffness uncertainty are increased compared to those in
matrices B, Bd and C remain unchanged.
492 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

Fig. 8. Time-domain performance comparison of GSPI, DAC, and adaptive control under 17 m/s turbulent eld. (a) Wind speed; (b) Generator speed; (c) Pitch angle; and (d)
Flapwise root moment.

increased when the mean wind speed changes from 12 m/s to 17 m/


s and thus the larger wind speed will lead to increased ap DEL.

5. Concluding remarks

In this research an adaptive control with disturbance rejection


and load mitigation capability is studied, aiming at improving the
life span of wind turbines operating in high wind speeds. The
proposed algorithm is applied to the NREL offshore 5-MW baseline
wind turbine. It is identied that the load mitigation performances
of the adaptive control and DAC are comparable which are both
better than the baseline GSPI. Furthermore, for generator speed
regulation, adaptive control performs better than DAC. With the
possible existence of structural uncertainty such as stiffness
reduction in wind turbine blades, GSPI will be sensitive to model
Fig. 9. Performance comparison of GSPI, DAC and adaptive control under 17 m/s tur- uncertainty. The adaptive control, on the other hand, can deal with
bulent eld.
such uncertainty. The results indicate that adaptive control can
both enhance the load mitigation over GSPI and improve speed
regulation over DAC. The formulation developed in this research is
the nominal condition. On the other hand, the stiffness uncertainty
general, and can be extended to mitigating loads on other
does not affect the generator speed error. The ap DEL values are
components.
Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494 493

Fig. 10. A summary of the apwise moment DEL and generator speed RMS error for 20 blade stiffness reduction sets under 12 m/s, 15 m/s and 17 m/s turbulent elds.

Acknowledgment References

This research is supported by National Science Foundation un- [1] L.Y. Pao, K.E. Johnson, A tutorial on the dynamics and control of wind turbines
and wind farms, in: American Control Conference, St. Louis, MO, June 10-12,
der grant CMMI e 1300236. 2009, pp. 2076e2089.
[2] J.M. Jonkman, S. Buttereld, W. Musial, G. Scott, Denition of a 5-MW
Reference Wind Turbine for Offshore System Development, Technical Report
No. NREL/TP-500e38060, National Renewable Energy Laboratory, Golden, CO,
Appendix 2009.
[3] A.D. Wright, Modern Control Design for Flexible Wind Turbines, Technical
Report No. NREL/TP-500e35816, National Renewable Energy Laboratory,
K refers to the nominal stiffness matrix in the reduced-order Golden, CO, 2004.
model without uncertainty. K~ refers to the stiffness matrix under [4] A.D. Wright, M.J. Balas, Design of state-space-based control algorithms for
uncertainty. In this research we let the mean stiffness change wind turbine speed regulation, J. Sol. Energy Eng. 125 (4) (2003) 386e395.

be 0.05 and the variance be 105 . K and an example of K ~ are [5] S.A. Frost, M.J. Balas, A.D. Wright, Direct adaptive control of a utility-scale
wind turbine for speed regulation, Int. J. Robust Nonlinear Control 19 (1)
shown below for comparison, i.e., (2009) 59e71.
[6] B. Connor, S.N. Iyer, W.E. Leithead, Control of a horizontal axis wind turbine
2 3 using H innity control, Control Applications, in: First IEEE Conference on,
334:77 343:22 932:55 932:81 932:73 Dayton, OH, September 13e16, 1992, pp. 117e122.
6 334:77 8:68  108 932:55 932:81 932:77 7
6 7 [7] K.A. Stol, M.J. Balas, Periodic disturbance accommodating control for blade
K6
6 27:06 27:12 1:93  104
0:01 0:0363 7
7 load mitigation in wind turbines, J. Sol. Energy Eng. 125 (4) (2003) 379e385.
4 26:89 26:75 0:0511 1:93  104 0:02 5 [8] M.M. Hand, M.J. Balas, Blade load mitigation control design for a wind turbine
4 operating in the path of vortices, Wind Energy 10 (4) (2007) 339e355.
27:22 27:03 0:0202 0:05 1:93  10 [9] M. Soliman, O.P. Malik, D.T. Westwick, Multiple model MIMO predictive
control for variable speed variable pitch wind turbines, in: American Control
Conference, Baltimore, MD, June 30- July 02, 2010, pp. 2778e2784.
and [10] S. Bououden, M. Chadli, S. Filali, A. El Hajjaji, Fuzzy model based multivariable
predictive control of a variable speed wind turbine: LMI approach, Renew.
2 3 Energy 37 (1) (2012) 434e439.
302:87 17:72 1:20  103 1:20  103 1:20  103
6 302:87 8:68  108 1:20  103 1:20  103 37
1:20  10 7
[11] H. Sanchez, T. Escobet, V. Puig, P.F. Odgaard, Health-aware model predictive
6 control of wind turbines using fatigue prognosis, IFAC PapersOnLine 48 (21)
~ 6 28:20
K 28:48 1:85  104 2:40  103 0:05 7
6 7 (2015) 1363e1368.
4 28:40 28:69 0:0267 1:85  104 1:8  103 5 [12] P.F. Odgaard, L.F. Larsen, R. Wisniewski, T.G. Hovgaard, On using Pareto
optimality to tune a linear model predictive controller for wind turbines,
28:50 28:66 0:0202 0:03 1:85  104 Renew. Energy 87 (2016) 884e891.
[13] P.F. Odgaard, On usage of pareto curves to select wind turbine controller
tunings to the wind turbulence level, in: European Control Conference, Linz,
494 Y. Yuan, J. Tang / Renewable Energy 105 (2017) 483e494

Austria, July 15-17 July, 2015, pp. 1534e1539. operation, in: 49th AIAA Aerospace Sciences Meeting Including the New
[14] P.F. Odgaaard, T. Knudsen, A. Overgaard, H. Steffensen, M. Jrgensen, Impor- Horizons Forum and Aerospace Exposition, Orlando, FL, January 4e7, 2011.
tance of dynamic inow in model predictive control of wind turbines, IFAC [26] K.T. Magar, M.J. Balas, S.A. Frost, Direct adaptive torque control for maximizing
PapersOnLine 48 (30) (2015) 90e95. the power captured by wind turbine in partial loading condition, Wind Energy
[15] E. Bossanyi, Individual blade pitch control for load reduction, Wind Energy 6 19 (5) (2016) 911e922.
(2) (2003) 119e128. [28] P.J. Moriarty, A.C. Hansen, AeroDyn Theory Manual, Technical Report No.
[16] E. Bossanyi, Further load reductions with individual pitch control, Wind En- NREL/TP-500e36881, National Renewable Energy Laboratory, Golden, CO,
ergy 8 (4) (2005) 481e485. 2005.
[17] K.A. Stol, W. Zhao, A.D. Wright, Individual blade pitch control for the controls [29] J.M. Jonkman, M.L. Buhl Jr., FAST Users Guide, Technical Report No. NREL/EL-
advanced research turbine (CART), J. Sol. Energy Eng. 128 (4) (2006) 498e505. 500-38230, National Renewable Energy Laboratory, Golden, CO, 2005.
[18] H.M. Hassan, A.L. ElShafei, W.A. Farag, M.S. Saad, A robust LMI-based pitch [31] C. Johnson, Theory of disturbance-accommodating controllers, Control Dyn.
controller for large wind turbines, Renew. Energy 44 (2012) 63e71. Syst. 12 (1976) 387e489.
[19] C.L. Bottasso, P. Pizzinelli, C.E.D. Riboldi, L. Tasca, LiDAR-enabled model pre- [32] M.J. Balas, Y.J. Lee, L. Kendall, Disturbance tracking control theory with
dictive control of wind turbines with real-time capabilities, Renew. Energy 71 application to horizontal axis wind turbines, in: Proceedings of 1998 ASME
(2014) 442e452. Wind Energy Symposium, Reno, NV, January 12-15, 1998, pp. 95e99.
[20] F. Dunne, L.Y. Pao, A.D. Wright, B. Jonkman, N. Kelley, Adding feedforward [33] J.T.Y. Wen, M.J. Balas, Robust adaptive control in Hilbert space, J. Math. Anal.
blade pitch control to standard feedback controllers for load mitigation in Appl. 143 (1) (1989) 1e26.
wind turbines, Mechatronics 21 (4) (2011) 682e690. [34] Y. Lee, M.J. Balas, Adaptive control of persistent disturbance based on direct
[21] N. Wang, K.E. Johnson, A.D. Wright, FX-RLS-based feedforward control for model reference adaptive control, in: 36th Structures, Structural Dynamics
LIDAR-enabled wind turbine load mitigation, IEEE Trans. Control Syst. Tech- and Materials Conference, New Orleans, LA, April 10-13, 1995, pp.
nol. 20 (5) (2012) 1212e1222. 3466e3473.
[22] V. Petrovic, M. Jelavi
c, M. Baoti
c, Advanced control algorithms for reduction of [35] J.M. Jonkman, TurbSim Users Guide: Version 1.50, Technical Report No. NREL/
wind turbine structural loads, Renew. Energy 76 (2015) 418e431. TP-500-46198, National Renewable Energy Laboratory, Golden, CO, 2009.
[23] A.D. Wright, L.J. Fingersh, Advanced Control Design for Wind Turbines. Part I: [36] G.J. Hayman, M. Buhl Jr., Mlife Users Guide for Version 1.00, Technical Report,
Control Design, Implementation, and Initial Tests, Technical Report No. NREL/ National Renewable Energy Laboratory, Golden, CO, 2012.
TP-500e42437, National Renewable Energy Laboratory, Golden, CO, 2008. [37] G. Hayman, MLife Theory Manual Version 1.00, Technical Report, National
[24] K.E. Johnson, L.J. Fingersh, Adaptive pitch control of variable-speed wind Renewable Energy Laboratory, Golden, CO, 2012.
turbines, J. Sol. Energy Eng. 130 (3) (2008) 031012. [38] H. Kwakernaak, R. Sivan, Linear Optimal Control Systems, Wiley-Interscience,
[25] M. Balas, K.T. Magar, Q. Li, Adaptive disturbance tracking control for large New York, NY, 1972.
horizontal axis wind turbines with disturbance estimator in region ii