Professional Documents
Culture Documents
Abstract
We propose a novel adaptive learning-based model predictive control (MPC) scheme for interconnected systems which can
be decomposed into several smaller dynamically coupled subsystems with uncertain coupling. The proposed scheme is mainly
divided into two main online phases; a learning phase and an adaptation phase. Set membership identification is used in the
learning phase to learn an uncertainty set that contains the coupling strength using online data. In the adaptation phase,
rigid tube-based robust MPC is used to compute the optimal predicted states and inputs. Besides computing the optimal
trajectories, the MPC ingredients are adapted in the adaptation phase taking the learnt uncertainty set into account. These
MPC ingredients include the prestabilizing controller, the rigid tube, the tightened constraints and the terminal ingredients.
The recursive feasibility of the proposed scheme as well as the stability of the corresponding closed-loop system are discussed.
The developed scheme is compared in simulations to existing schemes including robust, adaptive and learning-based MPC.
2
terminal set satisfying the state and input constraints. OCP (4) can in principle be replaced by a distributed
The scalar ξ is used to ensure that the artificial equi- OCP. While this may alleviate the conservatism of the
librium lies in the interior of the constraint sets. This developed scheme, the computation and communication
requirement is needed later in the stability proof. Note complexity will increase significantly since distributed
that the global prestabilzing control gain is denoted by optimization is required.
K = diag(K1 , ..., KM ). Similarly, the global system ma-
trices A and B can be obtained in the obvious way. Fol- In the learning phase, we aim to reduce the volume of the
lowing Riverso et al. (2013), the control input vi (t) = uncertainty set Wi in (2) by learning the bounds amin ij
ui (0|t) + Ki (zi (t) − xi (0|t)) is applied to (1) to ensure and amax
ij of the uncertain parameter aij as more data
robust constraint satisfaction. is collected online. Hence, we develop a distributed set
membership identification approach which learns these
Interconnected systems with uncertain coupling (1) can bounds by sharing information among the neighbours
be exploited in various scenarios with varying-topology only once at each time instant. While the set Wi is a
networks and large-scale systems. For example, varying- function of amax
ij only, amin
ij is also updated since it can
topology networks comprise subsystems which can join speed up the shrinkage of Wi by finding smaller upper
and leave the network. While accurate models can be bounds. In the adaptation phase, we aim to adapt all
identified offline for the new subsystems, the coupling MPC ingredients depending on the learnt uncertainty set
strength between the network and these subsystems and compute the local optimal control inputs taking the
can only be identified online after the plug-in operation. updated uncertainty set into consideration. The adapted
Practical applications comprising such systems include MPC ingredients include the rigid tube represented by
microgrids, power systems and vehicle platoons. Note ZKi , the prestabilizing controller Ki , the tightened state
that considering the whole coupling term as a distur- constraint set Xi and the tightened input constraint set
bance might be conservative. This, however, leads to de- Ui . Note that these ingredients are updated at each time
centralized OCPs rather than distributed ones. Decen- instant t based on the learnt uncertainty set Wi which is
tralized OCPs are usually more robust against failures a function of the updated parameters amax ij and amin
ij at
and requires less communication and computation. This this time instant. Hence, these MPC ingredients together
is more appealing for safety-critical systems with fast with the uncertainty set are now functions of time.
dynamics. Moreover, the inclusion of parametric uncer-
tainties within an additive term might be also conser- In the sequel, we use the ellipsoidal rigid tube
vative. An alternative consideration involves employing ZKi (t) = {ei : e⊤ 2
i Zi ei ≤ αi (t)} and the linear presta-
adaptive MPC tailored for parametric uncertainties, as bilizing controller κi (xi ) = Ki (t)xi . We assume that
applied in monolithic systems (Lorenzen et al. (2019)). Ki (t) = Koi Ti (t) where Koi is a full-rank stabilizing
However, the complexity arises in interconnected sys- control gain and Ti (t) ∈ Sn++ i
. Recall also that the
tems when uncertainties are embedded within coupling tightened polytopic state and control constraint sets
terms, necessitating the formulation of a distributed are given, respectively, by Xi (t) = {xi : Gi xi ≤ gi (t)}
OCP, which imposes substantial communication and and Ui (t) = {ui : Hi xi ≤ hi (t)}. In addition to the
computational burdens. Deceralized OCPs have been decision variables in (4), αi (t), Ti (t), gi (t) and hi (t) are
widely used with interconnected systems (Riverso et al. also considered as decision variables in the developed
(2013), Betti et al. (2013)). Finally, note that coupling scheme in order to update the MPC ingredients on-
constraints among neighbours could in principle be line; Koi , on the other hand, is assumed to be known
incorporated in decentralized OCPs by finding a de- in advance, computed, for example, with methods as
centralized inner-approximation of the constraint sets. those in Aboudonia et al. (2021). When updating these
This, however, adds to the conservatism of decentralized ingredients at each time instant, one should ensure the
OCPs. following conditions,
I. The set ZKi (t) is an RPI set for the dynamics (3).
In this paper, we make use of the OCP (4) to develop
II. The local matrices Ai + Bi Ki (t) are Schur.
the proposed scheme which, at each time instant, per-
III. The global matrix AL + BK(t) is Schur.
forms two phases: a learning phase to learn the model
IV. The condition Xi (t) L ZKi (t) ⊆ Zi holds.
uncertainty and an adaptation phase to compute the
V. The condition Ui (t) Ki (t)Zki (t) ⊆ Vi holds.
optimal control action. We assume that every subsys-
tem can communicate with its neighbours at each time The aforementioned conditions are ensured by adding
instant. While this assumption is not required for the extra constraints in the OCP of the adaptation phase
adaptation phase which relies on the decentralized OCP while preserving its convexity. These additional con-
(4) that requires no communication, it is still required straints are derived in the next section. Although the
for the learning phase. This is because the model uncer- constraint sets are polytopic, the rigid-tube is chosen to
tainty in (2) relies on the states of the neighbours. Hence, be ellipsoidal. As shown in the next section, this leads
communication is required among neighbours only once in the adaptation phase to an OCP which considers
at each time instant in the learning phase. Since com- the prestabilizing controller as a decision variable while
munication is allowed by assumption, the decentralized maintaining the convexity of the OCP by making use of
3
the S-lemma (Boyd et al. (1994)). Recall also the scal- of the uncertain parameters. Hence, a better estima-
ability properties of ellipsoidal sets in previous related tion of these bounds leads to a better description of this
works when controlling networked dynamical systems set. Thus, we develop here a distributed set-membership
(Conte et al. (2016), Parsi et al. (2022)). identification technique to improve the estimates of these
bounds. Although such techniques are widely used with
The system (1) can be described by a graph G(M, E, A) monolithic systems (e.g. see Lorenzen et al. (2019)), we
where M = {1, ..., M } represents the set of nodes, E ⊆ extend this technique here to interconnected systems
(M × M) the set of edges and A = {aij ∈ R, (i, j) ∈ E} with uncertain couplings where the identification is not
the set of weights. Each subsystem is represented by a carried out by a central entity.
node. An edge exists from the j-th node to the i-th node
if the j-th subsystem belongs to the set of in-neighbours Assume that the bounds of the uncertain parameter aij
of the i-th subsystem. We define the adjacency matrix at the (t−1)-th time instant are amin max
ij (t−1) and aij (t−
Ag ∈ RM×M of the graph G such that the element in its 1). Our goal is to improve the estimates of these bounds
i-th row and j-th column is given by Agij = aij if i ∈ Nj+ at the t-th time instant. For this purpose, we construct
and zero otherwise. The degree matrix Dg ∈ DM ++ of
for every aij the feasible parameter set defined by the
the graph G is defined such that the diagonal element in inequality Fij aij ≤ fij (t − 1) where Fij = [1 − 1]⊤
⊤
P
the i-th row is given by j∈N − aij and the Laplacian
i
and fij (t − 1) = [amax min
ij (t − 1) aij (t − 1)] . We then
matrix is given by Lg = Dg − Ag ∈ SM define the vector θi which comprises all the uncertain
+ . These graph
theory concepts are required to locally ensure Condition parameters to be identified by the i-th subsystem (i.e.
III that the global matrix A+BK(t) is Schur. The global aij , j ∈ Ni− ). In this case, the feasible parameter set for
dynamics is then given by, the i-th subsystem is given by Fi θi ≤ fi (t − 1) where
Fi is a block diagonal matrix with the matrices Fij for
z(t + 1) = Az(t) + Bv(t) + d(t), (5) all j ∈ Ni− on the diagonal and fi is a vector vertically
concatenating the subvectors fij for all j ∈ Ni− .
where z(t) = [z1⊤ (t), . . . , zM ⊤
(t)]⊤ , v(t) = [v1⊤ (t), . . . ,
vM⊤
(t)]⊤ , d(t) = [d⊤ (t), . . . , d⊤ ⊤
M (t)] , Ad = diag(A1 , . . . ,
Recall that each subsystem can communicate with its
N1 neighbours in the learning phase. Hence, the states of
AM ), Ãg = Ag Im0 , A = Ad + E Ãg C, B = diag(B1 ,
. . . , BM ), C = diag(C1 , . . . , CM ) and E = diag(E1 , . . . , the neighbours of the i-th subsystem at the (t − 1)-th
EM ). Note that the proposed method is mainly devel- time instant are available for the i-th subsystem at the
oped for sparse graphs where only neighbours can com- t-th time instant. Furthermore, the states and inputs
municate. While this method can be used with fully of the i-th subsystem at the (t − 1)-th time instant are
connected graphs where all subsystems can share infor- available. Finally, the states of the i-th subsystem at
mation with each other, other efficient methods (e.g. the t-th time instant can be measured. Recall also that
Lorenzen et al. (2019)) can be used where all subsystems Di = {di : Ξi di ≤ 1}. Defining χi (t − 1) to be a matrix
share information with a central entity that does all the horizontally concatenating the vectors Cj zj (t − 1) for all
computations instead. In the sequel, we make use of the j ∈ Ni− , Ωi (t − 1) = −Ξi Ei χi (t − 1) and ωi (t − 1) =
following definitions. 1+Ξi Ai zi (t−1)+Ξi Bi vi (t−1)−Ξi zi (t), we can construct
for the i-th subsystem the non-falsified parameter set
Definition 1 (Kouvaritakis & Cannon (2016)) ∆i = {θi : Ωi (t − 1)θi ≤ ωi (t − 1)}.
A set ZK ⊂ Rn is robustly positively invariant under
z + = Az + Bv, v = Kz and z ∈ Z, v ∈ V, if and only if In order to update the bounds of the uncertain param-
z + ∈ Zk , z ∈ Z, v ∈ V for all z ∈ Zk . eters, the i-th subsystem should solve 2|Ni− | optimiza-
Definition 2 (Aliyu (2017)) The system z + = Az + tion problems at the t-th time instant where | · | denotes
Bv, y = Cz+Dv is strictly passive if there exist functions set cardinality. Each optimization problem updates one
Vp : Z → R and γp : Z → R+ so that Vp (0) = 0 and of these bounds. For each j ∈ Ni− , the lower and upper
Vp (z + ) − Vp (z) ≤ y ⊤ v − γp (z) for all v ∈ V. bounds of the parameter aij are given respectively by,
" # " #
min ⊤ Fi fi (t − 1)
3 Adaptive Learning-based MPC aij (t) = min eij θi s.t. θi ≤ , (6a)
aij Ωi (t − 1) ωi (t − 1)
" # " #
In this section, we start by discussing the distributed set F i f i (t − 1)
membership identification approach used to learn the amax
ij (t)=max e⊤
ij θi s.t. θi ≤ , (6b)
aij Ωi (t − 1) ωi (t − 1)
uncertainty set online in the learning phase. We then
derive the constraints added to the local online OCP of where eij is a unit vector which selects the parameter
each subsystem in the adaptation phase. First, we dis- aij in the vector θi . Note that computing the new lower
cuss the learning phase where we aim to learn the uncer- (upper) bound is a minimization (maximization) prob-
tainty set Wi (t) in real time, as more data is collected on- lem because we aim to find the smallest (largest) value of
line. According to (2), this set is a function of the bounds the corresponding parameter that lies inside the feasible
4
parameter set and the non-falsified set. In conclusion, We start with the condition which ensures the robust
the i-th subsystem learns the uncertainty set Wi (t) in positive invariance of the set ZKi (t). First, we define for
(2) by solving at the t-th time instant the optimization the i-th subsystem the output vector yi (t) = Ci zi (t) and
problems (6) which identify the bounds of the uncer- the constraint set Yi , obtained from Zi in the obvious
tain parameter. These bounds are communicated among way, such that yi ∈ Yi . We then compute a minimum
the in-neighbours and out-neighbours to be used in the volume ellipsoidal set Ŷi = {yi : kyi − cyi k2Yi ≤ 1} that
adaptation phase. We also highlight the dependency of contains the set Yi . For this purpose, the vertex descrip-
the upper bound amax min
ij (t) on the lower bound aij (t− 1) tion of Yi is needed. Note that this is in general com-
which is part of the vector fi (t−1) in (6b) to support the binatorial in the linear inequalities needed to define Yi .
fact that updating the lower bound might improve the However, the sets Yi of all subsystems do not change
updates of the upper bound. Note that using asymmet- over time and hence, their minimum volume ellipsoidal
ric duality of linear programs, (6) can be written com- sets Ŷi can be computed once offline by solving a con-
pactly as a single optimization problem. We guarantee vex optimization problem for each.L Finally, we construct
that the true parameters always lie within the updated the outer approximation Ŵi (t) = j∈N − amax ˆ
ij (t)Yj of
bounds in the following theorem. L i
max
the uncertainty set Wi (t) = j∈N − aij (t)Yj . Note
i
tainty sets. We now move to discussing the adaptation αi (t)si , the invariance condition becomes k(Ai αi (t) +
phase. As mentioned in Section 2, the online OCP is Bi Koi Ti (t)αi (t))si + Ei wi + di k2Zi ≤ α2i (t) for all si sat-
solved in the adaptation phase to adapt the MPC ingre- isfying ksi k2Zi ≤ 1, wij satisfying kwij −amax 2
ij (t)cyj kYj ≤
dients based on the learnt uncertainty set and compute amax
2
(t) for all j ∈ Ni− and di satisfying kdi k2Ξ̂ ≤ 1. We
the optimal control inputs. To adapt the MPC ingre- ij i
dients online while ensuring robust constraint satisfac- now define the block diagonal matrix Yλi (t) with the sub-
tion, the five conditions mentioned in Section 2 should matrices λij (t)Yj for all j ∈ Ni− on the diagonal and the
be satisfied. We derive here additional constraints which vector τi (t) which vertically concatenates the subvectors
ensure the satisfaction of these conditions while main- amax
ij (t)cyj . Using the S-lemma and applying the Schur
taining the convexity of the OCP (4). complement yields (7) where Ini is an identity matrix of
5
λi Zi 0 0 0 Ai αi + Bi Koi Ti α⊤
i
∗
Yλi 0 −Yλi τi Ei⊤
∗ ∗ λdi Ξ̂i 0 Ini ≥ 0. (7)
2
kcyj k2Yj )amax
P
∗ ∗ ∗ αi − λi − λdi − j∈Ni− λij (1 − ij 0
∗ ∗ ∗ ∗ αi Zi−1
size ni and Ei is a matrix horizontally concatenating the αi (t) > 0 such that,
matrix Ei for |Ni− | times.
(Pi − Γi )αi 0.5Ci⊤ αi (Ai αi + Bi Koi Ti αi )⊤
∗ Di αi Ei⊤ αi ≥ 0. (8)
∗ ∗ Pi−1 αi
Note that the dimension of the constraint in (7) increases
linearly with respect to the number of neighbours. Rigid
tubes are known for their little computation burden in PROOF. To ensure strict passivity, we have to satisfy
comparison to homothetic tubes (Raković et al. (2012)). the condition in Definition 2 for all wi ∈ Ŵi (t). Recall
Since we use time-varying rigid tubes, one might think also that the storage function Vpi (zi ) is defined over the
that rigid tubes lose their main computational advantage set Ẑi . Hence, the noise-free dynamics of the i-th subsys-
and could be better replaced by homothetic tubes. This tem under the controller ui (t) = Koi Ti (t)xi (t) is strictly
is, however, not the case because when using homothetic passive according to Definition 2 if kzi+k2Pi − kzi k2Pi ≤
tubes, the nominal dynamics would be replaced by tube
ỹi⊤ wi − kzi k2Γi for all zi satisfying kzi − czi k2Ẑ ≤ 1 and
dynamics. Since the prestabilizing controller is a decision 2
i
2
variable and the tube is ellipsoidal, the tube dynamics is wij satisfying kwij −amax
ij (t)cyj kYj ≤ aij
max
(t) for all j ∈
represented by one matrix inequality for each prediction −
Ni . By substituting for the dynamics, the strict passiv-
step. In this case, the problem becomes computationally
ity condition is given by kzi k2L1 + 2wi⊤ L2i zi + kwi k2L3 ≥
challenging as the prediction horizon gets longer. i i
0 for all zi satisfying kzi − czi k2Ẑ ≤ 1 and wij satisfying
i
2
kwij − amax 2 max
ij (t)cyj kYj ≤ aij (t) for all j ∈ Ni− where
Next, we consider Conditions II and III in Section 2 by
deriving constraints ensuring that Ai + Bi Ki (t) for all L1i = Pi − Γi − (Ai + Bi Koi Ti (t))⊤ Pi (Ai + Bi Koi Ti (t)),
i ∈ {1, . . . , M } and A + BK(t) are Schur. For this pur- L2i = 0.5Ci − Ei⊤ Pi (Ai + Bi Koi Ti (t)) and L3i = Di −
pose, we rely on passivity arguments for discrete-time Ei⊤ Pi Ei . We now define the block diagonal matrix Yσi (t)
interconnected systems. First, we derive a constraint en- with the submatrices σij (t)Yj for all j ∈ Ni− on the di-
suring that the noise-free dynamics of the i-th subsys- agonal and the vector τi (t) which vertically concatenates
−
tem (i.e. (1) with di (t) = 0 for all t) under the controller the subvectors amax
ij (t)cyj for all j ∈ Ni . We also recall
ui (t) = Koi Ti (t)xi (t) is strictly passive. We show that that the ellipsoidal set Ŷi contains the origin in its inte-
this implies that Ai + Bi Ki (t) is Schur. We then derive rior and hence, kcyi k2Yi ≤ 1. Using the S-lemma, apply-
other constraints ensuring that the strict passivity of all ing the Schur complement and multiplying the resulting
subsystems implies the asymptotic stability of the whole matrix inequality by αi (t) thus leads to (8).
system ensuring that A+BK(t) is Schur. To ensure strict
passivity of the i-th subsystem, we define the virtual out-
put ỹi (t) = Ci zi (t) + Di wi (t) where Di ∈ Dm++ is a de-
0
Note that (8) implies that Pi −(Ai +Bi Koi Ti (t))⊤ Pi (Ai +
cision variable. We also assume that a matrix Pi ∈ Sn++ i
Bi Koi Ti (t)) ≥ Γi (t) where Γi (t) ∈ Dn++ i
and hence,
is computed offline together with Koi , for example as in Ai + Bi Ki (t) is Schur. One source of conservatism
Aboudonia et al. (2021), so that Vpi (zi ) = kzi k2Pi is a in Proposition 5 is that the local controller is de-
storage function over the set Zi . We finally define a min- signed to not only stabilize the local dynamics, but
imum volume ellipsoidal set Ẑi = {zi : kzi − czi kẐi ≤ 1} also to passivize it. The benefit of passivizing the
for the state constraint set Zi . The dependence on time local dynamics will become obvious in the next
in (8) is omitted in the interest in space. proposition. We now consider Condition III in Sec-
tion 2 by deriving constraints which ensure that the
strict passivity of all noise-free subsystems implies
Proposition 5 The noise-free dynamics of the i-th sub- the asymptotic stability of the overall noise-free sys-
system zi+ = (Ai + Bi Koi Ti (t))zi + Ei wi is strictly pas- tem. For this purpose, we consider the noise-free
L
sive with respect to wi ∈ Ŵi (t) and ỹi ∈ Ci Ẑi Di Ŵi (t) global dynamics z(t + 1) = Acl (t)z(t) + Ew(t) where
if there exist Γi (t) ∈ Dn++
i
, Di (t) ∈ Dn++i
, Ti (t) ∈ Sn++
i
, Acl (t) = diag(A1 +B1 K1 (t), . . . , AM +BM KM (t)), E =
6
diag(E1 , . . . , EM ) and w(t) = [w1⊤ (t), . . . , wM
⊤
(t)]⊤ . We for all j ∈ {1, ..., ni } and [Di (t)−1 ]j ≥ |Wi |j for all
also define the Lyapunov function V (z) = kzk2P where j ∈ {1, ..., m0 }. Multiplying the former by αi (t) and the
P = diag(P1 , ..., PM ). Furthermore, we define the matri- latter by αi (t)−1 and inverting, these inequalities be-
ces Γ(t) = diag(Γ1 (t), ..., ΓM (t)), C = diag(C1 , ..., CM ) come ([Γi (t)]j − ǫ)αi (t) ≥ |Ui |j αi (t) + |Vi |j αi (t) for all
and D(t) = diag(D1 (t), ..., DM (t)). Recall that Ag j ∈ {1, ..., ni } and [Di (t)]j αi (t) ≤ αi (t)/|Wi |j for all
and Dg are the adjacency and degree matrices, respec- j ∈ {1, ..., m0 } such that |Wi |j 6= 0. Note that the re-
tively. The global virtual output vector is given by sulting inequalities are functions of the uncertain pa-
ỹ(t) = Cz(t) + Dw(t) and the global coupling vector is rameters. More conservative inequalities can be obtained
given by w(t) = Ãg Cz(t). by replacing the uncertain parameters with their upper
bounds, leading to (9).
Finally, we construct Amax g (t) and Dgmax (t) by re-
placing all uncertain parameters in Ag and Dg
with their upper bounds. Consequently, we define In Proposition 6, (9) ensures that the Lyapunov condi-
U = diag(U1 , ..., UM ) = C ⊤ Dg C, V = [V1⊤ , ..., VM
⊤ ⊤
] = tion is satisfied for the noise-free global dynamics z(t +
⊤ ⊤ ⊤
C Ãg and W = [W1 , ..., WM ] ⊤ ⊤
= Ãg C where 1) = Acl (t)z(t) + Ew(t) or equivalently, z(t + 1) = (A +
Ui ∈ Rni ×ni , Vi ∈ Rni ×m0 and Wi ∈ Rm0 ×ni . We also BK(t))z(t) and hence, A + BK(t) is Schur. Sources of
construct U max (t), V max (t) and W max (t) by replacing conservatism in this proposition include implying ỹ ⊤ w−
all uncertain parameters in U , V and W with their kzkΓ(t) < 0 instead of V (z + ) − V (z) < 0, ignoring
upper bounds. In the sequel, we denote the diagonal the term C ⊤ Lg C, using diagonal dominance and replac-
element in the j-th row of a matrix by [·]j and the sum- ing the true values of the parameters with their upper
mation of the absolute values of the entries in the j-th bounds. These sources are utilized to eventually reach
row of a matrix by | · |j . The dependence on time in (9) local convex constraints. We now move to Conditions IV
is omitted in the interest in space. and V in Section 2. In the sequel, we denote the j-th row
(j) (j)
of the matrices Gi and Hi by Gi and Hi respectively.
Proposition 6 The strict passivity of the noise-free dy- The dependence on time is omitted in (13) and (14) in
namics of all subsystems under the corresponding con- the interest in space.
troller ui (t) = Koi Ti (t)xi (t) (i.e. zi (t + 1) = (Ai +
Bi Koi Ti (t))zi (t) + Ei wi (t)) implies the asymptotic sta- Proposition 7 The Minkowski sum of Xi (t) and ZKi (t)
bility of the noise-free dynamics of the overall system is contained in Zi if there exist µij (t) ≥ 0 for all j ∈
(i.e. z(t + 1) = Acl (t)z(t) + Ew(t)) if, {1, ..., qi } such that,
7
gi (t) and the off-diagonal elements to be equal to zero conclusion, the OCP of the i-th subsystem becomes,
and hence, (12) reduces to (11).
∀k ∈ {0, ..., N − 1},
zi (t) − xi (0|t) ∈ ZKi (t),
Proposition 8 The Minkowski sum of Ui (t) and
xi (k + 1|t) = Ai xi (k|t) + Bi ui (k|t),
Koi Ti (t)Zki (t) is contained in Vi if there exist νij (t) > 0
for all j ∈ {1, ..., ri } and Mi (t) ∈ Sp++ xi (k|t) ∈ Xi (t), ui (k|t) ∈ Ui (t),
i
such that,
xi (N |t) = xei (t) = Ai xei (t) + Bi uei (t),
min Ji s.t. xei ∈ ξXi (t), uei ∈ ξUi (t),
(j)⊤
Mi−1 Ki⊤ αi Mi−1 0.5Mi−1 Hi
≥ 0, ≥ 0. (13) (7), λi (t), λdi (t), λij (t) ≥ 0, ∀j ∈ Ni− ,
∗ νij Zi ∗ 1 − νij − hi
(8), (9), (11), µij (t) ≥ 0 ∀j ∈ {1, ..., qi },
(13), νij (t) > 0 ∀j ∈ {1, ..., ri },
M (t) ∈ Sn++ , Ti (t)αi (t) ∈ Sn++ ,
i i
i
ni
Γi (t)αi (t) ∈ D++ , Di (t)αi (t) ∈ Dn++
PROOF. First, we define Mi (t) ∈ Sp++ .
i
i
and νij (t) > 0
−1
such that Mi (t) ≤ νij (Koi Ti (t)αi (t)Zi αi (t)Ti⊤ (t)Ko⊤i )−1 (15)
for all j ∈ {1, ..., ri }. Since Ti (t) ∈ Sn++
i
, αi (t) > 0 and The decision variables are xi (k|t) and ui (k|t) for all k ∈
Koi is full-rank, this inequality is invertible and hence, {0, ..., N }, xei (t), uei (t), gi (t), hi (t), αi (t), Ti (t)αi (t),
Mi (t)−1 −νij (t)−1 Koi Ti (t)αi (t)Zi−1 αi (t)Ti⊤ (t)Ko⊤i αi (t) ≥ Γi (t)αi (t), Di (t)αi (t), Mi (t)−1 , λi (t), λdi (t), λij (t) for
0 for all j ∈ {1, ..., ri }. Applying the Schur Complement all j ∈ Ni− , µ(t)ij for all j ∈ {1, ..., qi } and ν(t)ij for all
yields the first condition in (13). The Minkowski sum j ∈ {1, ..., ri }. Although the OCP might seem noncon-
of Ui (t) and Koi Ti (t)ZKi (t) is contained in Vi if and vex using these decision variables, one can turn it into
(j) an SDP by redefining the decision variables and, if de-
only if for all j ∈ {1, ..., ri }, Hi (uei + usi ) ≤ 1 for all
sired, reversing the transformation to recover the origi-
2
uei satisfying kuei k(K T (t)Z −1 T ⊤ (t)K ⊤ )−1 ≤ α2i (t) and
oi i i i oi nal variables from the solution. As the number of neigh-
usi satisfying Hi usi ≤ hi (t). By defining si and ti such bours of each subsystem increases, the dimensions of all
that uei = si and usi = hi (t)ti , we reach the condition constraints remain unchanged except for (7) which in-
(j) (j)
Hi si + Hi hi (t)ti ≤ 1 for all j ∈ {1, ..., ri }, si and ti creases linearly and hence, the computation time is not
satisfying s⊤ −1 ⊤ ⊤ −1
si ≤ 1 significantly altered. Similar to the OCP (4), the online
i (Koi Ti (t)αi (t)Zi αi (t)Ti (t)Koi ) OCP (15) is solved at each time instant and the optimal
and Hi ti ≤ 1. Using the S-lemma, we reach for all
j ∈ {1, ..., ri }, control input vi (t) = ui (0|t) + Koi Ti (t)(zi (t) − xi (0|t))
is applied to the i-th subsystem.
(j)⊤
νij ψi−1 0 −0.5Hi
P ri (l)⊤ (j)⊤
4 Feasibility and Stability
∗ 0 0.5νijl Hi − 0.5Hi ≥ 0,
hi (t)
l=1
P ri
∗ ∗ 1 − νij − l=1 νijl In this section, we prove that the online OCP (15) is
(14) recursively feasible and the resulting closed-loop sys-
tem is input-to-state stable. In the sequel, we use the
where ψi (t) = Koi Ti (t)αi (t)Zi−1 Ti⊤ Ko⊤i (t)αi (t). Note
superscript (·)∗ to refer to the optimal solution. The
Pri (l) (j)⊤
that (14) holds if l=1 0.5νijl (t)Hi = 0.5Hi hi (t). proof is inspired from Kouvaritakis & Cannon (2016),
By stacking this equality for all j ∈ {1, ..., ri }, we reach Limón et al. (2008) and is divided into three main steps.
hi (t)Hi = νi (t)Hi where νijl (t) is the entry of νi (t) in First, we discuss the convergence of the nominal state
⊤ ⊤
the j-th row and l-th column. To satisfy this equality, x∗ (t) = [x∗1 (t), . . . , x∗M (t)]⊤ to the artificial equilib-
we choose the diagonal elements of νi (t) to equal hi (t) ⊤ ⊤
rium x∗e (t) = [x∗e1 (t), . . . , x∗eM (t)]⊤ . Then, we discuss
and the off-diagonal elements to equal zero. Recalling
the convergence of the artificial equilibrium x∗e (t) to the
that Mi (t) ≤ νij (t)(Ki (t)αi (t)Zi−1 Ki⊤ (t)αi (t))−1 , (14)
target point xr = [x⊤ ⊤ ⊤
r1 , . . . , xrM ] . Finally, we discuss
reduces to a matrix, that if pre and post-multiplied by
the stability of the error dynamics between the actual
diag(Mi (t)−1 , 1), yields the second condition in (13). ⊤
∗⊤
state z ∗ (t) = [z1∗ (t), . . . , zM (t)]⊤ and the nominal state
∗
x (t). We start with the first step in Lemma 9.
8
target point xr . For this purpose, we define (x̃r ,ũr ) to the cost corresponding to the optimal solution and the
be the projection of (xr ,ur ) onto the set of equilibrium feasible solution aiming to converge to x̄e and note that
points contained in the set (ξXi (t),ξUi (t)). We also de- J ∗ < J.¯ Note also that J¯ ≤ kx∗ − x̄e k2 + kx̄e − x̃r k2 due
P̃ S
fine the matrix P̃ such that P̃ − (A + BK(t))⊤ P̃ (A + to the Lyapunov inequality that P̃ satisfies. Since x∗ ∈
BK(t)) − Q − K ⊤ (t)RK(t) ≥ 0. Note that such P̃ E(x∗e , β6 ), then kx∗ − x̄e k2P̃ + kx̄e − x̃r k2S ≤ kx∗e − x̃r k2S .
exists since K(t) asymptotically stabilizes the actual But, kx∗e − x̃r k2S ≤ kx∗e − xr k2S ≤ J ∗ which contradicts
global dynamics as shown in Proposition 6. Finally, we the optimality of J ∗ .
define E(c, s) = {x : kx − ck2P̃ ≤ s}. For brevity, we
omit the dependence on t in the proof of Lemma 10.
This proof is inspired from Limón et al. (2008), how- Finally, we discuss the input-to-state stability of the
ever, we prove that limt→∞ (x∗ (t) − x∗e (t)) = 0 implies global closed-loop system under the MPC scheme (15).
limt→∞ (x∗e (t) − xr ) = 0 instead of x∗ (t) − x∗e (t) = 0 Recall from (5) that the global system is given by z(t +
implying x∗e (t) − xr = 0.
1) = (Ad +E Ãg C)z(t)+Bv(t)+d(t). Recall also that the
optimal input is given by v ∗ (t) = u∗ (t) + K ∗ (t)(z ∗ (t) −
Lemma 10 If x∗ (t) and x∗e (t) are such that limt→∞ (x∗ (t)− x∗ (t)) where K ∗ (t) = diag(Ko1 T1∗ (t), . . . , KoM TM ∗
(t))
x∗e (t)) = 0, then limt→∞ (x∗e (t) − x̃r ) = 0. ∗ ∗ ∗⊤ ∗⊤ ⊤
and u (t) = u (0|t) = [u1 (0|t), . . . , uM (0|t)] . Hence,
the closed-loop system is given by z ∗ (t + 1) = Ad z ∗ (t) +
PROOF. According to the OCP (15), (x∗e , u∗e ) ∈ Bu∗ (t) + BK ∗ (t)(z ∗ (t) − x∗ (t)) + E Ãg Cz ∗ (t) + d(t). By
int(X × U). Hence, there always exists E(x∗e , β1 ) such defining e∗ (t) = z ∗ (t) − x∗ (t) and η(t) = d(t) − x∗ (t +
that (x, K(x − x∗e ) + u∗e ) ∈ γX × γU for all x ∈ E(x∗e , β1 ) 1) + Ax∗ (t) + Bu∗ (t), the closed-loop system becomes,
where γ ∈ (0, 1). Note that E(x∗e , β1 ) is invariant since
it is defined using the level sets of a Lyapunov func- e∗ (t + 1) = (A + BK ∗ (t))e∗ (t) + η(t). (16)
tion. We now define an equilibrium point (x̄e , ūe ) ∈
int(X × U) such that kx̄e − x̃r k2S = β2 kx∗e − x̃r k2S and While e∗ (t) is the state of this system, x∗ (t) and u∗ (t),
kx̄e − x∗e kP̃
2
= (1 − λ)kx̃r − x∗e k2P̃ where β2 ∈ (0, 1) together with d(t), are exogenous signals to this system.
and λ ∈ (0, 1) is chosen sufficiently large such that While d(t) is bounded by assumption, x∗ (t) and u∗ (t)
(0, K(x̄e − x∗e ) + ūe − u∗e ) ∈ (1 − γ)X × (1 − γ)U. are bounded since they are the outputs of the proposed
Next, we define E(x̄e , β3 ) ⊂ E(x∗e , β1 ) such that MPC problem (15) which consider compact constraint
x∗e ∈ int(E(x̄e , β3 )). This set is invariant since it is sets. Hence, we prove input-to-state stability with re-
defined using the level sets of a Lyapunov function. spect to these bounded exogenous signals. Note that
Moreover, for all x ∈ E(x̄e , β3 ), (x, K(x − x̄e ) + ūe ) = x∗ (t+1) is not necessarily equal to Ax∗ (t)+Bu∗ (t) where
(x, K(x − x∗e ) + u∗e ) + (0, K(x̄e − x∗e ) + ūe − u∗e ) ∈ A = diag(A1 , . . . , AM ) + EAg C. This is because x∗ (·)
(γX × γU) ⊕ ((1 − γ)X × (1 − γ)U) = X × U. and u∗ (·) are the nominal states and inputs which satisfy
x∗ (t + 1) = diag(A1 , . . . , AM )x∗ (t) + Bu∗ (t). Moreover,
the considered system is switched since K ∗ (t) is adapted
We now define β4 > 0 such that S > β4 P̃ and the set at each time instant. Hence, the closed-loop system un-
X̂ = {x : kx − x∗e k2P̃ ≤ β5 kx∗e − x̃r k2P̃ & (x∗e − x̄e )⊤ P̃ |x − der the proposed MPC scheme can be considered as an
x∗e | ≤ β5 kx∗e − x̃r k2P̃ } where 0 < β5 < β4 (1 − β2 )/3. uncertain switched system with bounded disturbances.
Hence, kx − x̄e k2P̃ = kx − x∗e k2P̃ + 2(x∗e − x̄e )⊤ P̃ (x − x∗e ) +
(1 − λ)kx∗e − x̃r k2P̃ ≤ kx∗e − x̃r k23β P̃ +(1−λ)P̃ for all x ∈ Theorem 11 The MPC scheme (15) is recursively fea-
5 sible and the global closed-loop system (16) under this
E(x∗e , β6 ) ⊂ X̂ . Note that E(x∗e , β6 ) can be also chosen to scheme is input-to-state stable with respect to η(t).
lie inside E(x̄e , β3 ) ⊂ E(x∗e , β1 ) and hence is invariant.
Note also that kx − x̄e k2P̃ + kx̄e − x̃r k2S ≤ kx∗e − x̃r k2S for
all x ∈ E(x∗e , β6 ) if (1 − β2 )S − (1 − λ + 3β5 )P̃ ≥ 0 or PROOF. Recursive feasibility is proven using stan-
equivalently, β4 (1 − β2 ) − (1 − λ + 3β5 ) ≥ 0. This can dard MPC arguments (Kouvaritakis & Cannon (2016))
be achieved if λ ≥ 1 − (β4 (1 − β2 ) − 3β5 ). Since β5 < by making use of the tail sequence. Regarding sta-
β4 (1 − β2 )/3, this condition on λ does not contradict the bility, we first consider the disturbance-free sys-
aforementioned conditions on λ. tem e∗ (t + 1) = (A + BK ∗ (t))e∗ (t). The matrix
A + BK ∗ (t) is always Schur for any K ∗ (t) since
We conclude the proof using contradiction. Assume that the constraints derived in Propositions 5 and 6 en-
limt→∞ (x∗ (t) − x∗e (t)) = 0 and limt→∞ (x∗e (t) − x̃r ) 6= 0. sure that P − (A + BK ∗ (t))⊤ P (A + BK ∗ (t)) ≥
Hence, it is always possible to pick an arbitrarily large Γ(t) − C ⊤ Ãg C − C ⊤ LD(t)LC ≥ ǫIn at the t-th time
t such that x∗ ∈ E(x∗e , β6 ) ⊂ (x̄e , β3 ). In addition, there instant for all t. Since this condition is always satisfied
exists β7 such that kx∗e − x̃r k ≥ β7 for infinitely many using the same matrix P , then V (e∗ ) = ke∗ k2P is a
t. Since E(x∗e , β6 ) and E(x̄e , β3 ) are both invariant and Lyapunov function for the disturbance-free system in-
satisfy the state and input constraints, let J ∗ and J¯ be dependent of K ∗ (t). Now, we consider the closed-loop
9
system (16). Using the Lyapunov function V (e∗ ) and
BK ∗ (t))e∗ (t)kP + kη(t)k2P where 0 < δ < ǫ. To 0 100 200 300 400 500
Timestep
600 700 800 900 1000
Parameter a31
δke∗ (t)k22 ≥ 2kη(t)kP k(A + BK ∗ (t))e∗ (t)kP + kη(t)k2P Upper Bound
Lower Bound
10
4000 distributed model predictive control for linear systems’,
Closed-loop Cost
ROB
ADP
LRN
Automatica 69, 117–125.
2000 ADP+LRN
Hewing, L., Wabersich, K. P., Menner, M. & Zeilinger,
M. N. (2020), ‘Learning-based model predictive control:
0
0 100 200 300 400 500 600 700 800 900 1000 Toward safe learning in control’, Annual Review of Con-
Timestep
4000
trol, Robotics, and Autonomous Systems 3, 269–296.
Closed-loop Cost
ROB
LRN
Kouvaritakis, B. & Cannon, M. (2016), ‘Model predictive
2000
ADP
ADP+LRN
control’, Switzerland: Springer International Publishing
38.
0
-3 -2 -1 0
Limón, D., Alvarado, I., Alamo, T. & Camacho, E. F. (2008),
10 10 10 10
Computational Time ‘MPC for tracking piecewise constant references for con-
strained linear systems’, Automatica 44(9), 2382–2387.
Fig. 2. Top: closed-loop cost, Bottom: computation time Lofberg, J. (2004), YALMIP: A toolbox for modeling and
mean (square) and standard deviation (bar). optimization in MATLAB, in ‘2004 IEEE international
conference on robotics and automation (IEEE Cat. No.
ADP+LRN) result in severe closed-loop performance 04CH37508)’, Taipei, Taiwan, pp. 284–289.
deterioration. On the other hand, all other schemes Lorenzen, M., Cannon, M. & Allgöwer, F. (2019), ‘Robust
have lower computation time due to their decentralized MPC with recursive model update’, Automatica 103, 461–
471.
nature which do not require using distributed opti-
Mayne, D. Q., Seron, M. M. & Raković, S. (2005), ‘Ro-
mization techniques such as ADMM. Since the OCP bust model predictive control of constrained linear systems
of ADP+LRN is a semidefinite program, ADP+LRN with bounded disturbances’, Automatica 41(2), 219–224.
requires slightly higher computation compared to LRN Mesbah, A. (2016), ‘Stochastic model predictive control: An
and ROB whose OCPs are quadratic programs. overview and perspectives for future research’, IEEE Con-
trol Systems Magazine 36(6), 30–44.
Müller, M. A. & Allgöwer, F. (2017), ‘Economic and dis-
6 Conclusion tributed model predictive control: Recent developments
in optimization-based control’, SICE Journal of Control,
We propose an adaptive learning-based MPC scheme Measurement, and System Integration 10(2), 39–52.
Parsi, A., Aboudonia, A., Iannelli, A., Lygeros, J. & Smith,
for uncertain interconnected systems. The developed
R. S. (2021), A distributed framework for linear adaptive
scheme is found to yield a better trade-off between MPC, in ‘2021 60th IEEE Conference on Decision and
computation complexity and closed-loop performance Control (CDC)’, Austin, TX, USA, pp. 460–465.
compared to existing schemes. Future work include ap- Parsi, A., Iannelli, A. & Smith, R. S. (2022), ‘Scalable tube
plying the developed scheme to real-world applications. model predictive control of uncertain linear systems us-
ing ellipsoidal sets’, International Journal of Robust and
Nonlinear Control .
Acknowledgements Raković, S. V., Kouvaritakis, B., Findeisen, R. & Cannon,
M. (2012), ‘Homothetic tube model predictive control’,
The authors are grateful to Prof. Roy Smith, Prof. An- Automatica 48(8), 1631–1638.
Ren, Y. M., Alhajeri, M. S., Luo, J., Chen, S., Abdullah, F.,
drea Iannelli, Dr. Goran Banjac and Dr. Anil Parsi for Wu, Z. & Christofides, P. D. (2022), ‘A tutorial review of
the fruitful discussions on the topic. neural network modeling approaches for model predictive
control’, Computers & Chemical Engineering p. 107956.
Riverso, S., Farina, M. & Ferrari-Trecate, G. (2013), ‘Plug-
References and-play decentralized model predictive control for lin-
ear systems’, IEEE Transactions on Automatic Control
Aboudonia, A., Martinelli, A. & Lygeros, J. (2021), 58(10), 2608–2614.
‘Passivity-based decentralized control for discrete-time Saltık, M. B., Özkan, L., Ludlage, J. H., Weiland, S. &
large-scale systems’, IEEE Control Systems Letters Van den Hof, P. M. (2018), ‘An outlook on robust model
5(6), 2072–2077. predictive control algorithms: Reflections on performance
Aliyu, M. (2017), Nonlinear H-infinity control, Hamiltonian and computational aspects’, Journal of Process Control
systems and Hamilton-Jacobi equations, CRC Press. 61, 77–102.
Aswani, A., Gonzalez, H., Sastry, S. S. & Tomlin, C. (2013),
‘Provably safe and robust learning-based model predictive
control’, Automatica 49(5), 1216–1226.
Betti, G., Farina, M. & Scattolini, R. (2013), Distributed
mpc: A noncooperative approach based on robustness con-
cepts, in ‘Distributed model predictive control made easy’,
Springer, pp. 421–435.
Boyd, S., El Ghaoui, L., Feron, E. & Balakrishnan, V. (1994),
Linear matrix inequalities in system and control theory,
SIAM.
Conte, C., Jones, C. N., Morari, M. & Zeilinger, M. N.
(2016), ‘Distributed synthesis and stability of cooperative
11