A gain-scheduled multivariable LQR controller for

Permanent Magnet Synchronous Motor

Michał Brasel
Faculty of Electrical Engineering
West Pomeranian University of Technology, Szczecin
Szczecin, Poland
[email protected]

Abstract — The paper presents a gain-scheduled LQR control nonlinear plant. This approach can result in a switched
system for a nonlinear model of permanent magnet synchronous structure with a cluster of linear controllers [7, 8] or one linear
motor (PMSM). The most popular cascade FOC (Field Oriented adaptive controller with stepwise varying parameters [9, 10,
Control) structure including a few single-loop PI control systems 11].
is replaced with a single multivariable state feedback controller.
An optimal LQR controller designed for a plant, which is
Due to a nonlinear nature of the PMSM dynamics equations the
developed gain-scheduled controller is tuned in relation to the
linearized at the selected operating point, only provides optimal
changeable operating conditions of the motor. In order to ensure control in the immediate neighborhood of this point. Therefore,
zero steady-state speed error the system synthesis is carried out in the case of nonlinear plant or various operating point of the
by means of the LQ optimal control method with state vector plant on-line linearization around the current operating point is
augmentation. The presented simulation results show that the use required. The paper presents a gain-scheduled LQR controller
of gain-scheduled controller may improve control quality. tuned in accordance with changing PMSM operating points in
comparison with a LQR controller with fixed parameters
Keywords — Multivariable Control Systems; Nonlinear designed for a chosen operating point of the motor.
Systems; Gain-scheduled LQR controller; Permanent Magnet
Synchronous Motor II. NONLINEAR STATE SPACE MODEL OF PMSM
A nonlinear dynamic model of a PMSM in the d-q
I. INTRODUCTION
transformed rotor reference frame can be presented in the state-
In general, permanent magnet synchronous machines are space form as follows:
nonlinear multivariable dynamic plants. Due to nonlinear
dynamic interrelations occurring among mechanical and ⎧d Rs Lq 1
electromagnetic quantities it is useful to describe a permanent ⎪ id = − id + pωiq + ud
⎪ dt L L L
magnet synchronous motor in the state space form as d d d

a nonlinear MIMO (Multi-Input Multi-Output) dynamic ⎪⎪ d Rs Ld pωψ f 1

system. For implementation of PMSM vector control the ⎨ iq = − iq − pω id − + uq (1)
⎪ dt Lq Lq Lq L q
cascade FOC control structures are still the most commonly
⎪d
used. The single-loop control systems in the FOC structure
synthesized independently does not seem to take into account
⎪ ω=
⎪⎩ dt
3p
2J
( ( ) ) fc
ψ f iq + Ld − Lq id iq − ω − TL
J
1
J
a multivariable nature of the plant.
where: id, iq and ud, uq are currents and voltages in d and q axis
Multivariable state feedback control is an effective alternative respectively, Ld, Lq are inductances in d and q axes, Rs is
for a popular cascade FOC structure including conventional PI resistance of the stator, ψf is permanent magnetic flux linkage,
controllers. In [1] performance of the LQR control system was p is number of pole pairs, ω is rotor angular speed, J is rotor
compared with the classical cascade control structure. The inertia, fc is viscous friction coefficient, and TL is torque load.
disadvantages of using the state feedback LQR control method These equations can be written in a vector form as follows:
mentioned in [1] can be eliminated through state vector
augmentation by means of additional integral state variables ⎪⎧ x = f ( x, u, TL )
[2, 3, 4]. Primarily the classical state feedback control ⎨ (2)
⎪⎩ y = Cx
methods are based on plant linear models [5, 6]. One of many
possible solutions used to control nonlinear plants includes where the state, input and output vectors and matrix C are
adaptive control methods based on linearization of the defined as follows:
nonlinear model for various operating regimes of the plant.
The controller designed in such a way is tuned in accordance
with changeable dynamic properties of linear models of the

978-1-4799-5081-2/14/$31.00 ©2014 IEEE 722

 ⎡id ⎤ In order to ensure zero steady-state speed error the system
⎡u d ⎤ ⎡i ⎤ ⎡1 0 0 ⎤ synthesis is carried out by means of the LQ optimal control
x = ⎢⎢ iq ⎥⎥ , u = ⎢ ⎥ , y = ⎢ d ⎥ , C = ⎢ . (3)
⎣ uq ⎦ ⎣ω⎦ ⎣ 0 0 1 ⎥⎦ method with state vector augmentation [2, 3, 4].
⎢⎣ω ⎥⎦
The augmented state vector is defined as follows:
The nominal steady-state operating regimes of the PMSM
described by (2) can be calculated from the set of nonlinear ⎡ x (t ) ⎤
xa ( t ) = ⎢ ⎥ (9)
algebraic equations: ⎣⎢ xi ( t ) ⎦⎥
⎧⎪0 = f ( x n , u n , TL ) where:
⎨ (4)
⎪⎩ y n = Cx n .
(
xi ( t ) = C∫ x ref − x ( t ) dt. ) (10)
As a result of linearization performed in the whole range of
Assuming availability of the state variables the control signals
the nominal state vector xn=[idn iqn ωn]T linear state-space
vector is defined as follows:
models of the PMSM are obtained:
u ( t ) = −F1x ( t ) − F2 xi ( t ) . (11)
⎧⎪ x = A ( x − x n ) + B ( u − u n )
⎨ (5) The result is augmented linear state-space models of the
⎪⎩ y − y n = C ( x − x n ) PMSM:
where:
⎡ x ( t ) ⎤ ⎡ A 0 ⎤ ⎡ x ( t ) ⎤ ⎡ B 0 ⎤ ⎡u(t ) ⎤ ⎡G ⎤
⎢ ⎥=⎢ ⎥⎢ ⎥+⎢ ⎥⎢ ⎥ + ⎢ ⎥ TL (12)
⎡∂ T ⎤
T ⎣⎢ x i ( t ) ⎦⎥ ⎣-C 0 ⎦ ⎢⎣ xi ( t ) ⎦⎥ ⎣ 0 C ⎦ ⎣ x ref ⎦ ⎣ 0 ⎦
A = ⎢ f ( x, u ) ⎥ =
⎣ ∂x ⎦ x=x n
where: G=[0 0 -1/J]T and matrices A, B, C are defined in (6)
and (7).
⎡ − Rs Lq Lq ⎤
⎢ pωn piqn ⎥ The obtained linear models (12) with the defined
Ld Ld Ld
⎢ ⎥ (6) parameters are the starting point for applying many known
⎢ L − Rs Ld pψ f ⎥ methods for linear control system design. In the case of the
⎢ − d pωn − pidn − ⎥ linear MIMO systems multivariable LQR/LQG or possibly
⎢ Lq Lq Lq Lq ⎥
⎢3 p ⎥ modal controllers are usually designed [5, 6, 9]. The LQ-
3 p f
⎢
⎢⎣ 2 J
( Ld − Lq ) iqn 2J
(( L d )
− Lq idn + ψ f ) − c
J
⎥
⎥⎦
optimal state-feedback control law has been applied in the
developed control system. This control law minimizes the
quality criterion formulated in the infinite horizon quadratic
⎡1 ⎤ continuous-time cost functional:
⎢L 0 ⎥
⎢ d ⎥ ∞
⎛ [ x a (t ) − x aref ]T Q[ x a (t ) − x aref ] + ⎞
⎢ 1⎥ ⎡1 0 0⎤ 1
B = ⎢0 J = ∫⎜ ⎟⎟dt (13)
⎥, C = ⎢ ⎥. (7)
2 t = 0 ⎜⎝ +[u (t ) − u ref ]T R[u (t ) − u ref ]
⎢ Lq ⎥ ⎣0 0 1 ⎦ ⎠
⎢0 0 ⎥
⎢ ⎥ where: the weight matrices Q∈R5x5 and R∈R2x2 are positive
⎣ ⎦ semi-definite matrix Q>0 and positive definite matrix R≥0,
The resulting models (5) are linear approximations of respectively. The weight matrices Q and R are chosen so as to
equations (2) around the nominal operating points of the maintain a proper relationship between input and output signals
PMSM. for various operating regimes of the system (12).
The optimal state-feedback matrix F, which allows to
III. SYNTHESIS OF THE GAIN-SCHEDULED LQR minimize the functional (13), can be calculated for the given
CONTROL SYSTEM weight matrices Q and R and parameters of a local linear
model (12) by means of the following formula:
The primary objective for vector control systems is rotor
angular speed control. A typical vector control scheme with F = [ F1 F2 ] = R −1BT P, (14)
keeping the d axis current component which equals zero is
assumed in the study. Then the reference state, input and output where: the matrix P∈R5x5 is the symmetric and positive semi-
vectors are as follows: definite solution of the Algebraic Riccati Equation [6].
⎡ 0 ⎤ The results of the control system synthesis performed in the
⎢ ⎥ ⎡udref ⎤ ⎡ 0 ⎤ whole range of the nominal state vector xn=[0 iqn ωn]T, for the
x ref = ⎢ iqref ⎥ , u ref = ⎢ ⎥ , y ref = ⎢ω ⎥ . (8)
u weight matrices chosen as Q=10diag[1, 1, 2, 20, 20] and
⎢ωref ⎥ ⎣ qref ⎦ ⎣ ref ⎦
⎣ ⎦ R=diag[1, 1], are shown in Fig. 1. These graphs show how the
elements of the state-feedback matrix F depend on the values
of the nominal state vector xn.

723
 around the current operating point of the motor. In this way the
gain-scheduled LQR controller is tuned in accordance with the
F(1,1) F(2,1)
changing PMSM operating points.
3 0.2
2.8 0
2.6 -0.2 IV. RESULTS OF SIMULATION TESTS
40 4000 40 4000
20
iq [A] 0 0
2000
speed [rpm]
20
iq [A] 0 0
2000
speed [rpm]
The study focuses on the control system reaction to a
F(1,2) F(2,2) change in the set value of the reference rotor angular speed as
0.2 3.4
well as ability to maintain the speed reached at a stepwise
0 3.2 changing torque load. However, the changing torque load is
-0.2 3 treated as a non-measureable stepwise disturbance. In the
40
20 2000
4000 40
20 2000
4000 simulation tests a gain-scheduled LQR controller is compared
0 0 0 0
iq [A] speed [rpm] iq [A] speed [rpm] with a LQR controller with fixed parameters designed for the
F(1,3) F(2,3)
chosen PMSM operating point determined as: idn=0,
20 5
iqn=21.47A, ωn=3000rpm. The motor parameters used in the
0 0
-20 -5
simulation tests are as follows: Ld=2.5⋅10-4H, Lq=3⋅10-4H,
40
20 2000
4000 40
20 2000
4000 Rs=0.2Ω, ψf=0.1551Wb, p=6, J=0.0825Nm, fc=10-4Nm⋅s/rad.
iq [A] 0 0 speed [rpm] iq [A] 0 0 speed [rpm]
F(1,4) F(2,4) The simulation tests were performed in the following
20 0
manner. At the beginning of the simulation the motor was in
0 -10 the equilibrium point with the torque load TL=5Nm and the
-20 -20 initial state variables: idn=0, iqn=3.58A, ωn=200rpm. At the time
40 40
20
0 0
2000
4000 20
0 0
2000
4000 point of t=0.5s the reference rotor angular speed changed to
iq [A] iq [A]
F(1,5)
speed [rpm]
F(2,5)
speed [rpm]
ωref=1500rpm. On the basis of Fig. 3 it can be stated that
quality of getting to the reference speed for both controllers
20 20
0 0
was approximately the same.
-20 -20
40 4000 40 4000
20 2000 20 2000 1800
iq [A] 0 0 speed [rpm] iq [A] 0 0 speed [rpm]
1600

1400

Fig. 1. Entries of the state-feedback matrix F.

1200

Fig. 1 demonstrates that the most significant changes in the

speed [rpm]

1000
values of matrix F occur most often for a relatively low rotor
speed. The higher the rotor speeds the smaller the changes in 800

matrix F. Furthermore, it can be stated that in this specific case 600

current iqn has a small and approximately linear impact on the
elements of matrix F. Note that the control system synthesis is 400

performed for the specific weight matrices Q and R and idn=0. reference speed [rpm]
200
The results of the studies showed that for different values of the speed [rpm], LQR
speed [rpm], g-s-LQR
weight matrices the graphs in Fig. 1 have different shapes. 0
0 1 2 3 4 5 6 7 8
Therefore, the question of selecting optimal weight matrices Q t [s]
and R for nonlinear MIMO plants still remains to be answered.
It explicitly points to the requirement of performing Fig. 3. Rotor angular speed.
multicriteria analysis of the control system operation quality.
However, in the case of the gain-scheduled controller
The structure of the developed gain-scheduled LQR control reached the reference angular speed with a lesser id current
system is presented in Fig. 2. deviation, which is shown in Fig. 4. For both controllers the iq
yref u(t) y(t)
current control quality was similar.
1/s F2 Inverter PMSM
- - In the subsequent simulation stage an impact of stepwise
change of the torque load on the angular motor speed was
studied. At the time point of t=3s the torque load went up to
x(t) TL=40Nm, whereas at t=3s it went down to TL=5Nm, which is
F1 shown in Fig. 5.

Fig. 2. Block diagram of a gain-scheduled LQR control system.

The state-feedback matrix F=[F1 F2] is calculated from (14)

by means of a local linear model of the PMSM (5) linearized

724
 speed and better compensation of the speed deviation during
20
the stepwise torque load disturbance.
10

V. CONCLUSIONS
id [A]

0
reference id [A] In the paper a gain-scheduled LQR control system for a
-10 id [A], LQR
id [A], g-s-LQR
nonlinear model of PMSM is presented. It is possible that if the
-20 motor operates within a wide range of nominal operating points
0 1 2 3 4 5 6 7 8
t [s] a LQR controller designed for the selected operating point does
40
not provide optimal control quality in other operating points.
reference iq [A] Therefore, the synthesis of a gain-scheduled LQR controller is
30 iq [A], LQR based on linearization of a nonlinear PMSM model around the
iq [A], g-s-LQR
current operating points of the motor. The presented simulation
iq [A]

20
tests results show that adaptation of the multivariable LQR
10 controller parameters is able to improve control quality by
0
means of reducing the integral control quality criterion.
0 1 2 3 4 5 6 7 8 Depending on the selected weight matrices different graphs
t [s]
representing the state-feedback matrix dependent on nominal
Fig. 4. Current components in d and q axes. operating points are obtained. An appropriate choice of weight
matrices for a specific vector control scheme realization still
50 remains an open issue. It points to the requirement of
performing multicriteria analysis of the control system
40
operation quality.
30
TL [Nm]

REFERENCES
20
[1] J. Quirion, E. Gunn, J. Gu, “Optimal control of permanent magnet
10 motors using dynamic programming”, IEEE Conference on Robotics,
5 Automation and Mechatronics, Vol. 1, 2004, pp. 364-369.
0
0 1 2 3 4 5 6 7 8 [2] S. Hassaine, B. Sari, S. Moreau, B Mazari, “Rapid Prototyping of a
t [s] Multivariable Control with Pole Placement by State Feedback of a
Fig. 5. Torque load. PMSM: LMI Approach”, 38th Annual Conference of the IEEE
Industrial Electronics Society, 2012.
The impact of the torque load stepwise change on the [3] K.T. Chang, T. S. Low ; T. H Lee, “An optimal speed controller for
permanent-magnet synchronous motor drives”, IEEE Transactions on
angular speed manifests in a rotor speed deviation from the set Industrial Electronics, Vol. 41, Issue: 5, 1994, pp. 503-510.
point value. Fig. 3 illustrates that the gain-scheduled controller
[4] L.M. Grzesiak, T. Tarczewski, “Permanent magnet synchronous motor
is better at compensating stepwise torque load disturbances. discrete linear quadratic speed controller”, IEEE International
Symposium on Industrial Electronics, 2011, pp. 667-672.
*10^3 [5] G. F. Franklin, J. D. Powell, A. Emami-Naeini, “Feedback Control of
Dynamic Systems”, 5th Edition, Prentice Hall, 2006.
184
176 [6] S. Bańka: “Multivariable control systems: A polynomial approach” (in
J(g-s-LQR) < J(LQR) Polish), Szczecin University of Technology Press, Szczecin, 2007.
150 [7] S. Bańka, P. Dworak, M. Brasel, “Linear Multi-controller Structure for
Control of a Nonlinear MIMO Model of a Drill Ship”, 13th IFAC
Symposium on Large Scale Complex Systems: Theory and Applications,
7-10 July, Shanghai, China, 2013.
100 [8] S. Bańka, M. Brasel, P. Dworak, K. J. Latawiec, “A Switched Structure
J

of Linear MIMO Controllers for Positioning of a Drillship on a Sea

Surface”, 15th International Conference on Methods and Models in
Automation and Robotics, 2010.
50 [9] S. Bańka, P. Dworak and K. Jaroszewski, “Linear adaptive structure for
control of a nonlinear MIMO dynamic plant”, International Journal of
Applied Mathematics and Computer Science, 23(1), pp. 47-63, 2013.
J, LQR
J, g-s-LQR [10] P. Dworak, M. Brasel, “Improving quality of regulation of a nonlinear
MIMO dynamic plant”, Elektronika Ir Elektrotechnika, Vol. 19, No. 7,
0
0 1 2 3 4 5 6 7 8 2013, pp. 3-6.
t [s]
[11] M. Brasel, “Adaptive LQR Control System for the Nonlinear 4-DoF
Model of a Container Vessel”, 18th International Conference on
Fig. 6. Linear quadratic quality criterion. Methods and Models in Automation and Robotics, Międzyzdroje,
Poland, 26-29 August, 2013.
Integral control quality criterion (13) for both controllers is
shown in Fig. 6. The gain-scheduled controller reached a lesser
value of the control quality criterion. Obviously, it resulted
from a lesser id current deviation at getting to the reference

725