This article was downloaded by: [Universite Laval]

On: 16 July 2014, At: 01:17

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

Linear quadratic output tracking and disturbance

rejection
a a a a
Masoud Karimi-Ghartemani , S. Ali Khajehoddin , Praveen Jain & Alireza Bakhshai
a
Department of Electrical and Computer Engineering , Queen's University , Kingston, ON
K7L 3N6, Canada
Published online: 11 Aug 2011.

To cite this article: Masoud Karimi-Ghartemani , S. Ali Khajehoddin , Praveen Jain & Alireza Bakhshai (2011)
Linear quadratic output tracking and disturbance rejection, International Journal of Control, 84:8, 1442-1449, DOI:
10.1080/00207179.2011.605908

To link to this article: http://dx.doi.org/10.1080/00207179.2011.605908

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 International Journal of Control
Vol. 84, No. 8, August 2011, 1442–1449

Linear quadratic output tracking and disturbance rejection

Masoud Karimi-Ghartemani*, S. Ali Khajehoddin, Praveen Jain and Alireza Bakhshai
Department of Electrical and Computer Engineering, Queen’s University, Kingston, ON K7L 3N6, Canada
(Received 28 February 2011; final version received 13 July 2011)

This article introduces the problem of linear quadratic tracking (LQT) where the objective is to design a
closed-loop control scheme such that the output signal of the system optimally tracks a given reference signal and
rejects a given disturbance. Different performance indices that have been used to address the tracking problem
are discussed and an appropriate new form is introduced. It is shown that a solution to the proposed optimality
index exists under very mild conditions of stabilisability and detectability of the plant state-space equations. The
solution is formulated based on converting the LQT problem to a standard linear quadratic regulation problem.
The method is applied to two examples, a first-order plant and a third-order plant, and their simulation results
are presented and discussed.
Downloaded by [Universite Laval] at 01:17 16 July 2014

Keywords: LQR; linear quadratic output tracking; linear quadratic disturbance rejection; optimal tracking

1. Introduction (ii) (C, A) being detectable where Q ¼ CTC (Kalman

Linear quadratic regulation (LQR) is a well-known 1960; Anderson and Moore 1989, p. 48). These two are
and highly useful concept in optimal control theory. indeed sufficient and necessary conditions.
For the linear system with the state-space equation Closely related to the regulation problem is the
x_ ¼ Ax þ Bu, the LQR offers the feedback gains K in concept of tracking problem where it is desired that
the control law of u ¼  Kx, the output signal of the closed-loop system y track a
R 1which minimises the
quadratic functional JðuÞ ¼ 0 ðxT Qx þ uT RuÞdt: In desired reference r. The optimal tracking index is not,
the above, x is the system’s state vector, u is the however, straightforward to define because direct
control, T stands for vector transposition, Q is a extensions of the LQR index will either become
positive semi-definite matrix and R is a positive definite unbounded or suffer from nonexistence of a stabilising
matrix. The solution is K ¼ R1BTF where F is solution.
obtained from the algebraic Riccati equation (ARE) In this article, we select a closed-loop structure
ATF þ FA þ Q  FBR1BTF ¼ 0. The solution to finite- based on the internal model (IM) principle in order to
time LQR is also available (Anderson and Moore ensure that the actual value of the reference is achieved
1989) and is obtained from a differential Riccati in the steady state and the disturbance is completely
equation (DRE). rejected. We then introduce a function in the form of
The LQR design offers good gain margin and phase Z1
margin, and good tolerance of nonlinearities, when the J¼ ðqe2 þ v2 Þdt, ð1Þ
0
plant states are measurable (Anderson and Moore
1989). Such robustness properties can be frequently where e ¼ r  y is the tracking error, q is a positive
achieved even when state estimation is required scalar and v is a transformed input. The transformed
(Anderson and Moore 1989). The LQR design input is appropriately defined to reflect the transient
provides high robustness with respect to system behaviours of the control signal and to approach zero
parameters uncertainties which translate into 60 of as time tends to infinity. It will then be proved that a
phase margin and infinite gain margin (Safonov and solution to the above problem exists under mild
Athans 1977). There is a close relationship between conditions of stabilisability and detectability of the
robust and LQR designs (Lin 2000, 2007; Tan, Shu, plant. Such a solution is formulated using the LQR
and Lin 2009). technique.
Existence of the LQR solution and asymptotic This article is organised as follows. A review of
stability of the closed-loop system are ensured under existing treatments of the LQT problem is presented in
very mild conditions of (i) (A, B) being stabilisable and Section 2, and the problem is formulated in Section 3.

*Corresponding author. Email: [email protected]

ISSN 0020–7179 print/ISSN 1366–5820 online

ß 2011 Taylor & Francis
DOI: 10.1080/00207179.2011.605908
http://www.informaworld.com
 International Journal of Control 1443

that is a well-defined function. The complete solution to

minimise this cost function is formulated in Willems and
Mareels (2004).
The following statements where some of them
reflect serious drawbacks, apply to the methods
Figure 1. The optimal tracking problem addressed in
Anderson and Moore (1989).
discussed above: (i) the reference state trajectory xr
must be available (or estimated), while in practice only
the output trajectory may be known; (ii) the depen-
The proposed method is described in Section 4, and
dence of feed-forward term (Kr in Figure 1) on the
numerical examples are presented in Section 5.
plant parameters can cause robustness problems; (iii)
the closed-loop system does not offer disturbance
2. Literature review rejection and (iv) the reference trajectory may or may
not be achieved in the steady state.
An indexRT for finite-time optimal tracking is It is often required in practice that a reference
JðuÞ ¼ 0 f ½ð y  rÞT Q1 ð y  rÞ þ uT Rudt where Tf is signal is exactly achieved in the steady state by the
the final time (Anderson and Moore RT 1989). It can be output (not the states), and that robust properties
transformed into ~ T Q2 ðx  xÞ
JðuÞ ¼ 0 f ½ðx  xÞ ~ þ
including disturbance rejection are obtained, and that
u Rudt where x~ ¼ Lr, L ¼ C ðCC Þ1 . Assuming
T T T
these robust properties are obtained in an optimal way.
that the reference is generated by x_ r ¼ Ar xr ,
Downloaded by [Universite Laval] at 01:17 16 July 2014

An IM controller is used in Krikelis and Papadopoulos

r ¼ Cr xr , then x~ ¼ Lr ¼ LCr xr and the problem can (1982) to ensure zero steady-state error and uses
be converted to a finitetime regulation problem by R1
the functional JðuÞ ¼ 0 ½ðx  xss ÞT Q2 ðx  xss Þ þ
defining the augmented state vector z ¼ ½xT , xTr T . The ðu  uss Þ2 dt, where the subscript ss refers to steady
solution is expressed as u ¼  Kx  Krxr, as shown in state. It shows that the controller gains can be obtained
Figure 1, where K and Kr are obtained from DREs from an ARE. No discussion is presented regarding
(Anderson and Moore 1989, p. 73), and an estimator is conditions for the existence of a stabilising solution.
required to estimate the reference state variables from We will show in this article that this cost function is not
the reference output signal. This formulation cannot be a desirable function; a fact that is also confirmed in
extended to infinite-time because the index itself Willems and Mareels (2004).
becomes infinite, and even if some solution is obtained
for large Tf, ‘the label optimal is a misnomer’
(Anderson and Moore 1989, p. 85).
One way to overcome the unboundedness of the 3. Problem formulation
above performance index for infinite time is the concept Consider a linear single-input single-output (SISO)
of overtaking criterion (Gale 1967; Artestein and system with the representation
Leizarowits 1985). Based on this concept, the control
x_ p ¼ Ap xp þ Bp u þ Bpd d, y ¼ Cp xp ð2Þ
obtained for large time intervals overtakes that
obtained for small time intervals. It is shown in where the over-dot shows time-derivative, the index p
Artestein and Leizarowits (1985) and Bernstein and is used to denote the plant, xp is the n-dimensional
Haddad (1987) that the same structure of Figure 1 is state vector, u is a one-dimensional control and d is the
optimal in the overtaking sense. Moreover, the gain disturbance signal. Matrix Ap is n  n, and Bp and Bpd
vectors Kp and Kr are obtained as Kp ¼ R1BTF and are n  1 and Cp is 1  n dimensional vectors. The
Kr ¼ R1BT(A  BR1BTF )TQ where F is the solution objective is to design a closed-loop controller such that:
of the ARE. In other words, the overtaking method (i) the reference signal r is exactly achieved by the
states that the optimal gains for the infinite-time case output in the steady state; (ii) the disturbance signal is
can be obtained from the ARE associated with the DRE perfectly rejected from the output in the steady state;
of finite-time solution. and (iii) the performance index (1) is minimised. In (1),
A solution to the infinite time interval LQ problem e(t) ¼ r(t)  y(t) is the tracking error, q is a positive
with constant state tracking is presented in Willems and scalar, and v is a pseudo-input properly defined to
Mareels (2004). In Willems and Mareels (2004), the convey transient properties of u without causing the
problem is resolved into two subproblems that address cost index to become unlimited. Rigorous definition of
steady-state and transient stages separately. The steady- v is given below. The index (1) combines optimal
state solution (, u) is obtained by minimising tracking of the command signal r by the output y and
TQ þ uTRu where  ¼ x  xd and xd is the desired optimal rejection of the disturbance d from the
state. This steady-state
R 1 solution is used to refine the cost output y, and avoids excessive energy consumption
function to  ¼ 0 ðT Q þ uT Ru  T Q  uT Ru Þdt at v.
 1444 M. Karimi-Ghartemani et al.

Assuming that the SFC and the OFC gain vectors are
Kp and Kc, the control law will be
 
x
u ¼ Kp xp  Kc xc ¼ ðKp , Kc Þ p ¼ Kx ð8Þ
xc

Figure 2. The proposed closed-loop structure. where K is the overall controller gain vector.
Theorem 1: The pair (A, B) is stabilisable.
We assume that r and d are generated from
Proof of Theorem 1: It is sufficient to prove that the
finite-dimensional linear models given by
following matrix is full rank for all  2 Cþ (where Cþ is
x_ r ¼ Ar xr , xr ð0Þ ¼ xr0 , r ¼ Cr xr , the closed right half of the complex plane.)
ð3Þ
x_ d ¼ Ad xd , xd ð0Þ ¼ xd0 , d ¼ Cd xd :  
I  Ap 0 Bp
ðI  A, BÞ ¼ : ð9Þ
Assumption I: The state-space representation of (2) is Bc Cp I  Ac 0
stabilisable and detectable.
Based on Assumption I, the matrix (I  Ap, Bp) has
full rank for all  2 Cþ. This fact together with the
specific block structure of matrix (9) means that
4. Proposed method (I  A, B) has full rank for all  2 Cþexcept possibly
Downloaded by [Universite Laval] at 01:17 16 July 2014

4.1 Proposed closed-loop structure for eigenvalues of Ac. Now, assume that for a
The proposed closed-loop structure is shown in  2 Cþbelonging to the set of eigenvalues of Ac, the
Figure 2. It consists of a state-feedback controller matrix (I  A, B) is not full rank. This implies that
(SFC) plus an output feedback controller (OFC). there exists a non-zero vector v such that
Structure of the OFC is specified based on the IM vT(I  A, B) ¼ 0. Partitioning v into vT ¼ ðvT1 , vT2 Þ,
principle to ensure that the reference signal r is tracked we obtain
and the disturbance d is rejected (Francis and Wonham vT1 ðI  Ap Þ þ vT2 Bc Cp ¼ 0,
1976). Assume that the OFC representation is ð10Þ
vT1 Bp ¼ 0, vT2 ðI  Ac Þ ¼ 0:
x_ c ¼ Ac xc þ Bc e, ð4Þ
Now, we consider the proof in the two following cases.
where xc is the m-dimensional state vector. The . Case 1:  is not an eigenvalue of Ap. Then the
characteristic polynomial of the OFC is first equation in (10) results in vT1 ¼
Dc ðsÞ ¼ detðsI  Ac Þ, ð5Þ vT2 Bc Cp ðI  Ap Þ1 . Substituting this into
the second equation of (10) results in
and to satisfy the IM principle, Dc(s) must (at least) vT2 Bc ½Cp ðI  Ap Þ1 Bp  ¼ 0. Now, based on
contain Dr(s) ¼ det(sI  Ar) and Dd(s) ¼ det(sI  Ad). Assumption III, Cp(I  Ap)1Bp6¼0.
Two more assumptions are made as follows. Therefore vT2 Bc ¼ 0. This fact together with
Assumption II: The pair (Ac, Bc) in (4) is controllable. the third equation of (10) results in
vT2 ðI  Ac , Bc Þ ¼ 0. But this contradicts
Assumption III: None of the poles of OFC (eigenva- Assumption II.
lues of Ac) is equal to a zero of the plant. . Case 2:  is an eigenvalue of Ap. Then there
It will be shown shortly that, given the stated exist a non-zero vector v3 such that
assumptions, there exists a set of solutions for OFC (I  Ap)v3 ¼ 0. This equation and the first
and SFC gains to make the closed-loop stable. The equation of (10) results in that vT2 Bc Cp v3 ¼ 0:
optimal solution is obtained among the stabilising set Now, vT2 Bc 6¼ 0, because vT2 Bc ¼ 0 with the
of gains. third equation in (10), this will contradict
If we define x ¼ ½xTp , xTc T , then the closed-loop Assumption II. Therefore, Cpv3 ¼ 0, contra-
system is dicting the detectability condition in
Assumption I.
x_ ¼ Ax þ Bu þ Br r þ Bd d, y ¼ Cx, ð6Þ
     
Ap 0 Bp T CTp
A¼ B¼ , , C ¼ , 4.2 Proposed solution
Bc Cp Ac 0 0
    Define the operator Dc(D) ¼ det(DI  Ac) ¼
0 Bpd
Br ¼ , Bd ¼ : ð7Þ Dm þ a1Dm1 þ    þ am1D þ am where D is the differ-
Bc 0 entiating operator, and I is the identity matrix.
 International Journal of Control 1445

Obviously, Dc(D)r ¼ Dc (D) d ¼ 0. Apply the operator If Cpv1 ¼ 0, then this along with the first equation
to both sides of (6) and (8) to arrive at of (15) contradicts the detectability in Assumption I.
On the other hand, Cpv1 6¼ 0 with the second equation
z_ ¼ Az þ Bv, v ¼ Kz, ð11Þ
of (15) mean that  is a zero of the system x_ c ¼ Ac xc þ
where the transformed (or pseudo) state vector z and Bc e, yc ¼ Cc xc . But such a system has no zero because
the transformed (or pseudo) input v are its transfer function is Cc ðsI  Ac Þ1 Bc ¼ Dc1ðsÞ (given
the controllable form).
z ¼ Dc ðDÞx, v ¼ Dc ðDÞu: ð12Þ
Remark 2: The proposed method also offers the
And z ¼ ½zTp , zTc T , zp ¼ Dc(D)xp and zc ¼ Dc(D)xc. possibility of including time-derivatives of the tracking
Remark 1: The equations in (11) show a standard error up to order m  1 in the cost function.
state-feedback law with no reference signal. Moreover, Considering the controllable canonical form for Ac
the pair (A, B) is stabilisable. Therefore, for any and Bc, it can be shown that zc(m) ¼ e, zc(m  1) ¼
Q ¼ HTH that makes (H, A) detectable, there exists a e_ , . . . , zc(1) ¼ e(m  1). By selecting Q ¼ diag{0, 0, . . . , 0,
solution to the following LQR problem: qm, qm1, . . . , q1}, the cost index will be extended to
Z1 Z1
2
J¼ ðzT Qz þ v2 Þdt ð13Þ J¼ ðq1 e2 þ q2 e_2 þ    þ qm eðm1Þ þ v2 Þdt: ð16Þ
0 0
Downloaded by [Universite Laval] at 01:17 16 July 2014

that makes the closed-loop asymptotically stable. This index can be used to reduce undesired oscillations
Theorem 2: The matrices Ac and Bc can be properly (Example 2 in Section 5).
selected to ensure that z(m þ n) ¼ zc(m) ¼ e, where Remark 3: For Q ¼ diag{qmþn, . . . , qmþ1, qm, . . . , q1},
z(m þ n) is the last element of z. (16) further extends to
Proof of Theorem 2: The OFC state vector xc Z1
2
satisfies (DI  Ac)xc ¼ Bce. Therefore, zc ¼ Dc(D)xc ¼ J¼ ðq1 e2 þ q2 e_2 þ    þ qm eðm1Þ
adj(DI  Ac)Bce, where adj() denotes the matrix 0

adjugate satisfying adj(X )X ¼ det(X ). Now consider þ qmþ1 zp ðnÞ2 þ    þ qmþn zp ð1Þ2 þ v2 Þdt: ð17Þ
the controllable canonical form representation of
(Ac, Bc). The last element of zc, i.e. zc(m) will be equal The transformed state zp ¼ Dc(D)xc reflects the tran-
to the (m, 1) element of the adjugate matrix multiplied sient of the original state xc. Therefore, this extended
by e. The (m, 1) element of the adjugate matrix will be functional can be used to improve those transient
equal to the determinant of the submatrix that is behaviours (Example 2 in Section 5).
obtained by removing (1, m) element, and it is equal to Remark 4: When the OFC is a simple first-order
(1)(m1)(1)(mþ1) ¼ 1. Therefore, zc(m) ¼ e. integrator (I), i.e. Dc(s) ¼ s, then the index can only
The result of Theorem 2 is that, for the selection of contain the tracking error and higher order derivatives
Q ¼ diag{0, 0, . . . , 0, q}, (13) becomes of error are excluded. The OFC can be extended to a PI
Z1 Z1 structure to include the first-order derivative of the
J¼ ðzT Qz þ v2 Þdt ¼ ðqe2 þ v2 Þdt, ð14Þ error function. Assume that the OFC’s structure is 1þs
0 0 s
where  is a known positive scalar. Thus, (1)
that is, exactly equal to the desired cost function (1). changes to
The following Theorem proves that for such a selection Z1
of matrix Q, a solution exists such that the closed-loop
J¼ _ 2 þ v2 dt:
½qðe þ eÞ ð18Þ
system is asymptotically stable. 0

Theorem 3: Define the m-dimensional row vector If we define x_ c ¼ e þ e_ ¼ r  Cp xp  Cp ðAp xp þ
Cc ¼ (0, . . . , 0, 1) and the n þ m dimensional row vector Bp u þ Bpd d Þ, then A and B in (7) change to
H ¼ (0, Cc). Then, the pair (H, A) is detectable.
! !
Proof of Theorem 3: If (H, A) is not detectable, then Ap 0 Bp
A¼ , B¼ : ð19Þ
there exists a scalar  2 Cþ and a non-zero vector Cp  Cp Ap 0 Cp Bp
v ¼ ½vT1 , vT2 T such that
   It is straightforward to prove that the associated LQR
I  Ac Bc v2
ðI  Ap Þv1 ¼ 0, ¼ 0: problem has a stabilising solution for all 40. The
Cc 0 C p v1
same procedure that has been previously formulated
ð15Þ can then be followed to obtain the solution.
 1446 M. Karimi-Ghartemani et al.

Increasing  will improve the behaviour of the tracking The control law (25) comprises the term Kx þ u
error derivative at the expense of degrading the that is generally unknown as it depends on the
behaviour of the tracking error itself. This manifests disturbance signal. Such a term acts as a second
itself mainly in terms of slower transient response as disturbance signal if we consider the modified control
shown by an example in Section 5. law as u ¼  Kx. It is, however, noticed that the
minimum cost function value is Jmin ¼ xð0Þ ~ T Fxð0Þ,
~
which is independent of the disturbance signal.
Therefore, the minimal solution does not depend on
4.3 Importance of pseudo-input variable v the value of this term and we can remove it. Such a
The above derivations show that the cost function of solution makes the closed-loop stable as long as (Q, A)
(1) has a well-defined form for the proposed control is detectable.
structure of Figure 2. This is in the sense that it avoids In order to examine the index (20), we notice that
being unbounded and it offers a feedback solution that the tracking error is equal to e ¼ y. ~ Thus,
makes the loop asymptotically stable. One may think e2 ¼ ðCxÞ~ T ðCxÞ ~ Therefore, qe2 ¼ x~ T Qx~
~ ¼ x~ T CT Cx:
of other forms for a cost function such as where Q ¼ qCTC. With such selection for Q, the
index (24) coincides with (20). The most significant
Z 1 question now is whether the closed-loop system
J¼ ½qe2 þ ðu  u Þ2 dt, ð20Þ becomes asymptotically stable with such a selection.
0
Downloaded by [Universite Laval] at 01:17 16 July 2014

Theorem 4: The solution of (1) and the solution of (20)

where u is the steady-state value of the control signal
are both derived from the same ARE with the following
in Figure 2 in the presence of the reference r and the
two different Q matrices:
disturbance d. Here it is already assumed that a
stabilising control exists and converges to u as time Qð1Þ ¼ diagf0, 0, . . . , 0, qg,
tends to infinity. We show in this section that a ð26Þ
Qð20Þ ¼ qCT C, C ¼ ðCp , 0Þ:
stabilising solution to this optimal problem does not
exist. However, the solution to (20) does not stabilise the
Assume that the steady-state variables are denoted closed-loop system.
by xp , u, d and xc . The plant and the OFC equations Proof of Theorem 4: A proof for the first part of the
in the steady-state are statement is already given above. For the second part
of the theorem, we show that the pair (C, A) is not
x_ p ¼ Ap xp þ Bp u þ Bpd d , x_ c ¼ Ac xc ,
ð21Þ detectable. Now, we show that there is some  2 Cþ
y ¼ r ¼ Cp xp : such that the matrix [(I  A)T CT] is not full rank.
This means that there exists a non-zero vector v such
Define u~ ¼ u  u , x~ p ¼ xp  xp , x~ c ¼ xc  xc , y~ ¼ y  that
y ¼ y  r and d~ ¼ d  d , then 0 1
  I  Ap 0  
I  A @ A v1
x_~ p ¼ Ap x~ p þ Bp u~ þ Bpd d, ~ y~ ¼ Cp x~ p , v¼ Bc Cp I  Ac ¼ 0: ð27Þ
ð22Þ C v2
Cp 0
x_~ c ¼ Ac x~ c þ Bc e~ ¼ Ac x~ c  Bc Cp x~ p :
Equation (27) is satisfied for v1 ¼ 0 and v2 equals any
It can thus be concluded that eigenvectors of Ac. Notice that (at least some or all)
eigenvalues of Ac are located on the imaginary axis.
x_~ ¼ Ax~ þ Bu~ þ Bd d,
~ ~
e ¼ Cx, ð23Þ
Therefore, (C, A) is not detectable.
where x~ ¼ ½x~ Tp , x~ Tc T and the matrices A, B, C and Bd
are given in (7). Assuming that d~ ¼ 0, the following
LQR index seems to be well-defined and can be 5. Examples
minimised
5.1 Example 1
Z 1
Consider the plant GðsÞ ¼ 1s with the minimal realisation
J¼ ðx~ T Qx~ þ u~ 2 Þdt: ð24Þ of x_ p ¼ Ap xp þ Bp u þ Bp d, y ¼ Cp xp where Ap ¼ 0,
0
Bp ¼ 1, Cp ¼ 1. The objective is to optimally track step
The solution to (24) will be equal to u~ ¼ Kx~ where K commands and reject step disturbances. Thus, the OFC
is obtained from the associated ARE. And finally structure is also a simple integrator x_ c ¼ Ac xc þ Bc e
where Ac ¼ 0, Bc ¼ 1. To minimise the cost function (1),
u ¼ u~ þ u ¼ Kx þ Kx þ u : ð25Þ we choose Q ¼ diag{0, q}. The closed-loop response to a
 International Journal of Control 1447

Response to step command Response to unit step disturbance

(a) 1.1 (a)
1
q = 100
0.1 q = 1000
q= 100
q= 1000 q = 10,000
0.05
q= 10,000
0
0 0 1 2 3
0 1 2 3
Control signal
Control signal
(b) 0
(b)
4 q =100 q =100
q =1000 −0.5 q =1000
2 q = 10,000
q = 10,000
−1
0
0 1 2 3
0 1 2 3 Time (s)
Time (s)
Figure 4. Response of system of Example 1 to unit step
Downloaded by [Universite Laval] at 01:17 16 July 2014

Figure 3. Response of system of Example 1 to unit step disturbance: (a) output signal and (b) control signal.
command: (a) output signal and (b) control signal.

Table 1. Index data for Example 1 (step command). Table 2. Index data for Example 1 (step disturbance).

q 100 1000 10,000 q 100 1000 10,000

R1 2 R1 2
e dt 0.335 0.189 0.106 e dt 0.011 0.002 3.5  104
R01 2 R01 2
u dt 1.118 1.988 3.536 u dt 1 1 1
R01 2 R01 2
0 v dt 11.18 62.82 352.7 0 v dt 3.35 5.96 10.6

unit step command is shown in Figure 3(a) for three (a) 1

different values of q equal to 100, 1000 and 10,000. The
0.5
control signal is also shown in Figure 3(b). Increasing q
0 b=0 b =0.3 b =0.7
results in better tracking at the cost of larger control 0 1 2 3
signals. Table 1 shows the numerical values of tracking (b)
2
error and control signal energies. Performance of the 1
closed-loop system in rejecting the step disturbances is 0
0 1 2 3
shown in Figure 4, which indicates stronger rejection of (c) 0.1
the disturbance for larger values of q. Table 2 shows the 0.05
numerical values for the tracking error and control 0
signal energies for the disturbance response. (d) 0
0 1 2 3
The impact of  in functional (18) is shown in −0.5
Figure 5. In this figure, q is selected at 1000 and three −1
values of ¼0, 0.3 and 0.7 are examined. Larger  0 1 2 3
results in smoother but slower responses. Time (s)

Figure 5. Effect of  on the transient response of Example 1:

(a) closed-loop response to unit step command, (b) control
signal, (c) response to unit step disturbance and (d) control
5.2 Example 2 signal.
Consider a third-order plant described by the following
state-space matrices: and Cp ¼ (0, 0, 1). The plant transfer function has a
0 1 0 1 0 1 pole at origin and two others on the imaginary axis
0 a 0 a 0
(at j1414 for a ¼ 1000). This system is of the type
Ap ¼ @ a 0 a A, Bp ¼ @ 0 A, Bpd ¼ @ 0 A
occurring in many practical applications such as
0 a 0 0 a
grid-connection of voltage source power electronic
 1448 M. Karimi-Ghartemani et al.

Tracking error to unit sinusoidal reference (q2 = 0)

(a) 1 (a)
q1 =100 q1 =1000 q1 =10,000
0.5

(b)
−0.5
0 0.1 0.2 0.3 0.4 0.5
Tracking error to unit sinusoidal reference (q1= 10,000)
(b) 1

q2 =0 q2 =100 q2 = 1000 (c)

0.5

−0.5
0 0.1 0.2 0.3 0.4 0.5
Time (s) Figure 7. Response of system of Example 2 to unit sinusoi-
dal disturbance when q1 ¼ 10,000 and (a) q2 ¼ 0, (b) q2 ¼ 100
Downloaded by [Universite Laval] at 01:17 16 July 2014

Figure 6. Tracking error in Example 2 to a unit sinusoidal and (c) q2 ¼ 1000.

reference when (a) q1 ¼ 100, 1000, 10,000, q2 ¼ 0, and when
(b) q1 ¼ 10,000, q2 ¼ 0, 100, 1000.

R
(a)
converters (Wu and Lehn 2006; Huerta, Bueno,
Cobreces, Rodriguez, and Giron 2008). The objective
is to design an optimal controller for tracking
sinusoidal references and rejecting sinusoidal distur-
bances both at frequency !o ¼ 20. To this end, the
(b)
OFC matrices are selected as
   
0 !2o 1
Ac ¼ , Bc ¼ :
1 0 0
The overall closed-loop system matrices A and B are (c)
calculated from (7) and the LQR problem is solved for
three different values of q ¼ 100, 1000 and 10,000. The
closed-loop response to the unit sinusoidal reference is
shown in Figure 6(a) for all three values of q. It is
observed that the tracking features improve by
Figure 8. Response of system of Example 2 to unit sinusoi-
increasing q. Response of the system to a unit dal disturbance when q1 ¼ 10,000, q2 ¼ 1000 and (a) q3 ¼ 0
sinusoidal disturbance is shown in Figure 7(a) for (b) q3 ¼ 0.1 and (c)q3 ¼ 1.
q ¼ 10,000. The response to the disturbance has some
oscillations with very low damping. This behaviour
originates from the two undamped poles of the plant function of (17). Figure 8 shows the impact of three
on the imaginary axis. Such behaviour can be values of q3 ¼ 0, 0.1 and 1 when q1 and q2 are fixed at
10,000 and 1000 respectively. This shows that unde-
improved by incorporating the time-derivative of the
sired resonances can greatly be reduced by q3. Notice
tracking error in the cost function, as shown in
that q3 is the weight applied to zp(3) that is the
Remark 3, Equation (16). By increasing q2 from 0 to
transformed output. Increasing q3 thus improves the
100 and then to 1000, the undesired oscillations on the
transient behaviour of the output.
tracking error are reduced. Figure 6(b) shows the
tracking error when the unit sinusoidal reference is
applied. Figure 7(b) and (c) show the tracking error to
unit sinusoidal disturbance. 6. Conclusion
It is possible to make the output signal behave The LQT problem over infinite-time interval was
more smoothly by considering the extended cost addressed in this note. The cost function is based on
 International Journal of Control 1449

using a transformation on the state variables and Gale, D. (1967), ‘On Optimal Development in
control signal such that the steady-state response of the Multi-sector Economy’, Review of Economic Studies, 37,
system is eliminated and components of the transient 1–19.
Huerta, F., Bueno, E., Cobreces, S., Rodriguez, F., and
response penalised proportional to their ‘distance’
Giron, C. (2008), Control of grid-connected voltage source
from the steady-state response. This article derived
converters with LCL filter using a Linear
necessary and sufficient conditions for existence of the Quadratic servocontroller with state estimator, IEEE
solution. It is also shown that the direct extension of Power Electronics Specialists Conference PESC,
the regulation functional (LQR) does not guarantee pp. 3794–3800.
stability of the closed-loop system. Results confirmed Kalman, R. (1960), ‘Contributions to the Theory of Optimal
that the solution for the LQT problem (as formulated Control’, Bol. Sot. Matem. Mex., 5(2), 102–119.
here) is actually obtained from the same Riccati Krikelis, N., and Papadopoulos, E. (1982), ‘An Optimal
equation associated with the LQR problem. This is Approach for Tracking Problems and its Assessment
not unexpected due to the close relationship between Against Classical Controllers’, International Journal of
Control, 36, 249–265.
the two problems. However, the LQR formulation
Lin, F. (2000), ‘An Optimal Control Approach to Robust
does not provide an adequate guideline as to how to Control Design’, International Journal of Control, 73,
choose matrix Q to obtain good tracking properties. 177–186.
Our proposed method offers such an advantage. Lin, F. (2007), Robust Control Design: an Optimal Control
Approach, West Sussex, England: Wiley & Sons.
Downloaded by [Universite Laval] at 01:17 16 July 2014

Safonov, M., and Athans, M. (1977), ‘Gain and Phase

Margin for Multiloop LQG Regulators’, IEEE
References Transactions on Automatic Control, 22, 173–179.
Tan, H., Shu, S., and Lin, F. (2009), ‘An Optimal Control
Anderson, B., and Moore, J. (1989), Optimal Control, Linear Approach to Robust Tracking of Linear Systems’,
Quadratic Methods, Englewood Cliffs, NJ: Prentice-Hall. International Journal of Control, 82, 525–540.
Artestein, Z., and Leizarowits, A. (1985), ‘Tracking Periodic Willems, J., and Mareels, I. (2004), ‘A Rigorous
Signals with the Overtaking Criterion’, IEEE Transactions Solution of the Infinite Time Interval LQ Problem with
on Automatic Control, 30, 1123–1126. Constant State Tracking’, Systems and Control Letters, 52,
Bernstein, D., and Haddad, W. (1987), ‘Optimal Output 289–296.
Feedback for Non-Zero Set Point Regulation’, IEEE Wu, E., and Lehn, P. (2006), ‘Digital Current Control of a
Transactions on Automatic Control, 32, 641–645. Voltage Source Converter with Active Damping of LCL
Francis, B., and Wonham, W. (1976), ‘The Internal Model Resonance’, IEEE Transactions on Power Electronics, 21,
Principle of Control Theory’, Automatica, 12, 457–465. 1364–73.