ISA Transactions xxx (xxxx) xxx

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

Cooperative optimization-based distributed model predictive control

for constrained nonlinear large-scale systems with stability and
feasibility guarantees
∗
Ahmad Mirzaei, Amin Ramezani
Control Department, Faculty of Electrical and Computer Engineering, Tarbiat Modares University, Tehran, Iran

article info a b s t r a c t

Article history: This paper proposes a cooperative distributed model predictive control (DMPC) to control the con-
Received 16 September 2020 strained interconnected nonlinear large-scale systems. The main contribution of this approach is its
Received in revised form 10 January 2021 proposed novel cooperative optimization that improves the global cost function of any subsystem.
Accepted 11 January 2021
Each subsystem calculates its optimal control by solving the corresponding global cost function. For
Available online xxxx
each subsystem, the global cost function is defined based on a combination of cost functions of
Keywords: all subsystems. If the sampling time is selected appropriately, then the feasibility of the proposed
Distributed model predictive control approach will be guaranteed. Furthermore, the sufficient conditions for stability and consequently,
Interconnected nonlinear large-scale for the convergence of the whole system states towards the neighborhood of the origin’s positive
systems region are provided. The effectiveness and performance of the proposed approach are demonstrated
Cooperative optimization
via applying it to a nonlinear quadruple-tank system for both minimum-phase and nonminimum-phase
models.
© 2021 ISA. Published by Elsevier Ltd. All rights reserved.

1. Introduction for all of its neighbors, while in non-cooperative ones, it is ne-

glected. Third, unlike the non-cooperative distributed approaches,
Most practical systems like industrial plants, urban traffic, in cooperative ones all control inputs are optimized simulta-
power systems, and supply chain are large-scale systems and neously [3,4]. Dual-mode DMPC is known as a non-cooperative
have nonlinear dynamics consisting of several nonlinear intercon- DMPC method according to the proximity of the states to the
nected subsystems. Distributed model predictive control (DMPC) origin. When the states are in the neighborhood region of the
is one of the practical methods for dealing with nonlinear large- origin, a linear non-cooperative DMPC algorithm is used. On the
scale systems. DMPC approaches are categorized into cooperative other hand, when the states are far from the origin, a nonlinear
and non-cooperative [1]. In non-cooperative methods, each local one is applied [5]. A combination of distributed estimation and
controller computes its own optimal control input, whereas in DMPC architecture is proposed as another non-cooperative DMPC
cooperative methods, each local controller minimizes the central- approach for controlling nonlinear processes [6]. The gradient
ized global cost function and exchanges its information via an projection optimizer is a nonlinear nonconvex algorithm that
interconnection protocol [2]. modifies the objective function in cooperative DMPC approaches
Although the optimization process may be more time- without the need for a coordinator layer. However, there is
consuming and more complex in cooperative distributed ap- no guarantee to achieve convergence in a reasonable time [7].
proaches than in non-cooperative distributed ones, cooperative Nevertheless, some other DMPC methods require a coordination
distributed approaches are still more efficient than layer because local controllers are optimized independently and
non-cooperative ones in large-scale systems for three reasons. each local controller dispatches its information to its neighbors
First, in cooperative distributed approaches, interconnections be- via a coordinator [8,9]. On the contrary, in some other algo-
tween subsystems are considered in the optimization layer, while rithms, the interconnection between neighboring subsystems is
in non-cooperative ones they are ignored. Second, in cooperative considered as an internal constraint instead of being applied to
distributed approaches, the effect of control input of each sub- the coordinator layer [10]. The network-based cooperative DMPC
system is taken into account for the whole system or at least approach is a combination method with multi-rate sampling that
has been developed for a nonlinear uncertain large-scale system.
∗ Corresponding author. In this approach, local controllers are interconnected using a
E-mail addresses: [email protected] (A. Mirzaei), coordination layer in a network-based coordination platform that
[email protected] (A. Ramezani). applies an iterative algorithm to control the system [11]. The

https://doi.org/10.1016/j.isatra.2021.01.022
0019-0578/© 2021 ISA. Published by Elsevier Ltd. All rights reserved.

Please cite this article as: A. Mirzaei and A. Ramezani, Cooperative optimization-based distributed model predictive control for constrained nonlinear large-scale systems
with stability and feasibility guarantees. ISA Transactions (2021), https://doi.org/10.1016/j.isatra.2021.01.022.
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

communication delay is the key concern in methods with a 2. Formulation of CDMPC problem
coordination layer. Most of these methods consider the communi-
cation delay as an external constraint [12]. Robust DMPC methods 2.1. Notation
are useful approaches for the constrained interconnected large-
scale systems. If the subsystems are coupled through their cost Variable v ar (t |tk ) is the predicted value of variable v ar(t) with
functions, these interconnections will be known as internal con- the available information at time sample √ tk . For a given vector,
√
straints. A non-cooperative DMPC approach based on a two-layer x: xT is the transpose of x, ∥x∥l = xT lx and ∥x∥2 = xT x
robust method is a proper algorithm for these constrained large- are its l weighted Euclidean and Euclidean norms, respectively,
scale systems [13]. In the first layer, each subsystem receives and ∥x∥2l = xT lx is the square of l weighted Euclidean norm.
control information of its neighboring subsystems and optimizes For a given square matrix S: ev (S) and ev (S) are its minimum
its local objective function that contains coupling terms. The and maximum eigenvalues, respectively. The time sample tk+g =
convergence of the states is prepared in the first layer. A ro- tk + gts where, ts is sampling time and g is an integer. mine∈E J(e)
bust non-cooperative DMPC algorithm, which tolerates larger is the minimum value of J(e) for all e ∈ E.
disturbances using the shorter prediction horizon, is presented
in the second layer [14,15]. Robust DMPC approaches are useful 2.2. Problem statement
to control nonlinear large-scale systems whose subsystems are
exposed to external disturbances and constraints of control in- Consider a nonlinear large-scale system involving M intercon-
puts [16,17]. Sequential DMPC algorithms are other approaches nected constrained subsystems. The nominal dynamic of subsys-
for large-scale systems that can be designed in both cooperative tem i is:
and non-cooperative DMPC architectures. In these approaches, ẋi (t) = fi xi (t), ui (t), xij (t) + di (t),
( )
i = 1, 2, . . . , M (1)
each local controller optimizes its own objective function and
ni mi
dispatches its information through the communication channel in which, xi (t) ∈ R and ui (t) ∈ R ∈ Ui are the nominal
to attain the global optimality at each sampling time. The pro- dynamic and the constrained control signal of the subsystem
posed sequential DMPC operates similar to a centralized MPC. i, respectively. di (t) is the external disturbance of subsystem i
However, they are more useful for large-scale systems compared that is bounded
∑ in a compact close-set Υ , Υ = d: ∥d∥22 ≤ γ 2 .
to centralized approaches [18]. Stronger interconnection between j∈Ni nj
xij (t) ∈ R denotes the nominal dynamics of all subsystems
subsystems and larger sampling intervals can be taken into ac- j ∈ Ni that are the neighboring subsystems of subsystem i
count in large-scale systems by using the contraction theory in where, Ni is the set of the neighboring subsystems of subsystem
DMPC approaches. [19]. One of the useful DMPC methods for i. A subsystem j is a neighbor of subsystem i if the dynamic of
uncertain systems is the two-level DMPC algorithm that uses the subsystem i depends on the dynamic of subsystem j.
hierarchical framework. The control objective is compensating ac- The linearized dynamic of subsystem i in (1) is:
tuators’ faults, including uncertainties and time delays. At the first ∑
level, faults are retrieved to preserve the design specifications ẋi (t) = Aii xi (t) + Aij xij (t) + Bi ui (t) + di (t) (2)
for the subsystem. The process of retrievement is applied at the j∈Ni

second level by increasing the performance of the whole system. ∂ fi

which is linearized around the origin with Aii = ∂ xi
(0, 0), Aij =
Retrievement design specifications are satisfied by applying the ∂ fi ∂ fi
proposed method [20,21]. In [22], a cooperative DMPC method ∂ xj
(0, 0),
and Bi = (0, 0).
It is assumed in this paper that the
∂ ui
is developed to control linear systems. The cooperative DMPC nonlinear system has one equilibrium point.
approach is taken into account where a centralized global cost The overall nominal dynamic of the constrained nonlinear
function is optimized for each subsystem, which is defined based system is:
on the state and input information of all subsystems. Each control
ẋ(t) = f (x(t), u(t)) + d(t) (3)
input is calculated separately by solving a set of linear matrix
∑M ∑
inequalities. In the presented cooperative DMPC algorithm only where x(t) ∈ R i=1 ni and u(t) ∈ R i mi :
one control input is optimized at a time, which is a limitation
T T
x(t) = x1 (t) , . . . , xM (t) T
[ ]
for the presented method. This limitation is removed in the
cooperative DMPC proposed in this paper. ]T
u(t) = u1 (t)T , . . . , uM (t)T
[
In this paper, a novel cooperative DMPC (CDMPC) approach is
proposed to control constrained interconnected nonlinear large-
[ ( )T ) T ]T
f (x, u) = f1 x1 , u1 , x1j , . . . , fM xM , uM , xMj
(
scale systems. In this approach, a novel cooperative optimization
method that is its main contribution and improves the global and its overall nominal linearized dynamic is:
cost function of each local controller, is proposed. Since in the
proposed approach, the effect of each local controller on the cost ẋ(t) = Ax(t) + Bu(t) + d(t) (4)
function of other controllers is considered, the proposed approach where:
is proper for interconnected large-scale systems. Moreover, fea-
A11 A12 ··· A1M
⎡ ⎤
sibility, stability, and convergence of the proposed approach are
guaranteed. The performance and effectiveness of the proposed ⎢ A21 A22 ··· A2M ⎥
approach are investigated for the nonminimum phase model of a ⎣ ..
A=⎢ .. ..
. ⎦,
.. ⎥
. . .
large-scale system. AM1 AM2 ··· AMM
The remainder of paper is organized as follows. Section 2 dis-
B1 0 ··· 0
⎡ ⎤
cusses the formulation of a cooperative DMPC problem for non-
linear interconnected large-scale systems. The proposed cooper- ⎢0 B2 ··· 0⎥
ative optimization strategy is presented in Section 3. Section 4 ⎣ ..
B=⎢ .. ..
.
.. ⎥ (5)
. . .
⎦
deals with the feasibility of the proposed approach. The stability 0 0 ··· BM
of the proposed approach is discussed in Section 5. Section 6
demonstrates the simulation results for a nonlinear quadruple- The nominal nonlinear dynamics in (1) and (3), and also their
tank system. Finally, concluding remarks are presented in Sec- linearized dynamics in (2) and (4), comply with the following
tion 7. assumption.

2
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx
∑M
Assumption 1. The function fi in (1) for all subsystems i = Ui ⊂ Rmi and U i ∈ R i=1 mi are corresponding control constraints
1, 2, . . . , M and consequently the overall function f in (3) are of ui and ui , respectively which involves the origin.
continuous and twice differentiable and also satisfy fi (0, 0, 0) = After solving all OP i , i = 1, . . . , M, the optimal control tra-
∗
0 and f (0, 0) = 0. jectories, ui (t), i = 1, . . . , M, are obtained and each nominal
subsystem i in (1) is updated using u∗i (t), which is the first
In this paper, a novel CDMPC strategy is proposed that calcu- ∗
element of ui (t). Hence, the nominal overall dynamic in (3) is
lates the optimal controls for all subsystems by solving their cor-
updated using u∗ (t), which consists of the u∗i (t), i = 1, . . . , M, so
responding optimization problems. The proposed strategy con- ∗
that each u∗i (t) is the corresponding control element of ui (t) in
ducts the nominal nonlinear system’s state to the origin.
each OP i .
Definition 1. The optimization problem of subsystem i is denoted
Remark 2. In the approach proposed in this paper, the upper
by OP i hereinafter.
bound of the deviation between the overall nominal dynamic,
x(t), and each of the overall applied dynamic, xi (t), which is
Remark 1. Since in the proposed method in this paper and in the
applied in each OP i , is designed in Lemma 3. The designed upper
current time sample, tk :
bound guarantees the feasibility of the proposed approach for the
(a) In each OP i , only ui is optimized, nominal system in (3).
(b) From the point of view of a cooperative approach, in each
OP i , the overall dynamic and all other subsystems’ control 3. Proposed cooperative optimization protocol
inputs are taken into consideration,
(c) In each time sample of OP i , the control inputs of other Assuming that each subsystem i has its corresponding lo-
subsystems are considered fix and equal to their optimal cal controller i, at the current time sample, tk . Each local con-
values at the last time sample. troller i uses the following proposed cooperative strategy in its
corresponding OP i :
Hence, the optimal control inputs obtained from the proposed
method differ from those of typical CDMPCs [1,2,23] in which in 1- Each local controller i receives the corresponding overall
each OP i , instead of the control inputs of other subsystems being applied state measurement xi (tk ) from (8) using the opti-
considered fix at each time sample, they are considered fix at each mal control inputs calculated in the previous time sample,
iteration of the optimization process. So, the proposed method tk−1 .
has two important advantages over the typical CDMPCs: 2- In each iteration p, each local controller i evaluates its
own future input based on xi (tk ) and control inputs of all
1- In each time sample of OP i , the control inputs of other
other subsystems. Note that in all iterations of OP i , in time
subsystems are considered fix to their optimal values at the
sample tk , the control inputs of all other subsystems are
last time sample but not to their optimal values at the last
considered fix and equal to their optimal values at the last
iteration. Thus, the proposed method avoids the complexity
time sample, tk−1 .
of the iterative optimization and consequently, the overall
3- If the designed termination condition is satisfied, each lo-
computational burden is significantly reduced. ∗
cal controller i calculates the optimal control, ui (tk ), and
2- The feasibility and closed-loop stability are guaranteed ∗ ∗
extracts its first element, ui (t), and stores the ui (t) in the
with this assumption that the initial state of the nominal
overall control input matrix, u (tk ). If the designed termina-
system is feasible. Thus, in this situation, there is no need
tion condition is not satisfied, go back to step 2 and repeat
to design the auxiliary controllers [24].
OP i for the next iteration, p + 1.
As such, the overall control input that is used in each OP i is 4- When all OP i ∀i = 1, . . . , M are solved and all optimal
denoted by ui (t): control inputs, u∗i (t), i = 1, . . . , M, are obtained, extracted,
]T and stored, the overall optimal control input matrix, u (tk ),
ui (t) = ui (t) vij (t)
[
(6) will be obtained and consequently, the overall nominal
dynamic in (3) will be updated. Then, go back to step 1 and
where, vij (t), j = 1, . . . , M , j ̸ = i is the vector of control inputs of
repeat the process for the next time sample, tk+1 .
all other subsystems whose elements are fixed and equal to their
optimal values at the last time sample: Each local controller i solves the following OP i at each itera-
vij (t) = vij∗ (t − 1) tion p:
]T (7)
Ji tk , xi (tk )
( )
vij (t) = u1 (t) · · ·
[
ui−1 (t) ui+1 (t) ··· uM (t) min
ui (t |tk ),t ∈[tk ,tk +Np ]
Thus, the overall dynamic that is used in each OP i and is updated Ji tk , xi (tk )
( )
by the optimal value of ui (t), is denoted by xi (t). xi (t) is called an ∫ tk +Np (
overall applied dynamic and has the following representation:
 )
xi (t |tk )2 i + ∥ui (t |tk )∥2 + vij (t |tk )2 dt
 
= Q Ri R ij
ẋi (t) = f xi (t), ui (t)
tk
( )
(8)  ( )2
+ xi tk + Np |tk  i P
(10a)
and its corresponding overall applied linearized dynamic is:
Subject to (8) and:
ẋi (t) = Axi (t) + Bui (t) (9)
xi (tk |tk ) = xi (tk ) , t ∈ tk , tk + Np
[ ]
(10b)
vij (t |tk ) = vij (tk−1 ) , t ∈ tk , tk + Np
∗
[ ]
(10c)
Assumption 2. For any initial condition xi (t0 ) , i = 1, 2, . . . , M, ui (t |tk ) ∈ U i , t ∈ tk , tk + Np
[ ]
(10d)
each nominal dynamic i in (1) and consequently, the overall nom-
x (t |tk )2 i ≤ Np r i , t ∈ tk , tk + Np
 i 
inal dynamic in (3) have continuous solutions for any constrained
[ ]
P
(10e)
control input ui : [t0 , ∞) → Ui , and ui : [t0 , ∞) → U i where, t − tk−1

3
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

where, Q i > 0, Ri > 0, and Rij > 0 are overall applied dynamic xi (tk ) ∈ Λr i :
state, control input of subsystem i, and its neighboring subsys- ∫ tk +ts
tems’ control inputs weighting matrices, respectively; P i is the
(  )
xi (t)2 i + ∥ui (t)∥2 + vij (t)2 dt
 
Q Ri R
weighting matrix of the terminal condition on the corresponding tk
ij

overall applied state, applied in OP i , which will be designed in 2  2

≤ xi (tk )P i − xi (tk + ts )P i

(12)
Lemma 1; Np is the prediction horizon. The (10b) indicates the
initial value of[ the corresponding overall applied state over the d ( i 2 )
x (t) i ≤ − xi (t)2 i − ∥ui (t)∥2 − vij (t)2
   
(13)
time interval tk , tk + Np . The constraints on control inputs are
] Ri
P Q R
dt ij

considered in the constrained (10d). The predicted overall applied

dynamic is decreasingly bounded by the constraint imposed in Proof. (a) According to the Lyapunov stability criteria, by defini-
(10e) in which r i is a positive constant radius that specifies the tion, Q́ i > 0, and also from the definition of α i , it is concluded
neighborhood of the origin’s positive region Λr i . that AK i + α i I is stable. Hence, a positive definite matrix, P i , will
be rendered as a solution of the Lyapunov equation in (11).
Remark 3. In the proposed method, it is assumed that the nom- (b.1) The state-feedback controllers ui (t) = Ki xi (t) and vij (t) =
inal and applied dynamics start from initial states with feasible Kij xi (t) are linear mappings, and since the constraint U i includes
values. Thus, by satisfying all constraints of all optimization pro- the origin, a positive constant radius ŕ i exists such that K i xi (t) ∈
cesses, the feasibility and stability of the overall nominal system U i for all applied dynamics xi (t) ∈ Λr i in which r i ≤ ŕ i .
will be guaranteed.
(b.2) The differentiation of the terminal cost of applied dy-
namic leads to:
Remark 4. From a general point of view, the proposed method
is similar to the iterative CDMPC [23] with the difference that in d (  )
xi (t)2 i = xi (t)T AT i P i + P i AK i xi (t) + 2xi (t)T P i Γ i xi (t)
( ) ( )
the optimization process of each subsystem it is assumed that the P K
dt
control inputs of other subsystems in all iterations of the current
(14)
time sample are fixed to their optimal values at the last time
sample. Although the proposed method is much simpler than
where, Γ xi (t) denotes the difference between the linearized
i
( )
iterative CDMPC and significantly reduces the computational bur-
and nonlinear applied dynamics by applying the state-feedback
den. However, if sufficient conditions for feasibility and stability
are not designed properly, it will not be applicable. controllers:

Γ i xi (t) = f xi (t), K i xi (t) − AK i xi (t)

( ) ( )
i
Assumption 3. In a linearized version of applied dynamic x (t) in
(9), used in OP i , the (Ai , Bi ) is stabilizable, so there exists a matrix Assuming that:
K i such that AK i = Ai + Bi K i is stable:
2 2
Ci (t) = xi (t)Q i + ∥ui (t)∥2Ri + vij (t)R
   
ui (t) = K i xi (t) ij
 2  2  2
[ ]T = xi (t)Q i + xi (t)K T R K + xi (t)K T R
B = B1 B2 i i i ij ij Kij

ui (t) = Ki xi (t), and give that:

[ ] {
kii Oii Ki = [kii Oii ]
Ki = →
vij (t) = Kij x (t), Kij = Oij kij
i
[ ]
Oij kij
ev KiT Ri Ki  ev KiT Ri Ki  i 2
( ) ( )
i
2  i 2
where Oii and Oij are zero matrices with appropriate dimensions.
 x (t) Q i ≤ x (t) K T R K ≤
   ( ) x (t)Q i
ev Q i ev Q i
( )
i i i

ev KijT Rij Kij  i 2 ev KijT Rij Kij  i 2

( ) ( )
Remark 5. By designing OP i in (10), the effect of the control  i 2
 x (t) Q i ≤ x (t) K T R K ≤
   ( ) x (t)Q i
ev Q i ev Q i
( )
input of each subsystem is taken into account in the entire plant. ij ij ij

Lemma 1. For the overall applied dynamic in (8), which is used in so we have:
OP i , and with Assumption 3:
ev KijT Rij Kij
( ))
ev KiT Ri Ki
( ) (
 i 2
(a) A positive definite matrix P i exists that constitutes and solves − 1+ + x (t) i ≤ −Ci (t)
ev Q i ev Q i
( ) ( ) Q
the following Lyapunov equation:
ev KijT Rij Kij
( ))
ev KiT Ri Ki
)T ( ) (
AK i + α i I P i + P i AK i + α i I = −Q́ i
( ( )
(11)  i 2
≤− 1+ + x (t) i
ev Q i ev Q i
( ) ( ) Q
where:
ev KijT Rij Kij
( ))
ev KiT Ri Ki
( ) (
i
Thus, inequality (13) will be satisfied if the following inequality
Q́ = 1+ + Qi is satisfied:
ev ev
( ) ( )
Qi Qi
ev KijT Rij Kij
( ))
ev KiT Ri Ki
( ) (
and: d (  )
xi (t)2 i ≤ − 1 +
 i 2
( ) + x (t) i
ev Q ev Q
P
( ) Q
dt i i
α i ∈ [0, ev AK i )
( )

(15)
{ ∑M  2 }
(b) A corresponding region Λr i = xi ∈ R i =1 n i : xi P ≤ r i
i
exists in the neighborhood of the origin in which: The substitution of (14) in inequality (15):
(b.1) The state-feedback controllers ui (t) = Ki xi (t) and vij (t) =
xi (t)T ATK i P i + P i AK i xi (t) + 2xi (t)T P i Γ i xi (t)
( ) ( )
Kij xi (t) satisfy the control constraints U i for applied dynamic xi (t) ∈
Λr i . (
ev KijT Rij Kij
))
ev KiT Ri Ki
( ) (
(b.2) The following integral and differential 2 inequalities are es- ≤− 1+ + xi (t)T Q i xi (t) (16)
ev Q i ev Q i
( ) ( )
tablished for the terminal condition xi (tk )P i , in OP i , for applied


4
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx
( ( ) ( T
))
ev KiT Ri Ki ev Kiji Rij Kiji (b) A matrix P exists, which is positive definite and satisfies
and the substitution of Q́ i = 1+ + Q i in
ev ( )
Qi ( )
ev Q i the following Lyapunov equation:
inequality (16) leads to: (AK + α I )T P + P (AK + α I ) = −Q́ (20)
i
x (t) T
(
ATK i P i + P AK i x (t) + 2x (t) P Γ
i
) i i T i i
( i
)
x (t) ≤ −x (t) Q́ x (t) i T i i α ∈ [0, ev (AK ))
( ))
ev K T RK
(
(17) Q́ = 1 + Q
)2 2 ev (Q )
Since Γ i xi (t) P i → 0 as xi (t)P i → 0, so a neighbor-
 ( 
hood region around the origin exists such that the condition and a choice for R is:
2
 ( )
 i i
Γ x (t)  i
R = Diag {R1 , . . . , RM }
( )2
P
≤ α i is established.
∥xi (t)∥2P i
In addition, given the left-hand side of (11), the left-hand side Since the overall applied dynamic, xi (t), is applied in the corre-
of (17) can be represented as: sponding OP i , so Q , Q́ and P could be equal to Q i , Q́ i and P i ,
respectively for i = 1, . . . , M.
xi (t)T ATK i P i + P i AK i xi (t) + 2xi (t)T P i Γ i xi (t)
( ) ( )
(( )T )) As a result, both the weighting matrix P i and the neighborhood
= xi (t)T AK i + α i I P i + P i AK i + α i I xi (t) of the origin’s positive region Λr i for the overall applied dynamic
(
used in each OP i in (10) are obtained by the proposed method,
− 2xi (t)T α i P i xi (t) + 2xi (t)T P i Γ i xi (t)
( )
which is presented in Lemma 1. The neighborhood of the origin’s
= −xi (t)T Q́ i xi (t) + 2xi (t)T P i Γ i xi (t) − 2xi (t)T α i P i xi (t) positive region Λr for the nominal dynamic in (3) and with the
( )
(18)
control law u(t) = Kx(t) will be provided in Lemma 2.
The last two terms in (18) are upper bounded as:
The following useful assumption will be used hereinafter:
2xi (t)T P i Γ i xi (t) − 2xi (t)T α i P i xi (t)
( )
[ ]T [ ]
( i ) 12 ( i ) 12 Assumption 4. After solving each OP i and calculating the corre-
xi (t) Γ i xi (t) − 2α i xi (t)T P i xi (t)
( )
≤2 P P (19) sponding optimal control ui (t), we have:
 ( )2 The overall function f holds the following Lipschitz condition:
 i i
Γ x (t)  i

( )2
≤ α i , we have: f x (t |tk ) , ui (t |tk ) − f xi (t |tk ) , ui (t |tk ) 2
P
 ( ) ( )
From the condition
∥xi (t)∥2P i 2
2

≤ Lfi x (t |tk ) − xi (t |tk )2


 i ( i )2
Γ x (t)  i ( )2
≤ α i → Γ i xi (t) P i ≤ α i xi (t)P i
P
 ( )  
xi (t)2 i
 
Lemma 2. For the overall nominal dynamic in (3) and its linearized
P
version in (4), with u(t) = Kx(t) and considering Assumptions 1
and its substitution in inequality (19) leads to:
and
{ 2, a neighborhood of } the origin’s positive control region Λr =
2 xi (t)P i Γ i xi (t) P i − α i xi (t)T P i xi (t)
[   ( ) ] ∑M
x∈R i=1 ni : ∥x∥2P ≤ r exists for the nominal dynamic in (3).
The positive control region Λr can be determined as:
[  2 ]
≤ 2 α i xi (t)P i − α i xi (t)T P i xi (t) = 0
1- Consider the lower limit of positive constant radius, rmin =
Therefore, it is concluded that by satisfying the condition
 (
 i i
)
2 0 and its upper limit, rmax , to be a large enough value to
Γ x (t)  i ( )2
P
≤ α i , the inequality (17) will be satisfied and the comprise r, and also define the difference between lower and
∥xi (t)∥2P i upper limits, ∆r, as a criteria for termination condition to
proof is completed. determine the value of r;
2- The process of determining the value of r starts with an initial
Remark 6. Considering that r +r
guess, r = min 2 max , and goes on with the calculation of the
1- the overall applied dynamic in (8) is applied in each OP i , following cost-like function:
i = 1, . . . , M, ( )
2- based on Assumption 3, for each subsystem i, (i = 1, . . . , fmin = min x(t)T Q́ x(t) + 2α x(t)T Px(t) − 2x(t)T P Γ (x(t))
x(t)∈Λr
M) a matrix K i exists such that AK i = A + BK i is stable,
3- Lemma 1 is satisfied for all subsystems, thus for any applied 3- If fmin < 0, then r = rmax ; otherwise, r = rmin ;
dynamic xi (t) used in OP i , there is a positive definite 4- If rmax − rmin < ∆r, then the termination condition is
matrix P i for each OP i , which solves the corresponding satisfied and r = rmin . If the termination condition is not
Lyapunov equation in (11), satisfied, rmax − rmin > ∆r, go back to step 2 and repeat the
4- the nominal dynamic in (3) is updated by the optimal value optimization-like process with the latest received r.
of u(t) where each of its elements, ui (t), i = 1, . . . , M,
is calculated in each corresponding OP i using the corre- Proof. It can be shown that Λr is a neighborhood of the origin’s
sponding overall applied dynamic, xi (t), and positive region for the nominal linearized dynamic in (4), from
5- the deviation between the applied, xi (t), and nominal, x(t), the view of (13), and with the state-feedback control law u(t) =
dynamics is upper bounded for each OP i , which will be Kx(t), i.e., the following inequality is satisfied for x(t) ∈ Λr :
designed in Lemma 3,
d (
∥x(t)∥2P = x(t)T ATK P + PAK x(t) + 2x(t)T P Γ (x(t)) ≤ 0 (21)
) ( )
Thus, for the overall nominal dynamic in (3) and its linearized dt
version in (4) we have: where, Γ (x(t)) = f (x(t), Kx(t)) − AK x(t) interprets the difference
(a) A state-feedback matrix K exists so that u(t) = Kx(t) and between linearized and nonlinear nominal dynamics. Similar to
also AK = A + BK is stable. A choice for K is: the process used to prove Lemma 1, it is also analytically provable
[ ] that, for the nominal linearized dynamic in (4) and with consider-
K = K1 ··· KM ing Assumption 3 for the nominal linearized dynamic, there exists
5
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

∫ t( ) ∫ t
a matrix P that is positive definite and satisfies the following Lf i Np
equation: − ri dt + ∥d(t)∥22 dt
ev P i t − tk−1
( )
tk tk

(AK + α I )T P + P (AK + α I ) = −Q́ (22)

( )∫
t t
ri
∫
r Np
≤ Lfi − dt + γ 2 dt
ev (P) ev P i
( )
and: tk t − tk−1 tk

α ∈ [0, ev (AK )) Thus, we have:

x (t |tk ) − xi (t |tk )2
 
The right-hand side of (21) can be represented as follows by
2
substitution from (22): ( ( ) )
Lf i N p r ri
T
ATK P + PAK x(t) + 2x(t) P Γ (x(t)) T ≤ − +γ 2
(t − tk ) (28)
( )
x(t)
ev (P) ev P i
( )
ts
= −x(t) Q́ x(t) − 2α x(t) Px(t) + 2x(t) P Γ (x(t))
T T T
(23)
and the proof is completed.
Since ∥Γ (x(t))∥2P → 0 as ∥x(t)∥2P → 0, thus a positive constant
radius r > 0 can always exist such that if ∥x(t)∥2P ≤ r, then 4. Feasibility analysis
∥Γ (x(t))∥2P
∥x(t)∥2P
≤ α is satisfied. Hence, similar to the process used to
2
The feasibility of an optimization process means that its fea-
prove Lemma 1, it is concluded that: sibility at the current time tk demonstrates its feasibility at the
−x(t)T Q́ x(t) − 2α x(t)T Px(t) + 2x(t)T P Γ (x(t)) ≤ −x(t)T Q́ x(t) ≤ 0 next time tk+1 . Like other methods, the successful operation of the
proposed CDMPC method directly depends on the feasibility of its
so we have: optimization process. The sufficient conditions for the satisfaction
of the terminal conditions in (10e) and the feasibility of each OP i
−fmin ≤ 0 in (10) will be provided in Theorems 1 and 2, respectively. The
and r = rmin , and finally, the proof is completed. assumed overall dynamic, x̂i (t |tk ), is defined for each OP i as:
The optimal control inputs are obtained using the correspond-
x̂˙ i (t |tk ) = f x̂i (t |tk ) , ûi (t |tk )
( )
ing applied dynamics in (8), as a prediction model, for each OP i in
(10) via the proposed CDMPC strategy in this paper. However, the where, ûi (t |tk ) is assumed control trajectory defined as:
predicted trajectory of the applied dynamic, xi (t), which is used ]T
ûi (t |tk ) = ûi (t |tk ) v̂ij (t |tk )
[
in (10), is deviated from the nominal state in (3). This deviation
is restricted with the upper bound that is designed in Lemma 3. in which:
u∗i (t |tk−1 ) , t ∈ [tk , tk−1 + Np )
{
Lemma 3. For the overall applied dynamic, used in each OP i in ûi (t |tk ) =
(10), and with Assumption 4, the deviation between the predicted Ki x̂i (t |tk ) , t ∈ [tk−1 + Np , tk + Np ]
applied and the nominal dynamics is upper bounded by the following v̂ij (t |tk ) = vij∗ (t |tk−1 ) , t ∈ tk , tk + Np
[ ]
inequality:
and after solving all OP i , i = 1, . . . , M, we have:
x (t |tk ) − xi (t |tk )2 ≤ λ (t − tk )
 
(24) ∗
ui (t |tk−1 ) , t ∈ [tk , tk−1 + Np )
2
{
ûi (t |tk ) = (29)
where: K i x̂i (t |tk ) , t ∈ [tk−1 + Np , tk + Np ]
( )
Lf i N p r ri
λ= − + γ2 (25) Theorem 1. If the sampling time, ts , satisfies the following condi-
ev (P) ev P i
( )
ts
tions for the nominal dynamic in (3) and for each OP i in (10):
( ( ) )
Proof. After solving each OP i , the feasible vector of the control r ri ts r i
ev P + γ ts eLfi (Np −ts ) ≤
( i) 2
Lfi Np −
input, ui (t |tk ), and consequently, the corresponding applied dy- ev (P) ev
( )
Pi Np + ts
namic, xi (t |tk ), are determined and after substitution from (3) and
(8) and considering Assumption 4, the left-hand side of (24) leads (30)
to: Np
≥ e−θi ts (31)
x (t |tk ) − xi (t |tk )2
  Np + ts
2
∫ t(
2 ) where:
Lfi x (t |tk ) − xi (t |tk )2 + ∥d(t)∥22 dt

≤ (26) ev Q i ev KiT Ri Ki ev KijT Rij Kij
 ( ) ( ) ( )
tk θi = ( ) + + (32)
ev P ev P i ev P i
( ) ( )
i
Inequality (26) will be satisfied if the following inequality is
satisfied: then the assumed applied dynamic, x̂i (t ; tk ), will be conducted into
x (t |tk ) − xi (t |tk )2
  the neighborhood of the origin’s positive region Λ Np i , using the
Np +ts
r
2
∫ t(
2 ) assumed control input ûi (t |tk ) in (29).
Lfi ∥x (t |tk )∥22 − Lfi xi (t |tk )2 + ∥d(t)∥22 dt

≤ (27)
tk
Proof. From inequality (24), we have:
And after substitution from (10e) and considering Remark 6,
x (t |tk ) − xi (t |tk )2 i
 
inequality (27) leads to: P
2
≤ ev P i x (t |tk ) − xi (t |tk )2
( )
x (t |tk ) − xi (t |tk )2
 
2 ( ( ) )
( i ) Lfi Np ri
∫ t ( )
r
Lf i Np
≤ r dt ≤ ev P − ( i ) + γ (t − tk )
2
(33)
tk ev (P) t − tk−1 ts ev (P) ev P
6
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

By substitution t = tk+1 in inequality (33), it yields that: Given (13), it can be analytically shown that:
x (tk+1 |tk ) − xi (tk+1 |tk )2 i d ( i
   )
x (t |tk+1 )2 i ≤ −θi xi (t |tk+1 )2 i
 
P
P P
(41)
( ( ) ) dt
r ri
≤ ev P − ( i ) + γ ts
( i) 2
Lfi Np (34) By integrating (41) from tk + Np to t, we have:
ev (P) ev P (  )
2
∫ t d xi (t |tk+1 )P i ∫ t
and for the assumed overall dynamic x̂i (t) in tk+1 , we have: dt ≤ −θi dt
xi (t |tk+1 )2 i
 
 ∗
2 tk +Np tk +Np
x̂ (t |tk+1 ) − xi (t |tk ) i P
 i 
P
2 and:
( i)  ∗
≤ ev P x̂ (t |tk+1 ) − xi (t |tk ) x (t |tk+1 )2 i ≤ xi tk + Np |tk+1 2 i e−θi (t −tk −Np ) ,
 i   i   ( )
2 P P

t ∈ tk + Np , tk+1 + Np
( 2 [ ]
( i)  i∗ (42)
≤ ev P x̂ (tk+1 |tk+1 ) − x (tk+1 |tk )
 i 
2
) By combining inequalities (31), (39), and (42), it concludes that:
∫ t   2
∗
x̂ (s|tk+1 ) − xi (s|tk ) ds
 i Np
+ Lfi (35) x̂ tk+1 + Np |tk+1 2 i ≤
  i(
ri
)
2 P
(43)
tk+1 Np + ts
By applying inequality (34) and using the Gronwall–Bellman in- which demonstrates that the terminal value of the assumed over-
all ]dynamic, x̂i tk+1 + Np |tk+1 , in time interval t ∈ [tk+1 , tk+1 +
( )
equality, inequality (35) leads to:
 ∗
2 Np enters into the neighborhood of the origin’s positive region
x̂ (t |tk+1 ) − xi (t |tk ) i
 i 
Λ Np i , and consequently, the proof is completed.
( ( P ) ) Np +ts
r

ri The terminal condition for time interval t ∈ tk+1 , tk+1 + Np

[ ]
r
≤ ev P − ( i ) + γ ts eLfi (t −tk+1 )
( i) 2
Lfi Np (36)
ev (P) ev P is obtained by the substitution of t = tk+1 + Np in (40):
Np
x tk+1 + Np |tk+1 2 i ≤
 i(
By substitution ]of the terminal time, tk + Np , over the time interval ri
)
P
t ∈ tk , tk + Np in the constraint (10e), it obtains that: Np + ts
[
 ∗(
 i 2
) Np that verifies inequality (43).
x tk + Np |tk  i ≤ ri (37) Therefore, the assumed overall dynamic, x̂i (t), conducts to the
P Np + ts
neighborhood of the origin’s positive region via sufficient condi-
and for the assumed
[ overall dynamic x̂i (t), in tk + Np , over the tions provided by Theorem 1 for the assumed control ûi (t) and by
time interval t ∈ tk+1 , tk+1 + Np , we have:
]
the prediction horizon Np . The feasibility of the assumed control,
ûi (t |tk ), is investigated in Theorem 2 via designing sufficient
x̂ tk + Np |tk+1 2 i
 i( )
P conditions.
2
 ∗( )
 i
≤ x tk + Np |tk  Theorem 2. OP i in (10) is feasible with a feasible initial nominal
Pi
 ( ∗ ( 2
) state, if the sampling time, ts , satisfies the following condition in
+ x̂ tk + Np |tk+1 − xi tk + Np |tk 
 i )
i
addition to satisfying (30) and (31):
P
Np i Np θi − 1 ≥ 0 (44)
≤ r
Np + ts
Proof. To determine the feasibility of the OP i at time tk+1 ,
( ( ) )
r ri
+ ev P + γ ts eLfi (Np −ts ) (38)
( i) 2
Lfi Np − suppose that it is feasible at time tk , so according to
[ (10e), the
ev (P) ev P i
( )
terminal condition over the time interval, t ∈ tk , tk + Np ,
]

If condition (30) is satisfied, then it concludes that: satisfies the following inequality:

ri Np
Np x (t |tk )2 i ≤ r i, t ∈ tk , tk + Np
 i  [ ]
x̂ tk + Np |tk+1 2 i ≤ ts
 i(
ri
)
P
+ P
t − tk−1
Np + ts Np + ts
x̂ tk + Np |tk+1 2 i ≤ r i
 i( ) and
[ furthermore, ] the terminal condition over the interval, t ∈
(39)
P tk+1 , tk+1 + Np , is:
i
(which demonstrates that the assumed overall dynamic, x̂ Np
x (t |tk+1 )2 i ≤ r i, t ∈ tk+1 , tk+1 + Np
)  i  [ ]
tk + Np |tk+1 , of the applied system is conducted to the neigh- P
(45)
borhood of the origin’s positive region Λr i , and the state-feedback t − tk
control law, ui (t) = K i xi (t), is used. To investigate the feasibility of OP i in tk+1 , the assumed dynamic
The constraint (10e) for t ∈ tk+1 , tk+1 + Np is rewritten as: x̂i (]t |tk+1 ), should be analyzed in the time interval, [tk+1 , t]k+1 +
[ ]

] into two time intervals, tk+1 , tk + Np and

[
x (t |tk+1 )2 i ≤ Np r i , t ∈ tk+1 , tk+1 + Np
 i  [Np , which is divided
tk + Np , tk+1 + Np , like Theorem
[ ]
(40)
P
t − tk [ 1.
Consider the first interval, tk+1 , tk + Np . The combination of
]
The final condition for the first subset, t ∈ tk+1 , tk + Np , is
[ ]
(36) and (10e) gives:
obtained by the substitution of t = tk + Np in (40):
x̂ (t |tk+1 )2 i
 i 
P
x (t |tk+1 )2 i ≤ r i
 i 
 ∗ 2  2
P ∗
≤ x (t |tk ) + x̂i (t |tk+1 ) − xi (t |tk )
 i   
that verifies inequality (39). i P i P

] final condition for the second subset t ∈ tk + Np , tk+1

[
The Np r i
+Np is calculated as follows. ≤
t − tk−1
7
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx
( ( ) )
r ri By the substitution of the initial time, it achieves as:
+ e v P i Lf i N p + γ 2 ts eLfi (t −tk+1 )
( )
−
ev (P) ev P i Np eθi (tk +Np −tk −Np ) − tk + Np − tk
( ) ( )
∆Fi tk + Np =
( )
=0
(46) tk + Np − tk eθi (tk +Np −tk −Np )
( )
The second term on the right-hand side of inequality (46) is
represented as following inequality using (30): [ ∆Fi (t) starts from
so, zero in the second time interval, t ∈
tk + Np , tk+1[+ Np .
]
( ( ) )
ri Since t ∈ tk + Np , tk+1 + Np , so the denominator of ∆Fi (t) is
]
r
ev P i + γ 2 ts eLfi (t −tk+1 )
( )
Lfi Np −
ev (P) ev P i positive. Thus, whether function ∆Fi (t) is ascending or descend-
( )
ing is determined by its numerator, B∆Fi (t). The derivation of
ts r i
≤ ( ) eLfi (t −tk+1 +ts −Np ) (47) B∆Fi (t) is:
Np + ts
dB∆Fi (t)
and by substituting (47) in (46), we have: = Np θi eθi (t −tk −Np ) − 1
dt
Np r i ts r i θi (t −tk −Np )
) eLfi (t −tk+1 +ts −Np )
dB∆F (t)
x̂ (t |tk+1 )2 i ≤ [ Np θi e
 i  i
+( (48) If ]− 1 ≥( 0, then ) dt ≥ 0. Since t ∈
P
t − tk−1 Np + ts tk + Np , tk+1 + Np , then t − tk − Np ∈ [0, ts ]. So, inequality
The feasibility of OP i in (10) is guaranteed if: Np θi eθi (t −tk −Np ) − 1 ≥ 0 will be satisfied if (44) is satisfied. If (44)
dB∆F (t)
ts r i ts Np r i is satisfied, then inequality dti ≥ 0 will be satisfied, and since
e fi ( k+1 s p ) ≤
L t −t +t −N
(49) the denominator of ∆Fi (t) is positive, it is concluded that if (44)
Np + ts (t − tk ) (t − tk−1 )
By the substitution of (49) in (48), we have:
[ ∆Fi (t) ≥ 0 and
is satisfied then ] increases over the second time
interval, t ∈ tk + Np , tk+1 + Np :
i
x̂ (t |tk+1 )2 i ≤ Np r
 i 
Np
P
t − tk − e−θi (t −tk −Np ) ≥ 0
t − tk
which satisfies[the terminal]condition of OP i in (10) over the first
and the combination of (39) and (42) leads to:
time interval, tk+1[, tk + Np . ]
In the interval, tk+1 , tk + Np , it is clear that t − tk ≤ Np and i
x (t |tk+1 )2 i ≤ r i e−θi (t −tk −Np ) ≤ Np r
 i 
t − tk−1 ≤ Np + ts , so from these inequalities, we have: P
t − tk
ts Np r i ts Np r i ts r i
] control, û (t |tk+1 ), in
i
≥ )= Consequently, the feasibility
[ of the assumed
(t − tk ) (t − tk−1 ) the second time interval, tk+1 , tk + Np , is guaranteed, and the
(
Np Np + ts Np + ts
proof is completed.
thus, (49) will be satisfied if:
ts r i ts r i
e fi ( k+1 s p ) ≤ 5. Stability analysis
L t −t +t −N
(50)
Np + ts Np + ts
For the beginning and final times of the interval, tk+1 , tk + Np ,
[ ] To ensure stability, it should be proven that the cost function
we have: in (10) is a descending function, i.e., the optimal value of the cost
function in tk+1 is lower than its value in tk . Thus, the difference
t = tk+1 → e fi ( k+1 s p ) = e fi ( s p )
L t −t +t −N L t −N
between the optimal value of the cost function in two consecutive
t = tk + Np → e fi ( k+1 s p ) = 1 time samples, tk and tk+1 , should be investigated. If stability is
L t −t +t −N
ensured, it will be guaranteed that the overall nominal dynamic
Since ts − Np ≤ 0, then e fi ( s p ) ≤ 1 and it is clear that
L t −N
in (3) conducts to the neighborhood of the origin’s positive region
e fi ( k+1 s p ) ≤ 1 over the first time interval, tk+1 , tk + Np . Λr .
L t −t +t −N
[ ]
Thus, inequality (50) is satisfied and consequently, the feasibil- The sufficient conditions for each OP i in the proposed CDMPC
[ity of the assumed control, ûi (t |tk+1 ), in the first time interval, in (10) are presented here to ensure stability.
tk+1 , tk + Np , is guaranteed. The optimal value of the cost function in (10) in tk+1 , Ji∗ tk+1 , xi
] (

Now, the second time interval, tk + Np , tk+1 + Np is con- (tk+1 )), can be upper bounded as follows where the assumed
[ ]
sidered. By [ inspiring from ](42), which is defined for the time vector of the control input, ûi (t |tk+1 ), is used:
interval, tk + Np , tk+1 + Np , function Fi (t) is defined as:
Ji∗ tk+1 , xi (tk+1 ) ≤ Ji tk+1 , xi (tk+1 )
( ) ( )

Fi (t) = e−θi (t −tk −Np ) (51) ∫ tk+1 +Np (  )

x̂i (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 dt
   
= Q R R
In view of (31) and (51), and also by considering that the objective tk+1
i ij

is satisfying the terminal condition in (45), (31) is rewritten as:  ( )2

+ x̂i tk+1 + Np |tk+1  i P
Np
≥ e−θi (t −tk −Np ) (52) Also, the difference between the optimal value of the cost func-
t − tk
tion in two consecutive time samples, tk and tk+1 , can be upper
in which, by substituting the final time, t = tk+1 + Np in (52), (31) bounded as:
is achieved. As such, function Fi0 (t) is defined as:
I = Ji∗ tk+1 , xi (tk+1 ) − Ji∗ tk , xi (tk )
( ) ( )
Np
Fi0 (t) = (53) ≤ Ji tk+1 , xi (tk+1 ) − Ji tk , xi (tk ) ∗
( ) ( )
t − tk
For further derivations, we need the difference between functions = I1 + I2 + I3
Fi (t) and Fi0 (t), which is defined as ∆Fi (t):
∫ tk+1 +Np (
 )
x̂ (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 dt
 i    
=
Np eθi (t −tk −Np ) − (t − tk )
Q R Ri ij
B∆Fi (t) tk+1
∆Fi (t) = = )2
(t − tk ) eθi (t −tk −Np )
 (
A∆Fi (t) + x̂i tk+1 + Np |tk+1  i P

8
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

tk +Np
∫ ( 2 ) ( ( ) )
 i∗ 2  2 ( i ) Lfi Np r ri
x (t |tk ) i + u∗i (t |tk )Ri + vij∗ (t |tk )Rij dt

− ≤ ev P +γ 2
(t − tk )

− (61)
ev (P) ev P i
Q
( )
tk ts
2
 ∗( )
− xi tk + Np |tk 

i P
The integral term in (54) is expressed over the interval, t ∈
[tk , tk+1 ], which indicates the sampling time, ts , i.e., over the
interval, t ∈ [tk , tk+1 ], the prediction horizon can be consid-
tk+1
∫ ( 2 )
 i∗ 2  2 ered equal to the sampling time. If the overall nominal dynamic,
x (t |tk ) i + u∗i (t |tk )Ri + vij∗ (t |tk )R dt

I1 = −

tk Q ij x (t ; tk ), does not steer into the corresponding neighborhood of
the origin’s positive region over the time interval t ∈ [tk , tk+1 ],
(54) then the following conclusion will be drawn from (10e):
i
x (t |tk )2 i ≥ r , t ∈ [tk , tk+1 ]
∫ tk +Np
(  i 
x̂i (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 (62)
    
I2 = Q Ri Rij
P
2
tk+1
 ∗ 2 2  2
) By the substitution of (62) in (61), we have:
− xi (t |tk ) − u∗i (t |tk )R − vij∗ (t |tk )R dt

(55)
   ∗ 2
i Q i ij
− xi (t |tk )
 
Pi
2  2
tk+1 +Np ∗
≤ − xi (t |tk )P i + xi (t |tk ) − xi (t |tk )
∫ ( 
 )
x̂i (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 dt
     
I3 = Q R R Pi
i ij
tk +Np ( ( ) )
i i
r r r
2 ≤ − + ev P − ( i ) + γ 2 ts
)2  ) ( i)
 (  ∗( Lfi Np (63)
+ x̂i tk+1 + Np |tk+1 P i − xi tk + Np |tk  (56) ev (P)
i P 2 ev P
Theorem 3 establishes the upper bounds for (54)–(56). and by the substitution of (63) in (60), it is concluded that:
 ∗ 2
Theorem 3. According to the condition in (44) and with the sam- − xi (t |tk )
 
i
pling time, ts , which holds conditions (30) and (31), the upper ( i Q) 
bounds in (54)–(56) will be satisfied for the overall applied dynamic ev Q  ∗ 2
≤ − ( i ) xi (t |tk )

in (8) as: ev P Pi
tk+1 ( i) ( ( ( ) ))
ev Q
∫ ( 2 )
i
 i∗ 2  2 r r ri
x (t |tk ) i + u∗i (t |tk )Ri + vij∗ (t |tk )Rij dt

( ) − + ev P − ( i ) + γ ts
( i) 2
I1 = −

≤ Lfi Np
tk Q ev P i 2 ev (P) ev P
ev Q i
( )(
r i ts t 2 r i Lf (ts −Np )
)
≤ πi ( i ) − + s e i (57) ∫ tk+1
ev P
2
2 N p + ts
 ∗
I1 ≤ − πi xi (t |tk ) dt
 
Qi
∫ tk +Np (
tk
x̂ (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2
 i     
I2 =
ev Q
( i) ∫ t (
Q Ri Rij i
tk+1
k+1 r
≤ πi ( i ) −
ev P
2 )
2
 ∗ 2  2
− xi (t |tk ) − u∗i (t |tk )R − vij∗ (t |tk )R dt t

(k (
 
i i
Q ij
) ))
r ri
Lfi (ts −Np ) + ev P − ( i ) + γ 2 ts
( i)
ev Q Lfi Np dt
( i)
i
ts r 1−e
≤ ( ) πi (58) ev (P) ev P
ev P i N p + ts Lfi
ev Q i
( ) (∫ t
k+1 ri
= πi ( i )
∫ tk+1 +Np (
     )
x̂ (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 dt
 i − dt
I3 = Q R i R ij
ev P tk 2
tk +Np ( ( ) ) )
∫ tk+1
)2  2
) ts r i r ri
ev P − ( i ) + γ ts dt
 (  ∗( ( i) 2
+ x̂i tk+1 + Np |tk+1 P i − xi tk + Np |tk  ≤ (59) + Lfi Np
i P Np + ts tk ev (P) ev P
ev Q
( i) (
where: r i ts
= πi ( i ) −
ev KijT Rij Kij
( ))
ev KiT Ri Ki ev P
( ) (
2
πi = 1 + +
ev Q i ev Q i
( ) ( ) ∫ tk+1 ( ( ) ) )
r ri
ev P − ( i ) + γ ts dt (64)
( i) 2
+ Lfi Np
The structure of πi is derived from the cooperative policy proposed in tk ev (P) ev P
this paper. By considering πi , the interconnection between the sub-
By applying (30), the last term of (64) leads to:
systems in the global cost functions applied to each local controller
and also in the process of proving the stability is considered.
( ( ) )
r ri ts r i
ev P + γ 2 ts e fi ( s p )
L t −N
( i)
Lfi Np − ≤
ev (P) ev P i
( )
Proof. According to the procedure employed in the proof of Np + ts
Lemma 1, we have for the first integral term in (54): (65)
tk+1
∫ ( 2 )
 i∗ 2  2
x (t |tk ) i + u∗i (t |tk )Ri + vij∗ (t |tk )R dt Consequently, we have:

I1 = −

Q ij
ev Q i
tk
( )( ∫ tk+1
r i ts ts r i
)
I1 ≤ πi e fi ( s p ) dt
tk+1 L t −N
( ) − +
∫  ∗ 2
≤− πi xi (t |tk ) dt (60) ev P i 2 Np + ts
 
tk
i Q
tk
ev Q
( i) (
i 2 i
)
r ts ts r
I1 ≤ π i e fi ( s p )
L t −N
From (24), it concludes that: ( ) − + (66)
 2 ev P i 2 Np + ts
∗
x (t |tk ) − xi (t |tk ) i
 i 
which implies (57).
P
9
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

Similar to the first integral term in (54), for the second integral and from (30) we have:
term in (55), we have:
2 ts r i
 ∗( )
x̂ tk + Np |tk+1 2 i − 
 i(
xi tk + Np |tk  i ≤
)
∫ tk +Np
P
(
x̂i (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2
     P Np + ts
I2 = Q Ri Rij
tk+1
) and:
 ∗ 2
ts r i
2  2
− xi (t |tk ) − u∗i (t |tk )R − vij∗ (t |tk )R dt
  
Q i i ij I3 ≤ (72)
∫ tk +Np ( ) Np + ts
2
2  
 ∗
πi x̂i (t |tk+1 )Q i − xi (t |tk )

≤ dt which implies (59).
i Q
tk+1
∫ tk +Np
(
∗
2 ) Theorem 4. Consider the nominal dynamic in (3) and also applied
≤ πi x̂i (t |tk+1 ) − xi (t |tk ) dt
 
i
dynamics in (8), which is used in the proposed CDMPC controller in
Q
tk+1
(10) with a feasible initial dynamic. If a parameter ϑ exists such that
( 2 ) ϑ ≤ π2i , and if:
∗
x̂ (t |tk+1 ) − xi (t |tk ) i
 i
1 − e fi ( s p )
 ( )
ev Q i ev Q i
( ) ( ) L t −N
Q 1 (ts −Np )
1+ ( ) πi ts e i + ( ) πi
Lf

ev Q i ts + Np ev P i ev P i Lf i
( ) ( 2 )
i∗
x̂ (t |tk+1 ) − x (t |tk ) i
 i
≤

ev (Pi ) ev Q
( i)
P
2 ≤ϑ (73)
ev P i
  ( )
(36)  i ∗
→ x̂ (t |tk+1 ) − xi (t |tk )

Qi
( ( ) ) is satisfied, then the applied dynamic conducts to the neighborhood
ev Q
( i)
r ri of the origin’s positive region Λr i .
( ) ev P − ( i ) + γ ts eLfi (t −tk+1 )
( i) 2
≤ Lf i N p
ev P i ev (P) ev P
Proof. Combining (54)–(56) and (57)–(59) yields:
ev Q
( i) tk +Np
Ji∗ tk+1 , xi (tk+1 ) − Ji∗ tk , xi (tk )
∫ ( ) ( )
( ) πi ev P i
( )
→ I2 ≤
ev P i ev Q
( i)
tk+1 r i ts
≤− ( ) πi
ev
( ( ) )
r ri Pi 2
× Lfi Np − + γ 2 ts eLfi (t −tk+1 ) dt (67)
1 − e fi ( s p )
( )
ev (P) ev Q i ev Q i
( ) L t −N
( )
ev
i
( )
Pi ts r
+ ( ) πi + ( ) πi ts eLfi (ts −Np ) + 1
Np + ts ev P i Lfi ev P i
from (30) we have:
ev Q ev Q πi ) i ev Q i
( i) ( i) ( )
r i ts (
( ( ) ) ≤− ( ) πi + ϑ r i ts ( ) = ϑ− r ts (74)
r ri ts r i ev P i ev P i ev P i
( )
−Lfi (Np −ts ) 2 2
ev P i + γ 2 ts
( )
Lfi Np − ≤ e
ev (P) ev P i
( )
Np + ts so it is concluded that:
(68) Ji∗ tk+1 , xi (tk+1 ) − Ji∗ tk , xi (tk ) ≤ 0
( ) ( )

By combining (67) and (68), it obtains that: Thus, the applied dynamic xi (t) enters the neighborhood of the
ev Q origin’s positive region Λr i and the proof is completed.
( i) tk +Np
ts r i
∫
( ) πi e fi ( k+1 s p ) dt
L t −t +t −N
→ I2 ≤
ev Pi Np + ts tk+1 The sufficient condition for the overall state of the nominal
interconnected nonlinear constrained large-scale system in (3) to
ev Q 1 − e i ( s p)
( i) Lf
t −N
i
ts r enter to the robust region, Λr , that is the neighborhood of the
≤ ( ) πi (69)
ev Pi Np + ts Lf i origin, is provided by the following theorem.
which proves (58).
Theorem 5. The overall nominal state of the dynamic in (3) with
The third[ integral term in ] (59) is defined over the time inter- Assumption 1 and with the following disturbance condition:
val, t ∈ tk + Np , tk+1 + Np , which similar to the time interval,
t ∈ [tk , tk + ts ], indicates the sampling time, ts . According to (12)
( )
ev Q́ σ r
in Lemma 1, it follows for the third integral term that: ∥d(t)∥2 ≤ ), σ ∈ (0, 1) (75)
2ev (P ) ev P 0.5
(
∫ tk+1 +Np (  )
x̂i (t |tk+1 )2 i + ûi (t |tk+1 )2 + v̂ij (t |tk+1 )2 dt
   
I3 =
tk +Np
Q R i R ij converges to the robust neighborhood of the origin’s positive region,
Λr .
 ( )2   ∗( 2
)
+ x̂i tk+1 + Np |tk+1 P i − xi tk + Np |tk 
Pi
Proof. It demonstrates from Lemma 2 that:
 i( )2   i∗ ( 2
)
≤ x̂ tk + Np |tk+1 P i − x tk + Np |tk 
  (70) d
(V (x(t))) ≤ −x(t)T Q́ x(t) + 2 P 0.5 x(t)2 P 0.5 2 ∥d(t)∥2 (76)
i
   
P

By the substitution of t = tk + Np in (36) we have: dt

by the substitution from (75) and the appropriate selection of σ ,
2
 ∗( )
x̂ tk + Np |tk+1 2 i −  it is concluded from (76) that dtd (V (x(t))) ≤ 0 and the system is
 i( ) i
P
 x t k + N p | tk i
P asymptotically stable.
2
 ( )
 i i∗
) (
≤ x̂ tk + Np |tk+1 − x tk + Np |tk 
( ( ) P
i
) Remark 7. The approach proposed in this paper may be con-
r ri servative for constrained interconnected large-scale nonlinear
≤ ev P − ( i ) + γ ts eLfi (Np −ts )
( i) 2
Lfi Np (71) systems, which arises from the condition of the feasibility that
ev (P) ev P
fundamentally depends on how the states of subsystems of the
10
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

interconnected nonlinear large-scale system stay close to the

trajectories of subsystems of the linearized system.

6. Simulation results

This section verifies the performance and effectiveness of the

proposed CDMPC approach by applying it to a constrained in-
terconnected nonlinear quadruple-tank system consisting of two
coupled subsystems description presented in [5,25]. The feasibil-
ity of convergence and controller and also the stability of the
proposed CDMPC approach are verified by its application on a
quadruple-tank system aimed at converging the states towards
a neighborhood of the origin in the presence of disturbance.
Each subsystem comprises of two states, each of which de-
notes its corresponding tank’s height of water (kc = 1) [26].
The subsystems are described as:
Subsystem 1:
a1 a3
√ ( ) √ ( )
ẋ1 = − 2g x1 + hSP
1 + 2g x4 + hSP
3
A1 A1 Fig. 1. The states (water level) of subsystem 1 in the minimum-phase situation
γ1 k1 ( and in the presence of disturbance.
u1 + v SP
)
+ 1 + d1 (t)
A1
a4 (1 − γ1 ) k1 (
√
u1 + v1SP + d2 (t)
( ) )
ẋ2 = − 2g x2 + hSP
4 + thus, the constraints imposed on the control inputs and states
A4 A4
include:
Subsystem 2:
a2
√ ( ) a4
√ ( ) − 3 ≤ ui (t) ≤ 3V , i = 1, 2
ẋ3 = − 2g x3 + hSP
2 + 2g x2 + hSP
4 − 12.4 ≤ x1 (t) ≤ 7.6 cm
A2 A2
γ2 k2 ( − 1.4 ≤ x2 (t) ≤ 4.6 cm
u2 + v 2 SP
)
+ + d3 (t)
A2 − 12.7 ≤ x3 (t) ≤ 7.3 cm
a3 (1 − γ2 ) k2 ( − 1.8 ≤ x4 (t) ≤ 4.2 cm
√
u2 + v2SP + d4 (t)
( ) )
ẋ4 = − 2g x4 + hSP
3 +
A3 A3
The initial state variables are:
The above-mentioned continuous-time dynamics are
x4 (0) = −7.2 −0.9 −2.7 −1.5
[ ] [ ]
discretized by applying Euler derivative approximation with sam- x1 (0) x2 (0) x3 (0)
pling time Ts = 1(Sec).
Definitions of ai , Ai , (i = 1, . . . , 4), γi , ki (i = 1, 2) and g are The prediction
[ ] horizon is Np = 4, the weighting matrices are
presented in [5]. Experimental values of these parameters are 1 0
Q́ i = and Ri = 0.1, i = 1, 2, the eigenvalues of Ai + Bi K i
presented in [26] as: 0 1
are −1 and −0.9 and [ also α = 0.9.]The matrices
i
[ K , i = 1, 2 are]
i
A1 = A3 = 28 cm2 calculated as K 1 = −600. 1543.3 and K 2 = −454.5 557.3
A2 = A4 = 32 cm2 [ r = 0.9. The matrices P i , i = [1, 2 are calculated as
i
and also
283.98 −744.25 161.6 −205.27
] ]
a1 = a3 = 0.071 cm2 , a2 = a4 = 0.057 cm2 P1 = and P 2 =
−744.25 1950.84 −205.27 260.76
g = 981 cm/s 2
and also ∥d(t)∥2 ≤ 0.08.
(k1 , k2 ) = (3.33, 3.35) cm3 /V s The states of the subsystems in the minimum-phase situation
are illustrated in the presence of disturbance in Figs. 1 and 2,
(γ1 , γ2 ) = (0.7, 0.6)
respectively. Figs. 1 and 2 show that the states of the constrained
set-point values of h3 , h4 , v1 and v1 are as below: interconnected nonlinear quadruple-tank system are conducted
to the origin using the proposed CDMPC approach in the pres-
3 = 1.8 cm, h4 = 1.4 cm, v1 = v2 = 3V
hSP SP SP SP
ence of disturbance. The control inputs of the subsystems in
the presence and absence of disturbance and in minimum-phase
which correspond to the following set-point values of h1 and
situation are depicted in Figs. 3 and 4, respectively. By comparing
h2 : hSP
1 = 12.4 cm, hSP2 = 12.7 cm The control objective is
the control input values in Figs. 3 and 4, it is concluded that
to control water levels in tanks 1 and 2 (h1 and h2 ) with the
despite the effect of disturbance on the control input values,
above-mentioned set-points.
which is shown in Fig. 3, the feasibility of the proposed approach
The above-mentioned ki , γi and set-points are correspond-
is concluded in the presence of disturbance.
ing to the minimum-phase model of the system based on the
The cost function values of the subsystems with and without
experimental data presented in [26].
disturbance and in the minimum-phase situation are illustrated
The constraints imposed on the input voltages to the pumps
in Figs. 5 and 6, respectively. The maximum, average, and stan-
and water levels are:
dard deviation of the cost function values of the subsystems with
0 ≤ vi (t) ≤ 6 V, i = 1, 2 and without disturbance and in the minimum-phase situation are
presented in Tables 1 and 2, respectively. It is drawn from the
0 ≤ hi (t) ≤ 20 cm, i = 1, 2
cost function curves that the cost is increased when disturbance
0 ≤ hi (t) ≤ 6 cm, i = 3, 4 is applied.
11
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

Fig. 5. The cost function values of both subsystems in the minimum-phase

Fig. 2. The states (water level) of subsystem 2 in the minimum-phase situation
situation and in the presence of disturbance.
and in the presence of disturbance.

Fig. 6. The cost function values of both subsystems in the minimum-phase

Fig. 3. The control inputs (input voltages to the pumps) of both subsystems in situation and with no disturbance.
the minimum-phase situation and in the presence of disturbance.
Table 1
The maximum, average, and standard deviation of the cost function values of the
subsystem 1 with and without disturbance and in the minimum-phase situation.
Subsystem 1 With disturbance Without disturbance
Max 2309.001 2032.214
Avg 98.674 85.3695
Std 355.668 308.6312

Table 2
The maximum, average, and standard deviation of the cost function values of the
subsystem 2 with and without disturbance and in the minimum-phase situation.
Subsystem 2 With disturbance Without disturbance
Max 708.968 357.248
Avg 25.848 9.192
Std 96.394 41.996

Now, the proposed CDMPC approach is investigated when the

above-mentioned ki , γi and set-points are corresponding to the
nonminimum-phase model of the quadruple-tank system based
on the experimental data presented in [26] as:
Fig. 4. The control inputs of both subsystems in the minimum-phase situation
and with no disturbance. (k1 , k2 ) = (3.14, 3.29) cm3 / V s
12
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

Fig. 7. The states of subsystem 1 in the nonminimum-phase situation and in Fig. 8. The states of subsystem 2 in the nonminimum-phase situation and in
the presence of disturbance. the presence of disturbance.

(γ1 , γ2 ) = (0.43, 0.34)

3 = 4.8 cm, h4 = 4.9 cm, v1 = v2 = 3.15 V
hSP SP SP SP

1 = 12.6 cm, h2 = 13 cm
hSP SP

So, the constraints imposed on the control inputs and states are
as below:

− 3.15 ≤ ui (t) ≤ 2.85 V, i = 1, 2

− 12.6 ≤ x1 (t) ≤ 7.4 cm
− 4.9 ≤ x2 (t) ≤ 1.1 cm
− 13 ≤ x3 (t) ≤ 7 cm
− 4.8 ≤ x4 (t) ≤ 1.2 cm
The initial state variables are:

x4 (0) = −2.4 −0.45 −2.3 −0.5

[ ] [ ]
x1 (0) x2 (0) x3 (0)

Figs. 7 and 8 depict the results as to the states of the subsystems

Fig. 9. The control inputs of both subsystems in the nonminimum-phase
for nonminimum-phase model of the quadruple-tank system in
situation and in the presence of disturbance.
the presence of disturbance. They demonstrate that the states of
constrained interconnected nonlinear quadruple-tank system for
a nonminimum-phase model are conducted to the origin using
the proposed CDMPC approach in the presence of disturbance.
The control inputs of the subsystems with and without distur-
bance are displayed in Figs. 9 and 10, respectively. By comparing
the control input values in these two figures, it is concluded
that despite the effect of disturbance on the control input values,
which is shown in Fig. 9, the feasibility of the proposed approach
is supported in the presence of disturbance for the nonminimum-
phase model. The results of states and control inputs emphasize
the performance of the proposed CDMPC approach in dealing
with the nonminimum-phase specifications of the system.
The cost function values of the subsystems with and without
disturbance and in the nonminimum-phase situation are shown
in Figs. 11 and 12, respectively. The maximum, average, and stan-
dard deviation of the cost function values of the subsystems with
and without disturbance and in the nonminimum-phase situation
are presented in Tables 3 and 4, respectively. According to the
cost function curves, it is concluded that the cost is increased
when the disturbance is applied.
All the simulations are done using MATLAB on a personal com- Fig. 10. The control inputs of both subsystems in the nonminimum-phase
puter (CPU: Intel (R) Core (TM) i7-4790 K, 4.00 GHz; 4.00 GHz: situation and with no disturbance.

13
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

Table 4
The maximum, average, and standard deviation of the cost function values of
the subsystem 2 with and without disturbance and in the nonminimum-phase
situation.
Subsystem 2 With disturbance Without disturbance
Max 296.753 214.964
Avg 27.111 19.163
Std 51.894 36.849

Fig. 11. The cost function values of both subsystems in the nonminimum-phase
situation and in the presence of disturbance.

Fig. 13. The cost function values of both subsystems, obtained by the method
presented in [5], in the minimum-phase situation and with no disturbance.

that the method proposed here decreases the simulation time

in both minimum-phase and nonminimum-phase situations and
in the presence of disturbance. Moreover, [5] and [25] ignore
the disturbance and do not consider the nonminimum-phase
situation.
According to Figs. 1 and 2 in the minimum-phase situation and
Figs. 7 and 8 in the nonminimum-phase situation, it is concluded
that the convergence quality and time are improved in compari-
son with [5]. The convergence times of our proposed method and
the method proposed in [5] are shown in Table 6. It should be
Fig. 12. The cost function values of both subsystems in the nonminimum-phase noted that in [5], the disturbance and the nonminimum-phase
situation and with no disturbance.
situation are ignored.
The results of the simulation and convergence times, which
Table 3
are given in Tables 5 and 6 respectively, confirm that the compu-
The maximum, average, and standard deviation of the cost function values of
the subsystem 1 with and without disturbance and in the nonminimum-phase tation burden is reduced using the proposed method.
situation. To further emphasize the advantages of the proposed method
Subsystem 1 With disturbance Without disturbance in this paper, the cost function values of the method proposed
Max 329.505 228.995 in [5] are depicted in Fig. 13. The disturbance is ignored in this
Avg 17.94 12.101 method and the nonminimum-phase situation is not considered.
Std 46.143 30.974 The comparison of the results presented in Figs. 5 and 6 and the
results presented in Figs. 11 and 12 with the results presented in
Fig. 13 reveals that the cost function values are reduced using the
CDMPC method proposed in this paper.
RAM: 32.00 GB). The simulation times of the proposed coopera-
The maximum, average, and standard deviation of cost func-
tive DMPC method in both minimum-phase and nonminimum-
tion values, obtained by the method presented in [5], are pre-
phase situations and in the presence of disturbance are shown
sented in Table 7. By comparing Tables 1 to 4 with Table 7, it
in Table 5 along with the simulation times of the methods pre-
is concluded that the maximum, average, and standard deviation
sented in [5] and [25]. It is evident from the simulation times

Table 5
The simulation times of the proposed cooperative DMPC method and the presented methods in [5] and [23].
Proposed CDMPC in the Proposed CDMPC in the [25] [5]
minimum-phase situation nonminimum-phase situation
and with disturbance and with disturbance
Simulation time (s) 5.708 7.188 23.35 35.72

14
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

Table 6
The convergence times of the proposed cooperative DMPC and the method proposed in [5].
Proposed CDMPC in the Proposed CDMPC in the Reference
minimum-phase situation nonminimum-phase situation [5]
and with disturbance and with disturbance
x1 80 110 80
x2 100 105 190
x3 130 160 80
x4 140 140 150

Table 7
The maximum, average, and standard deviation of cost function values, obtained
by the method presented in [5].
Subsystem 1 Subsystem 2
Max 3457.607 876.495
Avg 171.6617 32.481
Std 538.138 112.236

Fig. 16. The standard deviation values of cost functions of both subsystems,
obtained by the CDMPC method proposed in this paper and the method
presented in [5].

input by solving a global cost function that considers the effect of

all other control inputs. The feasibility of the proposed approach
is presented and derived via simulation results. The closed-loop
Fig. 14. The maximum values of cost functions of both subsystems, obtained by stability of the system are provided via the presented sufficient
the CDMPC method proposed in this paper and the method presented in [5]. conditions. The performance and effectiveness of the proposed
approach are examined by applying it to the nonlinear quadruple-
tank system for both minimum and nonminimum-phase models.
Methods to address the conservativeness of the proposed ap-
proach could be addressed in future works. Another topic for
future research is to design a dual-mode cooperative optimization
strategy for the constrained interconnected nonlinear large-scale
systems that have more than one equilibrium point.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared
to influence the work reported in this paper.

References

[1] Christofides PD, Scattolini R, De la Pena DM, Liu J. Distributed model

Fig. 15. The average values of cost functions of both subsystems, obtained by
predictive control: A tutorial review and future research directions. Int
the CDMPC method proposed in this paper and the method presented in [5].
J Comput Chem Eng 2013;51:21–41.
[2] Santana DD, Martins MAF, Odloak D. An efficient cooperative-distributed
model predictive controller with stability and feasibility guarantees for
values are reduced using the CDMPC method proposed in this constrained linear systems. Int J Syst Control Lett 2020;141:1–11.
paper versus the method presented in [5]. The above mentioned [3] Bestler A, Graichen K. Distributed model predictive control for continuous-
time nonlinear systems based on suboptimal ADMM. Int J Opt Control Appl
comparison for maximum, average, and standard deviation of cost
Methods 2019;40(1):1–23.
function values are illustrated in Figs. 14–16 respectively. [4] Shen C, Shi Y. Distributed implementation of nonlinear model predictive
control for AUV trajectory tracking. Int J Autom 2020;115:1–9.
7. Conclusion [5] Grancharova A, Johansen TA, Olaru S. Dual-mode distributed model pre-
dictive control of a quadruple-tank system. Int J Chem Technol Metall
2018;53(4):674–82.
This paper proposes a novel cooperative architecture of DMPC [6] Pourkargar DB, Moharir M, Almansoori A, Daoutidis P. Distributed estima-
for constrained interconnected nonlinear large-scale systems. In tion and nonlinear model predictive control using community detection.
the proposed approach, each subsystem computes its control Int J Indust Eng Chem Res 2019;58(30):13495–507.

15
A. Mirzaei and A. Ramezani ISA Transactions xxx (xxxx) xxx

[7] Bai T, Li S, Zheng Y. Distributed model predictive control for networked [18] Zheng Y, Li SE, Li K, Borrelli F, Hedrick JK. Distributed model predictive
plant-wide systems with neighbourhood cooperation. Int J IEEE/CAA J control for heterogeneous vehicle platoons under unidirectional topologies.
Autom Sin 2019;6(1):108–17. IEEE Trans Control Syst Technol 2018;25(3):899–910.
[8] Sadati N, Rahmani M, Saif M. Two-level robust optimal control of [19] Long Y, Liu S, Xie L, Johansson KH. Distributed nonlinear model predictive
large-scale nonlinear systems. Int J IEEE Syst J 2015;9(1):242–51. control based on contraction theory. Internat J Robust Nonlinear Control
[9] Zhao M, Ding B. Distributed model predictive control for constrained 2018;28(2):492–503.
nonlinear systems with decoupled local dynamics. ISA Trans 2015;55:1–12. [20] Esfahani NR, Khorasani K. A distributed model predictive control (MPC)
[10] Liu X, Shi Y, Constantinescu D. Distributed model predictive control of fault reconfiguration strategy for formation flying satellites. Internat J
constrained weakly coupled nonlinear systems. Int J Syst Control Lett Control 2016;89(5):960–83.
2014;74:41–9. [21] Kong X, Liu X, Ma L, Lee KY. Hierarchical distributed model predictive
[11] Ravi A, Kaisare NS. Distributed output feedback control for multi-unit control of standalone wind/solar/battery power system. IEEE Trans Syst
system with delayed multirate measurements. Int J Comput Chem Eng Man Cybern Syst 2019;49(8):1570–81.
2020;139:1–13. [22] Liu Q, Abbas HS, Velni JM. An LMI-based approach to distributed model
[12] Yamchi MH, Esfanjani RM. Distributed predictive formation control of predictive control design for spatially-interconnected systems. Int J Autom
networked mobile robots subject to communication delay. Int J Robot 2018;95:481–7.
Autonom Syst 2017;91:194–207. [23] Stoyanova I, Gumrukcu E, Aragon G, Hidalgo-Rodriguez DI, Monti A,
[13] Liu X, Shi Y, Constantinescu D. Robust distributed model predictive Myrzik J. Distributed model predictive control strategies for coordination
control of constrained continuous-time nonlinear systems using two-layer of electro-thermal devices in a cooperative energy management concept.
invariant sets. Int J Dynam Syst Meas Control 2016;138(6):1–7. Int J Opt Control Appl Methods 2020;41(1):170–89.
[14] Ma A, Liu K, Zhang Q, Xia Y. Distributed MPC for linear discrete-time [24] Li A, Sun J. Stability of nonlinear system under distributed Lyapunov-
systems with disturbances and coupled states. Int J Syst Control Lett based economic model predictive control with time-delay. ISA Trans
2020;135:1–8. 2020;99:148–53.
[15] Guo Z, Li S, Zheng Y. Feedback linearization based distributed model [25] Mirzaei A, Ramezani A. Cooperative distributed constrained adaptive gen-
predictive control for secondary control of islanded microgrid. Int J Asian eralized predictive control for uncertain nonlinear large-scale systems:
J Control 2020;22(1):1–14. Application to quadruple-tank system. Int J Electr Comput Eng Innov
[16] Li H, Shi Y. Robust distributed model predictive control of constrained 2020;7(2):183–94.
continuous-time nonlinear systems: A robustness constraint approach. IEEE [26] Johansson KH. The quadruple-tank process: A multivariable labora-
Trans Automat Control 2014;59(6):1673–8. tory process with an adjustable zero. IEEE Trans Control Syst Technol
[17] Liu X, Constantinescu D, Shi Y. Robust model predictive control of con- 2000;8(3):456–65.
strained non-linear systems: adopting the non-squared integrand objective
function. Int J IET Control Theory Appl 2015;9(5):649–58.

16