2142 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO.

5, SEPTEMBER 2021

Distributed Optimization Using ALADIN for MPC

in Smart Grids
Yuning Jiang , Member, IEEE, Philipp Sauerteig, Boris Houska , and Karl Worthmann

Abstract— This article presents a distributed optimization compensation of these fluctuations requires, in general, costly
algorithm tailored to solve optimization problems arising in control energy, one of the grid operator’s main objectives is
smart grids. In detail, we propose a variant of the aug- flattening the aggregated power demand [6], e.g., by using
mented Lagrangian-based alternating direction inexact Newton
(ALADIN) method, which comes along with global conver- energy storage devices on a household level [7]. In [8]–[10],
gence guarantees for the considered class of linear-quadratic it is investigated how batteries can be exploited for peak
optimization problems. We establish local quadratic conver- shaving. Moreover, in [11], batteries are used for dispatching
gence of the proposed scheme and elaborate its advantages photovoltaic power, while, in [12], thermal energy storages
compared with the alternating direction method of multipli- are considered. In practice, flexible coordination mechanisms
ers (ADMM). In particular, we show that, at the cost of
more communication, ALADIN requires fewer iterations to are required in order to efficiently manage such systems
achieve the desired accuracy. Furthermore, it is numerically despite the fact that the (local) battery dynamics are, in gen-
demonstrated that the number of iterations is independent of eral, unknown to the grid operator. Therefore, the associated
the number of subsystems. The effectiveness of the proposed optimization problems have to scale with the number of
scheme is illustrated by running both an ALADIN and an subsystems.
ADMM-based model predictive controller on a benchmark
case study. Typically, problems, as described earlier, are embedded
in a model predictive control (MPC) framework [7], [13],
Index Terms— Distributed optimization, model predictive con- where, at each time instant, one optimization problem has to
trol (MPC), smart grid.
be solved. This makes the design of efficient online solvers
I. I NTRODUCTION inevitable. For this purpose, distributed optimization meth-
ods have been developed to allow for parallelizable online
T HE rapid uptake of renewable energy sources requires
a fundamental transition of power networks from cen-
tralized to decentralized power generation [1], [2]. Herein,
calculations. A classical approach is based on dual decom-
position, where gradient-based first-order methods [14], [15]
locally distributed residential energy systems—equipped with are used to solve the concave dual problem. Alternatively,
loads, generators, and energy storage devices—play a major semismooth Newton methods [16] can be applied. However,
role in order to successfully master the required paradigm such Newton-type methods are, in general, only convergent
shift. However, their incorporation in the electricity grid entails if they are equipped with additional smoothing heuristics
the intermittent generation and bidirectional power flow and, and line-search routines. Compared with this, the alternating
thus, creates challenges for network power quality and stability direction method of multipliers (ADMM) has more favorable
[3], [4]. In particular, the volatile local energy generation convergence properties [17], [18]. Recently, many variants
results in peaks in the aggregated power demand profile, which of ADMM were proposed that exploit inherent hierarchical
can cause bottlenecks (even outages) or overload [5]. Since the structures. One drawback of ADMM, however, is its scale
dependence [19], which is typically tackled via heuristic
Manuscript received March 20, 2020; revised August 23, 2020; accepted pre-conditioners in order to accelerate convergence. Another
October 19, 2020. Date of publication November 4, 2020; date of current
version August 5, 2021. Manuscript received in final form October 20, 2020. way to construct a parallelizable online solver for MPC is
The work of Yuning Jiang and Boris Houska was supported by ShanghaiTech to use classical Newton-type methods originally proposed
University under Grant F-0203-14-012. The work of Philipp Sauerteig was for nonlinear programming, such as the interior point (IP)
supported by the German Federal Ministry for Education and Research
(BMBF) under Grant 05M18SIA. The work of Karl Worthmann was supported method [20], which can parallelize most of its operations.
in part by the German Federal Ministry for Education and Research (BMBF) Although these methods converge much faster than ADMM
under Grant 05M18SIA and in part by the German Research Founda- and have the potential of parallel implementation, a global
tion (DFG) under Grant WO 2056/6-1. Recommended by Associate Editor
J. H. Braslavsky. (Corresponding author: Yuning Jiang.) line-search routine is required to control the step size [21].
Yuning Jiang and Boris Houska are with the School of Information Science Recently, the augmented Lagrangian-based alternating direc-
and Technology, ShanghaiTech University, Shanghai 201210, China (e-mail: tion inexact Newton (ALADIN) method was proposed in [22].
[email protected]; [email protected]).
Philipp Sauerteig and Karl Worthmann are with the Institute for Mathe- Similar to ADMM, it requires the local agents to solve
matics, Technische Universität Ilmenau, 98693 Ilmenau, Germany (e-mail: small-scale decoupled problems and the central entity (CE)
[email protected]; [email protected]). to solve a consensus problem in each iteration. However,
Color versions of one or more of the figures in this article are available
online at https://ieeexplore.ieee.org. in contrast to ADMM, ALADIN solves a coupled quadratic
Digital Object Identifier 10.1109/TCST.2020.3033010 programming (QP) in the consensus step enabling locally
1063-6536 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: DISTRIBUTED OPTIMIZATION USING ALADIN FOR MPC IN SMART GRIDS 2143

quadratic convergence if suitable Hessian approximations For a number p ≥ 1, the p-norm of x ∈ Rn is defined by
are used.  n  1p
For the particular problem of minimizing variability in   
x p = |x i | p
with x = x 1 · · · x n .
power demand, a distributed algorithm was proposed in [23].
i=1
As a follow-up, Braun et al. [24] further exploited the available
flexibility resulting from battery management and applied a Moreover, we use the notations
   
centralized consensus variant of ADMM to increase efficiency ½n = 1 · · · 1  ∈ Rn , 0n = 0 · · · 0  ∈ Rn
and privacy. Note that, in this context, distributed methods are
required not just for computational efficiency but even more and denote the unit matrix in Rn×n by In . For any integers
important for maintaining the privacy and enabling plug-and-  ≤ m, we define [ : m] := {,  + 1, . . . , m} ⊆ Z. The
play capability. In detail, the grid operator is not supposed to open ball with center x ∈ Rn and radius r is denoted by
know the battery dynamics of single households. However, Br (x) = {z ∈ Rn | z − x2 < r }. We call a function f :
due to its linear convergence [18], the proposed ADMM Rn → R ∪ {∞} strongly convex on a convex set  ⊆ Rn with
variant requires many communication rounds between the grid positive parameter m > 0 if the inequality
operator and the residential energy systems to achieve the m
desired accuracy. To this end, Baumann et al. [25] proposed f (t x + (1−t)y) ≤ t f (x)+(1−t) f (y)− t(1−t)x − y22
2
to replace the expensive optimization routine with surrogate holds for all x, y ∈  and all t ∈ [0, 1]. The Kronecker
models. product of two matrices A ∈ Rm×n and B ∈ R p×q is defined
In this article, we propose a tailored algorithm based by
on ALADIN for solving the optimization problem proposed ⎡ ⎤
in [26], which aims at flattening the aggregated power demand a11 B · · · a1n B
⎢ ⎥
profile in distribution grids. Due to the local quadratic conver- A ⊗ B = ⎣ ... . . . ... ⎦ ∈ Rmp×nq .
gence of the proposed algorithm under the assumption that am1 B · · · amn B
the optimal solution is regular1 as established in Theorem 2,
the number of iterations and, thus, the total communication Finally, for given functions f, g : Rn → Rm we write
effort are reduced significantly. First, we reformulate the f (x) = O(g(x)) if there exists a constant c ∈ R such that
problem such that the objectives and inequality constraints are lim x→x0 g(x)
f (x)
≤ c. Throughout this paper, x 0 is depending on
decoupled, while only the affine equality constraints remain the context, either 0 or ∞.
intertwined. Then, the proposed algorithm, similar to the
standard ALADIN method [22], alternates solving local small- II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
scale problems and solving a coupled QP. However, our In this section, we recall the dynamic model of residential
algorithm slacks the active local constraints in the coupled energy systems incorporating loads, energy generation, and
QP. Hence, the scale of the coupled QP is fixed. Moreover, storage devices [24]. Furthermore, we provide a mathematical
we establish global convergence by using an 1 -penalty and formulation of possible goals of both the local systems and the
provide guidelines on finding suitable tuning parameters. In grid operator, yielding an overall linear-quadratic optimization
addition, we numerically show that the required number problem.
of communication rounds is independent of the number of
subsystems—typically two are sufficient. Moreover, we design
an MPC scheme based on the proposed algorithm and illus- A. Residential Energy Systems
trate advantages compared with ADMM by applying it in a In this article, we consider a smart grid with I ∈ N
benchmark case study. residential energy systems, which are coupled via a CE,
The remainder is structured as follows. Section II recalls the grid operator (see Fig. 1). The latter has to compensate
the physical model of residential energy systems and a peak- for the need for and the surplus of energy. Note that we are
shaving problem based on optimal energy storage control. only interested in energy conservation on a residential level.
In Section III, we propose the variant of ALADIN tai- We do not incorporate the grid topology.
lored to the structured optimization problem and elaborate The battery dynamics of system i , i ∈ [1 : I], is given by
on the implementation in detail. Moreover, we establish  
x i (n + 1) = αi x i (n) + T βi u + −
i (n) + u i (n) (1a)
local quadratic convergence and global convergence. Then,
ALADIN is embedded within an MPC scheme. Based on a z i (n) = wi (n) + u + −
i (n) + γi u i (n) (1b)
benchmark case study, the numerical results in Section IV
show that ALADIN is scalable and outperforms ADMM. with state of charge (SoC) x i (n) ∈ R of the battery in kWh,
Notations: We use (Sn+ ) Sn++ to denote the set of symmetric power demand z i (n) ∈ R in kW at time instant n ∈ N0 , and
positive (semi) definite matrices in Rn×n , n ∈ N. For a given sampling time T > 0 in h. The control inputs
 ∈ Sn++ and x ∈ Rn , the notation x2 = x  x is used.  
u i (n) = u + −
i (n) u i (n) ∈ R2

1 An
represent the charging/discharging rate in kW, while the net
optimal solution is called regular if the linear independence constraint
qualification (LICQ), strict complementarity conditions (SCCs), and the consumption wi (n) ∈ R in kW is load minus generation,
second-order sufficient conditions (SOSCs) are satisfied [27, p. 591]. i.e., power demand neglecting the battery. The constants

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
2144 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO. 5, SEPTEMBER 2021

B. Objective: Load Shaping Subject to Local Costs

The grid operator is interested in flattening the aggregated
power demand profile. One way is to track the overall net
consumption ζ ∈ R N , given by

n I

1
ζ (n) = wi ( j )
N j =n−N +1 i=1

for n ∈ [k : k + N − 1], where k ≥ N − 1 denotes the

current time instant. The corresponding objective function can
be modeled as f0 : R N → R≥0
k+N −1 
σ0  z̄(n) − ζ (n) 2
f 0 (z̄) =
N n=k I
σ0
= z̄ − ζ 22 (3)
N · I2

Fig. 1. Electrical network consisting of several residential energy systems
connected to a CE.
with σ0 > 0 and z̄(n) = Ii=1 z i (n). In addition, we introduce
the local costs f i : R N × R2N → R≥0
σi  
f i (u i ) = z i − wi 22 + u i 22 (4a)
αi , βi , γi ∈ (0, 1] model efficiencies with respect to self- 2
σ
(1b) i     
discharge and energy conversion, respectively. =  I N ⊗ 1 γi u i 22 + u i 22 (4b)
At time k, we consider the initial condition 2
1
= u i 2Q i (4c)
x i (k) = x̂ i , i ∈ [1 : I] 2
 
with σi ≥ 0, z i = z i (k) · · · z i (k + N − 1) , and coefficient
of the i th battery with current measurements x̂ i and assume matrices
the future net consumption  
1 γi
  Q i = σi I2N + I N ⊗ (5)
γi γi2
wi = wi (k) · · · wi (k + N − 1)
for all i ∈ [1 : I]. As discussed in [24], the first term on
on the prediction window [k : k+N −1], N ∈ N≥2 , to be given. the right-hand side of (4a) ensures that each agent’s demand
The states and control inputs are subject to the constraints does not change drastically within the optimization, i.e., each
household does not end up with an entirely different demand.
0 ≤ x i (n) ≤ Ci , (2a) The second term penalizes the charging and discharging effort.
≤ u− Moreover, the parameters σ0 and σi , i ∈ [1 : I], represent
ui i (n) ≤ 0, (2b)
the weights of the global cost (3) and local objectives (4),
0 ≤ u+
i (n) ≤ ui , (2c)
respectively. The use of the i th battery is penalized more with
u−
i (n) u+
i (n)
0 ≤ ui
+ ui
≤ 1. (2d) increasing σi , i ∈ [1 : I].

where Ci ≥ 0 denotes the battery capacity of the i th system,

C. Problem Formulation
and ū i and u i are the bounds of the charging and discharging
rates, respectively. The considerations in Sections II-A and B motivate the
Remark 1: Due to the recursive structure of the individual optimization problem
system (1a), the future SoCs x i (k + n), n ∈ [1 : N], are I

determined by (1a), and the initial SoC x i (k), in particular min f 0 (z̄) + f i (u i ) (6a)
z̄,u
i=1

n−1
  I
x i (k +n) = αin x̂ i + T αin−1− βi 1 u i () 
=0
s.t. z̄ = w + Ai u i |λ (6b)
i=1
holds for all i ∈ [1 : I] and n ∈ [1 : N]. Hence, state and Di u i ≤ di , i ∈ [1 : I] (6c)
input constraints (2) of system i over the next N time steps I
can be summarized as polyhedral constraints with w = i=1 wi . The matrices Ai ∈ R N ×2N are given by
 
Ai = I N ⊗ 1 γi , i ∈ [i : I] (7)
Di u i ≤ d i
while constraint (6b) introduces the coupling between the CE
with matrix Di ∈ R8N ×2N and vector di ∈ R8N , i ∈ [1 : I]. and subsystems.

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: DISTRIBUTED OPTIMIZATION USING ALADIN FOR MPC IN SMART GRIDS 2145

Remark 2: The Ai expression (7) shows that the affine 1) Parallelizable Step: The small-scale QPs
equality constraint (6b) satisfies the LICQ condition, 1
i.e., matrix min f i (v i ) − (Ai v i ) λ + v i − u i 2Q i
vi 2
 
−I N A1 · · · AI s.t. Di v i ≤ di | κi (8)

has full row rank. are solved in parallel for all i ∈ [1 : I]. Here, (u i , λ)
Throughout this article, we write the Lagrangian multiplier denotes the current primal and dual iterates of ALADIN for
right after the constraints such that the notation “|λ” denotes solving 6. If we denote the primal-dual solution of (8) by
the Lagrangian multipliers of constraint (6b). (v i , κi ), the first-order optimality condition of (8) is given by
Next, we state sufficient conditions for uniqueness of the
0 = ∇ f i (v i ) + Di κi − A
i λ + Q i (v i − u i ).
optimal solution of problem (6).
Proposition 1: Let σ0 > 0 and Problem (6) be feasible. Hence, the modified gradient gi = ∇u i { f i (v i ) + κi Di v i } can
Then, the optimal power demand z̄ ∗ and corresponding dual be evaluated by
λ∗ are unique. If, in addition, σi > 0 for all i ∈ [1 : I], then
the optimal control u ∗ of (6) is also unique. gi = A 
i λ + Q i (u i − v i ) (9)
Proof: First,
 let σ0 > 0. Due to  the compactness of the in a derivative-free manner. As discussed in [22], this paral-
feasible set u i ∈ R2N  Di u i ≤ di of the i th system, the set
   lelizable step exploits the distributed structure of (6) by using
 z̄ = w + I A u dual decomposition and augmented Lagrangian.
N i=1 i i
Z = z̄ ∈ R  2) Consensus Step: The following equality constrained QP
 Di u i ≤ di , i ∈ [i : I]
is solved
is compact (see [28]). Since f 0 is strongly convex, the optimal  + + +
output z̄ ∗ corresponding to (6) and, hence, the dual λ∗ are z̄ , u , s
unique. This proves the first assertion. Now, let σi > 0 for all := arg min f 0 (z̄) (10a)
i ∈ [0 : I]. Then, problem (6) is a strongly convex quadratic z̄,u,s
I
 
program. Furthermore, the constraint Jacobian of (6b) has 1 μi
full row rank, indicating that the LICQ is satisfied. Hence, + u i − v i 2Q i + u
i gi + si 22
2 2
the optimal solution (z̄ ∗ , u ∗ , λ∗ ) is unique [29].  i=1
I

s.t. z̄ = w + Ai u i | λ+ (10b)
From now on, we assume that σi > 0 for all i ∈ [1 : I],
i=1
which implies that the matrices Q i are invertible. Solving
problem (6) by using a centralized method assumes the exis- Diact (u i − v i ) = si | κiQP , i ∈ [1 : I] (10c)
tence of an all-knowing entity, i.e., the grid operator knows the where matrix Diact denotes the active Jacobian matrix at the
dynamical model of each household and its predicted load and local solution v i , i.e., Diact v i = diact for all i ∈ [1 : I]. The
generation. This is impractical in the case of several thousand solution (u + , λ+ ) of (10) is used to update the primal and dual
households. Motivated by this, distributed or parallelizable iterates. The value of the slackness parameters μi reflects how
methods [23], [24], [26] have been developed to solve (6). much system i trusts the local active constraints. Later on, it is
However, as discussed in Section I, existing methods either shown that updating μi helps to improve the local convergence
have a slow convergence or require a global line-search rate.
routine, which needs many undesired communication rounds Remark 3: Compared with the standard ALADIN [22],
between the grid operator and the residential energy sys- the consensus QP (10) relaxes the decoupled equality con-
tems. In Section III, we propose a line-search-free distributed straints but not the coupled affine constraints. If we denote
method with fast local convergence to reduce the number of scaling matrices
communication rounds significantly.

Hi = Q i + μi Diact Diact , i ∈ [1 : I] (11)
III. ALADIN FOR S MART G RIDS
QP (10) can be rewritten in the form
In this section, we propose a variant of the ALADIN I 
algorithm to solve (6) in a distributed manner. Furthermore,  1 
min f0 (z̄) + u i − v i  Hi + u i gi
2
(12a)
we establish both global convergence and locally quadratic z̄,u 2
i=1
convergence for the proposed algorithm and show how it can I

be embedded efficiently in an MPC scheme. s.t. z̄ = w + Ai u i | λ. (12b)
i=1
A. Distributed Optimization with ALADIN The advantage is that the scale of QP (12) is fixed, i.e.,
This section presents a variant of the ALADIN method to the number of constraints is fixed and does not depend on
solve (6). Similar to the standard ALADIN method proposed the number of active constraints. However, in contrast to the
in [22], there are two main steps: a parallelizable step and a ADMM variant proposed in [24], the consensus QP (12) is still
consensus step. large scale due to its dependence on the number of systems.

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
2146 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO. 5, SEPTEMBER 2021

Moreover, as discussed in Remark 2, constraint (12b) Algorithm 1 Tailored ALADIN Method for Solving (6)
satisfies the LICQ such that QP (12) is always feasible
independently of the feasibility of (6). Note that as the sum
of a positive definite and a positive semidefinite matrix, Hi is
positive definite.

B. Algorithm with Implementation Details

In this section, we elaborate on the implementation details
(see Algorithm 1) of the distributed method proposed in
Section III-A.
Before we have a closer look at Algorithm 1, let us
introduce two important ingredients used within the algorithm.
First, since QP (12) only incorporates equality constraints,
the analytical solution is given by
 I

  
λ+ = −1 ζ − w + Ai Hi−1 gi − v i (13a)
i=1
 
u+
i = vi + Hi−1 A
i λ
+
− gi , i ∈ [1 : I] (13b)
N ·I + 2
z̄ + = ζ − λ . (13c)
σ0
Here, the scaling matrix  is defined as
 I
N · I2
 = IN + Ai Hi−1 A
i . (14)
2σ0 i=1

The derivation is a direct consequence of the fact that the

KKT conditions of (12) are linear equations. Note that the dual
update λ+ requires communication between local subsystems
and the CE, while the primal update u + i is parallelizable. As
mentioned in Section III-A, in order to improve the local
convergence, the parameters μi need to be updated online.
To this end, we introduce the 1 -penalty-based merit function
: R N × R2I N → R
I
 I

(z̄, u) = f 0 (z̄) + f i (u i ) + λ̄z̄ − w − Ai u i 1
i=1 i=1

with parameter λ̄ > 0. Here, we assume that λ̄ is sufficiently

large, which is important to establish the global convergence of
Algorithm 1. Whenever, the merit function fulfills the descent
condition (17), we set  = 1; otherwise,  = 0, indicating
the μi update.
In Algorithm 1, the integer  denotes the index of the iter-
ation. During the initialization phase, the CE precomputes the
scaling matrix −10 similar to (14) with μi = 0, i.e., Hi = Q i .
Here, we use the Sherman–Morrison–Woodbury formula [27,
Appendix A] to compute Q −1 i
 
1 1 1 γi
Q −1 = I − I ⊗
i
σi
2N
2 + γi2
N
γi γi2
 
1 1 1 γi
= I N ⊗ I2 −
2 + γi2 γi γi
2
σi
 
1 1 + γi2 − γi
=   IN ⊗ (18)
σi 2 + γi2 −γi 2

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: DISTRIBUTED OPTIMIZATION USING ALADIN FOR MPC IN SMART GRIDS 2147

which yields practical way to implement the descent condition (17) is to

    check
1 1 + γi2 − γi 1  
γi −γi 2 γi z̄  , v  ≤ ψ − ε̂
Ai Q −1 
i Ai = I N ⊗  
σi 2 + γi2
for a small constant ε̂ of order of the machine precision
(or local QP solver’s accuracy). This strategy has been used
1 + γi2
=   IN . (19) in some existing open-source toolkit, such as ACADO [30].
σi 2 + γi2 Moreover, the numerical results in the consecutive section (see
Fig. 2) illustrate that the locally quadratic convergence rate can
This indicates that computing −1
0 by (15) only requires the
be observed if this strategy is applied within Algorithm 1.
subsystems to send scalars
Although this can be interpreted as the weakest possible
1 + γi2 descent condition that can be checked with finite precision
  , i ∈ [1 : I]
σi 2 + γi2 arithmetic, it works well in practice.
Remark 5: The merit function only aims to control the
to the CE. update of μi either according to (16) or μi = 0. It never rejects
In the main loop, the subsystems first update u +1 i in steps or slows down the progress of the iterations in Algo-
parallel based on the current μi in Step 1a). Once the local rithm 1, which is different from globalization routines, such
QP (8) is solved and the gradient gi is evaluated in Step 1b), as line search methods, when using centralized approaches.
parameter μi and matrix Hi are computed in Step 1c). Then,
the subsystems send the local information constructed in
C. Convergence Analysis
Step 1d) to the CE. Here, the symmetric matrix Ai Hi−1 A i
can be computed as In this section, we analyze the theoretical convergence
properties of Algorithm 1 in two steps: local convergence
1 + γi2 μi and global convergence. Since σi > 0 for all i ∈ [0 : I],
Ai Hi−1 A
i =   · IN −  2
σi 2 + γi 2
σi2 2 + γi2 Proposition 1 guarantees the uniqueness of the optimal primal
 −1 (z̄ ∗ , u ∗ ) and dual solution λ∗ of (6). Moreover, we assume that
 act 
·Ai Diact In act
i
+ μDiact Q −1
i Di Diact A
i Algorithm 1 checks the Armijo condition (21) at Step 2b) to
(20) determine the update of μi .
Proposition 2: Let the optimal solution of Problem (6) be
where n act
i denotes the number of rows of Diact . The explicit regular and μi be updated such that
form (20) implies that sending Ai Hi−1 A i to the CE only  
1
requires N · n act
i floats. < O v i − u i 2 , i ∈ [1 : I]. (22)
After the CE collects the local information, Step 2a) checks μi
the terminal condition of Algorithm 1. If Then, there exist constants r, η > 0 with r η ≤ 1 such that all
v  ∈ Br (u ∗ ) satisfy
max v i − u i 1 < ε
i v +1 − u ∗ 2 ≤ ηv  − u ∗ 22 .
holds, the current iterate v  satisfies the stationary condition Proof: Locally, following the regularity assumption that
and primal condition of (6) up to an error of order O(ε), i.e., the active sets at the local solutions v i ∈ Br (u ∗ ) are fixed
I
 such that the local solution v  of (8) can be represented as
Ai v i − z̄  − w = O(ε) the image of (u, λ) under an affine map, i.e., there exist some
i=1 matrices T1 , T2 ∈ RI N ×I N such that
and gi + A     
i λ = O(ε), i ∈ [1 : I]. v  − u ∗ = T1 u  − u ∗ + T2 λ − λ∗ . (23)
Otherwise, Step 2b) is executed based on the merit function . Moreover, if the μi ’s satisfy the conditions (22), the coupled
In order to check the strict descent condition (17), the Armijo QP (12) yields the quadratic contractions
conditions [27, p. 540] defined by α
  u +1 − u ∗ 2 ≤ v  − u ∗ 22
z̄  , v  ≤ ψ + ηD (z̄, v) with η ∈ (0, 1) (21) 2
α
and λ+1 − λ∗ 2 ≤ v  − u ∗ 22 (24)
is used. Here, (z̄, v) denotes the difference between the 2
current iterates (z̄  , v  ) and the previous iterates associated with a positive constant α similar to the Newton-type meth-
with ψ, and D (z̄, v) denotes the directional derivative ods [22, Sec. 7]. Considering two consecutive iterations in
of in the direction (z̄, v). How to compute D (z̄, v) Algorithm 1, we get
has been elaborated in [27, Th. 18.2]. Then, in Step 2c),
the dual iterate is updated, and (λ+1 , ) is sent to the local v +1 − u ∗ 2
(23)
systems in Step 2d). ≤ T1 2 u +1 − u ∗ 2 + T2 2 λ+1 − λ∗ 2
Remark 4: In practice, the Armjio condition (21) is expen- (24) α
sive to check as it requires to compute D (z̄, v). One ≤ (T1 2 + T2 2 )v  − u ∗ 22
2

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
2148 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO. 5, SEPTEMBER 2021

which concludes the proof with η = (α/2)(T1 2 + T2 2 ).  Algorithm 2 ALADIN-Based Distributed MPC Scheme
This proposition establishes the local quadratic convergence
rate of the iterates v  while implying that the local convergence
progress benefits from the μ update.
Remark 6: If the iterates (v  )∈N converge locally with
quadratic rate, as discussed in Proposition 2, the descent
condition (21) always holds in a neighborhood of a regular
minimizer [27]. Hence, μi is updated in every step. In Algo-
rithm 1, the μi update is based on (16) ensuring that (22)
holds locally.
Concerning the global convergence analysis, we introduce
an auxiliary function L : R2I N × R N → R
I

L(u, λ) = λ − λ∗ 20 + u i − u ∗i 2Q i
i=1

to measure the distance of iterates (u, λ) to the opti-

mum (u ∗ , λ∗ ). In order to establish global convergence,
we introduce the following technical result.
Lemma 1: Let Problem (6) with σi > 0, i ∈ [1 : I], be globally to the optimal solution u ∗ . If, additionally, the optimal
feasible and μ = 0 at the current iteration of Algorithm 1. solution of (6) is regular, the iterates v  locally converge with
Then, there exists some m > 0 such that the iterates satisfy quadratic rate.
the inequality Proof: The first statement follows Lemma 1 and the fact
    that is an exact merit function. As Proposition 2 presented,
L u +1 , λ+1 ≤ L u  , λ − 4mv +1 − u ∗ 22 . (25) the second statement holds since v  converges with quadratic
rate after entering the local neighborhood Br (u ∗ ). 
Proof: According to Proposition 1, the choice of σi >
Remark 7: In practice, the rigorous construction of the
0 for all i ∈ [1 : I] indicates that Problem (6) is strongly
parameter λ̄ < ∞ in function requires metadata from the
convex. Then, the proof can be established following the proof
user in order to ensure that is an exact merit function.
of Theorem 1 in [31] step by step. 
However, state-of-the-art sequential QP (SQP) solvers [27]
Since function L is bounded from below by 0, Lemma 1
implement effective heuristics for choosing λ̄ online in order to
indicates that the sequence (L(u  , λ ))∈N converges as the
avoid this need for metadata. In this article, we update λ̄ based
algorithm progresses, i.e., v  converges to the optimal solution
on the heuristic used in the open-source toolkit ACADO [30].
if μi = 0 is fixed. However, this is in conflict with the local
If the descent condition of at Step 4b) holds, we update
quadratic convergence established in Proposition 2, which
requires to update μi . In order to overcome this problem λ̄ = 10λ ∞ .
and establish the global convergence of Algorithm 1 with
optionally updating μi , we introduce the following technical
result.
Proposition 3: Let the optimal solution (z̄ ∗ , u ∗ , λ∗ ) of Prob- D. Distributed MPC Scheme Using ALADIN
lem (6) be regular and consider In practice, the net consumption is not known in advance
min (z̄, u) s.t. Di u i ≤ di , i ∈ [1 : I]. (26) but can be predicted for a few hours based on standard load
z̄,u profiles and weather forecasts. However, these predictions
If λ̄ satisfies change continually. In particular, short-term predictions are
more reliable than long-term predictions. MPC is a state-of-
λ̄ > λ∗ ∞ = max |λ∗i | the-art technique to handle such predictions.
i
In the context of MPC, problem (6) is solved during each
with λ∗i
the i th element of λ , then∗
sampling time step k based on the current measurements,
 ∗ ∗   and then, each subsystem implements the first element of its
z̄ , u primal solution of (6) ⇔ z̄ ∗ , u ∗ solves (26)
control sequence, u ∗i (k), i ∈ [1 : I]. This process defines an
holds, i.e., is an exact merit function of Problem (6). implicit feedback law that is applied iteratively [26].
For details of the proof, we refer to [27, Th. 17.3]. As λ̄ is Algorithm 2 outlines an ALADIN-based distributed MPC
assumed to be sufficiently large, Proposition 3 indicates the scheme for smart grids. Similar to the proposed ADMM-
exactness of . Therefore, if the Armijo condition (21) holds, based MPC scheme in [24], Algorithm 2 does not assume
the convergence of v  can be preserved with μi update. This the existence of some all-knowing entity. Each local agent
result is summarized in the following theorem. only knows its own parameters and optimization problem.
Theorem 1: Let σi > 0, i ∈ [0 : I], and Problem (6) The CE collects only specific local information, as outlined
be feasible. If the Armijo condition (21) is used in Step 2d) in Algorithm 1, and broadcasts the updated dual variables and
of Algorithm 1, the iterates (v  )∈N in Algorithm 1 converge one float encoding whether μi is updated or not. Furthermore,

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: DISTRIBUTED OPTIMIZATION USING ALADIN FOR MPC IN SMART GRIDS 2149

TABLE I TABLE II
C ONVERGENCE AND C OMMUNICATION C OMPARISON PARAMETER VALUES

the control is initialized based on the optimal solution of the

previous MPC.
Although the reference trajectory in our setting is based on
the (predicted) future net consumption, one could consider ζ =
{ζ(n)}n∈[k:k+N −1] at time step k as a segment of a given
sequence such that we do not have to compute it in each
MPC step. In that case, the reference does not depend on
predicted values. As a result, the subsystems do not need
to send wi to the CE. Notice that the online choice of the
reference does not affect the established convergence results
applying for solving (6).

IV. C ASE S TUDY: ALADIN V ERSUS ADMM

In this section, we compare Algorithm 1 with the state-
of-the-art distributed optimization algorithm ADMM. Here,
we use the implementation of ADMM, as proposed in [24].
Fig. 2. Convergence comparison with I = 100 subsystems.
A. Theoretical Comparison
Regarding the theoretical convergence, both ALADIN subsystems and the CE is the most time-consuming part in the
and ADMM have global convergence guarantees. However, context of predictive control for smart grids. Speeding up the
ADMM only achieves linear convergence, while ALADIN convergence progress by reducing communication rounds is
converges locally with a quadratic rate. The comparison in the key measure to save run time in practice. In Section IV-B,
detail is listed in Table I (left). Here, the symbol  represents Algorithm 1 and ADMM are compared numerically.
that Algorithm 1 converges globally without mentioning the
convergence rate.
Table I (right) compares the communication overhead per B. Numerical Results
iteration using Algorithm 1 and ADMM. In this context The main advantage of ALADIN compared with state-of-
forward and backward are to be understood that one local the-art distributed optimization algorithms, such as ADMM,
agent uploads information to the CE, and vice versa, and n act i is the local quadratic convergence (see Theorem 2). In this
is the number of active constraints at the i th system. The table section, we show the practical importance of this fact based
shows that the communication overhead for both Algorithm 1 on numerical simulations.
and ADMM is O(N). In order to update Hi , the CE needs The parameter values used throughout all experiments are
to collect the information of the active Jacobian Diact for listed in Table II. Since the prediction of the future net
each agent i ∈ [1 : I] and broadcast . Thus, the forward consumption is out of the scope of this article, we use data
communication overhead is O(N). provided by an Australian grid operator [32].
Concerning the online computational efforts, both Algo- 1) Open-Loop Comparison: In order to illustrate the numer-
rithm 1 and ADMM execute two main steps per iteration: ical performance of Algorithm 1, we compare it to ADMM and
a parallelizable step and a consensus step. The complexity of an IP method for a benchmark with I = 100. In a result, each
the local problems solved by the subsystems in parallel is the local subsystem has 48 variables, and Problem (6) incorporates
same for both ADMM and Algorithm 1. The main difference 4824 variables in total. Here, we initialize all algorithms with
between these approaches is the consensus step performed by zero initial guesses.
the CE. Since the update of −1 is optionally required in The results in Fig. 2 confirm that Algorithm 1 outperforms
Algorithm 1, the cost of the ALADIN consensus step in worst ADMM [24] and the IP method [20]. Here, we solved Prob-
case is O(N 2 I), while ADMM only needs O(N). Note that lem (6) for fixed time step k and initial condition and plotted
the number of subsystems I, in practice, is much larger than the gap between the iterates u  and the optimal solution u ∗ .
the length of prediction horizon N. However, these minor dif- It can be seen that ADMM needs almost 20 iterations to
ferences in communication effort and computational complex- achieve an accuracy of 10−2 , while ALADIN converges within
ity do not dominate the run time. Communication between the six steps to an error less than 10−6 . Moreover, ALADIN

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
2150 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO. 5, SEPTEMBER 2021

TABLE III
N UMBER OF I TERATIONS FOR D IFFERENT M ETHODS

converges two times faster than the IP method if we choose

the stop tolerance 10−4 . This trend can also be observed
in Table III, where all methods were performed based on 100
randomly generated initial conditions. It is worth mentioning
that, in every single of these randomly generated test cases,
Algorithm 1 outperformed the other methods.
Hence, compared with the existing state of the arts, sig-
nificantly fewer communication rounds between the local Fig. 3. Impact of the number of systems on the number of iterations needed
to achieve pre-defined accuracy.
agents and the CE are needed in practice for Algorithm 1,
as discussed in Section III-B. Note that each iteration of the IP
method requires the CE to communicate with the local agents
additionally to apply a line-search routine, as discussed in [21],
even though most of its operations have been parallelized.
Therefore, the IP method needs more communication rounds
than ADMM. In the following, we focus on the comparison
between two distributed optimization approaches: Algorithm 1
and ADMM.
In our example, the complexity of the coupled problems
solved within each iteration of ALADIN and ADMM, respec-
tively, does not depend on I. However, the number of systems
affects the number of iterations of ADMM needed to solve
problem (6). Fig. 3 visualizes the impact of I on the number of
iterations of ALADIN and ADMM needed to achieve different
accuracies, in particular, {10−1 , 10−3 , 10−4 }. It can be seen
that the convergence speed of ALADIN does not depend on
the number of systems, while ADMM requires more iterations
if the number of agents increases.
2) Closed-Loop Comparison: Fig. 4 is dedicated to the
closed-loop (CL) performance of Algorithm 1 and ADMM
for 30 h, i.e., 60 MPC steps. More precisely, Fig. 4 (top)
compares the effort needed to achieve a predefined accuracy
using ALADIN and ADMM by counting the executed itera-
tions. Again, the local quadratic convergence rate of ALADIN Fig. 4. MPC closed loop: Number  of iterations needed to achieve some pre-
outperforms ADMM. ALADIN needs approximately six iter- defined accuracy (top) and actual optimization gap at iteration  (bottom). The
dashed black lines (top) represent the respective average number of iterations.
ations in average to achieve accuracy v  − u ∗ ∞ < 10−4 ,
while ADMM needs more than 20 and 30 iterations for
accuracy 10−3 and 10−4 , respectively. Both algorithms benefit
from warm start techniques used in the implementation when warm start. However, in a particular time window (18–22 h),
k ≥ N. Since the reference value ζ (k + 1) strongly differs the warm start does not work efficiently due to the big changes
from ζ (k) with k < N, the optimal dual variable might of the optimal active sets. The value of does not always
change abruptly within consecutive MPC iterations as well. decrease in such cases, which means that ALADIN needs
As a result, the warm start as outlined in Algorithm 2 does more steps to detect the correct active set before achieving
not yield a reasonable performance if k < N. To avoid this the local quadratic convergence progress. Note that, in some
effect, we only compare the CL performance after N steps. MPC iterations, ADMM needs fewer iterations since the initial
After that, updating ζ equals a simple shift in time enabling a guess based on the warm start is close to the optimum. In the
majority of the cases, the quadratic convergence of ALADIN
1 This notation means that Algorithm 1 takes on average ≈ 4 iterations to yields even better performance than aimed for, as visualized
achieve accuracy 10−2 , while the standard deviation is ≈ 2 iterations. in Fig. 4 (bottom). A potential future direction to reduce

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: DISTRIBUTED OPTIMIZATION USING ALADIN FOR MPC IN SMART GRIDS 2151

[7] K. Worthmann, C. M. Kellett, L. Grüne, and S. R. Weller, “Distributed

control of residential energy systems using a market maker,” IFAC Proc.
Volumes, vol. 47, no. 3, pp. 11641–11646, 2014.
[8] E. Telaretti and L. Dusonchet, “Battery storage systems for peak load
shaving applications: Part 1: Operating strategy and modification of the
power diagram,” in Proc. IEEE 16th Int. Conf. Environ. Electr. Eng.
(EEEIC), Jun. 2016, pp. 1–6.
[9] E. Telaretti and L. Dusonchet, “Battery storage systems for peak
load shaving applications: Part 2: Economic feasibility and sensitivity
analysis,” in Proc. IEEE 16th Int. Conf. Environ. Elect. Eng. (EEEIC),
Jun. 2016, pp. 1–6.
[10] Y. Shi, B. Xu, D. Wang, and B. Zhang, “Using battery storage for
peak shaving and frequency regulation: Joint optimization for super-
linear gains,” IEEE Trans. Power Syst., vol. 33, no. 3, pp. 2882–2894,
Fig. 5. Comparison between open and CL performance. May 2018.
[11] M. Brenna, F. Foiadelli, M. Longo, and D. Zaninelli, “Energy storage
control for dispatching photovoltaic power,” IEEE Trans. Smart Grid,
vol. 9, no. 4, pp. 2419–2428, Jul. 2018.
the number of communication rounds further would be to [12] P. Kohlhepp, H. Harb, H. Wolisz, S. Waczowicz, D. Müller, and V.
prematurely stop the algorithm [33]. Hagenmeyer, “Large-scale grid integration of residential thermal energy
storages as demand-side flexibility resource: A review of international
3) Open Loop Versus Closed Loop: In order to demonstrate field studies,” Renew. Sustain. Energy Rev., vol. 101, pp. 527–547,
the benefits of the CL implementation with respect to inac- Mar. 2019.
curate weather forecast, we perform a case study inspired [13] A. Parisio, E. Rikos, and L. Glielmo, “A model predictive control
by [26]. Let ξ(n) ∼ U([−0.1, 0.1]) denote independent approach to microgrid operation optimization,” IEEE Trans. Control
Syst. Technol., vol. 22, no. 5, pp. 1813–1827, Sep. 2014.
identically uniformly distributed random variables for all n ∈ [14] A. Rantzer, “Dynamic dual decomposition for distributed control,” in
[k : k + N − 1]. Then, we construct the disturbed predictions Proc. Amer. Control Conf., 2009, pp. 884–888.
at time k via [15] S. Richter, M. Morari, and C. N. Jones, “Towards computational com-
plexity certification for constrained MPC based on Lagrange relaxation

n
and the fast gradient method,” in Proc. IEEE Conf. Decis. Control Eur.
widis,k (n) = wi (n) + ξ ( j ), n ∈ [k : k + N − 1] Control Conf., Dec. 2011, pp. 5223–5229.
j =k [16] J. V. Frasch, S. Sager, and M. Diehl, “A parallel quadratic programming
method for dynamic optimization problems,” Math. Program. Comput.,
such that the variance increases in time. Fig. 5 demonstrates vol. 7, no. 3, pp. 289–329, Sep. 2015.
the superiority of the CL implementation with respect to [17] T. Goldstein, B. O’Donoghue, S. Setzer, and R. Baraniuk, “Fast alternat-
these disturbances compared with the open-loop (OL) solution. ing direction optimization methods,” SIAM J. Imag. Sci., vol. 7, no. 3,
pp. 1588–1623, Jan. 2014.
Here, we depicted the optimal solutions for 100 Monte Carlo [18] M. Hong and Z.-Q. Luo, “On the linear convergence of the alternating
simulations, as well as mean and standard deviation. Since direction method of multipliers,” Math. Program., vol. 162, nos. 1–2,
the CL solution incorporates updated predictions in each time pp. 165–199, Mar. 2017.
[19] B. O’Donoghue, E. Chu, N. Parikh, and S. Boyd, “Conic optimization
step, the standard deviation is smaller. via operator splitting and homogeneous self-dual embedding,” J. Optim.
Theory Appl., vol. 169, no. 3, pp. 1042–1068, Jun. 2016.
V. C ONCLUSION AND O UTLOOK [20] I. Necoara and J. A. K. Suykens, “Interior-point Lagrangian decomposi-
tion method for separable convex optimization,” J. Optim. Theory Appl.,
In this article, we proposed an optimization scheme based vol. 143, no. 3, pp. 567–588, Dec. 2009.
on ALADIN for the optimal control of locally distributed [21] A. Bitlislioğlu, I. Pejcic, and C. N. Jones, “Interior point decomposition
energy storage to achieve an overall goal. Due to its locally for multi-agent optimization,” IFAC-Papers OnLine, vol. 50, no. 1,
pp. 233–238, 2017.
quadratic convergence, it outperforms state-of-the-art methods, [22] B. Houska, J. Frasch, and M. Diehl, “An augmented Lagrangian based
such as ADMM. A numerical case study illustrates that algorithm for distributed nonconvex optimization,” SIAM J. Optim.,
our approach reduces the number of communication rounds vol. 26, no. 2, pp. 1101–1127, Jan. 2016.
[23] P. Braun, L. Grüne, C. M. Kellett, S. R. Weller, and K. Worthmann,
required to achieve a given accuracy tremendously. Future “A distributed optimization algorithm for the predictive control of smart
research will investigate the impact of flattening the aggregated grids,” IEEE Trans. Autom. Control, vol. 61, no. 12, pp. 3898–3911,
power demand on congestion in the grid. Dec. 2016.
[24] P. Braun, T. Faulwasser, L. Grüne, C. M. Kellett, S. R. Weller, and
K. Worthmann, “Hierarchical distributed ADMM for predictive control
R EFERENCES with applications in power networks,” IFAC J. Syst. Control, vol. 3,
pp. 10–22, Mar. 2018.
[1] A. Ipakchi and F. Albuyeh, “Grid of the future,” IEEE Power Energy
Mag., vol. 7, no. 2, pp. 52–62, Mar./Apr. 2009. [25] M. Baumann, S. Grundel, P. Sauerteig, and K. Worthmann, “Sur-
[2] Y. Guo, M. Pan, Y. Fang, and P. P. Khargonekar, “Decentralized rogate models in bidirectional optimization of coupled micro-
coordination of energy utilization for residential households in the smart grids,” at-Automatisierungstechnik, vol. 67, no. 12, pp. 1035–1046,
grid,” IEEE Trans. Smart Grid, vol. 4, no. 3, pp. 1341–1350, Sep. 2013. Nov. 2019.
[3] D. E. Olivares et al., “Trends in microgrid control,” IEEE Trans. Smart [26] K. Worthmann, C. M. Kellett, P. Braun, L. Grüne, and S. R. Weller,
Grid, vol. 5, no. 4, pp. 1905–1919, Jul. 2014. “Distributed and decentralized control of residential energy systems
[4] X. Yu and Y. Xue, “Smart grids: A cyber–physical systems perspective,” incorporating battery storage,” IEEE Trans. Smart Grid, vol. 6, no. 4,
Proc. IEEE, vol. 104, no. 5, pp. 1058–1070, May 2016. pp. 1914–1923, Jul. 2015.
[5] T. Morstyn, B. Hredzak, and V. G. Agelidis, “Control strategies for [27] J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed. Cham,
microgrids with distributed energy storage systems: An overview,” IEEE Switzerland: Springer, 2006.
Trans. Smart Grid, vol. 9, no. 4, pp. 3652–3666, Jul. 2018. [28] R. T. Rockafellar, Convex Analysis. Princeton, NJ, USA: Princeton Univ.
[6] S. Parhizi, H. Lotfi, A. Khodaei, and S. Bahramirad, “State of the art in Press, 2015.
research on microgrids: A review,” IEEE Access, vol. 3, pp. 890–925, [29] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
2015. Cambridge Univ. Press, 2004.

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.
2152 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 29, NO. 5, SEPTEMBER 2021

[30] B. Houska, H. J. Ferreau, and M. Diehl, “ACADO toolkit—An open- Boris Houska received the Diploma degree in math-
source framework for automatic control and dynamic optimization,” ematics from the University of Heidelberg, Hei-
Optim. Control Appl. Methods, vol. 32, no. 3, pp. 298–312, May 2011. delberg, Germany, in 2007, and the Ph.D. degree
[31] Y. Jiang, J. Oravec, B. Houska, and M. Kvasnica, “Parallel MPC for in electrical engineering from KU Leuven, Leuven,
linear systems with input constraints,” IEEE Trans. Autom. Control, p. 1, Belgium, in 2011.
2020, doi: 10.1109/TAC.2020.3020827. From 2012 to 2013, he was a Post-Doctoral
[32] E. L. Ratnam, S. R. Weller, C. M. Kellett, and A. T. Murray, “Residential Researcher with the Centre for Process Systems
load and rooftop PV generation: An australian distribution network Engineering, Imperial College London, London,
dataset,” Int. J. Sustain. Energy, vol. 36, no. 8, pp. 787–806, Sep. 2017. U.K. From 2013 to 2014, he was a Research Fac-
[33] A. Bestler and K. Graichen, “Distributed model predictive control ulty Member with Shanghai Jiao Tong University,
for continuous-time nonlinear systems based on suboptimal ADMM,” Shanghai, China. He has held visiting professor
Optim. Control Appl. Methods, vol. 40, no. 1, pp. 1–23, Jan. 2019. positions at the Freiburg Institute for Advanced Studies, Freiburg im Breisgau,
Germany, and the Institute for Microsystems Engineering, and University of
Freiburg, Freiburg im Breisgau, in 2014, and various shorter academic visiting
appointments, for example, at the University of California at Berkeley (UC
Berkeley), Berkeley, CA, USA, in winter 2017, and Imperial College London
Yuning Jiang (Member, IEEE) received the B.Sc. in summer 2018. He is currently an Associate Professor with the School of
degree in electrical engineering from Shandong Uni- Information Science and Technology, ShanghaiTech University, Shanghai. His
versity, Jinan, China, in 2014, and the Ph.D. degree research interests include numerical optimization and optimal control, robust
in information engineering from ShanghaiTech Uni- and global optimization, and fast model predictive control algorithms.
versity, Shanghai, China, and the University of Chi-
nese Academy of Sciences, Beijing, China, in 2020.
He has ever been a Visiting Scholar with the
University of California at Berkeley (UC Berkeley),
Berkeley, CA, USA, the University of Freiburg,
Freiburg im Breisgau, Germany, and Technische
Universität Ilmenau (TU Ilmenau), Ilmenau, Ger-
many, during his Ph.D. study. He is currently a Post-Doctoral Researcher with
the Laboratoire d’Automatique, École Polytechnique Fédérale de Lausanne
(EPFL), Lausanne, Switzerland. His research interests include distributed
optimization and model predictive control, particularly for power and energy
systems, intelligent transportation, stochastic control, and fast model predictive
control (MPC) algorithms. Karl Worthmann received the Diploma and Ph.D.
degrees in mathematics from the University of
Bayreuth, Bayreuth, Germany, in 2006 and 2012,
respectively.
In 2014, he was appointed as an Assistant Pro-
Philipp Sauerteig received the B.Sc. and M.Sc. fessor of differential equations at Technische Uni-
degrees in mathematics from Technische Universität versität Ilmenau, Ilmenau, Germany, where he was
Ilmenau, Ilmenau, Germany, in 2016 and 2017, promoted to Full Professor in 2019 after receiving
respectively, where he is currently pursuing the the Heisenberg-professorship “Optimization-based
Ph.D. degree with the Chair for Optimization-Based Control” by the German Research Foundation. His
Control. current research interests include control theory with
In 2019, he was with the School of Informa- a particular focus on nonlinear model predictive control, stability analysis, and
tion Science and Technology (SIST), ShanghaiTech sampled-data systems.
University, Shanghai, China, for a research stay. Dr. Worthmann has been appointed as a Junior Fellow of the Society of
He currently works as a Research Assistant at his Applied Mathematics and Mechanics (GAMM) in 2013, where he served
alma mater on distributed optimization and control as a speaker from 2014 to 2015. He was a recipient of the Ph.D. Award
of microgrids within the project “CONSENS: Consistent Optimization and from the City of Bayreuth and the stipend of the German National Academic
Stabilization of Electrical Networked Systems” funded by the Federal Ministry Foundation. In 2016, he received the Best Paper Award “Mathematics and
of Education and Research. His interests are distributed optimization, model Natural Sciences” of Technische Universität Ilmenau. He is also the Chairman
predictive control, multiobjective optimization, and surrogate models, such as “Mathematical Systems Theory” of the Interdisciplinary GAMM Activity
artificial neural networks. Group “Dynamics and Control.”

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on September 12,2021 at 19:46:13 UTC from IEEE Xplore. Restrictions apply.