Linear Quadratic Gaussian (LQG) Control of Wind

Turbines
Abdulrahman Kalbat
Electrical Engineering Department
Columbia University in the City of New York
New York, NY 10027
Email: ak3369@columbia.edu

Abstract—With the increasing deployment of wind energy presented and the limitations of the LQG controller will be
technologies, innovative modern control theories, which utilize discussed.
linear time invariant models, are being developed. The potential
for the success of utilizing these theories is better with the II. LQG G ENERAL F ORM
availability of concepts such as the variable speed wind trubines.
Linear Quadratic Gaussian (LQG) design problem is rooted
Such control theories are not only capable of increasing the
energy capture efficiency of the wind turbines, but also reducing in optimal stochastic control theory and has many applications
the number of measurements needed which in turn reduces the in the modern world which ranges from flight and missile
operation and maintenance costs associated with the system. navigation control systems, medical porocesses controllers
In this paper, the general formulation of the Linear Quadratic and even nuclear power plants. It combines both concepts of
Gaussian Control (LQG) along with the design procedures and
Linear Quadratic Regulators (LQR) for full state feedback and
the assumptions needed will be stated. Then, a numerical model
for the Controls Advacned Research Turbine (CART) that was Kalman Filters for state estimation. In this section, the state
created and field tested by the National Renewable Energy space equations defining the LQG regulator along with the
Laboratory (NREL) will be used to simulate LQG regulator statistical assumptions will be presented first. Then, the design
using MATLab script. At the end, the results obtained from steps for finding the optimal state feedback gain (Kf ) and
the simulation will be presented and the limitations of the LQG
optimal state estimation gain (Kk ) will be listed. Finally, the
controller will be discussed.
Index Terms—Linear Quadratic Gaussian, LQG, Kalman Fil- state space equations of the augmented system which closes
ter, Estimation, Wind Turbine. the loop will be presented.
The state space equations of the open loop plant for a
standard LQG problem is shown in 1.
I. I NTRODCUTION
The future plans for smart grid deployment , which mainly ẋ(t) = Ax(t) + Bu(t) + Gw(t)
(1)
depend on the concept of distributed generation, make wind y(t) = Cx(t) + v(t)
turbines more attractive solution, especially in regions with the
high average wind speed. The control methods of wind tur- where x(t) is the state vector, u(t) is the control input
bines started by PI controllers which imposed many limitations vector, y(t) is the measured output vector, w(t) and v(t) are
and operation and maintenance costs to the system. With the stochastic white noise processes assoctiated with the process
advancement of modern control theory, it is now possible to and the measurement, respectively. The matrices A(state ma-
use linear state space models of the wind turbines to implement trix), B(control input gain matrix), G(plant noise gain matrix),
control systems which are more powerful than PI controllers. C(measured state matrix) are all Linear Time Invariant (LTI).
These modern control methods also reduced the maintenance As shown in 2, both w(t) and v(t) are assumed to be white
and operation cost by the utilization of Kalman Filters for state gaussian noises with zero mean and the expected values of
estimation that reduced the number of the sensors necessary the initial values of the states x(0) are assumed to be equal to
to implement full state feedback controllers. x̂o . From 3, w(t) and v(t) have covariance matrices of W and
V, respectively, and the cross covariance R12 is assumed to
In section II of this paper, the general formulation of the
be zero (w(t) and v(t) are uncorrelated) in order to simplify
Linear Quadratic Gaussian Control (LQG) along with the de-
many expressions and derivations. Finally, the initial values of
sign procedures and the assumptions needed will be stated. In
the states x(0) are assumed to be uncorrelated with both w(t)
section III, National Renewable Energy Laboratory’s (NREL)
and v(t), as shown in 4.
Controls Advacned Research Turbine (CART) features and
capabilities will be presented. In section IV, a numerical model
for CART that was created and field tested by NREL will E[x(0)] = x̂o
be used to simulate LQG regulator using MATLab script. In E[w(t)] = 0 (2)
section V, the results obtained from the simulation will be E[v(t)] = 0
 engineering sense, or by using the method shown in 9 or by
T

W, if t = τ using Bryson’s Rule shown in 10. Bryson’s Rule related the
E[w(t)w(τ ) ] = reciprocal of the maximum squared values of the states with
0, if t 6= τ

V, if t = τ Qf and the reciprocal of the maximum squared values of the
T
E[v(t)v(τ ) ] = (3) control inputs with Rf .
0, if t 6= τ
 After finding the optimal gains, the closed loop system
T R12 , if t = τ
E[w(t)v(τ ) ] = could be created by augmenting the system with Kalman
0, if t 6= τ Filter resulting in the new model shown in 11. The augmented
system model which include the dynamics of the system and
T
E[x(0)w(t) ] = 0 the state estimation error e(t) [2].
T
(4)
E[x(0)v(t) ] = 0
Step 1: Optimal gain existance criteria:
In real world control design problems, it is rarely possible
to have access to all states of the system which are needed for (A, B) is Controllable
full state feedback. Instead, access is only possible to specific (6)
(A, C) is Observable
measured outputs of the system. If these measurements carry
enough information about the states of the system, then a
Step 2: Optimal State Estimation Gain Calculation:
state observer using Kalman Filter could be implemented to
estimate all states of the system. This observer is capable of
Jk = E (x − x̂)T (x − x̂)

(7a)
rejecting disturbances of the system by acting as a low pass
T T T −1
filter. The main inputs to the observer are the control input APk + Pk A + GW G − Pk C V CPk = 0 (7b)
T −1
(u(t)) and the system output (y(t)). The state space equations Kk = Pk C V (7c)
of the Kalman Filter are shown in 5. It should be noticed that
it uses the same state space matrices (A, B and C) as the main Step 3: Optimal State Feedback Gain Calculation and
system and the estimated states (x̂(t)) are used as the system Weighing Matrices Selection:
states [1].
Z T
Jf = (z T Qf z + uT Rf u)dt (8a)
˙
x̂(t) = (A − Kk C)x̂(t) + Bu(t) + Kk y(t) 0
(5)
ŷ(t) = C x̂(t) AT Pf + Pf A − Pf BRf−1 B T Pf + Qf = 0 (8b)
From Fig.1, it should be noticed that LQG is formed by Kf = Rf−1 B T Pf (8c)
connecting the system and the Kalman Filter through the
optimal state estimation gain (Kk ) and then creating full
state feedback by using the estimated states (x̂(t)) which
passed through the optimal feedback gain (Kf ). Because of Q = CT C
(9)
the stochastic seperation principle, the previoulsy mentioned R = ρI
gain could be designed individually.
The design process starts with checking controllability and
observability of the pairs (A, B) and (A, C), respectively, as 1
Qii =
shown in 6. These criterias are necessary for the existance of M ax (x2ii )
the solutions for the equations used to find the optimal gains. (10)
1
Then, the optimal state estimation gain (Kk ) is calculated as Rii =
M ax (u2ii )
shown in 7c, where Pk is a positive semi-definite mtarix and
the solution of the Filter Algebraic Riccati Equation (FARE) Step 4: Linear Quadratic Gaussian Regulator by combining
shown in 7b. This solution ensures a minimum value of Optimal State Estimation and Optimal State Feedback:
the cost function shown in 7a. After that, the optimal state
feedback gain (Kf ) is calculated as shown in 8c, where Pf is     
a positive semi-definite matrix and the solution of the Control ẋ(t) A − BKf BKf x(t)
=
Algebraic Riccati Equation (CARE) shown in 8b. This solution ė(t) 0 A − Kk C e(t)
ensures a minimum value of the cost function shown in 8a.
  
G 0 w(t)
Unlike FARE, which requires only noise covariances W +
G −Kk v(t) (11)
and V, CARE requires two weighing matrices: Qf (symmetric
positive semi-definite) and Rf (symmetric positive definite).    
 x(t)  w(t)
These weighing matrices provide a means to trade-off op-
 
y(t) = C 0 + 0 1
posing objectives: state regulation and control usage. They e(t) v(t)
could be selected based on trial and error approach with some
 w(t) Noisy System implement the individual pitching technique. In the collective
pitching technique each of the blades are picthed by the
same amount and this controlls the symmetric loading on the
v(t)
rotor. But in the individual pitching technique, each blade is
G individually pitched and this offers a control over both the
symmetric and asymmetric loading on the rotor [6]. Because
of the access to all sensors data and actuators and the flexible
u(t) + + ẋ(t) R x(t) + y(t) capabilities of CART, researchers at NWTC could implement
B C +
+ infinite combinations of control systems.
IV. N UMERICAL E XAMPLE U SING MATLAB
A In [7], several Linear Time Invariant (LTI) state space
+ models of CART were created which varied in their complex-
Kk ities between 1-state model to 9-state model. In this paper,
y(t) − ŷ(t) −
the 3-state model was considered because of its simplicity
+ ˙
x̂(t) x̂(t) ŷ(t) and because this model is both controllable and observable
+ R
B C (minimal) which are both conditions needed for designing
+
LQG controllers. The numerical model shown below was
created for CART by NWTC after linearizing the motion
A equations of the turbine’s considered states at a control design
Kalman-Bucy Filter point of 18 m/s for wind speed, 12 degress for rotor collective
Kf
pitch and 42 RPM for rotor speed. The main objective of the
controller is to operate the machine as a variable speed wind
Determinisric Optimal Controller
turbine in region 3 by applying constant torque to the generator
Fig. 1. LQG regulator block diagram
by maintaining a constant rotor speed through the collective
rotor blade pitching.
The control input u(t) of the considered model is the
III. C ONTROLS A DVACNED R ESEARCH T URBINE (CART) collective blade pitch angle and the states are the rotor speed
x1 (t), drive train (shaft) torsional spring force x2 (t) and the
The numerical example that will be discussed in the next generator speed x3 (t). The disturbances are the turbine system
section is based on a linear model of the Controls Advacned noise w(t) and the measurement noise v(t). For the sake of
Research Turbine (CART) which is used by the National simplicity, both turbine and measurement noises were assumed
Wind Technology Center (NWTC) that is operated by the to have the same magnitude as the wind but in the case of
National Renewable Energy Laboratory (NREL) and located turbine system noise, it was scaled by the matrix G. The
at Boulder, Colorado. Researchers at NWTC tackles wind in- non-zero first row of the control input gain B indicates that
dustry engineering challenges ranging from atmospheric fluid the control input (rotor collective pitch) affects the system by
mechanics and aerodynamics, dynamics, structures, fatigue, changing the rotor speed which affetcs the other states through
power systems and electronics and wind turbine engineering the coupling caused by the state gain matrix A. Also, from
applications [3]. matrix C, it should be noticed that the only measurement taken
CART is a modified Westinghouse WGT-600, 600-KW, is the generator speed.
2-bladed, horizontal-axis research wind turbine capable of The simualtion of LQG controller for CART was done using
operating in constant speed or variable speed mode [4]. It is MATLab by following the design steps shown in section II
used by the researchers at NREL’s NWTC as a test bed for and by using the numerical model’s matrices shown below.
exploring potential control innovations and field test advanced The design started by checking the controllability and the
control systems. These control designs could be implemented observability of the pairs (A, B) and (A, C), respectively.
by writing an ANSI-C based program on a dedicated personal Then, the optimal state estimation gain (Kk ) was calculated
computer that is connected to the turbine sensors and actuators. using the noise covariances shown below. Sizing the optimal
The turbine is instrumneted with more than 80 sensors with state feedback gain was found using a trial and error approach
measurements being recorded at a rate of 100 Hz. These by changing the weighing matrices (Qf and Rf ). The best
sensors are used to record the dynamic state of the turbine response of the generator speed at which it stabilized at steady
and they fall in one of three categories: Performance (torque state when perturbed by step changes in wind speed was found
and power sensors), loads (strain gauges, accelerometers and at the values of the weighing matrices shown below. If the
position sensors) or meteorological (wind speed, direction, maximum limits of all states and control inputs were available,
temperature and pressure) [5]. which was not the case for CART, then Bryson’s rule could
One of the unique capabilities of CART is that beside the be used to initialize the weighting matrices as shown below
conventional collective pitching technique, it is possible to [8]. After that, the closed loop system of the LQG regulator
 Wind Speed (m/sec)
was created by augmenting the turbine system with the state
20
estimator. The inputs to the closed loop regulating system
are turbine system noise w(t) and the measurement noise 15
0 10 20 30 40 50 60 70
v(t) which were generated as a normally distributed random
Generator Speed (RPM)
number by MATLab to represent the white gaussian noise. 50
The peroformance of the closed loop LQG system will be 40
30
discussed next. 20
0 10 20 30 40 50 60 70

Pitch Angle (degrees)

MATLab Simulation inputs for CART 15
  
x1 (t)  x1 (t) : is rotor speed 10

x(t) = x2 (t) where x2 (t) : is drive train torsion 5

0 10 20 30 40 50 60 70
Time sec
x3 (t) x3 (t) : is generator speed


Fig. 2. Wind turbine’s response to different wind speeds

  
w(t) w(t) : is turbine system noise Pole−Zero Map

u(t) = where 25
v(t) v(t) : is measurement noise
20

15

−1.4454x10−1 −3.1078x10−6
 
0.0 10
(12)
A= 2.6910x107 0.0 −2.6910x107 

Imaginary Axis
 5

0.0 1.5601x10−5 0.0 0

 T
B = −3.4559 0.0 0.0 −5

  −10
C= 0 0 1
−15
T
G = 7.8938x10−2 0.0 0.0

−20

T −25
W = E[w(t)w(τ ) ] = 0.1 (Turbine system noise covariance) −4.5 −4 −3.5 −3 −2.5 −2
Real Axis
−1.5 −1 −0.5 0

T
V = E[v(t)v(τ ) ] = 0.1 (Measurement noise covariance)
Fig. 3. Poles/Zeros plot of the closed loop system

Results reason might be models’ simplicity and uncertainty because


1 0 0
 of the neglected high frequecny dynamics of the turbine. In
Q = 0 1x10−13 0 R = 1 order to improve the robustness of the LQG regulator, one
0 0 1 could go further and design a Loop Transfer Recovery (LTR)
 −3 2
 (13)
−2 T
which recovers the robustness of Linear Quadratic Regulators
Kk = 7.6282x10 1.2663x10 6.2859x10 (LQR) and Kalman Filters.
Kf = −2.0336 −2.1225x10−7 6.6055x10−1
 
Fig. 3 shows the pole/zero map of the closed loop system.
Since all poles lies in the left half plane, then the closed
loop system is stable. From Fig. 4, one should notice how the
V. R ESULTS AND D ISCUSSION LQG regulator successfully reduced the effect of the system
Fig. 2 shows the generator speed and the pitch angle when white gaussian noise w(t) (input 1) on the closed loop system.
the wind speed is varying between 14 m/s to 20 m/s. As the Another observation is that the LQG at high frequencies the
wind speed is increasing, the pitch angle is also increasing to effect of the measurement white gaussian noise v(t) is higher
keep the generator speed at a constnt value of 42 RPM. From in the closed loop system than in open loop system. One
the figure, it is clear that the generator is operating at 42 RPM explanation might be because the v(t) is being multiplied by
only at 18 m/s and when the pitch angle is 12 degrees, the the optimal state estimation gain Kk in the final augmented
control design point. So, LQG regulator is not able to keep model shown in 11.
the generator speed at 42 RPM even when the wind speed is
varying. This example clearly shows the robustness problem of VI. C ONCLUSION
LQG controllers. According to [9] , control system robustness In this paper, the general formulation of the Linear
is defined as the ability to maintain satisfactory stability or Quadratic Gaussian Control (LQG) along with the design
performance characteristics in the presence of all conceivable procedures and the assumptions needed were stated. Then, a
system parameter variations. In other words,a robust control numerical model for the Controls Advacned Research Turbine
system works not only for the linear system which serves (CART) was used to simulate LQG controller using MATLab
as the plant model but it also works for the real physical script. Then, the peroformance of the LQG regulator was
system with minor performance degradation [10]. Another evaluated by plotting the time response of the system and the
 Closed loop)

From: In(1) From: In(2)

50

To: Out(1)
−50
Magnitude (dB) ; Phase (deg)

−100

−150
180
Open
90 closed
To: Out(1)

−90

−180

−270
0 0
10 10
Frequency (rad/sec)

Fig. 4. bode plot of the open and closed loop systems

pole/zero map of the closed loop system. Finally, the bode

plot of both the open loop and the closed loop systems were
compared.
R EFERENCES
[1] C.-H. Won and K. Gunaratne, “Performance study of lqg, mcv, and risk-
sensitive control methods for satellite structure control,” in American
Control Conference, 2002. Proceedings of the 2002, vol. 3, pp. 2481–
2486 vol.3.
[2] E. E. Ostertag, “Optimal stochastic control,” in Mono- and Multivariable
Control and Estimation, ser. Mathematical Engineering. Springer
Berlin Heidelberg, 2011, vol. 2, pp. 259–266. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-13734-1-5
[3] National wind technology center (nwtc). [Online]. Available:
http://www.nrel.gov/wind/pdfs/53405.pdf
[4] K. Johnson, L. Fingersh, and A. Wright, “Controls advanced research
turbine: Lessons learned during advanced controls testing,” National
Renewable Energy Laboratory, NREL/TP 500-38130, June 2005.
[5] L. Fingersh and K. Johnson, “Controls advanced research turbine
(cart) commissioning and baseline data collection,” National Renewable
Energy Laboratory, NREL/TP 500-32879, October 2002.
[6] K. A. Stol, W. Zhao, and A. D. Wright, “Individual blade pitch control
for the controls advanced research turbine (cart),” Journal of solar
energy engineering, vol. 128, no. 4, pp. 498 – 505, 2006.
[7] A. D. Wright, “Modern control design for flexible wind turbines,”
National Renewable Energy Laboratory, NREL/TP 500-35816, July
2004.
[8] . Oral, L. etin, and E. Uyar, “A novel method on selection of q and
r matrices in the theory of optimal control,” International Journal of
Systems Control, vol. 1, no. 2, pp. 84–92, 2010.
[9] R. Stengel and L. Ryan, “Stochastic robustness of linear time-invariant
control systems,” Automatic Control, IEEE Transactions on, vol. 36,
no. 1, pp. 82–87, 1991.
[10] U. Mackenroth, Control Systems: Basic Definitions and Concepts.
Springer Berlin Heidelberg, 2004, pp. 4–5.