Copyright (c) 2013 IEEE. Personal use is permitted.

For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
1
A distributed control strategy for reactive
power compensation in smart microgrids
Saverio Bolognani and Sandro Zampieri
AbstractWe consider the problem of optimal reactive power
compensation for the minimization of power distribution losses
in a smart microgrid. We rst propose an approximate model
for the power distribution network, which allows us to cast the
problem into the class of convex quadratic, linearly constrained,
optimization problems. We then consider the specic problem
of commanding the microgenerators connected to a microgrid,
in order to achieve the optimal injection of reactive power.
For this task, we design a randomized, leader-less, gossip-like
optimization algorithm. We show how a distributed approach
is possible, where microgenerators need to have only a partial
knowledge of the problem parameters and of the state, and can
perform only local measurements. For the proposed algorithm,
we provide conditions for convergence together with an analytic
characterization of the convergence speed. The analysis shows
that, in radial networks, the best performance is achieved
when we command cooperation among microgenerators that are
neighbors in the electric topology. Numerical simulations are
included to validate both the proposed model and the analytic
results about the performance of the proposed algorithm.
I. INTRODUCTION
M
OST of the distributed optimization methods focus on
the problem of how to dispatch parts of a large scale
optimization algorithm to different processing units [1]. When
the same methods are applied to networked control systems
[2], however, different issues arise. The way in which decision
variables are assigned to different agents is not part of the
designer degrees of freedom. Moreover, each agent has a
local and limited knowledge of the problem parameters and
of the system state. Finally, the information exchange between
agents can occur not only via a given communication channel,
but also via local actuation and subsequent measurement
performed on an underlying physical system. The extent of
these issues depends on the particular application. In this work
we present a specic scenario, belonging to the motivating
framework of smart electrical power distribution networks [3],
[4], in which these features play a central role.
In the last decade, a number of technological advances,
together with economical and environmental reasons, has been
driving the introduction of distributed microgeneration (solar
panels, combined heat-and-power plants, micro wind turbines,
etc.) in the electric power distribution network. This fact,
together with an increased demand and the need for higher
quality of service, motivated the design and the creation
of smart microgrids, i.e. portions of the low or medium
voltage power distribution network that are provided with a
large amount of information and communication technologies
S. Bolognani is with the Laboratory on Information and Decision Sys-
tems, Massachusetts Institute of Technology, Cambridge (MA), USA. Email:
saverio@mit.edu.
S. Zampieri is with the Department of Information Engineering, University
of Padova, Padova (PD), Italy. Email zampi@dei.unipd.it
(ICT), host a number of microgenerators, and are managed
autonomously from the rest of the network, in order to achieve
better quality of service, improve efciency, and pursue spe-
cic economic interests.
Microgenerators are connected to the microgrid via elec-
tronic interfaces (inverters), whose main task is to enable
the injection of the produced power [5], [6]. However, these
devices can also perform different other tasks, denoted as
ancillary services [7][9]: reactive power compensation, har-
monic compensation, voltage support.
In this work we consider the problem of optimal reactive
power compensation. Loads belonging to the microgrid may
require a sinusoidal current which is not in phase with voltage.
A convenient description for this, consists in saying that they
demand reactive power together with active power, associated
with out-of-phase and in-phase components of the current,
respectively. Reactive power is not a real physical power,
meaning that there is no energy conversion involved nor fuel
costs to produce it. Like active power ows, reactive power
ows contribute to power losses on the lines, cause voltage
drop, and may lead to grid instability. It is therefore preferable
to minimize reactive power ows by producing it as close as
possible to the users that need it. The electronic interfaces
of the microgenerators can be used for this task. Indeed,
the inverters of these units are generally oversized, because
most of the distributed energy sources are intermittent in time,
and the electronic interface is designed according to the peak
power production. When they are not working at the rated
power, these inverters can be commanded to inject a desired
amount of reactive power at no cost [10].
This idea has been recently investigated in the literature
on power systems [11][13]. However, these works consider
a centralized scenario in which the parameters of the entire
power grid are known, the controller can access the entire sys-
tem state, the microgenerators are in small number, and they
receive reactive power commands from a central processing
unit. In [14], a hierarchical and secure architecture has been
proposed to support such dispatching. The contribution of this
work consists in casting the same problem in the framework of
networked control systems and distributed optimization. This
approach allows the design of algorithms and solutions which
can guarantee scalability, robustness to insertion and removal
of the units, and compliance with the actual communication
capabilities of the devices. Up to now, the few attempts to
apply these tools to the power distribution networks have
focused on grids comprising a large number of mechanical
synchronous generators, instead of power inverters. See for
example the stability analysis for these systems in [15] and
the decentralized control synthesis in [16]. Seminal attempts to
distribute optimal reactive power compensation algorithms in a
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
2
large-scale power network, consisted in dividing the grid into
separate regions, each one provided with a supervisor [17],
[18]. Each of the supervisors can access all the regional mea-
surements and data (including load demands) and can solve
the corresponding small-scale optimization problem, enforcing
consistency of the solution with the neighbor regions. Simi-
larly, in the strategy recently proposed by [19], the problem of
optimal reactive power compensation is cast as an optimization
program, which is then distributed across all the microgrid
agents (including loads) assuming that all the reactive power
demands are known. Instead, preliminary attempts of design-
ing distributed reactive power compensation strategies that
only require the participation of the microgenerators, have
been performed only very recently. The available results in
this sense are mainly supported by heuristic approaches [20],
or they follow suboptimal criteria for sharing an aggregate
reactive power command [21], and in some cases do not
implement any communication or synergistic behavior of the
microgenerators [22]. In [23], a multi-agent-system approach
has been employed for commanding reactive power injection
in order to support the microgrid voltage prole.
The contribution of this paper is twofold. On one side,
we propose a rigorous analytic derivation of an approximate
model of the power ows. The proposed model can be
considered as a generalization of the DC power ow model
commonly adopted in the power system literature (see for ex-
ample [24, Chapter 3] and references therein). Via this model,
the optimal reactive power ow (ORPF) problem is cast into a
quadratic optimization (Section III). The second contribution
is to propose and to analyze a leader-less and asynchronous
strategy to command the reactive power injection of each
microgeneration unit, capable of optimizing reactive power
ows across the microgrid (Section IV). In Section V we
characterize convergence of the proposed algorithm to the
global optimal solution, and we study its performance. In
Section VI we show how the performance of this algorithm
can be optimized by a proper choice of the communication
strategy. In Section VII, we nally validate both the proposed
model and the proposed algorithm via simulations.
II. MATHEMATICAL PRELIMINARIES AND NOTATION
Let ( = (1, c, , ) be a directed graph, where 1 is the
set of nodes, c is the set of edges, and , : c 1 are
two functions such that edge e c goes from the source
node (e) to the terminal node (e). Two edges e and e

are
consecutive if (e), (e) (e

), (e

) is not empty. A
path is a sequence of consecutive distinct edges.
In the rest of the paper we will often introduce complex-
valued functions dened on the nodes and on the edges. These
functions will also be intended as vectors in C
n
(where n =
[1[) and C
|E|
. Given a vector u, we denote by u its (element-
wise) complex conjugate, and by u
T
its transpose.
Let moreover A 0, 1
|E|n
be the incidence matrix of
the graph (, dened via its elements
[A]
ev
=
_
_
_
1 if v = (e)
1 if v = (e)
0 otherwise.
If J is a subset of nodes, we dene by 1
W
the column
vector whose elements are
[1
W
]
v
_
1 if v J
0 otherwise.
Similarly, if w is a node, we denote by 1
w
the column vector
whose value is 1 in position w, and 0 elsewhere, and we denote
by 1 the column vector of all ones. If the graph ( is connected
(i.e. for every pair of nodes there is a path connecting them),
then 1 is the only vector in the null space ker A.
An undirected graph ( is a graph in which for every edge
e c, there exists an edge e

c such that (e

) = (e) and
(e

) = (e). In case the graph has no multiple edges and no

self loops, we can also describe the edges of an undirected
graph as subsets h

, h

1 of cardinality 2. Similarly, we
dene a hypergraph 1 as a pair (1, c) in which each edge is
a subset h

, h

, . . . of 1, of arbitrary cardinality [25].

III. PROBLEM FORMULATION
A. Model of a microgrid
We dene a smart microgrid as a portion of the power
distribution network that is connected to the power trans-
mission network in one point, and hosts a number of loads
and microgenerators, as described for example in [5], [6] (see
Figure 1, lower panel). For the purpose of this paper, we model
a microgrid as a directed graph (, in which edges represent the
power lines, and nodes represent the devices that are connected
to the microgrid (see Figure 1, middle panel). These include
loads, microgenerators, and also the point of connection of the
microgrid to the transmission grid (called point of common
coupling, or PCC).
We limit our study to the steady state behavior of the system,
when all voltages and currents are sinusoidal signals at the
same frequency. Each signal can therefore be represented via
a complex number y = [y[e
jy
whose absolute value [y[
corresponds to the signal root-mean-square value, and whose
phase y corresponds to the phase of the signal with respect
to an arbitrary global reference.
In this notation, the steady state of a microgrid is described
by the following system variables (see Figure 1, lower panel):
u C
n
, where u
v
is the grid voltage at node v;
i C
n
, where i
v
is the current injected by node v;
C
|E|
, where
e
is the current owing on the edge e.
The following constraints are satised by u, i and
A
T
+ i = 0, (1)
Au +Z = 0, (2)
where A is the incidence matrix of (, and Z = diag(z
e
, e
c) is the diagonal matrix of line impedances, z
e
being the
impedance of the microgrid power line corresponding to the
edge e. Equation (1) corresponds to Kirchhoffs current law
(KCL) at the nodes, while (2) describes the voltage drop on
the edges of the graph.
Each node v of the microgrid is also characterized by a law
relating its injected current i
v
with its voltage u
v
. We model
the PCC (which we assume to be node 0) as an ideal sinusoidal
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
3
E
l
e
c
t
r
i
c
n
e
t
w
o
r
k
G
r
a
p
h
m
o
d
e
l
C
o
n
t
r
o
l
u
v
i
v

e
z
e
v
(e) (e)
e
0
C
i
C
j
C
k
Figure 1. Schematic representation of the adopted microgrid model. In
the lower panel a circuit representation is given, where black diamonds are
microgenerators, white diamonds are loads, and the left-most element of the
circuit represents the PCC. The middle panel illustrates the adopted graph
representation for the same microgrid. Circled nodes represent compensators
(i.e. microgenerators and the PCC). The upper panel shows how the com-
pensators can be divided into overlapping clusters in order to implement the
control algorithm proposed in Section IV. Each cluster is provided with a
supervisor with some computational capability.
voltage generator at the microgrid nominal voltage U
N
with
arbitrary, but xed, angle
u
0
= U
N
e
j
. (3)
We model loads and microgenerators (that is, every node v
of the microgrid except the PCC) via the following law relating
the voltage u
v
and the current i
v
u
v

i
v
= s
v

u
v
U
N

v
, v 10, (4)
where s
v
is the nominal complex power and
v
is a char-
acteristic parameter of the node v. The model (4) is called
exponential model [26] and is widely adopted in the literature
on power ow analysis [27]. Notice that s
v
is the complex
power that the node would inject into the grid, if the voltage
at its point of connection were the nominal voltage U
N
. The
quantities
p
v
:= Re(s
v
) and q
v
:= Im(s
v
)
are denoted as active and reactive powers, respectively. The
nominal complex powers s
v
corresponding to microgrid loads
are such that p
v
< 0, meaning that positive active power
is supplied to the devices. The nominal complex powers
corresponding to microgenerators, on the other hand, are such
that p
v
0, as positive active power is injected into the
grid. The parameter
v
depends on the particular device.
For example, constant power, constant current, and constant
impedance devices are described by
v
= 0, 1, 2, respectively.
In this sense, this model is also a generalization of ZIP
models [26], which also are very common in the power system
literature. Microgenerators t in this model with
v
= 0, as
they generally are commanded via a complex power reference
and they can inject it independently from the voltage at their
point of connection [5], [6].
The task of solving the system of nonlinear equations given
by (1), (2), (3), and (4) to obtain the grid voltages and currents,
given the network parameters and the injected nominal powers
s
v
, v 10 at every node, has been extensively covered
in the literature under the denomination of power ow analysis
(see for example [24, Chapter 3]). In the following, we derive
an approximate model for the microgrid state, which will be
used later for the setup of the optimization problem and for
the derivation of the proposed distributed algorithm. To do so,
a couple of technical lemmas are needed.
Lemma 1. Let L be the complex valued Laplacian L :=
A
T
Z
1
A (also denoted at bus admittance matrix). There
exists a unique symmetric matrix X C
nn
such that
_
XL = I 11
T
0
X1
0
= 0.
(5)
Proof: Let us rst prove the existence of X. As ker L =
Image 1 = ker(I 11
T
0
), there exists X

C
nn
such that
X

L = (I 11
T
0
). Let X = X

(I 1
0
1
T
). Then
XL = X

(I 1
0
1
T
)L = X

L = I 11
T
0
,
X1
0
= X

(I 1
0
1
T
)1
0
= 0.
Existence is then guaranteed. To prove uniqueness, notice that
_
X 1
1
T
0
_ _
L 1
0
1
T
0
0
_
=
_
XL+11
T
0
X1
0
1
T
L 1
T
1
0
_
=
_
I 0
0 1
_
.
Therefore
_
X 1
1
T
0
_
=
_
L 1
0
1
T
0
0
_
1
,
and uniqueness of X follows from the uniqueness of the
inverse. Moreover, as L = L
T
, we have
_
X
T
1
1
T
0
_
=
_
L 1
0
1
T
0
0
_
T
=
_
L
T
1
0
1
T
0
0
_
1
=
_
X 1
1
T
0
_
and therefore X = X
T
.
The matrix X depends only on the topology of the micro-
grid power lines and on their impedance (compare it with the
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
4
denition of Green matrix in [28]). Indeed, it can be shown
that, for every pair of nodes (v, w),
(1
v
1
w
)
T
X(1
v
1
w
) = Z
eff
vw
, (6)
where Z
eff
vw
represents the effective impedance of the power
lines between node v and w.
All the currents i and the voltages u of the microgrid are
therefore determined by the equations
_

_
u = Xi + U
N
e
j
1
1
T
i = 0
u
v

i
v
= s
v

u
v
U
N

v
, v 10,
(7)
where the rst equation results from (1), (2), and (3) together
with Lemma 1, while the second equation descends from (1),
using the fact that A1 = 0 in a connected graph.
We can see the currents i and the voltages u as functions
i(U
N
), u(U
N
) of U
N
. The following proposition provides the
Taylor approximation of i(U
N
) and u(U
N
) for large U
N
.
Proposition 1. Let s be the vector of all nominal complex
powers s
v
, including
s
0
:=

vV\{0}
s
v
. (8)
Then for all v 1 we have that
i
v
(U
N
) = e
j
_
s
v
U
N
+
c
v
(U
N
)
U
2
N
_
u
v
(U
N
) = e
j
_
U
N
+
[X s]
v
U
N
+
d
v
(U
N
)
U
2
N
_ (9)
for some complex valued functions c
v
(U
N
) and d
v
(U
N
) which
are O(1) as U
N
, i.e. they are bounded functions for
large values of the nominal voltage U
N
.
The proof of this proposition is based on elementary mul-
tivariable analysis, but requires a quite involved notation. For
this reason it is given in Appendix A. The quality of the
approximation proposed in the previous proposition relies on
having large nominal voltage U
N
and relatively small currents
injected by the inverters (or supplied to the loads). This
assumption is veried in practice and corresponds to correct
design and operation of power distribution networks, where
indeed the nominal voltage is chosen sufciently large (subject
to other functional constraints) in order to deliver electric
power to the loads with relatively small power losses on the
power lines. Numerical simulation in Section VII will indeed
show that the approximation is extremely good when power
distribution networks are operated in their regular regime.
Remark. Notice that this approximation, in similar fashions,
has been used in the literature before in order to estimate
power ows on the power lines (see among the others [22],
[29] and references therein). It also shares some similarities
with the DC power ow model [24, Chapter 3], extending it
to the lossy case (in which lines are not purely inductive).
The analytic rigorous justication proposed here allows us to
estimate the approximation error, and, more importantly, will
provide the tools to understand what information on the system
state can be gathered by properly sensing the microgrid.
B. Power losses minimization problem
Similarly to what has been done for example in [30], we
choose active power losses on the power lines as a metric
for optimality of reactive power ows. The total active power
losses on the edges are given by
J
tot
:=

eE
[
e
[
2
Re(z
e
) =

i
T
Re(X)i, (10)
where we used (2) and (7), together with the properties (5) of
X, to express as a function of i.
In the scenario that we are considering, we are allowed to
command only a subset ( 1 of the nodes of the microgrid
(namely the microgenerators, also called compensators in this
framework). We denote by m := [([ its cardinality. Moreover,
we assume that for these compensators we are only allowed to
command the amount of reactive power q
v
= Im(s
v
) injected
into the grid, as the decision on the amount of active power
follows imperative economic criteria. For example, in the case
of renewable energy sources, any available active power is
typically injected into the grid to replace generation from
traditional plants, which are more expensive and exhibit a
worse environmental impact [10].
The resulting optimization problem is then
min
qv, vC
J
tot
, (11)
where J
tot
is a function of the decision variables q
v
, v (,
via (10) and via the implicit system of nonlinear equations (7).
In the usual formulations of the optimal reactive power com-
pensation problem, some constraints might be present. Typical
operational requirements constrain the voltage amplitude at
every node to stay inside a given range centered around the
microgrid nominal voltage
[u
v
[ [U
N
U, U
N
+ U] , v 1. (12)
Moreover, the power inverters that equip each microgenerator,
due to thermal limits, can only provide a limited amount of
reactive power. This limit depends on the size of the inverter
but also on the concurrent production of active power by the
same device [22]. These constraints can be quite tight, and
correspond to a set of box constraints on the decision variables,
q
v

_
q
min
v
, q
max
v

, v (0. (13)
Both these constraints have been relaxed for the analysis pre-
sented in this paper, and for the subsequent design of a control
strategy. This choice allowed to conduct an analytic study
of the performance of the proposed solution. The algorithm
proposed in this paper can be specialized in different ways,
in order to include such limits. The modied version of the
algorithm proposed in [31], for example, guarantees that the
hard constraints (13) are satised at every time, at the cost of
a slightly more complex communication protocol. In [32], on
the other hand, the voltage limits (12) have been incorporated
as soft constraints, via a dual decomposition of the problem.
In the following, we show how the approximated model
proposed in the previous section can be used to tackle the
optimization problem (11) and to design an algorithm for its
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
5
solution. By plugging the approximate system state (9) into
(10), we have
J
tot
=
1
U
2
N
s
T
Re(X)s +
1
U
3
N

J(U
N
, s)
=
1
U
2
N
p
T
Re(X)p +
1
U
2
N
q
T
Re(X)q +
1
U
3
N

J(U
N
, s)
where p = Re(s), q = Im(s), and where

J(U
N
, s) := 2 Re
_
s
T
Re(X)c(U
N
)

+
1
U
N
c
T
(U
N
) Re(X)c(U
N
)
is bounded as U
N
tends to innity. The term
1
U
3
N

J(U
N
, s) can
thus be neglected if U
N
is large.
Therefore, via the approximate model proposed in Proposi-
tion 1, we have been able to
approximate power losses as a quadratic function of the
injected power;
decouple the problem of optimal power ows into the
problem of optimal active and reactive power injection.
The problem of optimal reactive power injection at the com-
pensators can therefore be expressed as a quadratic, linearly
constrained problem, in the form
min
qv,vC
J(q), where J(q) =
1
2
q
T
Re(X)q,
subject to 1
T
q = 0
(14)
where the other components of q, namely q
v
, v 1(,
are the nominal amounts of reactive power injected by the
nodes that cannot be commanded, and the constraint 1
T
q = 0
directly descends from (8).
The solution of the optimization problem (14) would not
pose any challenge if a centralized solver knew the problem
parameters X and q
v
, v 1(. These quantities depend
both on the grid parameters and on the reactive power demand
of all the microgrid loads. In the case of a large scale system,
collecting all this information at a central location would
be unpractical for many reasons, including communication
constraints, delays, reduced robustness, and privacy concerns.
IV. A DISTRIBUTED ALGORITHM FOR REACTIVE POWER
DISPATCHING
In this section, we present an algorithm that allows the com-
pensators to decide on the amount of reactive power that each
of them have to inject in the microgrid in order to minimize the
power distribution losses. The proposed approach is based on a
leader-less strategy, meaning that it can be implemented by the
compensators without the supervision of any central controller.
In order to design such a strategy, the optimization problem
is decomposed into smaller, tractable subproblems that are
assigned to small groups of compensators (possibly even pairs
of them). We will show that the compensators can solve their
corresponding subproblems via local measurements, a local
knowledge of the grid topology, and a limited data processing
and communication. Moreover, they do not need to know
the demands of the loads, which then are not required to
have any communication or computation capability. We will
nally show that the repeated and asynchronous solution of
these subproblems yields the solution of the original global
optimization problem.
It is worth remarking that the decomposition methods pro-
posed in most of the literature on distributed optimization, for
example in [33], cannot be applied to this problem because the
cost function in (14) is not separable into a sum of individual
terms for each agent. Approaching the decomposition of this
optimization problem via its dual formulation (as proposed
in many works, including the recent [34]) is also unlikely
to succeed, as the feasibility of the system state must be
guaranteed at any time during the optimization process.
A. Optimization problem decomposition
Let the compensators be divided into possibly overlapping
subsets (
1
, . . . , (

with

r=1
(
r
= (. Nodes belonging to
the same subset (or cluster) are able to communicate each
other, and they are therefore capable of coordinating their
actions and sharing their measurements. Each cluster is also
provided with some computational capabilities for processing
the collected data. This processing is performed by a cluster
supervisor, capable of collecting all the necessary data from
the compensators belonging to the cluster and to return the
result of the data processing to them (see Figure 1, upper
panel). The cluster supervisor can possibly be one of the
compensators in the cluster. An alternative implementation
consists in providing all the compensators in the cluster with
identical instances of the same instructions. In this case, after
sharing the necessary data, they will perform exactly the same
data processing, and no supervising unit is needed.
The supervisor of each cluster (
r
activates at some (possibly
uneven) time instants, decided according to some individual
policy, and performs the following operations:
1) it gathers state information (the injected reactive powers
q
k
, k (
r
) and measurements from the compensators
belonging to the cluster (
r
;
2) based on these data, it determines the new feasible
state that minimizes the global cost J(q), while all the
compensators not in (
r
keep their state constant;
3) it dispatches the new state to the compensators in (
r
,
which update their states q
k
, k (
r
by injecting the new
reactive powers computed in the previous step.
In the following, we provide the necessary tools for imple-
menting these steps. In particular, we show how the supervisor
of a cluster (
r
can compute the new state that minimizes
the total power distribution losses, based only on the local
knowledge of the electrical network topology and on the
measurements performed by the the compensators in (
r
.
Consider the optimization (14), and let us distinguish the
controllable and the uncontrollable components of the vector q.
Assume with no loss of generality that the rst m components
of q are controllable (i.e. they describe the reactive power
injected by the compensators) and that the remaining n m
are not controllable. The state q is thus partitioned as
q =
_
q
C
q
V\C
_
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
6
where q
C
R
m
and q
V\C
R
nm
. According to this partition
of q, let us also partition the matrix Re(X) as
Re(X) =
_
M N
N
T
Q
_
. (15)
Let us also introduce some convenient notation. Consider
the subspaces
o
r
:=
_
_
_
q
C
R
m
:

jCr
[q
C
]
j
= 0 , [q
C
]
j
= 0 j , (
r
_
_
_
,
and the mm matrices
:=
1
2m

h,kC
(1
h
1
k
)(1
h
1
k
)
T
= I
1
m
11
T
,

r
:=
1
2[(
r
[

h,kCr
(1
h
1
k
)(1
h
1
k
)
T
= diag(1
Cr
)
1
[(
r
[
1
Cr
1
T
Cr
,
where the column vectors 1, 1
h
, 1
k
, 1
Cr
have dimension m,
and diag(1
Cr
) is the mm diagonal matrix whose diagonal
is the vector 1
Cr
. Notice that o
r
= Image
r
,
r
1 = 0, and
that
r
is a projection matrix, i.e.
2
r
=
r
.
If the system is in the state q =
_
q
C
q
V\C

and cluster (
r
is
activated, the corresponding cluster supervisor has to solve the
optimization problem
q
opt, r
C
:= arg min
q

C
J
__
q

C
q
V\C
__
subject to q

C
q
C
o
r
.
(16)
Using the standard formulas for quadratic optimization it can
be shown that
q
opt, r
C
= q
C
(
r
M
r
)

J, (17)
where means pseudoinverse and where
J = Mq
C
+ Nq
V\C
(18)
is the gradient of J
__
q
C
q
V\C
_
with respect to the decision
variables q
C
.
Given the denition of the projectors
r
and the properties
of the pseudoinverse, the following properties hold.
i) The matrices
r
,
r
M
r
, and (
r
M
r
)

have entries
different from zero only in position h, k with h, k (
r
.
ii) The matrices (
r
M
r
)

satisfy
ker(
r
M
r
)

= ker
r
(19)
and
Image(
r
M
r
)

= Image
r
. (20)
iii) The matrices (
r
M
r
)

satisfy
(
r
M
r
)

= (
r
M
r
)

r
.
By using property i) we have
_
q
opt, r
C

h
=
_
q
h

kCr
_
(
r
M
r
)

_
hk
[J]
k
if h (
r
q
h
if h / (
r
.
(21)
We now show how the supervisor of (
r
can perform an
approximate computation of
_
q
opt, r
C

h
, h (
r
, based on the
information available within (
r
.
B. Hessian reconstruction from local topology information
Compensators belonging to the cluster (
r
can infer some
information about the Hessian M. More precisely they can
determine the elements of the matrix (
r
M
r
)

that appear
in (21). Dene R
eff
as a m m matrix with entry in
h, k ( equal to Re
_
Z
eff
hk
_
, where Z
eff
hk
are the mutual
effective impedances between pairs compensators h, k. From
(6) and (15) we have that
[R
eff
]
hk
= Re
_
Z
eff
hk
_
= (1
h
1
k
)
T
M(1
h
1
k
) (22)
and so we can write
R
eff
=

h,kC
1
h
(1
h
1
k
)
T
M(1
h
1
k
)1
T
k
= diag(M)11
T
+11
T
diag(M) 2M
where diag(M) is the m m diagonal matrix having the
same diagonal elements of M. Consequently, since
r
1 = 0,
we have that
r
R
eff

r
= 2
r
M
r
and then

r
M
r
=
1
2

r
R
eff

r
. (23)
Because of the specic sparseness of the matrices
r
, we can
then argue that the elements of the matrix
r
M
r
can be com-
puted by the supervisor of cluster (
r
from the mutual effective
impedances Z
eff
hk
between compensators h, k both belonging to
the cluster (
r
. These impedances are assumed to be known by
the cluster supervisor. This is a reasonable hypothesis, since
the mutual effective impedances can be obtained for example
via online estimation procedures as in [35]. Moreover, in the
common case in which the low-voltage power distribution grid
is radial, the mutual effective impedances Z
eff
hk
correspond to
the impedance of the only electric path connecting node h to
node k, and can therefore be inferred from a priori knowledge
of the local microgrid topology, or via ranging technologies
over power line communications as suggested in [36].
Then the cluster supervisor needs to compute the pseudo-
inverse of
r
M
r
. Notice that all these operations have to
be executed ofine once, as these coefcients depend only on
the grid topology and impedances, and thus are not subject to
change.
C. Gradient estimation via local voltage measurement
Assume that nodes in (
r
can measure the grid voltage u
k
,
k (
r
, at their point of connection. In practice, this can
be done via phasor measurement units that return both the
amplitudes [u
k
[ and phases u
k
of the measured voltages (see
[37], [38]). Let us moreover assume the following.
Assumption 1. All power lines in the microgrid have the same
inductance/resistance ratio, i.e.
Z = e
j
Z
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
7
where Z is a diagonal real-valued matrix, whose elements are
[Z]
ee
= [z
e
[. Consequently, L = e
j
A
T
Z
1
A, and X :=
e
j
X is a real-valued matrix.
This assumption seems reasonable in most practical cases
(see for example the IEEE standard testbeds [39]). The effects
of this approximation will be commented in Section VII.
Similarly to the denition of q
C
, we dene by u
C
the vector
in C
m
containing the components u
k
with k (, i.e. the
voltages that can be measured by the compensators. Consider
now the maps K
r
: C
m
R
m
, r = 1, . . . , , dened
component-wise as follows
[K
r
(u
C
)]
k
:=
_
1
|Cr|

vCr
[u
v
[[u
k
[ sin(u
v
u
k
) if k (
r
0 if k , (
r
.
(24)
Notice rst that K
r
(u
C
) can be computed locally by the
supervisor of the cluster r, from the voltage measurements of
the compensators belonging to (
r
. The following result shows
how the function K
r
(u
C
) is related with the gradient of the
cost function, introduced in (18).
Proposition 2. Let K
r
(u
C
) be dened as in (24). Then
the elements [J]
k
, k (
r
, of the gradient (18) can be
decomposed as
[J]
k
= cos [K
r
(u
C
)]
k
+
r
+
1
U
N
_

K
r
(u
C
)
_
k
, (25)
where
_

K
r
(u
C
)
_
k
is bounded as U
N
, and
r
is a
constant independent of the compensator index k.
Proof: It can be seen that, for k (
r
,
[K
r
(u
C
)]
k
= Im
_
e
j
1
[(
r
[
_
1
T
Cr
u
_
u
k
_
.
By using (9), and via a series of simple but rather tedious
computation it can be seen that (25) holds with

r
=
cos
[(
r
[
Im
_
e
j2
_
1
T
Cr
Xs
_
+
cos
[(
r
[U
N
Im
_
e
j
_
1
T
Cr

d
_
sin cos U
2
N
and
_

K
r
(u
C
)
_
k
= cos Im
_
e
j
d
k

+
cos
U
N
[(
r
[
Im
_
e
j
_
1
T
Cr
Xs
_
[X s]
k

+
cos
U
2
N
[(
r
[
Im
_
e
j2
_
1
T
Cr
Xs
_
d
k

+
cos
U
2
N
[(
r
[
Im
__
1
T
Cr

d
_
[X s]
k

+
cos
U
3
N
[(
r
[
Im
_
e
j
_
1
T
Cr

d
_
d
k

,
which according to Proposition 1 is bounded when U
N
.
D. Description of the algorithm
By applying Proposition 2 to expression (21) for q
opt, r
C
we
obtain that, for h (
r
,
_
q
opt, r
C

h
= q
h
cos

kCr
_
(
r
M
r
)

_
hk
[K
r
(u
C
)]
k

cos
U
N

kCr
_
(
r
M
r
)

_
hk
_

K
r
(u
C
)
_
k
,
(26)
where the term
r
is canceled because of the specic kernel
of (
r
M
r
)

. Because
1
U
N

K
r
is innitesimal as U
N
,
then equation (26), together with identity (23), suggests the
following approximation for q
opt, r
C
_

q
opt, r
C
_
h
= q
h
+ 2 cos

kCr
_
_

r
R
eff

r
_

_
hk
[K
r
(u
C
)]
k
(27)
which can be computed for all h (
r
by the supervisor of
cluster r from the voltage measurements performed by the
compensators belonging to (
r
and from the mutual effective
impedances between them. Notice also that, via (20) and
(23), we have that

h
_
q
opt, r
C

h
q
h
= 0, and therefore the
constraint 1
T
q = 0 remains satised.
Based on this result, we propose the following iterative
algorithm, whose online part is also depicted in Figure 2 in
its block diagram representation.
OFFLINE PROCEDURE
The supervisor of each cluster (
r
gathers the system
parameter [R
eff
]
hk
, h, k (
r
;
each supervisor computes the elements
_
_

r
R
eff

r
_

_
hk
,
h, k (
r
.
ONLINE PROCEDURE
Every time a cluster (
r
activates, the following operation are
performed:
each compensator h (
r
measures its voltage u
h
;
each compensator h (
r
sends its voltage u
h
and its
state q
h
(the amount of reactive power it is injecting into
the grid) to the cluster supervisor;
the cluster supervisor computes K
r
(u
C
), via (24);
the cluster supervisor computes
_

q
opt, r
C
_
h
for all h (
r
via (27);
each compensator h (
r
receives the value
_

q
opt, r
C
_
h
from the cluster supervisor and updates its injection of
reactive power into the grid (i.e. q
h
) to that value.
Remark. It is important to notice that the distributed imple-
mentation of the algorithm is only possible because of the
specic physical system that we are considering. By sensing
the system locally, nodes can infer some global information
on the system state (namely, the gradient of the cost function)
that otherwise would depend on the states of all other nodes,
including loads. For this reason, it is necessary that, after
each iteration of the algorithm, the microgenerators actuate
the system by implementing the optimization step. It is also
necessary that the inter-time between consecutive iterations
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
8
Power
distribution
network
q
C
u
C
q
V\C
p
2 cos (
r
R
eff

r
)

u
V\C
K
r
(u
C
)
1
z 1
Discrete time
integrator
Figure 2. Block diagram representation of the proposed distributed algorithm.
The repeated solution of different optimization subproblems corresponds to
a switching discrete-time feedback system. Double arrows represent vector
signals. The power distribution network block receives, as inputs, the active
and reactive power injection of both compensators (nodes in C) and loads
(nodes in V\C). It gives, as outputs, the voltages at the nodes. The voltages
at the compensators (u
C
) can be measured, and provide the feedback signal for
the control laws. Notice that the control law corresponding to a specic cluster
r requires only the voltage measurements performed by the compensators
belonging to the cluster (u
k
, k Cr), and it updates only the amounts of
reactive power injected by the same compensators (q
k
, k Cr), while the
other compensators hold their reactive power injection constant.
is longer than the typical transient time of the system (in-
dicatively few cycles at the nominal frequency), in order for
the grid to settle at its steady state. In this way, subsequent
measurements will reect the state change and will provide
the necessary information for the next steps of the algorithm.
Such time separation between the system response and the
algorithm execution guarantees also that the stability of the
grid is not affected by the proposed closed loop control. This
approach resembles some other applications of distributed
optimization, e.g. the radio transmission power control in
wireless networks [40] and the trafc congestion protocol
for wired data networks [41], [42]. Also in these cases,
the iterative tuning of the decision variables (radio power,
transmission rate) depends on congestion feedback signals
that are function of the entire state of the system. Every
time the system is actuated (i.e. the decision variables are
updated), these signals can be detected locally by each agent
by measuring error rates, signal-to-noise ratios, or specic
feedback signals in the protocol.
V. CONVERGENCE OF THE ALGORITHM
In this section we will give the conditions ensuring that the
proposed iterative algorithm converges to the optimal solution,
and we will analyze its convergence rate.
Let T
t
, t = 1, 2, . . ., be the time instants in which cluster
supervisors activate, and let r(t) be the index of the cluster
supervisor that activates at time T
t
. According to (27), the
dynamics of the algorithm is then described by the iteration
q
C
(t + 1) = q
C
(t) cos (
r(t)
M
r(t)
)

K
r(t)
(u
C
(t)). (28)
where the voltages u
C
(t) are a function of the reactive power
q
C
(t), q
V\C
and of the active power p. We will consider instead
the iteration
q
C
(t+1) = q
C
(t)
_

r(t)
M
r(t)
_

(Mq
C
(t)+Nq
V\C
), (29)
which descends from (17) and which differs from (28) for the
term
cos
U
N
_

r(t)
M
r(t)
_

K
r(t)
(u
C
),
as one can verify by inspecting (26). This difference is
innitesimal as U
N
. Therefore in the sequel we will
study the convergence of (29), while numerical simulations in
Section VII will support the proposed approximation, showing
how both the algorithm steady state performance and the rate
of convergence are practically the same for the exact system
and the approximated model.
Remark. A formal proof of the convergence of the discrete
time system (28) could be done based of the mathematical
tools proposed in Sections 9 and 10 of [43] dedicated to the
stability of perturbed systems. However, in our setup there are
a number of complications that make the application of those
techniques not easy. The rst is that, since u
C
(t) is related
to q
C
(t) in an implicit way, then (28) is in fact a nonlinear
implicit (also called descriptor) system with state q
C
(t). The
second difculty comes from the fact that (28) is a nonlinear
randomly time-varying (also called switching) system.
In order to study the system (29), we introduce the auxiliary
variable x = q
C
q
opt
C
R
m
, where q
opt
is the solution of the
optimization problem (14), and q
opt
C
are the elements of q
opt
corresponding to the nodes in (. In this notation, it is possible
to explicitly express the iteration (29) as a discrete time system
in the form
x(t + 1) = F
r(t)
x(t), x(0) ker 1
T
, (30)
where
F
r
:= I E
r
E
r
:= (
r
M
r
)

M.
(31)
Notice that, because of (20), ker 1
T
is a positively invariant
set for (30), namely F
r
(ker 1
T
) ker 1
T
for all r.
It is clear that q(t) converges to the optimal solution q
opt
if and only if x(t) converges to zero for any initial condition
x(0) ker 1
T
. A necessary condition for this to happen is
that there are no nonzero equilibria in the discrete time system
(30), namely that the only x ker 1
T
such that F
r
x = x for
all r must be x = 0.
The following proposition provides a convenient character-
ization of this property, in the form of a connectivity test.
Let 1 be the hypergraph whose nodes are the compensators,
and whose edges are the clusters (
r
, r = 1, . . . , . This
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
9
hypergraph describes the communication network needed for
implementing the algorithm: compensators belonging to the
same edge of the hypergraph (i.e. to the same cluster) must
be able to share their measurements with a supervisor and
receive commands from it.
Proposition 3. Consider the matrices F
r
, r = 1, . . . , ,
dened in (31). The point x = 0 is the only point in ker 1
T
such that F
r
x = x for all r if and only if the hypergraph 1
is connected.
Proof: The proposition is part of the statement of
Lemma 5 in the Appendix B.
In order to prove that the connectivity of the hypergraph
1 is not only a necessary but also a sufcient condition for
the convergence of the algorithm, we introduce the following
assumption on the sequence r(t).
Assumption 2. The sequence r(t) is a sequence of indepen-
dently, identically distributed symbols in 1, . . . , , with non-
zero probabilities
r
> 0, r = 1, . . . , .
Remark. One possible way of satisfying Assumption 2 is
the following. Let each cluster supervisor be provided with
a timer, each one governed by an independent Poisson pro-
cess, which triggers the cluster after exponentially distributed
waiting times. If the time required for the execution of the
algorithm is negligible with respect to the typical waiting time
of the Poisson processes, this strategy allows to construct
the sequence of independent symbols r(t), t = 0, 1, 2, . . .
with non-zero symbol probability
r
for all r, in a purely
decentralized way (see for example [44]).
Let v(t) := E
_
x
T
(t)Mx(t))

= E [J(q(t)) J(q
opt
)] and
consider the following performance metric
R = sup
x(0)ker 1
T
limsupv(t)
1/t
(32)
which describes the exponential rate of convergence to zero
of v(t). It is clear that R < 1 implies the exponential mean
square convergence of q
C
(t) to the optimal solution q
opt
C
.
Using (30), we have
v(t) = E
_
x(t)
T
Mx(t)

= E
_
x(t)
T
Mx(t)

= E
_
x(t 1)
T
F
T
r(t1)
MF
r(t1)
x(t 1)
_
= x(0)
T
E
_
F
T
r(0)
F
T
r(t1)
MF
r(t1)
F
r(0)
_
x(0).
Let us then dene
(t) = E
_
F
T
r(0)
F
T
r(t1)
MF
r(t1)
F
r(0)
_
.
Via Assumption 2, we can argue that (t) satises the
following linear recursive equation
_
(t + 1) = /((t))
(0) = M,
(33)
where /() := E
_
F
T
r
F
r

. Moreover we have that

v(t) = x(0)
T
(t)x(0). (34)
Equations (33) and (34) can be seen as a discrete time linear
system with state (t) and output v(t).
Remark. The discrete time linear system (33) can be equiv-
alently described by the equation
vec ((t + 1)) = Fvec ((t)) ,
where vec() stands for the operation of vectorization and
where
F := E
_
F
T
r
F
T
r

R
m
2
m
2
.
Studying the rate R (which can be proved to be the slowest
reachable and observable mode of the system (33) with the
output (34), see [45]) is in general not simple. It has been done
analytically and numerically for some special graph structures
in [45]. It is convenient to analyze its behavior indirectly,
through another parameter which is easier to compute and to
study, as the following result shows.
Theorem 1. Consider the iteration (29), and assume that 1
describing the clusters (
r
is an arbitrary connected hyper-
graph dened over the nodes (. Let Assumption 2 hold. Then
the rate of convergence R dened in (32) satises
R < 1,
where
= max[[ [ (F
ave
), ,= 1,
with F
ave
:= E [F
r
].
Proof: The proof is given in Appendix B.
The following result descends directly from it.
Corollary. Assume that the hypergraph 1 is connected and
that Assumption 2 holds. Then the state of the iterative
algorithm described in Section IV-A converges in mean square
to the global optimal solution.
Remark. Observe that, by standard arguments based on Borel
Cantelli lemma, the exponential convergence in mean square
implies also almost sure convergence to the optimal solution.
The tightness of as a bound for R has been studied in [45],
by evaluating both and R analytically and numerically for
some special graph topologies. In the following we consider
as a reliable metric for the evaluation of the algorithm
performances.
The following result shows what is the best performance
(according to the bound on the convergence rate R) that the
proposed algorithm can achieve.
Theorem 2. Consider the iteration (29), and assume that 1
describing the clusters (
r
is an arbitrary connected hyper-
graph dened over the nodes (. Let Assumption 2 hold. Then
the bound introduced in Theorem 1 satises
1
_

r=1

r
[(
r
[
_
1
m1
(35)
where
r
is the probability of selecting cluster (
r
at each
iteration. In case all the sets (
r
have the same cardinality c,
namely [(
r
[ = c for all r, then
1
c 1
m1
. (36)
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
10
Proof: Let
E
ave
:= E [E
r
] =

r=1

r
E
r
.
It is clear that = 1

, where

= min[[ [
(E
ave
), ,= 0. We have that

j(Eave)

j
= Tr(E
ave
) = Tr
_

r=1

r
E
r
_
=

r=1

r
Tr (E
r
) .
It is shown in Lemma 4 in Appendix B that E
r
are projection
matrices and then have eigenvalues zero or one. Therefore, by
also using (20), we have
Tr (E
r
) = rank E
r
= dim(Image E
r
)
= dim
_
Image(
r
M
r
)

_
= dim(Image
r
)
= [(
r
[ 1.
This implies that

j(Eave)

j
=
_

r=1

r
[(
r
[
_
1,
and so

= min[[ [ (E
ave
), ,= 0

j(Eave)

j
m1
=
_

r=1

r
[(
r
[
_
1
m1
.
VI. OPTIMAL COMMUNICATION HYPERGRAPH FOR A
RADIAL DISTRIBUTION NETWORK
In this section we present a special case in which the optimal
convergence rate of Theorem 2 is indeed achieved via a
specic choice of the clusters (
r
. To obtain this result we need
to introduce the following assumption, which is commonly
veried in many practical cases (including the vast majority
of power distribution networks and the standard IEEE testbeds
[39]).
Assumption 3. The distribution network is radial, i.e. the
corresponding graph ( is a tree.
Let us now consider for any pair of nodes h, k the shortest
path T
hk
c (where, we recall, a path is a sequence of
consecutive edges) connecting h, k. Dene moreover
T
Cr
:=
_
h,kCr
T
hk
.
Lemma 2. Let Assumption 3 hold true. Then
i) (1
h
1
k
)
T
M(1
h
1
k
) =

eP
hk
P
h

Re(z
e
).
ii) E
r
(1
h
1
k
) = 0 if T
Cr
T
hk
= .
Proof:
i) Observe that (1
h
1
k
)
T
M(1
h
1
k
) = (1
h

1
k
)
T
X(1
h
1
k
), where with some abuse of notation
we have denoted with the same symbol 1
h
the vector in
R
m
and the corresponding vector in R
n
. Notice that, if
we have a current 1
h
1
k
in the network, then according
to (7) we get a voltage vector
u = Xi +1U
N
= X(1
h
1
k
) +1U
N
and so (1
h
1
k
)
T
X(1
h
1
k
) = u
h
u
k
. Observing
the following gure
h

k h
h

we can argue that u

h
u
k
= u
h
u
k
=

eP
h

z
e
and so (1
h
1
k
)
T
M(1
h
1
k
) =

eP
h

Re(z
e
).
ii) If T
Cr
T
hk
= , then

r
M(1
h
1
k
) =
1
2[(
r
[

,k

Cr
(1
h
1
k
)(1
h
1
k
)
T
M(1
h
1
k
) = 0
since T
hk
T
h

k
= for all h

, k

(
r
. We therefore
have that M(1
h
1
k
) ker
r
and thus, via (19), we
have that E
r
(1
h
1
k
) = (
r
M
r
)

M(1
h
1
k
) = 0.
We give now the key denition to characterize the optimality
of the communication hypergraph 1.
Denition. The hyperedges (
r
, r = 1, . . . , of the hyper-
graph 1 are edge-disjoint
1
if for any r, r

such that r ,= r

,
we have that
T
Cr
T
C
r

= .
In order to prove the optimality of having edge-disjoint
clusters, we need the following lemma.
Lemma 3. Assume that Assumptions 3 holds true and that the
clusters (
r
are edge-disjoint. Then E
r
E
r
= 0 for all r ,= r

.
Proof: Let x be any vector in R
m
. According to (20),
there exists z such that E
r
x =
r
z. We then have
E
r
E
r
x = E
r

r
z =
1
2[(
r
[

h,kC
r

E
r
(1
h
1
k
)(1
h
1
k
)
T
z = 0
where the last equality follows from Lemma 2 and from the
fact that T
Cr
T
hk
= for all h, k (
r
.
Lemma 4 in Appendix B shows that Image
r
= Image E
r
.
Therefore, by Lemma 3, we can argue that every vector in
Image
r
is eigenvector of E
ave
=

r=1

r
E
r
associated
with the eigenvalue
r
. On the other hand, because hypergraph
connectivity implies that ker 1
T
=

r=1
Image
r
(see
Lemma 5 in Appendix B) we have that, besides the zero
eigenvalue, the eigenvalues of E
ave
are exactly
1
, . . . ,

and
so
= 1 min
r=1,...,

r
.
1
This denition of edge-disjoint for this specic scenario resembles a
similar concept for routing problems in data networks, as for example [46].
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
11
PCC
Figure 3. Schematic representation of the IEEE37 testbed. Circled nodes
represent microgenerators taking part to the distributed reactive power com-
pensation. The thick curved lines represent clusters of size c = 2 (i.e. edges
of H) and thus connect pair of compensators that can communicate and
coordinate their behavior.
The bound is minimized by taking
r
= 1/ for all r,
obtaining
= 1
1

.
Notice nally that E
r
E
r
= 0 for all r ,= r

implies that

r=1
Image E
r
is a direct sum and then
m1 =

r=1
dimImage E
r
=

r=1
([(
r
[ 1) =

r=1
[(
r
[ .
If all the sets (
r
have the same cardinality c, then c = m +
1, and so
= 1
1

= 1
c 1
m1
which, according to Theorem 2, shows the optimality of the
hypergraph.
VII. SIMULATIONS
In this section we present numerical simulations to validate
both the model presented in Section III, and the randomized
algorithm proposed in Section IV.
To do so, we considered a 4.8 kV testbed inspired from the
standard testbed IEEE 37 [39], which is an actual portion of
power distribution network located in California. We however
assumed that load are balanced, and therefore all currents and
voltages can be described in a single-phase phasorial notation.
As shown in Figure 3, some of the nodes are microgen-
erators and are capable of injecting a commanded amount
of reactive power. Notice that the PCC (node 0) belongs to
the set of compensators. This means that the microgrid is
allowed to change the amount of reactive power gathered from
the transmission grid, if this reduces the power distribution
losses. The nodes which are not compensators, are a blend
of constant-power, constant-current, and constant-impedance
loads, with a total power demand of almost 2 MW of active
power and 1 MVAR of reactive power (see [39] for the testbed
data).
1 36
4,500
4,600
4,700
4,800
node v
|
u
v
|
[
V
]
1 36
3 10
3
2 10
3
1 10
3
0
node v

u
v
[
r
a
d
]
Figure 4. Comparison between the network state (node voltages) computed
via the exact model induced by (1), (2), (4), and (3) (indicated by ), and the
approximate model proposed in Proposition 1 (indicated by +).
The network topology is radial. While this is not a necessary
condition for the algorithm, this is typical in basically all
power distribution networks. Moreover, the choice of a radial
grid allows the validation of the optimality result of Section
VI for the clustering of the compensators.
The length of the power lines range from a minimum
of 25 meters to a maximum of almost 600 meters. The
impedance of the power lines differs from edge to edge (for
example, resistance ranges from 0.182 /km to 1.305 /km).
However, the inductance/resistance ratio exhibits a smaller
variation, ranging from z
e
= 0.47 to z
e
= 0.59. This
justies Assumption 1, in which we claimed that z
e
can
be considered constant across the network. The effects of this
approximation will be commented later.
Without any distributed reactive power compensation, distri-
bution power losses amount to 61.6 kW, 3.11% of the delivered
active power.
Given this setup, we rst estimate the quality of the linear
approximated model proposed in Section III. As shown in
Figure 4, the approximation error results to be negligible.
We then simulated the behavior of the algorithm proposed
in Section IV. We considered two clustering choices:
edge-disjoint gossip: motivated by the result stated in
Section VI, we enabled pairwise communication between
compensator in a way that guarantees that the clusters
(pairs) are edge-disjoint; the resulting hypergraph 1 is
represented as a thick line in Figure 3;
star topology: clusters are in the form (
r
= 0, v for
all v (0; the reason of this choice is that, as
0 is the PCC, the constraint 1
T
q = 0 is inherently
satised: whatever variation in the injected reactive power
is applied by v, it is automatically compensated by a
variation in the demand of reactive power from the
transmission grid via the PCC.
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
12
0 5 10 15 20 25 30
50,000
55,000
60,000
iteration
l
o
s
s
e
s
[
W
]
Figure 5. Power distribution losses resulting by the execution of the proposed
algorithm. A edge-disjoint hypergraph, yielding optimal convergence speed,
has been adopted. The dashed line represent the minimum losses that can be
achieved via centralized numerical optimization.
In Figure 5 we plotted the result of a single execution of the
algorithm in the edge-disjoint gossip case. The results achieved
by the proposed algorithm on this testbed are summarized in
the following table.
Losses before optimization 61589 W
Fraction of delivered power 3.11 %
Losses after optimization 50338 W
Fraction of delivered power 2.55 %
losses reduction 18.27 %
Minimum losses J
opt
50253 W
Fraction of delivered power 2.54 %
losses reduction 18.41 %
The minimum losses J
opt
are the solution of the original op-
timization problem (11), obtained by a state-of-the-art ORPF
numerical solver. They are adopted here as a baseline for the
evaluation of the performance of the proposed algorithm.
One can see that the algorithm converges quite fast, re-
ducing losses to a minimum that is extremely close to J
opt
.
The difference between the baseline J
opt
and the minimum
achieved by the algorithm proposed in this paper is partly due
to the approximation that we introduced when we modeled the
microgrid (assuming large nominal voltage U
N
), and partly
due to the assumption that is constant across the network.
Further simulative investigation on this testbed showed that the
effect of this last assumption is largely predominant, compared
to the effects of the approximation in the microgrid state
equations. Still, the combined effect of these two terms is
minimal, in practice.
In Figure 6 we compared the behavior of the two considered
clustering strategies. One can notice different things. First,
both strategies converge to the same minimum, which is
slightly larger than the minimum losses that could be achieved
by solving the original, nonconvex optimization problem. As
expected, the clustering choice do not affect the steady state
performance of the algorithm. Second, the performance of the
edge-disjoint gossip algorithm results indeed to be better than
the star topology in the long-time regime, as the analysis in
Section VI suggests, and its slope corresponds to the fastest
achievable rate of convergence, as predicted by Theorem 2.
Third, one can see that, while the performance metric that we
adopted is meaningful for describing the long-time regime, it
does not describe, in general, the initial stage (or short-time
0 10 20 30 40 50 60 70
100
1,000
10,000
iteration
E
[
J
(
q
(
t
)
)

J
o
p
t
]
[
W
]
Figure 6. Simulation of the expected behavior of the algorithm for different
clustering strategies. The difference between power losses at each iteration
of the algorithm (averaged over 1000 realizations) and the minimum possible
losses J
opt
, has been plotted for two different clustering choices: edge-disjoint
gossip (solid line) and star topology (dashed). The slope represented via a
dotted line represent the best possible rate of convergence that the algorithm
can achieve, according to Theorem 2.
behavior). Indeed, in the initial transient, the full dynamics of
the system (29) contribute to the algorithm behavior. This is
evident in the star-topology strategy. In the edge-disjoint case,
on the other hand, all the eigenvalues of the systems (29) result
to be the same, and thus the adopted performance metric (the
slowest eigenvalue) fully describes the algorithm evolution.
This fact suggests that a different metric can possibly be
adopted, in order to better characterize the two regimes.
Choosing the most appropriate metric requires a preliminary
derivation of a dynamic model for the variation of reactive
power demands of the loads in time. In the case of time-
varying loads, different strategies will corresponds to different
steady-state behaviors of the algorithm. We expect that an
L
2
type metric will be needed for the characterization of the
performance of different clustering choices, while the slowest
eigenvalue will not be informative enough. A preliminary
study in this direction has been proposed in [47].
VIII. CONCLUSIONS
The proposed formulation of the problem of optimal reactive
power compensation in smart microgrids exhibits two main
features. First, it can be cast into the framework of quadratic
optimization, for which robust solvers are available and the
performance analysis becomes tractable; second, it shows
how the physics of the system can be exploited to design a
distributed algorithm for the problem.
Based on the proposed model, we designed a distributed,
leaderless and randomized algorithm for the solution of the
optimization problem, requiring only local communication and
local knowledge of the network topology and state.
We then proposed a metric for the performance of the
algorithm. Based on it, we are able to prove which clustering
choice yields the best achievable performance. Interestingly,
the optimal strategy requires short-range communication: this
is a notable feature of this specic problem, as generic consen-
sus algorithms (which share many features with the proposed
algorithm) typically benet from long-range communication
that shorten the graph diameter. This approach seems also
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
13
preferable from the technological point of view, because it
avoids a number of issues that would arise in long distance
communication over the power grid (e.g. time synchronization,
limited range of power line communication technologies).
Motivated by the nal remarks of Section VII, we plan to
add two elements to the picture: the network dynamics (both
the dynamic behavior of the power demands, and the electrical
response of the grid when actuated), and the presence of oper-
ational constraints for the compensators, whose capability of
injecting reactive power is limited and possibly time varying.
APPENDIX A
PROOF OF PROPOSITION 1
We rst introduce the new variable := 1/U
N
. We also
assume, with no loss of generality, that the phase of the voltage
at the PCC is = 0. This way we can see the currents i and
the voltages u as functions i(), u() of , determined by the
equations
_

_
u = Xi +
1
1
1
T
i = 0
u
v

i
v
= s
v
[u
v
[
v
, v 10
(37)
We construct the Taylor approximation of u() and i() around
= 0.
Let us dene

v
() := i
v
()
1
s
v

v
() := u
v
()
1

2
[X s]
v
(38)
and substitute them into (37). We get a system of equations
in , , which can be written as
_
G
v
(, , ) = 0, v 1
F
v
(, , ) = 0, v 1
where
G
0
(, , ) :=
0
F
0
(, , ) :=

wV

w
and, for all v ,= 0,
G
v
(, , ) :=
v

wV
[X]
vw

w
,
F
v
(, , ) :=
_
1 +
2

v
+
2

wV
[X]
vw
s
w
_
(s
v
+

v
)
s
v

1 +
2

v
+
2

wV
[X]
vw
s
w

v
.
We have 2n complex equations in 2n+1 complex variables.
We interpret now the complex numbers as vectors in R
2
and G
v
, F
v
as functions from R
2n
R
2n
R
2
to R
2
.
Observe that G
v
(, , )
|=0,=0,=0
= 0 and
F
v
(, , )
|=0,=0,=0
= 0 for all v 1. By the implicit
function theorem, if the matrix
_

_
_
G
v

w
_
v,wV
_
G
v

w
_
v,wV
_
F
v

w
_
v,wV
_
F
v

w
_
v,wV
_

_
evaluated in = 0, = 0, = 0 is invertible, then (), ()
have continuous derivatives in = 0.
In order to determine this matrix we need to introduce the
following notation. If a C, and a = a
R
+ja
I
with a
R
, a
I

R, then we dene
a :=
_
a
R
a
I
a
I
a
R
_
With this notation observe that, if a, x C and we consider
those complex numbers as vectors in R
2
and if we consider
the function f(x) := ax as a function from R
2
to R
2
, then we
have that f/x = a. Notice moreover that the the function
g(x) := x has
g
x
= N :=
_
1 0
0 1
_
From these observations we can argue that
G
0

w
= 0,
G
0

w
=
_
I if w = 0
0 if w ,= 0
F
0

w
= I,
F
0

w
= 0
Instead for all v ,= 0 we have that
G
v

w
=
_
I if w = v
0 if w ,= v
while
G
v

w
= X
vw
w 1.
Observe nally that, if w ,= v, then F
v
/
w
= F
v
/
w
=
0 and that
F
v

v
=
_
1 +
2

v
+
2

wV
[X]
vw
s
w
_
N
which evaluated in = 0 yields the matrix N. On the other
hand,
F
v

v
=
2
(s
v
+

v
)s
v

1 +
2

v
+
2

wV
[X]
vw
s
w

v
where s
v
is a 2-dimensional column vector and
|1+v|
v
v
is
a 2-dimensional row vector. With some easy computations it
can be seen that
|1+v|
v
v
, evaluated in = 0, yields [0 0].
Hence in = 0 we have
F
v

v
= 0.
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
14
We can conclude that for = 0, = 0, = 0 we have
_

_
_
G
v

w
_
v,wV
_
G
v

w
_
v,wV
_
F
v

w
_
v,wV
_
F
v

w
_
v,wV
_

_
=
_

_
0 0 0 I 0 0
0 [X]
11
[X]
1,n1
0 I 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 [X]
n1,1
[X]
n1,n1
0 0 I
I I I 0 0 0
0 N 0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 N 0 0 0
_

_
which is invertible. By applying Taylors theorem, we can thus
conclude from (38) that
i
v
=
s
v
U
N
+
c
v
(U
N
)
U
2
N
u
v
= U
N
+
[X s]
v
U
N
+
d
v
(U
N
)
U
2
N
,
where c
v
(U
N
) and d
v
(U
N
) are bounded functions of U
N
.
APPENDIX B
PROOF OF THEOREM 1
In order to prove Theorem 1, we need the following
technical results. We start by introducing the matrix

M =
M+11
T
> 0. This allows us to state the following properties
of the matrices E
r
and F
r
dened in (31).
Lemma 4. Consider the matrices E
r
and F
r
, r = 1, . . . , ,
as dened in (31). The following properties hold.
i) E
T
r

M =

ME
r
and F
T
r

M =

MF
r
. Therefore the
matrices E
r
, F
r
are self-adjoint matrices with respect to
the inner product ,

M
, dened as x, y

M
:= x
T

My.
ii) The matrices E
r
, F
r
are projection operators, i.e. E
2
r
=
E
r
and F
2
r
= F
r
.
iii) Image E
r
= Image
r
.
Proof: Properties i) and ii) can be proved by inspection.
Property iii) descends from (20), using the fact that E
r
=
(
r
M
r
)

M = (
r
M
r
)

M.
We then have the following technical lemmas.
Lemma 5. Consider the matrices F
r
, r = 1, . . . , , as dened
in (31). The following statements are equivalent:
i) the point x = 0 is the only point in ker 1
T
such that
F
r
x = x for all r = 1, . . . , ;
ii) Image[
1
. . .

] = ker 1
T
;
iii) the hypergraph 1 is connected.
Proof: Let us rst prove that ii) implies i). Assume that
Image[
1
. . .

] = ker 1
T
and that x ker 1
T
is such
that F
r
x = x for all r. Then we can nd y
r
R
n
such
that x =

r

r
y
r
. Moreover, as F
r
x = x for all r, then
(
r
M
r
)

Mx = 0 and so, by properties (19), we have that

Mx ker
r
. Therefore also

Mx ker
r
, as 1 ker
r
.
We then have
x
T

Mx =

r
y
T
r

r

Mx = 0,
and so, since

M is positive denite, it yields x = 0.
We prove now that statement i) implies ii). The inclusion
Image[
1
. . .

] ker 1
T
is always true and follows from
the fact that Image
r
ker 1
T
. We need to prove only the
other inclusion. Suppose that x = 0 is the only point in ker 1
T
such that F
r
x = x for all i. This means that
ker 1
T
ker(E
1
) ker(E

) = 0,
which implies that
R
m
= Image 1 + Image(E
T
1
) + + Image(E
T

)
Take now any x ker 1
T
. Then there exist a R and
y
r
R
m
such that

Mx = 1 +

r=1
E
T
r
y
r
= 1 +

r=1
M(
r
M
r
)

y
r
Because of (19) we have that

M(
r
M
r
)

= M(
r
M
r
)

and therefore 1 =

Mw where
w = x

r=1
(
r
M
r
)

y
r
.
Using (20) we then have that w = x

r=1

r
z
r
for some
z
r
. From the fact that Image
r
ker 1
T
, we can argue that
w ker 1
T
and so it follows that w
T
1 = w
T

Mw = 0.
Finally, since

M is positive denite, we can conclude that
w = 0 and so that x Image[
1
. . .

].
We nally prove that ii) and iii) are equivalent. We consider
the matrix W R
mm
with entries [W]
hk
equal to the
number of the sets (
r
which contain both h and k. Let (
W
be
the weighted graph associated with W. It is easy to see that
the hypergraph having (
r
as edges is connected if and only if
(
W
is a connected graph.
Let us dene
Cr
: ( 0, 1 as the characteristic function
of the set (
r
, namely a function of the nodes that is 1 when the
node belongs to (
r
and is zero otherwise. Then the Laplacian
L
W
of (
W
can be expressed as follows
L
W
=

h,kC
(1
h
1
k
)(1
h
1
k
)
T
[W]
hk
=

h,kC
(1
h
1
k
)(1
h
1
k
)
T

i=1

Cr
(h)
Cr
(k)
=

r=1

h,kCr
(1
h
1
k
)(1
h
1
k
)
T
=

r=1
2[(
r
[
r
= [
1
. . .

] diag2[(
1
[I, . . . , 2[(

[I
_

1
.
.
.

_.
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
15
Notice that the graph connectivity is equivalent to the fact
that ker L
W
= Image 1 and so, by the previous equality, it is
equivalent to
ker
_

1
.
.
.

_ = Image 1
which is equivalent to ii).
Lemma 6. Let P, Q R
mm
and P Q. Then /
k
(P)
/
k
(Q) for all k Z
0
.
Proof: From the denition of /, we have
x
T
[/(P) /(Q)] x = x
T
_
E
_
F
T
r
PF
r

E
_
F
T
r
QF
r

x
= E
_
x
T
F
T
r
(P Q)F
r
x

0,
and therefore P Q implies /(P) /(Q). By iterating
these steps k times we then obtain /
k
(P) /
k
(Q).
Lemma 7. For any we have that /
k
() =
/
k
().
Proof: Proof is by induction. The statement is true for
k = 0, as
2
= . Suppose it is true up to k. We then have
/
k+1
() = /(/
k
()) = /(/
k
())
= /(/
k
()) = /
k+1
().
Lemma 8. All the eigenvalues of F
ave
are real and have ab-
solute value not larger than 1. If Image
_

= ker 1
T
and if Assumption 2 holds, then the only eigenvalue of F
ave
on the unitary circle is = 1, and it has multiplicity 1.
Proof: Recall that the matrices F
r
are projection oper-
ators, i.e. F
2
r
= F
r
and so they have eigenvalues 0 or 1.
Recall moreover that F
r
are self-adjoint matrices with respect
to the inner product ,

M
, dened as x, y

M
= x
T

My, This
implies that [[F
r
[[

M
1 where [[ [[

M
is the induced matrix
norm with respect to the vector norm [[x[[

M
:= x, x
1/2

M
. This
implies that
|F
ave
|

M
= |E [F
r
] |

M
E [|F
r
|

M
] 1
And so, since also F
ave
is self-adjoint, its eigenvalues are real
and are smaller than or equal to 1 in absolute value.
Assume now that F
ave
x = x, with [[ = 1. Then we have
|x|

M
= |F
ave
x|

M
= |E [F
r
] x|

M
E [|F
r
x|

M
] |x|

M
.
If Assumption 2 holds (and thus the probabilities
r
s are all
strictly greater than 0), then the last inequality implies that
|F
r
x|

M
= |x|

M
for all rs. As F
r
are projection matrices,
it means that F
r
x = x and so Mx ker
r
, r. Using the
fact that Image
_

= ker 1
T
, we necessarily have
Mx = 1 for some . By denition of

M, we also have

Mx =

1 for some

, and therefore the eigenvector x must

be parallel to

M
1
1.
We can now give the proof of Theorem 1.
Proof of Theorem 1: We rst prove that /(M)
M. Indeed, we have
x
T
/(M)x = E
_
x
T
F
T
r
MF
r
x

= E
_
x
T
F
T
r
MF
r
x

= x
T

M
1/2
E
_

M
1/2
F
r

M
1/2
_

M
1/2
x,
where we use the fact that F
r
= F
r
and F
T
r
MF
r
=
MF
r
=

MF
r
.
Notice also that E
_

M
1/2
F
r

M
1/2
_
=

M
1/2
F
ave

M
1/2
,
by Lemma 5 and Lemma 8, has only one eigenvalue on the
unit circle (precisely in 1), with eigenvector

M
1/2
1. As

M
1/2
x

M
1/2
1 for all x, we have
x
T
/(M)x x
T
Mx,
with = max[[ [ (F
ave
), ,= 1. From this result,
using Lemmas 6 and 7, we can conclude
/
t
(M) = /
t1
(/(M))
= /
t1
(/(M))
/
t1
(M)
= /
t1
(M)
t
M,
and therefore R .
REFERENCES
[1] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa-
tion: Numerical Methods. Athena Scientic, 1997.
[2] Special issue on technology of networked control systems, Proc. IEEE,
vol. 95, no. 1, Jan. 2007.
[3] E. Santacana, G. Rackliffe, L. Tang, and X. Feng, Getting smart. with a
clearer vision of the intelligent grid, control emerges from chaos, IEEE
Power Energy Mag., vol. 8, no. 2, pp. 4148, Mar. 2010.
[4] A. Ipakchi and F. Albuyeh, Grid of the future. are we ready to transition
to a smart grid? IEEE Power Energy Mag., vol. 7, no. 2, pp. 52 62,
Mar.-Apr. 2009.
[5] J. A. Lopes, C. L. Moreira, and A. G. Madureira, Dening control
strategies for microgrids islanded operation, IEEE Trans. Power Syst.,
vol. 21, no. 2, pp. 916924, May 2006.
[6] T. C. Green and M. Prodanovi c, Control of inverter-based micro-grids,
Electr. Pow. Syst. Res., vol. 77, no. 9, pp. 12041213, Jul. 2007.
[7] F. Katiraei and M. R. Iravani, Power management strategies for a
microgrid with multiple distributed generation units, IEEE Trans. Power
Syst., vol. 21, no. 4, pp. 18211831, Nov. 2006.
[8] M. Prodanovic, K. De Brabandere, J. Van den Keybus, T. Green, and
J. Driesen, Harmonic and reactive power compensation as ancillary
services in inverter-based distributed generation, IET Gener. Transm.
Distrib., vol. 1, no. 3, pp. 432438, 2007.
[9] E. Tedeschi, P. Tenti, and P. Mattavelli, Synergistic control and coop-
erative operation of distributed harmonic and reactive compensators, in
Proc. 39th IEEE Power Electronics Specialists Conf. (PESC), 2008.
[10] A. Rabiee, H. A. Shayanfar, and N. Amjady, Reactive power pricing:
Problems and a proposal for a competitive market, IEEE Power Energy
Mag., vol. 7, no. 1, pp. 1832, Jan. 2009.
[11] H. Yoshida, K. Kawata, Y. Fukuyama, S. Takayama, and Y. Nakanishi,
A particle swarm optimization for reactive power and voltage control
considering voltage security assessment, IEEE Trans. Power Syst.,
vol. 15, no. 4, pp. 12321239, Nov. 2000.
[12] B. Zhao, C. X. Guo, and Y. J. Cao, A multiagent-based particle swarm
optimization approach for optimal reactive power dispatch, IEEE Trans.
Power Syst., vol. 20, no. 2, pp. 10701078, May 2005.
[13] J. Lavaei, A. Rantzer, and S. H. Low, Power ow optimization using
positive quadratic programming, in Proc. 18th IFAC World Congr.,
2011.
[14] K. M. Rogers, R. Klump, H. Khurana, A. A. Aquino-Lugo, and T. J.
Overbye, An authenticated control framework for distributed voltage
support on the smart grid, IEEE Trans. Smart Grid, vol. 1, no. 1, pp.
4047, Jun. 2010.
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
16
[15] F. Dorer and F. Bullo, Synchronization and transient stability in power
networks and non-uniform Kuramoto oscillators, in Proc. American
Control Conf. (ACC), 2010, pp. 930937.
[16] Y. Guo, D. J. Hill, and Y. Wang, Nonlinear decentralized control of
large-scale power system, Automatica, vol. 36, no. 9, pp. 12751289,
Sep. 2000.
[17] N. I. Deeb and S. M. Shahidehpour, Decomposition approach for
minimising real power losses in power systems, IEE Proc. C, Gener.
Transm. Distrib., vol. 138, no. 1, pp. 2738, Jan. 1991.
[18] B. H. Kim and R. Baldick, Coarse-grained distributed optimal power
ow, IEEE Trans. Power Syst., vol. 12, no. 2, pp. 932939, May 1997.
[19] A. Y. Lam, B. Zhang, A. Dominguez-Garcia, and D. Tse, Optimal
distributed voltage regulation in power distribution networks, ArXiv e-
prints, vol. 1204.5226 [math.OC], 2012.
[20] P. Tenti, A. Costabeber, P. Mattavelli, and D. Trombetti, Distribution
loss minimization by token ring control of power electronic interfaces
in residential microgrids, IEEE Trans. Ind. Electron., vol. 59, no. 10,
pp. 38173826, Oct. 2012.
[21] B. A. Robbins, A. D. Dominguez-Garcia, and C. N. Hadjicostis, Con-
trol of distributed energy resources for reactive power support, in Proc.
of North American Power Symp., Boston, MA, Aug. 2011.
[22] K. Turitsyn, P. ulc, S. Backhaus, and M. Chertkov, Options for control
of reactive power by distributed photovoltaic generators, Proc. IEEE,
vol. 99, no. 6, pp. 10631073, Jun. 2011.
[23] M. E. Baran and I. M. El-Markabi, A multiagent-based dispatching
scheme for distributed generators for voltage support on distribution
feeders, IEEE Trans. Power Syst., vol. 22, no. 1, pp. 5259, Feb. 2007.
[24] A. Gmez-Expsito, A. J. Conejo, and C. Caizares, Electric energy
systems. Analysis and operation. CRC Press, 2009.
[25] V. Voloshin, Introduction to Graph and Hypergraph Theory. Nova
Science Publishers, 2009.
[26] IEEE Task Force on Load representation for dynamic performance,
Load representation for dynamic performance analysis, IEEE Trans.
Power Syst., vol. 8, no. 2, pp. 472482, May 1993.
[27] M. H. Haque, Load ow solution of distribution systems with voltage
dependent load models, Electr. Pow. Syst. Res., vol. 36, pp. 151156,
1996.
[28] A. Ghosh, S. Boyd, and A. Saberi, Minimizing effective resistance of
a graph, SIAM Rev., vol. 50, no. 1, pp. 3766, Feb. 2008.
[29] M. E. Baran and F. F. Wu, Optimal sizing of capacitors placed on a
radial distribution system, IEEE Trans. Power Del., vol. 4, no. 1, pp.
735743, Jan. 1989.
[30] A. Cagnano, E. De Tuglie, M. Liserre, and R. Mastromauro, On-line
optimal reactive power control strategy of PV-inverters, IEEE Trans.
Ind. Electron., vol. 58, no. 10, pp. 45494558, Oct. 2011.
[31] S. Bolognani, A. Carron, A. Di Vittorio, D. Romeres, and L. Schenato,
Distributed multi-hop reactive power compensation in smart micro-
grids subject to saturation constraints, in Proc. 51st IEEE Conf. on
Decision and Control (CDC), 2012.
[32] S. Bolognani, G. Cavraro, and S. Zampieri, A distributed feedback
control approach to the optimal reactive power ow problem, in
Proceedings of the Workshop on Control of Cyber-Physical Systems,
ser. Lecture Notes on Control and Information Sciences. Springer,
Mar. 2013.
[33] A. Nedic and A. Ozdaglar, Distributed subgradient methods for multi-
agent optimization, IEEE Trans. Autom. Control, vol. 54, no. 1, pp.
4861, Jan. 2009.
[34] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, Distributed
optimization and statistical learning via the alternating direction method
of multipliers, Foundations and Trends in Machine Learning, vol. 3,
no. 1, pp. 1122, 2011.
[35] M. Ciobotaru, R. Teodorescu, P. Rodriguez, A. Timbus, and F. Blaabjerg,
Online grid impedance estimation for single-phase grid-connected
systems using PQ variations, in Proc. 38th IEEE Power Electronics
Specialists Conf. (PESC), Jun. 2007.
[36] A. Costabeber, T. Erseghe, P. Tenti, S. Tomasin, and P. Mattavelli,
Optimization of micro-grid operation by dynamic grid mapping and
token ring control, in Proc. 14th European Conf. on Power Electronics
and Applications (EPE), Birmingham, UK, 2011.
[37] A. G. Phadke, Synchronized phasor measurements in power systems,
IEEE Comput. Appl. Power, vol. 6, no. 2, pp. 1015, Apr. 1993.
[38] A. G. Phadke and J. S. Thorp, Synchronized phasor measurements and
their applications. Springer, 2008.
[39] W. H. Kersting, Radial distribution test feeders, in IEEE Power
Engineering Society Winter Meeting, vol. 2, Jan. 2001, pp. 908912.
[40] S. A. Grandhi, R. Vijayan, and D. J. Goodman, Distributed power
control in cellular radio systems, IEEE Trans. Commun., vol. 42, no.
234, pp. 226228, 1994.
[41] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan, Rate control for commu-
nication networks: shadow prices, proportional fairness and stability, J,
Oper. Res. Soc., vol. 49, no. 3, pp. 237252, Mar. 1998.
[42] S. H. Low, A duality model of TCP and queue management algorithms,
IEEE/ACM Trans. Netw., vol. 11, no. 4, pp. 525536, Aug. 2003.
[43] H. K. Khalil, Nonlinear systems, 3rd ed. Prentice Hall, 2002.
[44] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, Randomized gossip
algorithms, IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 25082530,
Jun. 2006.
[45] S. Bolognani and S. Zampieri, A gossip-like distributed optimization
algorithm for reactive power ow control, in Proc. 18th IFAC World
Congr., Milano, Italy, Aug. 2011.
[46] A. Srinivasan, Improved approximations for edge-disjoint paths, un-
splittable ow, and related routing problems, in Proc. 38th Annu. Symp.
on Foundations of Computer Science, 1997.
[47] S. Bolognani and S. Zampieri, A distributed optimal control approach
to dynamic reactive power compensation, in Proc. 51st IEEE Conf. on
Decision and Control (CDC), 2012.
Saverio Bolognani received the B.S. degree in infor-
mation engineering, the M.S. degree in automation
engineering, and the Ph.D. degree in information
engineering from the University of Padova, Italy, in
2005, 2007, and 2011, respectively.
From 2006 to 2007, he was a visiting graduate
student at the University of California at San Diego.
In 2011 and 2012 he was a postdoctoral associate
at the Department of Information Engineering of
the University of Padova, and since 2013 he is a
postdoctoral associate at the Laboratory for Infor-
mation and Decision Systems of the Massachusetts Institute of Technology in
Cambridge (MA). His research interests include the application of networked
control system theory to smart power distribution networks, distributed
control, estimation, and optimization, and cyber-physical systems.
Sandro Zampieri received the Laurea degree in
electrical engineering, and the Ph.D. degree in sys-
tem engineering from the University of Padova, Italy,
in 1988 and 1993, respectively.
In 1991-92, 1993 and 1996 he was visiting scholar
at the Laboratory for Information and Decision Sys-
tems of the Massachusetts Institute of Technology in
Cambridge (MA). He has held visiting positions also
at the Department of Mathematics of the University
of Groningen and at the Department of Mechanical
Engineering of the University of California at Santa
Barbara. Since 2002 he is Full Professor in Automatic Control at the
Department of Information Engineering of the University of Padova. His
research interests include automatic control and dynamical systems theory,
and in particular distributed control and estimation, networked control, and
control under communication constraints.
Prof. Zampieri has published more than 100 journal and conference papers.
He has delivered several invited seminars and he was member of the Technical
Program Committee for several international conferences. He was general
chair of the 1st IFAC Workshop on Estimation and Control of Networked
Systems in 2009, program chair of the 3rd IFAC Workshop on Estimation
and Control of Networked Systems in 2012, and publication chair of the
IFAC World Congress in 2011. He served as an Associate Editor of the Siam
Journal on Control and Optimization in 2002-2004, and as the chair of the
IFAC technical committee Networked systems in 2005-2008. Since 2012 he
is serving as an Associate Editor of IEEE Transactions of Automatic Control.
Limited circulation. For review only
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2013 15:13:10 PST