Article

Transactions of the Institute of

Measurement and Control
Receding horizon HN control of 1–10
Ó The Author(s) 2014

time-delay systems Reprints and permissions:

sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0142331214538089
tim.sagepub.com

Fatma Yıldız Tasxc

xıkaraoglu, Levent Ucun and
xükdemiral
Ibrahim B Küc

Abstract
This paper deals with the disturbance rejection problem for discrete-time linear systems having time-varying state delays and control constraints. The
study proposes a novel receding horizon HN control method utilizing a linear matrix inequality based optimization algorithm which is solved in each
step of run-time. The proposed controller attenuates disturbances having bounded energies on controlled output and ensures the closed-loop stability
and dissipation while meeting the physical control input constraints. The originality of the work lies on the extension of the idea of the well-known HN
receding horizon control technique developed for linear discrete-time systems to interval time-delay systems having time-varying delays. The efficiency
of the proposed method is illustrated through simulation studies that are carried out on a couple of benchmark problems.

Keywords
Time delay, predictive control, HN optimization

Introduction consideration and the controller parameters are updated in

every step. During the last four decades, the studies on RHC
Time-delay systems (TDSs) have been a popular and attrac- were mostly focused on ordinary linear time-invariant sys-
tive topic in the literature for many years, since the phenom- tems without delays. However, in recent years, some results
enon can be considered as one of the main reasons for on RHC for systems having time delays started to appear in
instability and poor performance in many physical and dyna- the literature, such as Houda et al. (2013), Jeong and Park
mical systems. Due to this, many researchers have paid atten- (2005), Kwon et al. (2003), Shi et al. (2008) and Zhang et al.
tion to the stabilization and control of this type of system. (2014), in which the proposed techniques are all based on
The time delay usually affects the system states, control delay-independent schemes. Therefore, they are considered to
inputs and state derivatives, which leads to different classes of be highly conservative.
time-delay systems (Hale, 1977; Mahmoud, 2000; Richard, The receding horizon HN control method was first intro-
2003). duced by Chen and Scherer (2006), in which an additional
The existing results in the literature are often grouped into constraint for the optimization problem based on linear
two classes, due to their dependence on the size of the delay, matrix inequalities (LMIs) was introduced in order to ensure
i.e. delay-independent methods and delay-dependent methods the dissipativity of the closed-loop system. This method was
where the latter is usually less conservative than the former. adapted to a time-delay system by Mei et al. (2009) in which
Therefore, most of the work in literature is devoted to delay- the time delay was assumed to be constant and a delay-
dependent controller design problems (Fridman and Shaked, independent controller design method was proposed.
2003; Huang and Xiang, 2014; Mahmoud and Ismail, 2005; Inspired by the lack of delay-dependent receding horizon
Moon et al., 2001; Parlakc xı and Kücxükdemiral, 2011). HN controllers for time-delay systems having time-varying
On the other hand, the receding horizon control (RHC) delays in the literature, a delay-dependent receding horizon
method, also known as model predictive control (MPC), is HN controller design method is proposed in this study, where
known to be an efficient control method when dealing with the time delay is assumed to be time-varying over a known
linear systems having input–output constraints and distur- interval. Moreover, the closed-loop stability, dissipativity and
bances. It is an advanced method of process control that has
been in use in the process industries in chemical plants and oil
refineries since the 1980s. Carried out for a finite control hori- Department of Control and Automation Engineering, Yıldız Technical
University, Turkey
zon, this celebrated method is used to solve optimization
problems at every step of the run-time by utilizing the current Corresponding author:
plant measurements, and the current dynamic state of the Levent Ucun, Department of Control and Automation Engineering, Yıldız
process, together with the mathematical model of the system. Technical University, Davutpasa, Istanbul 34220, Turkey.
Hence, the varying parameters in the system are all taken into Email: [email protected]

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

2 Transactions of the Institute of Measurement and Control

HN performance of the time-delay system is guaranteed by Receding horizon HN control of time-

the proposed scheme. On the other hand, control constraints delay systems
are also taken into account during the controller design.
The rest of the paper is organized as follows: The control In order to construct the optimization problem for receding
problem is stated in Section ‘‘Problem Definition’’. The main horizon HN control of time-delay systems, we shall present
outcomes of the study about receding horizon HN control are the asymptotic stability condition for the system in equation
presented in detail in Section ‘‘Receding Horizon HN Control (4), the bounding constraint for state trajectories and the dis-
of Time-delay Systems’’. In Section ‘‘Simulation Studies’’, the sipation condition, respectively. Finally, an algorithm which
efficiency of the proposed theory and algorithm is demon- can be used in real-time control is also proposed.
strated through a couple of benchmark problems which con-
sist of retarded-type time-delay systems. Finally, the results
are discussed and possible future studies are addressed in Stability in receding horizon control
Section ‘‘Conclusion’’.
The theorem associated with the HN state feedback control
law u(k) = Kx(k) which assures the asymptotic stability of
Problem definition the closed-loop time-delay system given in equation (4) is
introduced next.
Let us consider a class of linear discrete time-delay system of
the form Theorem 1. At each sample k of run-time, given positive inte-
gers dmin, dmax and the positive real number g, the HN state
xðk + 1Þ = Axðk Þ + Ad xðk  d ðk ÞÞ + Bw wðk Þ + Bu uðk Þ feedback control law u(k) = K(k)x(k) with K(k) =
zðk Þ = Cxðk Þ + Cd xðk  d ðk ÞÞ + Dw wðk Þ + Du uðk Þ ð1Þ L(k)Y21(k) asymptotically stabilizes the system in equation
xðk Þ = fðk Þ, k 2 ½dmax , 0 (4) with a disturbance attenuation level of g, if there exists
matrices Y (k) = YT (k), W(k) = WT (k) . 0 and L(k) with
where x(k) 2 Rn , u(k) 2 Rm , z(k) 2 Rp , F(k) 2 Rn and appropriate dimensions which satisfy
v(k) 2 Rl stand for the state vector, control signal, controlled 2 3
output, the vector of initial conditions, and disturbance, Y (k)     
respectively. The control signal applied to the system also ful- 6 0 W (k)     7
6 7
6 0 0 g2 I    7
fils the constraint given as 6 7.0
6 AY (k) + Bu L(k) Ad W (k) Bw Y (k)   7
6 7
4 CY (k) + Du L(k) Cd W (k) Dw 0 I  5
kui (k)k‘  u2i, max , 8k  0, i = 1, . . . , m ð2Þ 1
Y (k) 0 0 0 0 dm W (k)
where ui,max is a known magnitude bound on the control ð5Þ
effort in the ith control channel. Also, the time-varying delay
d(k) satisfies where dmbdmax 2 dmin + 1.
dmin  d(k)  dmax 8k  0 ð3Þ Proof. Let us choose a Lyapunov–Krasovskii functional can-
didate as
where dmin and dmax are the known lower and upper bound of
the delay, respectively. V (x(k)) = V1 (k) + V2 (k) + V3 (k) ð6Þ
Here, the known interval for the time delay term corre-
sponds to the upper and lower bounds of the time delay. The where
time-delay interval including the upper and lower bounds is
highly related to the dynamics of the system. By using V1 (x(k)) = xT (k)P(k)x(k),
advanced system identification and modelling techniques, it is
possible to calculate the lower and upper bounds of the time X
k1
V2 (x(k)) = xT (s)Q(k)x(s),
delay, which leads us to a pre-defined time-delay interval. s = kd(k)
Our design goal is to find a feedback control law of the dX
min + 1 X
k 1
form, u(k) = K(k)x(k) for each step k, which ensures the min- V3 (x(k)) = xT (s)Q(k)x(s)
imum achievable HN gain for the closed system i = dmax + 2 s = k + i1

xðk + 1Þ = Acl xðk Þ + Ad xðk  d ðk ÞÞ + Bw wðk Þ Considering the closed-loop system in equation (4), the for-
zðk Þ = Ccl xðk Þ + Cd xðk  d ðk ÞÞ + Dw wðk Þ ð4Þ ward difference of energy functional V (x(k)) at step k is
xðk Þ = 0, k 2 ½dmax , 0 defined as

D
where Acl = A + BuK(k), Ccl = C + DuK(k) and guaran- DV (k) ¼ DV1 (k) + DV2 (k) + DV3 (k) ð7Þ
tees the stability of the closed-loop system under the con-
straints in equations (2) and (3). where

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

Tasxxıkaraoglu
c et al. 3

P kdmin
DV1 (k) = xT (k + 1)P(k)x(k + 1)  xT (k)P(k)x(k) and ss = T
= kdmax + 1 x (s)Q(k)x(s) can be used in equation (9) in
= xT (k)(ATcl P(k)Acl  P(k))x(k) order to replace and bound the term

+ 2xT (k)ATcl P(k)Ad x(k  d(k))

X
k X
k
+ 2xT (k)ATcl P(k)Bw w(k) xT (s)Q(k)x(s)  xT (s)Q(k)x(s)
s = kd(k + 1) + 1 s = kd(k) + 1
+ x (k  d(k))ATd P(k)Ad x(k  d(k))
T

+ 2xT (k  d(k))ATd P(k)Bw w(k) which leads us to equation (10).

+ wT (k)BTw P(k)Bw w(k) ð8Þ Finally
!
and dX
min +1 X
k X
k1
T T
DV3 = x (s)Q(k)x(s) x (s)Q(k)x(s)
X
k i=dmax +2 s=k +i s=k +i1
T
DV2 (k) = x (s)Q(k)x(s) dX
min +1  
s = kd(k + 1) + 1 = xT (k)Q(k)x(k))xT (k +i1)Q(k)x(k +i1)
X
k1 i=dmax +2
 xT (s)Q(k)x(s)
=(dmax dmin )xT (k)Q(k)x(k)
s = kd(k)
X
kdmin
= x (k)Q(k)x(k)  xT (k  d(k))Q(k)x(k  d(k))
T
 xT (s)Q(k)x(s) ð12Þ
X
k1 i=kdmax +1
+ xT (s)Q(k)x(s)
s = kd(k + 1) + 1
is obtained. In view of equations (8), (9) and (12), one can
X
k1
write
T
 x (s)Q(k)x(s) ð9Þ
s = kd(k) + 1
DV (k) + zT (k)z(k)  g2 wT (k)w(k)
ð13Þ
However, DV2(k) relies on the exact knowledge of the  (k) + zT (k)z(k)  g2 wT (k)w(k)
 DV
delay term d(k+1). Therefore, in order to get rid of the term
d(k+1) in equation (9), one can use the following bound where
DV2 (k)  xT (k)Q(k)x(k)  xT (k  d(k))Q(k)x(k  d(k))
 (k) = xT (k)(AT P(k)Acl  P(k))x(k)
DV cl
X
kdmin
ð10Þ
+ T
x (s)Q(k)x(s) + 2xT (k)ATcl P(k)Ad x(k  d(k))
s = kdmax + 1
+ 2xT (k)ATcl P(k)Bw w(k)
Note that this inequality is obtained with the help of the fol- + xT (k  d(k))ATd P(k)Ad x(k  d(k))
lowing identity + 2xT (k  d(k))ATd P(k)Bw w(k)
X
k X
k + wT (k)BTw P(k)Bw w(k)
xT (s)Q(k)x(s) = xT (s)Q(k)x(s) + xT (k)Q(k)x(k)  xT (k  d(k))Q(k)x(k  d(k))
s = kd(k + 1) + 1 s = kdmin + 1
X
kdmin
X
kd min
+ xT (s)Q(k)x(s)
+ xT (s)Q(k)x(s)
s = kdmax + 1
s = kd(k + 1) + 1

X
k + (dmax  dmin )xT (k)Q(k)x(k)
 xT (s)Q(k)x(s) X
kdmin

s = kd(k) + 1  xT (s)Q(k)x(s) ð14Þ

s = kdmax + 1
X
kd min

+ xT (s)Q(k)x(s)
s = kdmax + 1 is an upper bound for DV(k). Hence, the control law
ð11Þ satisfying

Therefore  (k) + zT (k)z(k)  g 2 wT (k)w(k)\0

DV ð15Þ

X
k X
k
for each step k  0 step, asymptotically stabilizes the system
xT (s)Q(k)x(s)  xT (s)Q(k)x(s)
in equation (4) for v(k) = 0 and satisfies the
s = kd(k + 1) + 1 s = kd(k) + 1
kz(k)k22  g 2 kv(k)k22 performance when v(k) 2 l2[0, N).
s =X
kdmin
 xT (s)Q(k)x(s) Rearranging equation (15), one obtains
s = kdmax + 1

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

4 Transactions of the Institute of Measurement and Control
2 3
xT (k)
6 T 7
4 x (k  d) 5
wT (k)
2 T 32 3
Acl P(k)Acl  P(k) + dm Q(k) + CclT Ccl ATcl P(k)Ad + CclT Cd ATcl P(k)Bw + CclT Dw x(k)
6 76 7
4 ATd P(k)Acl + CdT Ccl ATd P(k)Ad  Q(k) + CdT Cd ATd P(k)Bw + CdT Dw 54 x(k  d) 5\0 ð16Þ
BTw P(k)Acl + DTw Ccl BTw P(k)Ad + DTw Cd BTw P(k)Bw + DTw Dw  g2 w(k)

Then, employing the Schur complement on the multiplier if there exists matrices Y(k) = YT (k) . 0 and W(k) = WT
of equation (16) allows us to obtain (k) . 0 with appropriate dimensions which solve
2 3
r        
6 x(k) Y (k) 0  0  0 0 0 7
6 7
6 x(k  1) 0 1
dm W (k)    0  0 0 0 7
6 7
6 .. .. .. .. .. .. .. .. .. 7
6 . . . . . . . . . 7
6 7
6 x(k  dmin )  1
 0 7
6 0 0 dm W (k) 0 0 70 ð20Þ
6 .. .. .. .. .. .. .. .. .. 7
6 . . . . . . . . . 7
6 7
6 x(k  d + 2) 0 0  0  31 W (k) 0 0 7
6 max 7
4 x(k  d + 1) 0 0  0  0 21 W (k) 0 5
max
x(k  dmax ) 0 0  0  0 0 W (k)

2 3 where dmbdmax 2 dmin + 1.

dm Q(k)  P(k)    
6 0 Q(k)    7
6 7 Proof. Employing the Lyapunov–Krasovskii functional intro-
6 g2 I   7
6 0 0 7\0 ð17Þ duced in equation (6), the ellipsoid in equation (19) can be
4 Acl Ad Bw P1 (k)  5
Ccl Cd Dw 0 I rewritten as

Ultimately, replacing Acl with A + BuK, Ccl with X

k1

C + DuK, utilizing the definitions Y (k) := P21(k), W(k) := r  xT (k)P(k)x(k)  xT (s)Q(k)x(s)

s = kd(k)
Q21(k) and L(k) := K(k)P21(k) in (17) and pre- and post- ð21Þ
dX
min + 1 X
k1
multiplying equation (17) with diag {P21(k), Q2 1(k), I, I, I} T
 x (s)Q(k)x(s)  0
leads to
i = dmax + 2 s = k + i1
2 3
dm Y T (k)W 1 (k)Y (k)  Y (k)    
6 0 W (k)    7 Employing a bounding such as
6 7
6 g 2 I  7
6 0 0  7\0
4 AY (k) + Bu L(k) Ad W (k) Bw Y (k)  5 P
k1 P
k1
CY (k) + Du L(k) Cd W (k) Dw 0 I xT (s)Q(k)x(s)  xT (s)Q(k)x(s)
s = kd(k) s = kdmin
kdP
ð22Þ
min 1

where the Schur complement formula is employed for the + xT (s)Q(k)x(s)

s = kdmax
dmYT (k)W21(k)Y (k) 2 Y (k) term, which leads to equation
(5). This concludes the proof.
and plugging into equation (21) results in

Bounding the state trajectories X

k 1 X
kdmin 1

r  xT (k)P(k)x(k) xT (s)Q(k)x(s) + xT (s)Q(k)x(s)

s = kdmin s = kdmax
Lemma 1. Let us assume that there exists a solution (g, Y, W, dX
min + 1 X
k 1
X) to the stability condition given in equation (5) and the input  xT (s)Q(k)x(s)  0
constraint such as given in Boyd et al. (1994) i = dmax + 2 s = k + i1

  ð23Þ
1=rX L
 0, Xii  u2i, max ð18Þ
LT Y
which is a sufficient condition for equation (21) and lets us
avoid the use of the unmeasurable term, d(k). Then, the LMI
Given the positive constant r, the state trajectory starting from
given in equation (20) is readily obtained if the Schur comple-
x(0) always lies within the ellipsoid
ment formula is employed successively on the inequality in
e(P, r) :¼ fx 2 <n jV (x(k))  rg 8 k  0 ð19Þ equation (23). h

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

Tasxxıkaraoglu
c et al. 5

Dissipation condition p(k) = p(k  1)  xT (k)½P(k  1)  P(k)x(k)

The dissipativity of a system is related to the input–output X
k1
relationship and can be defined as follows: q(k) = q(k  1)  xT (s)Q(k  1)x(s)
s = kdmin

Definition 1. If a system with supply function

X
k1
 xT (s)Q(k)x(s)
2 2 2
s(w(k), z(k)) :¼ g kw(k)k  kz(k)k kdmin

dX
min + 1 X
k1
has a storage function V (x(k)) which satisfies  xT (s)½Q(k  1)  Q(k)x(s) ð27Þ
j = dmax + 2 s = j + k1
V (x(k)) + (s(w(k), z(k))  V (x(k + 1)) 8k  0 ð24Þ
Proof. By rearranging the dissipation inequality with respect
then the system is called dissipative and equation (24) is called
to i = 0, 1, ., k, one can obtain
the dissipation inequality.
Theorem 2. Let us assume that there exists a solution (g, Y, V (i + 1) + kzi k2  g 2 kwi k2 + V (i) ð28Þ
W, X) such that the stability condition, input constraint and
state constraint given in equations (5), (18) and (20) are all On the other hand, above-mentioned Lyapunov–Krasovskii
satisfied. Then, the HNfeedback control law u(k) = functional can be rewritten as
L(k)Y21(k)x(k) guarantees the dissipativity of the system in
equation (4) with a disturbance attenuation level of g and holds X
i1
the dissipation inequality V (i) = xT (i)P(i)x(i) + xT (s)Q(i)x(s)
s = id(i)
k 
X 
kz(i)k2  g 2 kw(i)k2  xT (0)P(0)x(0) dX
i=0
min + 1 X
i1
+ xT (s)Q(i)x(s) ð29Þ
X
1 dX
min + 1 X
1
j = dmax + 2 s = i + j1
T T
+ x (s)Q(0)x(s) + x (s)Q(0)x(s)
s = d0 j = dmax + 2 s = j1
If the equation (29) is used in equation (28) for i = 0, 1, ., k
ð25Þ
and the negative terms in the inequality are neglected due to
the assumptions P(k) . 0 and Q(k) . 0, one can obtain
if there exists positive matrices Y(k) = YT (k) . 0,
W(k) = WT (k) . 0 and L(k) with appropriate dimensions
that solve
2 3
O(k)       
6 x(k) Y (k) 0  0  0 7 0
6 7
6 x(k  1) 0 1
dm W (k)  0  0 7 0
6 7
6 .. .. .. .. .. .. .. 7 ..
6 . . . . . . . 7 .
6 70 ð26Þ
6 x(k  dmin ) 0 0 1
   dm W (k)  0 0 7
6 7
6 .. .. .. .. .. .. .. .. 7
6 . . . . . . . . 7
6 7
4 x(k  d + 2) 0 0  0  21 W (k) 0 5
max
x(k  dmax + 1) 0 0  0  0 W (k)

k 
X 
where dmbdmax 2 dmin + 1 and
kz(i)k2  g2 kw(i)k2  xT (0)P(0)x(0)
i=0
D T
O(k) ¼ p(0) + q(0)  p(k  1)  q(k  1) + x (k)P(k  1)x(k) X
k
 xT (i)½P(i  1)  P(i)x(i)
X
k1
i=1
T
+ x (s)Q(k  1)x(s)
s = kdmax X
1 X
k X
i1
+ xT (s)Q(0)x(s)  xT (s)½Q(i  1)  Q(i)x(s)
dX
min + 1 X
k1
s = d0 i = 1 s = idi
+ xT (s)Q(k  1)x(s)
dX
min + 1 X
1
j = dmax + 2 s = j + k1
+ xT (s)Q(0)x(s)
j = dmax + 2 s = j1
in addition to the dissipation level variables which are given X
k dX
min + 1 X
i1
inside O in equation (26) and defined as  x(s)T ½Q(i  1)  Q(i)x(s) ð30Þ
i = 1 j = dmax + 2 s = j + i1

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

6 Transactions of the Institute of Measurement and Control

Hence, in view of and the inequality in equation (35), it is necessary to expand

the summations in equation (35). With the help of the expan-
X
k sion, it is revealed that each term obtained by expanding the
xT (i)(P(i  1)  P(i))x(i) summation terms in equation (35) is exactly same as the quad-
i=1
ratic terms which are obtained by applying the Schur comple-
X
k X
i1
ment formula to equation (26). This concludes the proof. h
+ xT (s)½Q(i  1)  Q(i)x(s) ð31Þ
i = 1 s = idi Theorem 3. Solving the following optimization problem at each
X
k dX
min + 1 X
i1 step, allows us to calculate the optimal HNcontroller for the
+ xT (s)½Q(i  1)  Q(i)x(s)  0 system given in equations (1)–(3).
i = 1 j = dmax + 2 s = j + i1

min g 2
the dissipativity of the closed-loop system is assured. L, Y , X , W
ð36Þ
However, the inequality in equation (31) depends on the exact subject to equations (5), (18), (20) and (26)
knowledge of the delay term d(i). In order to avoid the depen-
dence, we can use a lower bound of the form Then the feedback gain K(k) = L(k)Y21(k) is determined
by solving the optimization problem given in equation (36) at
X
k X
i1 each step of run-time. The control signal u(k) = K(k)x(k)
xT (s)½Q(i  1)  Q(i)x(s) which is obtained by using the feedback gain K(k) is applied
i = 1 s = idi to the system, and the values of the p(k) and q(k) variables
ð32Þ
X
k X
i1 are calculated and then used at the next step. Note that the
 xT (s)½Q(i  1)  Q(i)x(s) gain of the feedback controller varies at every step, due to the
i = 1 s = idmin
x(k) term that exists in equations (20) and (26). The LMI in
Therefore, equation (31) can be rearranged as equation (26), which involves the Y (k) and W(k) variables
and the values of P(k 2 1) and p(k 2 1), is neglected at the
initial step, k = 0. The derivation of the LMI in equation
X
k
(18) is omitted in this study since it has already been intro-
xT (i)(P(i  1)  P(i))x(i)
i=1 duced in previous studies such as Boyd et al. (1994) and Chen
X
k X
i1 and Scherer (2006). The derivation of the LMI in equation
+ xT (s)½Q(i  1)  Q(i)x(s) ð33Þ (5) satisfying the stability condition, the LMI in equation (20)
i = 1 s = idmin which bounds the state trajectories, and the LMI in equation
X
k dX
min + 1 X
i1 (26) that guarantees the dissipativity of the closed-loop sys-
+ xT (s)½Q(i  1)  Q(i)x(s)  0 tem are given in Sections 3.1, 3.2 and 3.3, respectively. Next
i = 1 j = dmax + 2 s = j + i1 we introduce an algorithm which can be used in a real-time
application.
With the help of the definition given in equation (27), we
obtain
Algorithm
X
k 1
Step 1: Given r0, dmax and dmin, optimal g is obtained by sol-
p(k  1) = p(0)  xT (i)½P(i  1)  P(i)x(i)
i=1 ving the optimization problem in equation (36) in view of the
X
k X
i1 LMIs in equations (5), (18) and (20). First, a sufficiently high
q(k  1) = q(0)  xT (s)½Q(i  1)  Q(i)x(s) initial value is assigned to g and the value of g is updated as
i = 1 s = idi g = g 2 dg as long as the optimization problem in equation
X
k dX
min + 1 X
i1 (36) has a feasible solution set. Then, the minimum feasible
 xT (s)½Q(i  1)  Q(i)x(s) value of g is assigned as gopt. Utilizing the feasible solution
i = 1 j = dmax + 2 s = j + i1 set to the optimization problem, the LMI variables P(0),
ð34Þ Q(0), the feedback controller gain K(0) = L(0)Y (0), dissipa-
tion levels p(0) = x(0)TP(0)x(0) and
If the Schur complement formula is applied to equation (26)
X
1 dX
min + 1 X
1
p(0)  p(k  1) + xT (k)½P(k  1)  P(k)x(k) q(0) = xT (s)Q(0)x(s) + xT (s)Q(0)x(s)
s = dmin s = dmax + 2 s = j1
X
k 1
+ q(0)  q(k  1) + xT (s)½Q(k  1)  Q(k)x(s)
s = kdmin are calculated in order to be used in the next step. Jump to
dX
min + 1 X
k1 Step 3.
+ xT (s)½Q(k  1)  Q(k)x(s)  0 ð35Þ
j = dmax + 2 s = j + k1 Step 2: Given r0, dmax and dmin, optimal g is obtained by sol-
ving the optimization problem in equation (36) with the
is deduced, and utilizing the terms given in equation (34) in LMIs in equations (5), (18), (20) and (26). The value of g is
equation (35), we obtain equation (33). Here, in order to point decreased as g = g 2 dg as long as the optimization prob-
out the equivalence of the Schur complement of equation (26) lem in equation (36) can be solved. The value of r is increased

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

Tasxxıkaraoglu
c et al. 7

in the case where there is no solution for the optimization

50
problem in equation (36) and then g minimization is re-initi-
ated. g(k) = g + dg is obtained for the first value of g 40
where there is no solution to the optimization problem in
30
equation (36). Finally, P(k), Q(k), the feedback controller

γ
gain K(k) = L(k)Y (k), dissipation levels p(k) and q(k) are 20
calculated by the help of obtained g(k). Then, go to Step 3.
10
Step 3: Control signal u(k) = K(k)x(k) is applied to the sys-
tem. Then, go to Step 2. 0
0 10 20 30 40 50 60 70 80 90
k

Simulation studies Figure 1. The variation of the disturbance attenuation level of the
In order to demonstrate the application of the proposed algo- system.
rithm, two numerical examples are introduced in this section.
The simulations were carried out under MATLAB using
YALMIP with the SEDUMI solver. 6

Example 1. Let us consider the following linear time-delay sys- 4

tem given in Tang et al. (2008)
2

e(k)
x(k + 1) = Ax(k) + Ad x(k  d) + Bw w(k) + Bu u(k)
0
y(k) = Cx(k) ð37Þ
−2
x(k) = f(k), k 2 ½dmax , 0
−4
where 0 10 20 30 40 50 60 70 80 90
k
     
1 0:1 0:1 0 0
A= ; Ad = ; Bu = Figure 2. The variation of the controlled output defined to be the
0:2 0:9 0 0:1 0:1 error between the desired reference and system output.
 
0:1
Bw = ; C = ½ 0 20 ; f(k) = ½ 0 0:5 T
0:1

Different from the proposed method in this note, the considered

system is subject to time-invariant state delay. On the other
hand, the disturbance vector w(k) has the dynamics

y ðk + 1Þ = Gy ðk Þ
ð38Þ
wðk Þ = Dyðk Þ

where
 
0:95 0:18
G= ; D = ½ 1 0:1 ; y(0) = ½ 0 1 T
0:18 0:9 Figure 3. The control signal u(k) applied to the system.

Also, the desired reference output yref(k) is given by the dynamic and umax is selected as 2. By choosing r0 = 34 and dmax =
system governed by dmin = 8, the optimization algorithm for the proposed control-
ler is carried out and g is calculated as 47.01 at the first step.
t ðk + 1Þ = Ft ðk Þ Then, by solving the optimization problem in equation (36)
ð39Þ
yref ðk Þ = Htðk Þ with the LMIs in equations (5), (18), (20) and (26) recur-
sively for the given example, the controlled output z and control
with signal u are obtained. Also, the variation of the disturbance
attenuation level g is found to be as illustrated in Figure 1. The
   T
0:9 0:1 0:2 controlled output and control signal are illustrated in Figures 2
F= ; H= ; t(0) = ½ 1 0 T
0:2 0:8 0 and 3, respectively. The plots confirm that the proposed reced-
ing horizon HN control (RHHC). method has a satisfactory
The controlled output z(k) is defined as the output trajectory performance for the system given above. Figure 2 shows that
error e(k) converges to zero very rapidly with the constrained control
effort shown in Figure 3. Hence, it can be concluded that the
e(k) = yref (k)  y(k)

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

8 Transactions of the Institute of Measurement and Control

Figure 4. Schematic diagram of the industrial electric heater.

RHHC algorithm is very effective for state-delayed systems

having state and input constraints.
Example 2. An industrial electric heater which is defined with a
fifth-order state space model is considered in this example. The
model of the heater is given in detail in Chu (1995) and Leite
et al. (2009) and involves five heating zones, each equipped with
their own thermocouple and electric heater. The temperatures in
each zone are state variables (x1,.,x5) and the electrical cur-
rent signals applied to each zone of the heater are control inputs
(u1,.,u5), which are shown in Figure 4.
A state-delayed nominal discrete model for this system was
obtained in Chu (1995) in the form

x(k + 1) = Ax(k) + Ad x(k  d) + Bu u(k) ð40Þ

Figure 5. The time-varying delay d(k).
where the system matrices are calculated as
2 3
0:97421 0:15116 0:19667 0:05870 0:07144 2 3T
6 0:01455 0:88914 0:26953 0:11866 0:22047 7 0:2 0 0 1 0
6 7 6 7
A=6
6 0:06376 0:12056 1:00049 0:03491 0:02766 7
7 Bw = 4 0 1 0 0 0:2 5 ;
4 0:05084 0:09254 0:28774 0:82569 0:02570 5 0 0 1 0:2 0
0:01723 0:01939 0:29285 0:03544 0:87111 2 3
0:03 0:03 0 0 0
2 3 6 7
0:01000 0:08837 0:06989 0:18874 0:20505 C=4 0 0 0:03 0:03 0 5;
6 0:02363 0:03384 0:05282 0:09906 0:00191 7 0 0 0 0 0:03
6 7 2 3
Ad = 6
6 0:04468 0:00798 0:05618 0:00157 0:03593 7
7 0:3 0:3 0 0 0
4 0:04082 0:01153 0:07116 0:16472 0:00083 5 6 7
0:02537 0:03878 0:04683 0:05665 0:03130 Cd = 4 0 0 0:3 0:3 0 5;
0 0 0 0 0:3
2 3 2 3
0:53706 0:11185 0:09978 0:04652 0:25867 0 0 0:01 0 0:01
6 0:51718 0:73519 0:57518 0:40668 0:12472 7 6 7
6 7 Du = 4 0:1 0 0 0 0 5;
Bu = 6
6 0:29469 0:31528 1:16420 0:29922 0:23883 7
7
4 0:20191 0:19739 0:41686 0:66551 0:11366 5 0 0:01 0 0 0
0:11835 0:16287 0:20378 0:23261 0:36525 Dw = diagf0:1, 0:2,  0:4g
and the state and control vectors are defined as is employed in this study in order to prove the efficiency of the
 T
proposed controller. Different from Chu (1995) and Leite et al.
x = T1  T 1 T2  T 2 T3  T 3 T4  T 4 T5  T 5 (2009), the time-delay in the states given in Figure 5, is
assumed to be time-varying such as dmin  d(k)  dmax.
u = ½ u1  u1 u2  u2 u3  u3 u4  u4 u5  u5 T On the other hand, the disturbance vector is defined as
w(k) = ½w1 (k) w2 (k) w3 (k)T where
where Ti and ui (i = 1,2,3,4,5) are the temperature and control
electric current in each zone, and Ti and ui are their operating 0:1 when 10  k . 0
points. In addition to the nominal discrete model given in Chu w1 (k) =
0 when k . 10
(1995), Leite et al. (2009) take the presence of disturbance and
uncertainties due to changes in the operation points into consid- 0:2 when 10  k . 0
eration. Similar to Leite et al. (2009), a performance output w2 (k) =
0 when k . 10
z(k)which is in the form of equation (1) with

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

Tasxxıkaraoglu
c et al. 9

Figure 6. The states x(k) of the system. Figure 8. The control signal u(k) applied to the system.

state-delayed systems has been addressed in this study. The

main contributions of this note to the literature can be stated
as follows: First, different from the existing ones, the pro-
posed method is based on delay-dependent LMI conditions
and is therefore less conservative than its counterparts.
Secondly, different from the existing literature, this note con-
siders discrete-time state-delayed systems having time-varying
and interval delays, which is much more realistic than
constant-delayed systems. Moreover, thanks to the iterative
structure of the proposed method, the controller can adopt to
varying parameters such as disturbance and control input
constraints and effectively control time-delay systems in the
presence of constrained inputs and disturbances, especially
when the system is subject to a sufficiently large disturbance.
Figure 7. The performance outputs z(k) of the system.
The efficiency of the proposed approach has been validated
on a couple of numerical benchmark examples borrowed
from literature. The authors believe that the proposed
0:05 when 10  k . 0 method can be expanded to time-delay systems having para-
w3 (k) =
0 when k . 10 metric uncertainties as a future study, which are highly
related to the robustness of the method.
The control signal u(k) is bounded as jui(k)j  1, 8k  0,
i = 1,2,.,5 and the initial conditions are assumed to be all
equal to zero. Choosing r0 = 1, g(k) = 1.8708 is obtained for
dmin = 10 and dmax = 20. The simulation results of the states Funding
x(k), performance outputs z(k) and control inputs u(k) are given This research was supported by Yıldız Technical University
in Figures 6–8, respectively. Scientific Research Projects Coordination Department (grant
Similar to the previous example, the plots of x(k), z(k) and number 2011-04-04-KAP01).
u(k) confirm that the proposed RHHC method has a satisfac-
tory performance for the industrial electric heater system.
Different from the previous example, the proposed method is References
employed for the control of a system with time-varying delay.
Boyd S, Ghaoui LE, Feron E and Balakrishnan V (1994) Linear
Figures 6 and 7 clearly demonstrate the efficiency of the pro- Matrix Inequalities in System and Control Theory. Philadelphia:
posed RHHC algorithm when the system is subject to distur- SIAM Publishing.
bance w(k) which is given above. Moreover, the results illustrate Chen H and Scherer CW (2006) Moving horizon HN control with
that the proposed method results in satisfactory performance performance adaptation for constrained linear systems. Automa-
for the systems involving time-varying delay. Also note that the tica 42(6): 1033–1040.
control effort applied to the system is constrained, as illustrated Chu J (1995) Application of a discrete optimal tracking controller to
in Figure 8. an industrial electric heater with pure delays. Journal of Process
Control 5(1): 3–8.
Fridman E and Shaked U (2003) Delay-dependent stability and HN
Conclusion control: Constant and time-varying delays. International Journal
of Control 76(1): 48–60.
A systematic approach to obtaining a delay-dependent reced- Hale J (1977) Theory of Functional Differential Equations. Berlin:
ing horizon HN controller design method for discrete-time Springer.

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016

10 Transactions of the Institute of Measurement and Control

Houda BM, Nizar A and Said NA (2013) Predictive sliding mode Mei L, Chengbo J and Huihe S (2009) Receding horizon HN control
control for perturbed discrete delay time systems: Robustness for constrained time-delay systems. Journal of Systems Engineering
analysis. In: 2013 international conference on electrical engineering and Electronics 20(2): 363–370.
and software applications (ICEESA), Hammamet, Tunisia, 21–23 Moon YS, Park P, Kwon W and Lee Y (2001) Delay-dependent
March 2013, pp. 1–6. Piscataway: IEEE Press. robust stabilization of uncertain state-delayed systems. Interna-
Huang S and Xiang Z (2014) Robust reliable control of uncertain 2D tional Journal of Control 74(14): 1447–1455.
discrete switched systems with state delays. Transactions of the xı A and Küc
Parlakc xükdemiral IB (2011) Robust delay-dependent HN
Institute of Measurement and Control 36(1): 119–130. control of time-delay systems with state and input delays. Interna-
Jeong SC and Park P (2005) Constrained MPC algorithm for uncer- tional Journal of Robust and Nonlinear Control 21(9): 974–1007.
tain time-varying systems with state-delay. IEEE Transactions on Richard JP (2003) Time-delay systems: an overview of some recent
Automatic Control 50(2): 257–263. advances and open problems. Automatica 39(10): 1667–1694.
Kwon WH, Kang JW, Lee YS and Moon YS (2003) A simple reced- Shi Y, Chai T and Yue H (2008) Robust one-step model predictive
ing horizon control for state delayed systems and its stability cri- control for discrete time-delay systems. In: 17th world congress of
terion. Journal of Process Control 13(6): 539–551. the international federation of automatic control (IFAC), Seoul,
Leite VJS, Tarbouriech S and Peres PLD (2009) Robust HN state Korea, 6–11 July 2008, vol. 17, pp. 6160–6165. Laxenburg: IFAC.
feedback control of discrete-time systems with state delay: an LMI Tang G, Sun H and Pang H (2008) Approximately optimal tracking
approach. IMA Journal of Mathematical Control and Information control for discrete time-delay systems with disturbances. Progress
26(3): 357–373. in Natural Science 18(2): 225–231.
Mahmoud M (2000) Robust Control and Filtering for Time-Delay Sys- Zhang L, Wang J, Ge Y and Wang B (2014) Constrained distributed
tems. New York: Marcel Dekker. model predictive control for state-delayed systems with polytopic
Mahmoud MS and Ismail A (2005) New results on delay-dependent uncertainty description. Transactions of the Institute of Measure-
control of time-delay systems. IEEE Transactions on Automatic ment and Control. Epub ahead of print 7 April 2014. DOI:
Control 50(1): 95–100. 10.1177/0142331214528970.

Downloaded from tim.sagepub.com at PENNSYLVANIA STATE UNIV on May 10, 2016