See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/287198985

State Space and Internal Models for Linear Quadratic Regulator Design

Conference Paper · September 1997

CITATION READS
1 457

1 author:

Ryszard Gessing
Silesian University of Technology
297 PUBLICATIONS 499 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ryszard Gessing on 01 June 2016.

The user has requested enhancement of the downloaded file.

 STATE SPACE AND INTERNAL MODELS FOR LINEAR QUADRATIC
REGULATOR DESIGN
Ryszard Gessing
Silesian Technical University, Institute of Automatic Control,
ul. Akademicka 16, 44–101 Gliwice, Poland
fax: +4832371165, email: gessing @ ia.gliwice.edu.pl

Abstract— State space models are proposed for direct imple- model correctors corresponding to constant and/or sinusoidal
mentation of linear-quadratic regulator (LQR) with measured excitations. Further on, the modified LQR design is proposed
not all the state variables but only the output of the plant. making it possible to obtain a partially prescribed roots locus.
Using this and an appropriately chosen performance index the
regulator transfer function is derived which gives a partially Finally, the method is illustrated in an example.
prescribed roots locus of the closed loop (CL) system. It is The contribution of the paper is in the state space model
noted that usual LQR stabilization problem with zero set point proposal which together with applied internal model correc-
does not take into account the steady state error, when is tor and an appropriate performance index create the modified
used in the system with non-zero constant set point. Further LQR design method giving a partially prescribed roots locus
on, the case of nonzero both the set point and disturbance is
considered by using an appropriate internal model corrector. of the CL system.
Including this corrector to the augmented plant the considered
case is transformed to usual LQR stabilization problem with II. F IRST S TATE S PACE M ODEL
zero set point. Using proposed the state determination and
internal model corrector the modified LQR design technique, Consider the plant described by the strictly proper
giving a partially prescribed roots locus, is described. Finally, transfer function (TF):
an example illustrates the method.
Keywords. Linear-quadratic regulator; state space models; Y (s) b0 sl + b1 sl−1 + ... + bl
G(s) = = n (1)
internal models; system design; correctors. U (s) s + a1 sn−1 + ... + an
where l < n, Y (s) = L[y(t)], U (s) = L[u(t)], L is the
I. I NTRODUCTION symbol of the Laplace transform; y(t) and u(t) are the output
The linear quadratic regulator (LQR) is now a classical and input signals, and an 6= 0. Assume that the numerator
problem being the subject of many papers and books e.g. and denominator of (1) are relatively prime polynomials.
we remind here only two erly papers of Kalman (1960) and Determine the state variables in the form
of Letov (1960) and two contemporary books: of Anderson x1 = y (n−m−1) − γ0 u(l−m−1) − ... − γl−m−1 u
and Moore (1990) and of Dorato et al (1995). A compact x2 = y (n−m−2) − γ0 u(l−m−2) − ... − γl−m−2 u
recapitulation of the state of art concerning this problem is ...............................................................
given in the book of Dorato et al (1995). For regulator design xl−m = y (n−l) − γ0 u
the steady state solution of LQ problem corresponding to xl−m+1 = y (n−l−1)
infinite horizon is used, frequently (e.g. MATLAB lqr and ................................................................ (2)
lqry functions).
xn−m = y
It should be realized that usual LQR stabilization problem
xn−m+1 = y (−1) + δ1 u(−1)
takes only into account the transients resulting from initial
xn−m+2 = y (−2) + δ1 u(−2) + δ2 u(−1)
states with no external excitation. This means that the zero
................................................................
set point is then considered. The so called LQ tracking
xn = y (−m) + δ1 u(−m) + ... + δm u(−1)
problems addmit external excitations (non-zero set point or
disturbance) but have unrealistic assumption that they are
where m is an integer y (i) , u(i) are i-th derivative for i > 0
known in the whole control horizon.
and (−i)-th multiple integrals for i < 0 e.g. for (−i) < 0
In the present paper the original state space models are
proposed and used for derivation of the transfer function (TF) Z t Z t1 Z ti−1
of the regulator with using LQR technique. It is shown that u(−i) (t) = dt1 dt2 ... u(ti )dti . (3)
0 0 0
the solution of usual LQR stabilization problem, when used
to the case of non-zero constant set point does not take into Strictly speaking the derivatives y , u(i) may appear in the
(i)

account the steady state errors. The modified LQR problem first (n−m−1) rows, while the integrals y (−j) , u(−j) - in the
is formulated in which the internal model correctors are used. last m rows of (2). The constants γi , δi will be determined
It is shown the the integrator and/or oscillator are internal further on.
 After differentiating both sides of (2), using notation (2) where pi , i = 1, 2, ...m are dependent upon γi , δi and ki .
as well as resulting from (1) equation For the case 0 ≤ m < l the TF R(s) takes the form (12) in
which the term sm in the denominator disappears and m is
y (n−m) + ... + an−m y + an−m+1 y (−1) + ... replaced by l. In the latter case the TF (12) can be improper,
+an y (−m) = b0 u(l−m) + b1 u(l−m−1) + ... + bl u(−m) (4) e.g. if l = n − 1 and 0 ≤ m < l it has a derivative term
(−k1 /p1 )s. Note, that if l < n − 1 and 0 ≤ m < l then
we obtain the first state space model, n-dimensional in the the TF R(s) contains derivatives y (i) (t) of higher than first
form order. The implementation of this kind of regulator could be
ẋ = Ax + Bu, y = Cx (5) difficult. In this case, applying the state (2) with m > l and
integrals of y and u, we can obtain an implementable R(s)
where x = [x1 , x2 , ..., xn ]T , u, y are scalars,
even in the form of proper TF. For instance using the state
determined by (2) with m = n − 1 we obtain R(s) in the
 
−a1 −a2 ... −an−1 −an
 1
 0 ... 0 0 
 form of the proper TF (with (n − 1)-th order polynomials
A=  0 1 ... 0 0 
 (6) both in numerator and denominator).
 ... ... ... ... ...  The closed loop (CL) system composed of the plant (5)
0 0 ... 1 0 and feedback (11) is optimal in steady state and has n-
th order characteristic equation. However, then all the state
and B and C are column and row vectors, respectively, with
variables must be measured which usually is not possible.
components Bi and Ci , i = 1, 2, ..., n determined by
Note, that the CL system composed of the plant (1) and
Bi = γl−m+1−i , for 1 ≤ i ≤ l − m + 1 regulator R(s) obtained from substitution (2) into (11) is
Bi = 0, for l − m + 2 ≤ i ≤ n − m (7)
usually of higher order than n (if on the right-hand side of
Bi = δi−n+m , for n − m + 1 ≤ i ≤ n (2) the terms u(i) (t) appear). Therefore the latter, second
Cn−m = 1 and Ci = 0 for i 6= n − m system has usually worse performances than the first one.
The constants γi and δi are chosen so that the derivatives The first CL system is optimal in steady state (tf → ∞).
and integrals of u(t) are eliminated in (5). After derivations Then, there arises the question: if and in which sense the
we obtain second system is optimal? The response to this question will
be given in the two next sections.
γi = bi − a1 γi−1 − a2 γi−2 − ... − ai γ0 ,
IV. S ECOND S TATE S PACE M ODEL
i = 1, 2, ..., l − m, γ0 = b0 (8)
Consider the plant described by the TF (1). The
1 assumption an 6= 0 is now not needed. Determine the second
δi = − (bl−i+1 + an−i+1 δ1 + an−i+2 δ2 + ... state variables in the form
an
bl x1 = y (n−m−1) x2 = y (n−m−2) ... xn−m = y
+an−1 δi−1 ), i = 2, 3, ..., m, δ1 = − (9) xn−m+1 = y (−1) xn−m+2 = y (−2) ... xn = y (−m)
an
xn+1 = u(−1) xn+2 = u(−2) ... xn+m = u(−m)
III. R EGULATOR -O BSERVER TF.
(13)
From solving LQR problem described by (5)-(9) and
the performance index where, now m ≥ l and the notation used is the same as
Z tf previously.
I= (xT Qx + ru2 )dt (10) After differentiating both sides of (13), using notation
0 (13) and the equation (4), now we obtain the second
(where Q is a symmetric semipositive weighting matrix of state space model (5), (n + m)-dimensional with x =
the state, r is a small positive number and tf → ∞) the [x1 , x2 , ..., xn+m ]T , for which the formulas determining the
following feedback relation is obtained matrix A and vectors B, C are dependent upon relation
between m and l. For instance, for the case m = l it is
u = −kx = −k1 x1 − k2 x2 − ... − kn xn (11)  
−a1 −a2 ... −an b1 ... bl−1 bl
Here k is n-dimensional vector with constant components  1 0 ... 0 0 ... 0 0 
ki , i = 1, 2, ..., n which may be calculated using lqr
 
A= 0  1 ... 0 0 ... 0 0 
MATLAB function.

 ... ... ... ... ... ... ... ... 
The regulator-observer equation and corresponding TF can 0 0 ... 0 0 ... 1 0
be derived by substituting (2) to (11) and rearranging the
terms containing u(i) (t). For instance, for the case m ≥ l (14)
the regulator-observer TF takes the form Generally, in the case m ≥ l the elements Ai,j of the (n +
m) × (n + m) matrix A are determined as follows:
U (s) k1 sn−1 + k2 sn−2 + ... + kn A1,j = −aj , for 1 ≤ j ≤ n;
R(s) = − =− m (12)
Y (s) s + p1 sm−1 + ... + pm A1,j = 0, for n + 1 ≤ j ≤ n + m − l − 1,
 n + 1 ≤ n + m − l − 1; state LQR solution of the problem (5)-(9), (10). But, both
Ai,j = bj−n−m+l , for n + m − l ≤ j ≤ n + m, state space models, the first (5)-(9) and the second (5),
m > l; (14)-(16) describe the same plant (1). Let us distinguish,
Ai,j = 0, for 2 ≤ i ≤ n, j 6= i − 1; for a while, some notations used in both the problems,
Ai,i−1 = 1, for 2 ≤ i ≤ n; (15) namely, let x̌, Q̌, m̌ and x̂, Q̂, m̂ denote x, Q, m in the first
An+1,j = 0, for 1 ≤ j ≤ n + m; and second state space models and corresponding LQR
Ai,j = 0, for n + 2 ≤ i ≤ n + m, j 6= i − 1; problems (5)-(9), (10) and (5), (14)-(16), (10), respectively.
Ai,i−1 = 1, for n + 2 ≤ i ≤ n + m. For m̌ = m̂ = m ≥ l the regulators (12) and (18) have
The column and row vectors B and C with components Bi the same structure i.e. the numerator and denominator
and Ci , i = 1, 2, ..., n + m, respectively, are determined by polynomials of (12) and (18) have the same orders. If
additionally x̌T Q̌x̌ = x̂T Q̂x̂ then for both the problems the
B1 = b0 , f or m = l; B1 = 0, f or m > l plant, the performance index and the structure of resulting
Bi = 0, f or i ≥ 2, i 6= n + 1; Bn+1 = 1; regulators are the same.

Cn−m = 1; Ci = 0, f or i 6= n − m (16) Conclusion 2. If x̌T Q̌x̌ = x̂T Q̂x̂ and m̌ = m̂ = m ≥ l,

Note that both the state space models first (n-dimensional then both the regulators (12) and (18) are the same and
described in Section 2) and second (n + m-dimensional realize the optimal steady state LQR solution of the problem
described in this Section) have the same TF (1). The first (5), (14)-(16), (10) with given m̂ = m.
model is controllable and observable, while the second
controllable but not observable. Conclusion 3. If 0 ≤ m̌ ≤ l, m̂ = l and x̌T Q̌x̌ = x̂T Q̂x̂
From solving LQR problem described by (5), (14)-(16) , then both the regulators (12) and (18) are the same and
and (10) we obtain similarly as (11) realize the optimal steady state LQR solution of the problem
(5), (14)-(16), (10) with m̂ = l.
u = −kx = −k1 x1 − k2 x2 − ... − kn+m xn+m (17)
The second state space model is more friendly from
The regulator-observer TF results from substitution (13)
designing point of view since in it, more the state space
into (17) and has the form
components are determined directly by the output y its
k1 sn−1 + k2 sn−2 + ... + kn derivatives y (i) and integrals y (−j) . Therefore, further on if
R(s) = − (18) nothing opposite will be said, the second state space model
sm + kn+1 sm−1 + ... + kn+m
will be used with the state denoted simply by x.
A. Roots Locus Choice
Thus, for m = n − 1 the TF R(s) is proper, while for By means of the appropriate choice of the matrix Q
m = n − 2 it has additionally a derivative −k1 s. for the second state space model it is possible to establish
V. D ISCUSSION OF THE S OLUTIONS . an appropriate locus of n roots of CL system. Really, if we
Note that the CL system composed of the plant choose the matrix Q so that
described by the second state space model (5), (14)-(16) Q = f T f, f = [1, f1 , f2 , ..., fn , 0, ..., 0] (19)
and the feedback (17) realizes the optimal steady state
(where dimf = n + m + 1 and fi , i=1, 2, ..., n are
LQR solution. The system is of (n + m)-th order, i.e. its
appropriately chosen) then the performance index (10) takes
characteristic equation is of (n + m)-th order. However,
the form
the feedback (17) can be realized when all (n + m) Z tf
state variables can be measured. From the other side, the I= (ε2 + ru2 )dt, tf → ∞ (20)
regulator (18) results from substitution (13) to (17), and the 0
CL system composed of the plant (1) and regulator (18) is where
also of (n + m)-th order. The advantage is that the feedback ε = f x, ε2 = xT Qx (21)
regulator (18) has only one input signal y(t).
Substituting (13) and (19) into (21) and then differentiating
Conclusion 1. The CL system composed of the plant m-times both sides of (21) we obtain
(1) and regulator (18) realizes optimal steady state LQR y (n) + f1 y (n−1) + ... + fn y = ε(m) (22)
solution of the problem (5), (14)-(16), (10).
Minimisation of (20) especially for r → 0 means that ε(t)
and also ε(m) (t) tends to zero when t → ∞ . This means
Another situation is in the case of the CL system
that when r tends to zero then n roots of the characteristic
composed of the plant (1) and the regulator (12). Though,
equation of the CL system determined by the steady state
the latter regulator results from substitution (2) to (11) the
LQR solution of the problem (5), (14)-(16), (20) tend to the
resulting CL system is of (n + m)-th order, while the CL
roots of the equation
system (5)-(9), (11) is of n-th order. This means that the
CL system (1), (12) does not realizes the optimal steady sn + f1 sn−1 + ... + fn = 0 (23)
Thus, by means of an appropriate choice of the coefficients where x̄ corresponds to the constant input ū.
fi , i = 1, 2, ..., n, we can place n roots of the CL system, In the case of non zero set point w 6= 0 the performance
freely. index of corresponding LQR problem which we would like
to minimize takes the form (10)
Conclusion 4. The solution of the steady state LQR Z tf
problem (5), (14)-(16), (20) for r → 0 determines the CL I1 = (x∗T Qx∗ + ru2 )dt (25)
system for which n roots of the characteristic equations 0
tend to the roots of the equation (23). where, the components x∗i , i = 1, 2, ..., n + m, of the state
x∗ are determined by the formulas (13) in which the output
In applications one can use an appropriate lower order y is now replaced by the error e = y − w.
polynomial (22) e.g. of the second order establishing only It should be stressed that the use of the regulator resulting
two roots of the characteristic equation. from solving the steady state LQR problem (5), (14)-(16),
The stability of the CL system resulting from solving (20) with zero set point to the system with non zero set
the steady-state LQR problem (5), (14)-(16), (20) is a point is not justified. Really, substracting both side of the
necessary demand. For some reason we formulate the equations (20) from those of the equations (5) we obtain
stability condition exactly.
x̃˙ = Ax̃ + B ũ, ỹ = C x̃ (26)
Stability condition. The CL system resulting from where x̃ = x − x̄, ỹ = y − ȳ and ũ = u − ū. Consider the
solving the steady-state LQR problem (5), (14)-(16), (20) is performance index
asymptotically stable if the pair (A, B) is controllable and Z tf
the pair (A, f ) is observable [1, p, 2112 ].
I2 = (x̃T Qx̃ + rũ2 )dt, tf → ∞ (27)
0
We can say that the system (5), (14)-(16) is observable
The components x̃i , i=1, 2, ..., n+m, of the state x̃ are
by the control index (20) if and only if the pair (A, f ) is
determined by the formulas (13) in which the output y and
observable.
input u is replaced by the difference y − ȳ and u − ū ,
respectively.
Conclusion 5. The CL system resulting from solving the
The steady state LQR problem (26), (27), (19) is the same
steady-state LQR problem (5), (14)-(16), (20) is stable if the
as the problem (5), (10), (19). This means that the solution
system (5), (14)-(16) is observable by the control index (20).
of the problem (5), (10), (19) with w = 0 applied to the case
of w = const 6= 0 does not minimize the index (25), (19)
VI. M ODIFIED LQR P ROBLEM but the index (27), (19). Thus, if the solution of the steady
state LQR problem (5), (10), (19) is applied to a CL system
Till know, the CL system with zero set point was
with w̄ = const 6= 0 then, generally, the steady state error
considered. This results from the fact that usual steady-state
ē = ȳ − w̄ is not taken into account in the corresponding
LQR problem takes only into account the transients resulting
performance index (27) and should be checked, separately.
from initial states with zero external excitation.
B. LQR Problem with Internal Model
A. The Case of Non Zero Set Point
Consider the CL system shown in Fig. 1 in which the Consider the CL system shown in Fig. 1. Assume that
set point w̄ = const 6= 0 and the disturbance z = 0. The the signals w(t) and z(t) are different, but both fulfil the
plant is described by (5), (14)-(16). Assume, that for given following differencial equation
w̄ = const 6= 0 in steady state a constant output ȳ and input ũ(p) + d1 ũ(p−1) + ... + dp ũ = 0 (28)
ū appear which fulfil the plant state equations i.e.
This means that the equation (28) is fulfilled if we substitude
ũ = w(t) or ũ = z(t). The fact that the signals w and z
are different results from different initial conditions of the
equation (28) used for generating both the signals.
If for instance w1 = w̄ = const and z1 = Asin(ωt + ϕ)
then the equation (28) takes the form
∼(3) ∼(1)
u +ω 2 u =0 (29)
The signals w1 and z1 with any w̄ , A and ϕ are the solutions
of the equation (29) with appropriate initial conditions.
Consider the corrector described by the TF
U (s) 1
= C(s) = p (30)
x̄˙ = Ax̄ + B ū, ȳ = C x̄ (24) V (s) s + d1 sp−1 + ... + dp
which in denominator has the polynomial corresponding It is easy to note that for m = n − 1 the TF R(s) is proper
to the equation (28); V (s) = L[v(t)] is the input of the with polynomials of (n + p − 1) order both in numerator and
corrector. denominator, while for m = n − 2 R(s) has additionally a
The augmented plant described by the TF derivative −k1 s.
E(s)
= C(s)G(s) (31) VII. M ODIFIED LQR D ESIGN
V (s)
is of (n + p) order and has all the modes of the plant (1) and In the process of design the following steps are
of the corrector (30). Thus, the error e = y1 +z−w appearing important:
in the CL system shown in Fig.1 can be obtained directly as • the choice of the internal model (28) well suited to
the output of the augmented plant (31) for some appropriate the expected signals representing the set point and
initial conditions. This fact justifies the used notation E(s) = disturbance;
• the choice of the number m in the state x determination
∗
L[e(t)] for the augmented plant output.
Let by means of the modified formulas (13);
ẋ∗ = A∗ x∗ + B ∗ v, e = C ∗ x∗ (32) • the choice of the vector f which inluences the shape of
the performance index (33) and partially establishes the
be the state equation of the augmented plant (31) derived roots locus of the CL system.
by using the formulas (13)-(16). It means that the state
components x∗i are determined by the formulas (13) in which The internal model equation should be not to high order
the variables xi , y, u and n are replaced by x∗i , e, v and n+p, since it influences the order of the CL system. For different
respectively, and the elements of the (n+p+m)×(n+p+m) m from the interval (0, n − 1) the state x∗ determination
matrix A∗ and of the (n + p + m) vector B ∗ are determined with using the modified formulas (13) gives regulators of the
by the appropriately modified formulas (15), (16). different structures. Usually lower m gives lower order CL
Using the performance index system and its better performance connected unfortunately
Z tf with increasing order of derivative terms in the regulator.
I3 = (x∗T Q∗ x∗ + rv 2 )dt, tf → ∞ (33)
0 A. Plants with No Zeros in TF
T Consider the plant TF (1) which has all bi = 0 ,
where Q = f f , by the appropriate choice of the (n + p),
components of the (n + p + m) vector f we can establish except bn 6= 0. After including the corrector (30) to the
maximum (n + p − 1) roots of the CL system. augmented plant, creating the state space model (32) with
It is easy to note that the formulas (32), (33) determine m = 0, using LQR technique to the problem (32), (33), and
the usual steady state LQR stabilization problem with zero then including the corrector to the regulator, the following
set point. TF of the regulator is obtained
k1 sn+p−1 + k2 sn+p−2 + ... + kn+p
Conclusion 6. The problem with non zero both, the set R(s) = − (37)
point w and disturbance z which fulfil the internal model sp + d1 sp−1 + dp
equation (28) may be transformed to the usual steady state It is not an easy task to implement this kind of the
LQR stabilization problem (32), (33) with zero set point by regulator for higher order plants. However, note that for the
means of inclusion of the internal model corrector (30) into second order plant (n = 2) and C(s) = 1/s (this kind
the augmented plant (31) and using the second state space of internal model corrector corresponds to the case when
model (13). w = const and z = const) the PID regulator is obtained. For
any n, using the state x∗ determination (13) with m = n − 2
Thus, the considered problem with non zero w and z can an implementable regulator R(s) is obtained with only first
be solved by using lqr MATLAB function created for the order derivative.
case of zero w and z. Applying this function to the matrices
A∗ , B ∗ , Q∗ and r we obtain similarly as (11) B. Plants with Cancellable Zeros
v = −kx = ∗
−k1 x∗1 − k2 x∗2 − ... − kn+p+m x∗n+p+m (34) Consider the plants described by the TF (1) with stable
and not to oscillatory zeros which may be cancelled at the
The regulator-observer TF results from substitution the state
process of design. Including the following corrector
components x∗i determined by modified formulas (13) into
(34) 1
U (s) = C(s) V (s) (38)
V (s) k1 s n+p−1
+ k2 s n+p−2
+ ... + kn+p b0 sl + b1 sl−1 + ... + bl
R1 (s) = =− m
E(s) s + kn+p+1 sm−1 + ... + kn+p+m to the augmented plant, using LQR technique to this plant
(35) (which for m = 0 is the same as in Subsection 7.1) and then
Finally the regulator for the CL system which contains the including the corrector to the regulator we obtain
observer and the internal model corrector takes the form
U (s) k1 sn+p−1 + k2 sn+p−2 + ... + kn+p
R(s) = = C(s)R1 (s) (36) R(s) = −C(s) (39)
E(s) b0 sl + b1 sl−1 + ... + bl
 C. Plants with Non-Cancellable Zeros
Let N (s) = N + (s)N − (s) be the numerator of the
plant (1) and N + (s) and N − (s) are the polynomials con-
taining cancellable and non-cancellable zeros, respectively.
In this case the corrector included to the augmented plant
can have the form
1
U (s) = C(s) + V (s) (40)
N (s)
The augmented plant (with the corrector) has the TF with
numerator N − (s). Using state x∗ determination (13) with
appropriately chosen m and applying LQR technique to (32),
(33) we obtain the regulator R(s).
VIII. E XAMPLE
Consider the plant described by
10
G(s) = 2 (41) IX. C ONCLUSIONS
s + 3s + 2
We would like to design the regulator for which the CL The quadratic performance index applied in usual LQ
system shown in Fig.1 works accurately for any nonzero problems does not take into account the steady state errors
constant set point w = const 6= 0, and the sinusoidal appearing in the systems with non zero poles and non zero
disturbance z = Asin(2t + ϕ). constant set point and/or disturbance.
The internal model equation for these signals w and z The proposed state space models make it possible to derive
takes the form (29) with w = 2. The corresponding internal the regulator transfer function implementable in feedback
model corrector (30) takes the form control with measured not all the states but only the output
variable.
1
C(s) = 3 (42) The appropriate choice of the weighting matrix Q makes
s + 4s it possible to establish some of the roots of the characteristic
Including the corrector (42) to the plant (41), using the state equation of the CL system.
space model (32) corresponding to m = 0, assuming after The problem with non zero both the set point and the
trials f = [0, 0, 1, 1, 0.5] , r = 0.001 (which determines two disturbance fulfilling the internal model equation, may be
roots s1 = −0.5 + 0.5i, s2 = −0.5 − 0.5i) we obtain the transformed to the usual steady state LQR stabilization
regulator in the form problem with zero set point, by including the internal model
k1 s4 + k2 s3 + k3 s2 + k4 s + k5 corrector to the augmented plant and using the proposed
R(s) = − (43) second state space model. By this manner the solution of
s3 + 4s
latter problem can be used for designing the regulator for
where k1 = 1.9536, k2 = 24.94, k3 = 149.7064, k4 =
the system with nonzero set point and disturbance.
137.3628, k5 = 63.2456 were calculated using lqr MATLAB
The integrator and/or oscillator corrector is a special
function. The regulator (43) can be also described by
case of the internal model corrector. LQR technique with
R(s) = −1.9536s − 24.94−
using integrator-oscillator correctors makes it possible to
141.8928s2 + 37.6036s + 63.2456 design regulators working accurately in steady state for any
− (44)
s3 + 4s constant and sinusoidal set point and/or disturbance with
i.e. it has a derivative of the first order. The roots of the given frequency.
characteristic equation of the CL system (4), (44) are: s1 = Similar, modified LQR design method can be applied for
−10.7716, s2 = −5.3814 + 9.4060i , s3 = −5.3814 − discrete-time systems.
9.4060i, s4 = −0.5000 + 0.5000i, s5 = −0.5000 − 0.5000i. R EFERENCES
In Fig.2 the results of simulaitons performed with using
[1] Anderson, B.D.O. and J.B. Moore. (1990). Optimal Control: Linear
SIMULINK program are shown. In Fig. 2c the output Quadratic Methods. Englewood Cliffs, NJ: Printice Hall.
response to the set point w(t) = 1(t) (1(t) = 0 for [2] Dorato, P., C. Abdallah and V. Cerone. (1995) Linear-quadratic
t < 0, 1(t) = 1 for t ≥ 0) and the disturbance z(t) = Control, An Introduction. Printice Hall, New Jersey.
[3] Kalman, R.E. (1960) Contribution to the theory of optimal control
1(t − 3)Asin(2(t − 3)) is shown. In Fig 2a and b the Bol. Soc. Matem. Mex 5, pp. 102-119.
disturbance z(t) and the control u(t) are plotted, respectively. [4] Letov, A.M. (1960) Analytic controller design I, II. Autom. Remote
It is shown that beyond the transient interval the output y is Contr. 21, pp. 303-306.
[5] MATLAB (1993) User’s Guide, Natick, MA, Mathworks.
very accurate and the sinusoidal disturbance is compensated [6] SIMULINK (1993) Dynamic System Simulation Software, Users
by the control. In the transient interval the output is also Guide. The Math Works Inc.
satisfactory.

View publication stats