2016 American Control Conference (ACC)

Boston Marriott Copley Place

July 6-8, 2016. Boston, MA, USA

Distributed Model Predictive Control of Constrained Spatially-Invariant

Interconnected Systems in Input-Output Form∗
Qin Liu, Hossam S. Abbas, Javad Mohammadpour, Simon Wollnack, Herbert Werner

Abstract— This paper proposes a non-iterative, Lyapunov- requires the online measurement or estimation of the states.
based distributed model predictive control (MPC) design for In practice, an online state estimator is often constructed
invariant spatially-interconnected systems comprised of a net- separately from MPC; it can lead to a non-trivial stability
work of subsystems with coupled dynamcis and subject to
local state and/or input constraints. Considered here is the proof. Furthermore, the state-space based controller design
distributed MPC design for linear systems in input-output form normally results in controllers of full order (equal or larger
with fixed-structure local state-feedback-like control law. The than the plant order), whereas the input-output framework al-
proposed distributed MPC design approach ensures asymptotic lows to design fixed-structure controllers that have predefined
stability and recursive feasibility, and its online implementation orders [8]. The advantage of considering input-output form
can be formulated as solving a linear matrix inequality (LMI)
problem defined at the subsystem level. Simulation results becomes more obvious when dealing with linear parameter-
using a heat equation in one-dimensional space demonstrate varying systems [9].
the merits and effectiveness of the proposed approach. Inspired by the results reported in [6] and [7], this
paper studies the distributed MPC design for spatially-
I. INTRODUCTION interconnected systems in input-output form, where we con-
In practice, nearly every application is subject to various sider Lyapunov-based stability analysis and a local controller
design and operational constraints. Model predictive control design of fixed structure. We consider the class of spatially-
(MPC) has a long history in industry to deal with system distributed systems, where the attachment of actuators and
constraints. Not restricted to lumped systems, distributed sensors induces the spatial discretization of the overall
MPC has become an active research area in the last decade system into a network of interconnected subsystems, with
to cope with large, networked systems, which can either be each subsystem interacting with its nearest neighbours. Ex-
physically coupled, e.g., distributed parameter systems [1], amples include partial differential equation (PDE)-governed
or consist of a network of systems with decoupled dynamics, distributed systems, environmental systems [10], and canal
but coupled via constraints or a common objective, e.g., systems [11] among many others. The distributed MPC
multi-agent systems [2]. Distributed MPC allows information should preserve the localized structure of the plant, with
exchange among a small number of subsystems, as opposed stability and recursive feasibility guaranteed.
to a centralized MPC scheme. A review of recent results in This paper is organized as follows: Section II reviews
the design of distributed MPC can be found in [3] and [4]. the input-output representation for the class of spatially-
The vast majority of the published literature, either on interconnected systems considered in this paper. In Section
control of unconstrained distributed systems or control of III, a novel distributed MPC design approach in input-output
constrained systems using anti-windup or MPC approach, form is presented. A numerical example is used in Section IV
rely on state-space representation of plant dynamics, and thus to illustrate its closed-loop performance. Finally, conclusions
analysis and synthesis conditions are derived in state-space are drawn in Section V.
form e.g., [5]. Only very few address these issues based on
input-output model, e.g., [6] for unconstrained distributed II. PRELIMINARIES
controller design and [7] for MPC design in lumped systems. In this paper, we consider the class of spatially-
The importance of controller design in input-output form interconnected systems whose subsystems exhibit linear
is related to the fact that the experimentally identified models time- and space-invariant (LTSI) properties. This section
are in input-output form. It is then natural to solve for the reviews the two-dimensional difference equation that defines
controller based on the input-output dynamics and implement the distributed dynamics at a single subsystem, and presents
it digitally. MPC design based on state-space formulation an implicit representation of the closed-loop system.
∗ This work was supported by the German Research Foundation (DFG)
A. Input-Output Representation
through the research fellowship Li 2763/1-1.
Q. Liu, J. Mohammadpour are with the Complex Systems Control An LTSI system of one spatial dimension consists of Ns
Laboratory, College of Engineering, University of Georgia, Athens, GA identical subsystems, the dynamics of each governed by a
30602, USA {qinliu, javadm}@uga.edu two-dimensional transfer function
H. S. Abbas is with the Electrical Engineering Department,
Faculty of Engineering, Assiut University, Assiut 71515, Egypt
hossam.abbas@aun.edu.eg G(qt , qs ) : A(qt , qs )y(k, s) = B(qt , qs )u(k, s), (1)
S. Wollnack, H. Werner are with the Institute of Control Sys-
tems, Hamburg University of Technology, 21075 Hamburg, Germany (s = 1, . . . , Ns ), where qt and qs are forward temporal
{simon.wollnack, h.werner}@tuhh.de and spatial shift operators, respectively, y(k, s) ∈ Rny and

978-1-4673-8682-1/$31.00 ©2016 AACC 3600

u(k, s) ∈ Rnu are the output and input of subsystem s at Fixed-structure controller design (in the sense of fixed
time instant k, respectively. The polynomials A and B are temporal and spatial orders) allows us to determine the
defined as orders of the controller and weighting filters independently
via the choices of the input and output masks (MuK , MyK ),
a(ik ,is ) qt−ik qs−is ,
X
A(qt , qs ) = 1 + (2a)
(MuWs , MyWs ) and (MuWk , MyWk ), respectively. Considered
(ik ,is )∈My
here are strictly proper plant models, whereas the controller
b(jk ,js ) qt−jk qs−js ,
X
B(qt , qs ) = (2b) and weighting filters can in general be either bi-proper or
(jk ,js )∈Mu strictly proper.
It is desired that the controller inherits the spatial structure
where Mu and My are input and output masks, respectively,
of the plant, i.e., a distributed controller of an LTSI plant
determining how temporally- and spatially-shifted inputs and
is assumed to be an LTSI system as well. A distributed
outputs contribute to the dynamics of subsystem s, a(ik ,is )
controller takes the form
and b(jk ,js ) are coefficients weighting the contribution of
time and space samples. An example of masks is shown K(qt , qs ) : AK (qt , qs )u(k, s) = BK (qt , qs )e(k, s). (3)
in Fig. 1, where black dots indicate contributing input and The transfer functions of weighting filters Ws (qt , qs ) and
output samples. This choice of masks indicates that output Wk (qt , qs ) can be constructed in a similar way.
of subsystem s is directly determined by its past input and
outputs of itself and neighbouring subsystems s−1 and s+1. B. Implicit Representation
Without loss of generality, we consider single-input and
single-output subsystem G(qt , qs ). Let Ā and B̄ denote
Mu js My is vectors containing polynomial coefficients of A(qt , qs ) and
B(qt , qs ), respectively, ξ y (k, s) ∈ Rnξy and ξ u (k, s) ∈ Rnξu
jk ik the conformable temporally- and spatially-shifted outputs
and inputs, respectively. With the coefficient and signal
vectors for operators K, Ws and Wk defined accordingly,
Fig. 1. Input and output masks as an example. an implicit representation of the closed-loop system in Fig. 2
can be formulated as
In this paper, we assume that all subsystems interact with 
Ā −B̄ 0 0 0
 y
ξ (k, s)

neighbouring subsystems in the same pattern, i.e., the input  B̄K ĀK 0 0 −B̄K   u
 ξ z (k, s) :=Hξ(k, s) = 0,

and output masks, as well as the corresponding coefficients,

B̄Ws 0 ĀWs 0 −B̄Ws  ξ (k, s) 
are identical for all subsystems. 0 B̄Wk 0 −ĀWk 0 ξ w (k, s)
(4)
Definition 1 (Neighbourhood Ni ) Subsystem j exchanges
with ξ ∈ Rnξ . Note that to ensure the size compatibility
information with subsystem i (i 6= j) if j ∈ Ni , where Ni
of matrix H in case of different operator orders, the miss-
denotes the directly interconnected neighbouring subsystems
ing polynomial terms are filled with zero coefficients; this
of subsystem i.
implies that nξy = nξu = nξz = nξw [6].
A system with masks defined as Fig. 1 has a neighbour- In (4), each signal vector ξ i (k, s) (i = y, u, z, w) can be
hood Ns = {s− 1, s+ 1} of size nNs = 2. Here we focus on partitioned into three parts as
undirected communication between subsystems, i.e., i ∈ Nj T
ξ i (k, s) = ξtiT (k, s) ξ+
iT iT

(k, s) ξ− (k, s) , (5)
and j ∈ Ni .
In this paper, we consider the control problem of tracking where ξti (k, s) denotes the temporal variables, ξ+ i
(k, s) and
i
reference signal w(k, s). The closed-loop system at sub- ξ− (k, s) the positive and negative spatial variables, respec-
system s with performance channels incorporated is shown tively. Subsystems exchange information through spatial
in Fig. 2, where Ws (qt , qs ) and Wk (qt , qs ) are weight- states. Furthermore, a state-like vector x(k, s) ∈ Rnx can
ing filters for shaping the sensitivity and controller sen- be extracted from ξ(k, s) as
   
sitivity of the closed-loop system, respectively, z(k, s) = xt (k, s) xt (k + 1, s)
[zsT (k, s), zkT (k, s)]T is the corresponding fictitious perfor- x(k, s) = x+ (k, s) , ∆x(k, s) = x+ (k, s + 1) ,
mance output. x− (k, s) x− (k, s − 1)
where the augmented shift operator ∆ is defined as
zs (k, s)
Ws (qt , qs ) ∆ = diag{qt , qs , qs−1 }. Permutation matrices Π1 =
zk (k, s) diag{Π1t , Π1+ , Π1− } and Π2 = diag{Π2t , Π2+ , Π2− } are
Wk (qt , qs )
properly selected such that
w(k, s) y(k, s) x(k, s) = Π1 ξ(k, s), ∆x(k, s) = Π2 ξ(k, s).
K(qt , qs ) G(qt , qs )
e(k, s) u(k, s)
Denote the closed-loop system by L. A spatially-
interconnected system in one-dimensional space is shown in
Fig. 2. Generalized plant for subsystem s. Fig. 3.

3601
 z(k, s − 1) w(k, s − 1) z(k, s) w(k, s) z(k, s + 1) w(k, s + 1)
references at subsystem s, as well as its neighbourhood Ns ,
x+ (k, s) x+ (k, s + 1) as performance measure to account for the influence of the
input trajectory u(k + i, s) to s and Ns . Thus a single
L L L
subsystem does not compete for its own minimum cost at the
x− (k, s − 1) x− (k, s)
price of an increased cost at other subsystems. A collective
improvement can be expected through cooperation among
Fig. 3. Spatially-interconnected system in one-dimensional space.
localized subsystems.
The distributed MPC problem for a spatially-
III. D ISTRIBUTED MPC D ESIGN interconnected system in the presence of input constraint
can be formulated at subsystem s as
When dealing with a spatially-interconnected system, a
centralized MPC scheme requires that each subsystem ex- Ps : min VNs (x, u, w) (9a)
u
changes information with all other subsystems—it renders s. t. Hξ(k, s) = 0, (9b)
both the computation and implementation intractable. Al- s
though a distributed MPC approach is often more conserva- u(k + i, s) ∈ U , ∀i = 0, 1, . . . , N − 1 (9c)
tive than the central one, by exploiting the localized nature xt (k + N + 1, s) ∈ Xsf , (9d)
of subsystems, a distributed MPC based on information where Xsf is the terminal set containing the origin, whose
received from neighbouring subsystems is of smaller order significance will become clear soon. Similar to the input
and much easier to handle. constraint, the global terminal set can be defined as the
The following assumptions are made for the distributed Cartesian product of subsystem ones
MPC design:
i) There are no disturbance and noise present in the Xf = X1f × X2f × . . . XN
f .
s
(10)
closed-loop system and no modelling error of the plant. To ensure stability and recursive feasibility when solving
ii) For the sake of presentation simplicity, only input finite horizon optimization problem, the selection of the
constraints are considered here. Furthermore, we as- terminal cost Vf (·), the terminal set Xsf , and a local control
sume input constraints imposed at all subsystems are law valid in Xsf plays an important role. In the rest of the
identical, i.e., U1 = U2 = . . . = UNs , with the compact section, the offline computation of a local control law and the
set Us (s = 1, . . . , Ns ) defined as terminal set, and the formulation of the online optimization
Us = {u(k, s) ∈ R| − umax ≤ u(k, s) ≤ umax }, (6) problem Ps (9) in terms of linear matrix inequalities (LMIs)
are discussed.
where umax is the maximum capacity of a physical
A. Local Control Law and Terminal Cost
actuator.
The global control constraint is thus denoted as the Carte- For a constrained system, a global Lyapunov function can
sian product of all subsystem constraints hardly be found which fulfils Lyapunov-based stability con-
ditions [12]. Instead, a dual mode control is often employed:
U = U1 × U2 × . . . UN s . (7) within the prediction horizon, the predicted inputs are op-
timization variables; over the remaining infinite horizon of
Although only the input constraints are addressed, the ap-
mode 2, inputs are determined by a stabilizing control law.
proach presented in this paper can be easily extended to more
The objective of the MPC design is to steer the system state
general cases accounting for state or combined constraints.
into the terminal set at the end of the prediction horizon,
Taking the closed-loop system in Fig. 2 into account, we
where a local control law takes over. The computation of the
propose a cost function at subsystem s as
terminal set will be address in Section III-B. In this section,
N
X −1 analysis and synthesis conditions for a stabilizing control law
VNs (x, u, w) = l(ẽ(k + i + 1, s), u(k + i, s)) are provided.
i=0 Consider a Lyapunov function candidate as
+ Vf (xt (k + N + 1, s)), (8)
V (xt (k, s)) = xt (k, s)T Pt xt (k, s), Pt > 0. (11)
where N is the prediction horizon, control inputs u(k + i, s)
are the decision variables, ẽ(k + i + 1, s) ∈ RnNs +1 is a We are interested in controller design which establishes
vector formed by stacking up the error signals e(k +i+1, j), asymptotic stability, as well as certain performance criterion
j ∈ Ns ∪ s. The functions l(·) and Vf (·) are the so-called of the closed-loop system. Take the induced L2 norm as
stage cost and terminal cost, respectively. We define the stage performance measure. The closed-loop system is said to have
cost function as a positive definite function a quadratic performance γ if
Ns XNt   
X  T I 0 z(k, s)
qj e2 (k + i + 1, j) + ru2 (k + i, s),
X
l(·) = ∗ < 0. (12)
j∈Ns ∪s
0 −γ 2 I w(k, s)
s=1 k=0

with weighting factors qj > 0 and r > 0. The local Analysis conditions for the closed-loop system inside the
stage cost function takes the deviations between outputs and terminal set, i.e., Hξ = 0, xt (k, s) ∈ Xsf , can be derived by

3602
extending results for unconstrained LTSI systems reported After the temporal Lyapunov matrix Pt and a controller in
in [6], where distributed controller design has the size of a the form of (3) have been obtained, the terminal cost Vf (·)
single subsystem. in (8) is often chosen to be the Lyapunov function [12], i.e.,
Vf (xt (k, s)) = xT
t (k, s)Pt xt (k, s). (19)
Theorem 1 Assuming the well-posedness of the closed-loop
system (4), it is asymptotically stable and has a quadratic Note that the controller can be rewritten in a state-
performance γ (γ > 0) inside the terminal set, if there exist feedback-like control law
a matrix F and a symmetric matrix P , such that u(k, s) = Kfs ξ(k, s), (20)
 
Pt where the feedback gain ∈R Kfs 1×nξ
is a function of ĀK
P = , Pt > 0, det(Ps ) 6= 0, (13a)
Ps and B̄K .
U (P ) + M (γ) + F H + H T F T < 0, (13b) Remarks:
• The spatial state vectors x+ (k, s) and x− (k, s) and the
with
spatial Lyapunov matrix Ps are neither included in the
U (P ) = ΠT T T Lyapunov function, nor in the terminal cost. This is
4 P Π4 − Π3 P Π3 + Π5 SΠ5 ,
due to the spatial non-causality: the spatial dynamics
M (γ) = diag{0, I, −γ 2I},
cancel out after summed up over spatial coordinate, see
and [14] [6] for more details. Furthermore, the indefiniteness
T T of the spatial Lyapunov matrix Ps violates the positive
Π3 = ΠT ΠT ΠT Π4 = ΠT ΠT ΠT
 
1t 1+ 2− , 2t 2+ 1− . definiteness of a Lyapunov function.
• Energy dissipation of the global system implies stability.
Proof: It has been proven in [6] that an unconstrained Conservatism may be induced when energy decreased
LTSI system is asymptotic stable with monotonously at each subsystem is required. Possible
solutions for relaxing this constraint can be found in
V (xt (k + 1, s)) − V (xt (k, s)) < 0 (14)
[15].
fulfilled and has quadratic performance γ (12), if there exist B. Terminal Set
P and F such that (13a) and
Stability requires that the overall terminal set Xf (10)
ξ T (ΠT T T T
4 P Π4 −Π3 P Π3 +M (γ)+F H +H F )ξ < 0. (15)
is control invariant [12] for the controlled system under
constraints. In this work, by allowing certain conservatism,
To include the stage cost l(·) in (14) [12] such that we ensure the control invariance of the subsystem terminal
set Xsf , i.e., the existence of Kfs for all xt (k, s) ∈ Xsf ,
V (xt (k + 1, s)) − V (xt (k, s)) ≤ −l(ẽ(k − 1, s), u(k − 1, s)), such that Kfs ξ(k, s) ∈ Us , and xt (k + 1, s) ∈ Xsf . A
a permutation matrix Π5 is introduced with common choice for Xsf is a suitably small sub-level set of
  Vf (xt (k, s)), i.e., Xsf is an ellipsoidal set [12] defined as
ẽ(k − 1, s)
Π5 ξ(k, s) = . (16) Xsf := {xt (k, s)|Vf (xt (k, s)) ≤ αs }, αs > 0. (21)
u(k − 1, s)
The terminal sets Xsf can be computed offline by maxi-
A weighting matrix S = diag{. . . , qj , . . . , r}, j ∈ Ns ∪ s
mizing αs , such that the input constraints are satisfied inside
with predetermined weighting factors is constructed in a
the terminal region, i.e., Kfs ξ(k, s) ∈ Us , ∀xt (k, s) ∈ Xsf .
conformable way. The stage cost function can then be written
The computation of Xsf is formulated as
in a quadratic form as
max αs (22a)
l(ẽ(k − 1, s), u(k − 1, s)) = ξ T ΠT
5 SΠ5 ξ. (17) s
α ,ξ

s. t. xT s
t (k, s)Pt xt (k, s) ≤ α , (22b)
The application of the S-procedure [13] to (15) and (17)
results in (13b). |Kfs ξ(k, s)| ≤ umax . (22c)
Provided the analysis condition, controller synthesis con- Inequality (22b) can be rewritten as a function of ξ(k, s) as
dition can be accordingly derived by solving
ξ T (k, s)ΠT s
1t Pt Π1t ξ(k, s) ≤ α . (23)
min γ, (18)
ĀK ,B̄K ,F,P By eliminating ξ(k, s) in (22c) and (23), (22) can be refor-
mulated as
such that (13a) and (13b) hold.
Note that for controller synthesis, (13b) turns into a max
s
α̃s (24a)
α̃
bilinear matrix inequality (BMI) condition due to the fact that −1 sT
s. t. α̃s Kfs (ΠT
1t Pt Π1t ) Kf ≤ u2max , (24b)
the matrix H is a function of the decision variables ĀK and
s s 2 s s
B̄K . A DK-iteration based approach has been demonstrated with α̃ = (α ) . After the maximal α̃ (or α ) has been
in [6] to solve (13) effectively. computed, the local terminal set Xsf is determined.

3603
 Remarks: Given the definition of the ellipsoidal terminal set (21),
• Recursive feasibility has been guaranteed due to the the constraints (9c)-(9d) are characterized in LMI form as
control invariance of the terminal set. Once the terminal − umax ≤ u(k + i, s) ≤ umax , i = 0, . . . , N − 1, (27)
state enters the terminal set, the stabilizing local control  −1 
law ensures that the closed-loop state trajectory stays Pt xt (k + N + 1, s)
> 0. (28)
inside, and drives the state converging either to the ∗T αs
origin (in regulator problem) or a steady state (reference The online implementation of the proposed distributed
tracking). MPC can be summarized as follows:
s
• For reference tracking, the steady-state set X̄ needs
s s
to be a subset of the terminal set X̄ ⊂ Xf to ensure Algorithm 1 Distributed MPC Design (online)
feasibility, i.e., X̄s = {x̄t (k, s)|x̄T
t (k, s)Pt x̄t (k, s) < 1: for k = 1 to Nt
αs }, where x̄t (k, s) denotes the temporal state vector 2: solve (25)-(28) for up (k, s), s = 1, . . . , Ns , in parallel
in steady state. If this condition is violated, the desired at all subsystems
steady states are unreachable. To cope with it, the 3: apply u(k, s) as control action to subsystem s
concept of pseudo setpoints has been employed in [16] 4: end
to handle infeasible references.
Provided the previous discussions on the offline compu- Remarks:
tation of the local control law, the terminal cost, and the • The deviation ẽ in (26) involves the computation of
terminal set, stability and performance conditions inside the predicted outputs at subsystems s and Ns . The decision
terminal set can be summarized as follows. variables up (k, s) at subsystem s and spatial informa-
tion received from its neighbourhood Ns contribute
Theorem 2 An LTSI system whose subsystem dynamics are to the predicted output y(k + i, s), while the updated
defined by (4) under constraint u(k, s) ∈ Us (s = 1, . . . , Ns ) y(k + i, s) in turn influences the predicted outputs
is asymptotically stable and has a quadratic performance γ y(k + i + 1, j), j ∈ Ns , due to the fact that we con-
inside the terminal set (21), if there exist a local Lyapunov sider undirected communication between subsystems.
function (11) and a local control law (20) satisfying (13) for To keep the computation at the subsystem level, during
all xt (k, s) ∈ Xsf . the implementation of (9) at subsystem s, we assume
C. Online Optimization Problem that information that the neighbourhood Ns receives
from their own neighbourhoods remains the same as
After the offline computation of a local control law Kfs (·),
the last time instant before the new optimization starts,
the terminal cost Vf (·) and local terminal set αs , the opti-
except the updated y(k + i, s).
mization problem (9) is online implemented simultaneously
• As the decision variables, a trajectory of inputs up (k, s)
at all subsystems at each time instant. In this section, the
is to be optimized. Nevertheless, the computation of the
online convex optimization problem (9) is to be characterized
terminal state xt (k + N + 1, s) in (26) and (28) requires
in terms of LMIs.
the value of input at time k + N (one step after the
The minimization of the cost function VNs (x, u, w) equal-
prediction horizon), which can be computed using the
izes the minimization of a positive scalar β, such that
local control law valid in the terminal set as (20).
VNs (x, u, w) < β. By rewriting (8) in a quadratic form,
• The shaping filters Ws and Wk can be selected either
with the application of Schur complement, the minimization
as temporal or spatio-temporal systems. The predicted
problem (9a) can then be formulated as an LMI condition
fictitious outputs z(k + N, s) in (26) and (28) can be
min β, (25) computed in the same way as the predicted output y(k+
β,u
 −1 N, s) using their governing transfer functions.
Q ẽp (k, s)

 R−1 up (k, s)  IV. S IMULATION R ESULTS
s.t.  Pt−1 xt (k + N + 1, s) > 0, (26)

 To demonstrate the performance of the proposed approach,
∗ T
∗ T
∗T β the example used in [6] – heat equation that describes the
distribution of heat in a region – is borrowed here. Its
where governing PDE is given by
   
ẽ(k + 1, s) u(k, s)
.. .. ∂y(k, s) ∂ 2 y(k, s)
ẽp (k, s) :=

, up (k, s) :=
 
,
 −κ = u(t, s), (29)
. . ∂t ∂x2
ẽ(k + N − 1, s) u(k + N − 1, s) where the positive constant κ denotes the thermal diffusivity.
xt (k + N + 1, s) = [y(k + N, s), u(k + N, s), Let the sampling time and sampling space be chosen as
Tt = 0.01 s and Ts = 0.25 m, respectively, and κ = 1.
z(k + N, s), w(k + N, s)]T ,
After applying the finite difference method to discretize (29)
Q and R are diagonal matrices containing qj (j ∈ Ns ∪ s) both in time and space with Ns = 21, the resulting two-
and r, respectively, Pt has been solved from (13). dimensional difference equation has the input and output

3604
masks as shown in Fig. 1, and the vectors of polynomial 2
coefficients

y(k, 11)
    1 reference
Ā = 1 −1 + 2ν −ν −ν , B̄ = 0 Tt 0 0 , unconstrained
constrained w/o MPC
0
with ν = Tt /Ts2 . constrained w/ MPC
The structure of the local controller law is fixed as
h i
ĀK = 1 aK a K
a K
(1,0) (1,−1) (1,1) ,
1.5
umax

u(k, 11)
h i
B̄K = bK b K
b K
b K
(0,0) (1,0) (1,−1) (1,1) .
1

0.5
A generalized plant shaping the sensitivity of the closed-loop
system is considered, with the weighting filter Ws chosen as 0
5 10 15 20 25
a temporal system time samples
   
ĀWs = 1 −0.999 0 0 , B̄Ws = 0 0.02 0 0 . Fig. 4. Comparisons of the closed-loop response (upper) and control
input (lower) at subsystem 11: unconstrained closed-loop system (blue),
The choices of the controller and the weighting filter lead to and constrained systems without (red) and with (black) distributed MPC.
the vector ξ ∈ R16 , with the extracted temporal state vector
 T
xt (k, s) = y(k − 1, s) u(k − 1, s) z(k − 1, s) w(k, s − 1) . R EFERENCES
[1] R. R. Negenborn and B. de Schutter, “A Distributed Model Predictive
Control Approach for the Control of Irrigation Canals,” in Proc. 1st
An input constraint umax = 1 is imposed on all sub- Int. Conf. on Infrastructure Syst. and Services: Building Networks for
systems. The local control law and Lyapunov function are a Brighter Future, 2008.
computed by solving (13). The performance of the proposed [2] M. A. Müller, M. Reble, and F. Allgöwer, “Cooperative control
of dynamically decoupled systems via distributed model predictive
distributed MPC approach is evaluated by tracking a unit step control,” Int. J. Robust. Nonlinear Control, vol. 22, no. 12, pp. 1376–
reference at subsystem 11 from time instant 10 to 30, i.e., 1397, 2012.
w(10 : 30, 11) = 1. The steady-state set with respect to the [3] P. D. Christofides, R. Scattolini, Muñoz de la Peña, David, and J. Liu,
“Distributed Model Predictive Control: A Tutorial Review and Future
given reference is computed by ᾱs = x̄T t (k, s)Pt x̄t (k, s). Research Directions,” Computers & Chemical Engineering, vol. 51,
That ᾱs < αs with αs computed from (24) implies the pp. 21–41, 2013.
feasibility of tracking the given reference. [4] R. Scattolini, “Architectures for Distributed and Hierarchical Model
Predictive Control – A Review,” Journal of Process Control, vol. 19,
The prediction horizon is chosen as N = 5. Fig. 4 shows no. 5, pp. 723–731, 2009.
the comparisons of closed-loop response and control input [5] Q. Liu and H. Werner, “Distributed Anti-Windup Compensation for
at subsystem 11. Both the blue and red curves are generated Spatially-Interconnected Systems,” in Proc. American Control Conf.,
2015, pp. 1947–1952.
by implementing the control law derived in Section III- [6] S. Wollnack and H. Werner, “Distributed Fixed-Structure Control of
A without and with input constraint imposed, respectively. Spatially Interconnected Systems,” in Proc. American Control Conf.,
After imposing the input constraints (realized by connecting 2015, pp. 1953–1958.
[7] H. S. Abbas, R. Tóth, N. Meskin, J. Mohammadpour, and J. Hanema,
a saturation block between the controller and the plant), the “An MPC Approach for LPV Systems in Input-Output Form,” in Proc.
rise time (red) increases due to the slightly saturated control 54th IEEE Conf. Decision Control, 2015, pp. 91–96.
input. The distributed MPC controller (black), taking the [8] D. Henrion, “LMI Optimization for Fixed-Order H∞ Controller
Design,” in Proc. 42nd IEEE Conf. Decision Contr., 2003, pp. 0191–
constraints into account, guarantees stability of the controlled 2216.
system and achieves an effective reference tracking, without [9] S. Wollnack and H. Werner, “On the application of Finsler’s Lemma
the imposed constraints violated. to LPV controller synthesis,” in Proc. Amer. Control Conf., 2014.
[10] G. Belforte, F. Dabbene, and P. Gay, “LPV approximation of dis-
tributed parameter systems in environmental modelling,” Environmen-
V. C ONCLUSION tal Modelling & Software, vol. 20, no. 8, pp. 1063–1070, 2005.
[11] O. A. Akan, Open Channel Hydraulics. Amsterdam: Elsevier, 2006.
In this paper, we considered a distributed MPC design for [12] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory
and Design. Madison and Wis: Nob Hill Pub., 2009.
spatially-interconnected systems whose dynamics are gov- [13] K. Derinkuyu and M. Ç. Pınar, “On the S-procedure and Some
erned by a two-dimensional difference equation at subsys- Variants,” Math. Meth. Oper. Res., vol. 64, no. 1, pp. 55–77, 2006.
tems and can be represented in an implicit input-output form. [14] F. Wu, “Distributed Control for Interconnected Linear Parameter-
Dependent Systems,” IEE Proc., Control Theory Appl., vol. 150, no. 5,
We showed the computation of the local control law, terminal pp. 518–527, 2003.
cost and terminal set to ensure Lyapunov-based stability and [15] C. Corte, “Stability and Computations in Cooperative Distributed
recursive feasibility. The proposed MPC design conforms Model Predictive Control,” Ph.D. dissertation, Eidgenössische Tech-
nische Hochschule ETH Zürich, 2014.
to the localized property of subsystems, that information [16] D. Simon, J. Löfberg, and T. Glad, “Reference Tracking MPC using
exchange takes place only between neighbouring subsystems. Terminal Set Scaling,” in Proc. 51st IEEE Conf. Decision Control,
Closed-loop performance was demonstrated using a numer- 2012, pp. 4543–4548.
ical example.

3605