IEEE TRANSACTIONS ON SMART GRID 1

Distributed Optimal Power Flow

for Smart Microgrids
Emiliano Dall’Anese, Member, IEEE, Hao Zhu, Member, IEEE, and Georgios B. Giannakis, Fellow, IEEE

Abstract— Optimal power flow (OPF) is considered for micro- at the distribution substation [5]. To this end, optimal power
arXiv:1211.5856v5 [math.OC] 25 Jan 2014

grids, with the objective of minimizing either the power distribu- flow (OPF) approaches are increasingly advocated also in this
tion losses, or, the cost of power drawn from the substation and context to ensure efficient operation of smart microgrids and
supplied by distributed generation (DG) units, while effecting
voltage regulation. The microgrid is unbalanced, due to unequal effect strict voltage regulation [3], [5], [6], [7].
loads in each phase and non-equilateral conductor spacings on OPF problems are deemed challenging because they re-
the distribution lines. Similar to OPF formulations for balanced quire solving nonconvex problems. Nonconvexity stems from
systems, the considered OPF problem is nonconvex. Nevertheless, the nonlinear relationship between voltages and the complex
a semidefinite programming (SDP) relaxation technique is advo- powers demanded or injected at the nodes. In the context
cated to obtain a convex problem solvable in polynomial-time
complexity. Enticingly, numerical tests demonstrate the ability of of transmission networks, the Newton-Raphson method [8]
the proposed method to attain the globally optimal solution of the is traditionally employed to obtain a possibly suboptimal
original nonconvex OPF. To ensure scalability with respect to the solution of these nonconvex problems. In distribution systems
number of nodes, robustness to isolated communication outages, however, its convergence is challenged by the high resistance-
and data privacy and integrity, the proposed SDP is solved in to-reactance ratio of distribution lines. Alternative approaches
a distributed fashion by resorting to the alternating direction
method of multipliers. The resulting algorithm entails iterative include sequential quadratic optimization, steepest descent-
message-passing among groups of consumers and guarantees based methods [5], fuzzy dynamic programming [9], and
faster convergence compared to competing alternatives. particle swarm optimization [7]. However, these methods
Index Terms— Microgrids, distribution feeders, optimal power generally return suboptimal load flow solutions, and may be
flow, semidefinite relaxation, distributed optimization. computationally cumbersome [10]. To alleviate these concerns,
a relaxed semidefinite programming (SDP) reformulation of
the OPF problem for balanced transmission systems was
I. I NTRODUCTION
proposed in [11] and [12], where global optimality can be
Microgrids are portions of a power distribution network assessed by checking the rank of the obtained voltage-related
located downstream of the distribution substation that supply matrix. The relaxed SDP approach was extended to balanced
a number of industrial and residential loads, and may include distribution systems in [13] and [6]. Notably, for networks with
distributed generation (DG) and energy storage devices [1]. a tree topology, [6] and [13] established sufficient conditions
A microgrid can operate in either grid-connected, islanded, under which a globally optimal solution is attainable provided
or hybrid modes. Deployment of microgrids promises drastic the original OPF problem is feasible.
performance enhancement of the distribution grid in terms of Distribution networks are inherently unbalanced because:
efficiency and stability, along with increased network scalabil- i) unequal single-phase loads must be served, and ii) non-
ity and resilience to outages. equilateral conductor spacings of three-phase line segments
Besides bringing power generation closer to the end user, are involved [14]. Further, single-phase DG units may worsen
DG units offer environment-friendly advantages over conven- the network imbalance. As a consequence, optimization ap-
tional generation [1], may provide ancillary services such proaches can not rely on single-phase equivalent models as in
as reactive and harmonic compensation [2], [3], and enable e.g., [3], [5], [6], [7], [13]. For the unbalanced setup, an OPF
DG owners to actively participate in grid operations through framework was proposed in [15], where commercial solvers
supply contracts and pricing schemes. On the other hand, their of nonlinear programs were adopted, and in [16], where quasi-
operation must be carefully controlled in order to prevent Newton methods were utilized in conjunction with load flow
abrupt voltage fluctuations, which stem from the well known solvers. However, since these methods are inherently related to
sensitivity of voltages to variations of power injections, node gradient descent solvers of nonconvex programs, they inherit
over- and under-voltages [4], and drops of the power factor the limitations of being sensitive to initialization, and do not
guarantee global optimality of their solutions.
Manuscript submitted on October 1, 2012; revised January 12, 2013; ac- The first contribution of the present paper consists in perme-
cepted February 16, 2013. This work was supported by the Inst. of Renewable ating the benefits of SDP relaxation techniques [17] to OPF
Energy and the Environment (IREE) grant no. RL-0010-13, Univ. of Min-
nesota. E. Dall’Anese and G. B. Giannakis are with the Digital Technology problems for microgrids operating in an unbalanced setup.
Center and the Dept. of ECE, University of Minnesota, 200 Union Street SE, This optimization tool has three main advantages: i) it offers
Minneapolis, MN 55455, USA. Hao Zhu is with the Information Trust Institute the potential of finding the globally optimal solution; ii) its
at the University of Illinois at Urbana-Champaign, 1308 West Main St,
Urbana, IL 61801, USA. E-mails: {emiliano, georgios}@umn.edu, worst-case computational complexity is quantifiable; and, iii)
[email protected] it can accommodate additional thermal and quality-of-power
IEEE TRANSACTIONS ON SMART GRID 2

constraints without exacerbating the problem complexity [17]. ploying the ADMM to devise a distributed SDP solver.
Global optimality not only reduces power distribution losses, The rest of the paper is organized as follows. Section II
but also leads to higher monetary savings compared to subop- recapitulates the OPF formulation for microgrids, and Sec-
timal OPF solutions. tion III develops its centralized SDP solver. The distributed
The OPF problem is solved by a microgrid energy manager algorithm is presented in Section IV, while numerical tests
(MEM), which cooperates with local controllers (LCs) located are reported in Section V. Finally, concluding remarks are
throughout the network. Microgrids can vary in scope, size, provided in Section VI1 .
and ownership [18]. For those of medium and large size (of
a distribution feeder), solving the OPF problem centrally at
II. M ODELING AND PROBLEM FORMULATION
the MEM may become computationally prohibitive [10]. In
fact, interior point SDP solvers do not generally scale well Consider a microgrid comprising N nodes collected in the
with the problem size [6], [19], [17]. For a real-time network set2 N := {0, 1, . . . , N }, and overhead or underground lines
management, it is generally required to find a new network represented by the set of edges E := {(m, n)} ⊂ N × N .
operational setup rapidly (e.g., in a few seconds or minutes) Let node 0 represent the point of common coupling (PCC),
in order to promptly respond to abrupt load variations and to taken to be the distribution substation. Define as Pmn ⊆
cope with the intermittent power generation that is typical of {amn , bmn , cmn } and Pn ⊆ {an , bn , cn } the phases of line
renewable-based DG units. It is then of paramount importance (m, n) ∈ E and node n ∈ N , respectively. Let Vnφ ∈ C
to solve the SDP-based OPF problem in a distributed manner, be the complex line-to-ground voltage at node n ∈ N of
by decomposing the main problem into multiple sub-instances phase φ ∈ Pn , and Inφ ∈ C the current injected at the same
that can be solved efficiently and in parallel. A distributed node and phase. As usual, the voltages v0 := [V0a , V0b , V0c ]T
algorithm is desirable also when (a group of) customers do are taken as reference for the phasorial representation. Lines
not share data with the MEM due to privacy concerns, or (m, n) ∈ E are modeled as π-equivalent components [14,
because they wish to manage autonomously their DG units in Ch. 6], and the |Pmn | × |Pmn | phase impedance and shunt
order to pursue individual economic interests [18]. Finally, admittance matrices are denoted as Zmn ∈ C|Pmn |×|Pmn |
(s)
a distributed algorithms involves a modest communication and Ymn ∈ C|Pmn |×|Pmn | , respectively. Three- or single-
overhead compared to its centralized counterpart, as it does phase transformers (if any) are modeled as series components
not require to pool line, generator, and load data at the MEM, with transmission parameters that depend on the connection
and subsequently disseminate the OPF solution. type [14, Ch. 8], [15].
φ
Decentralized OPF approaches were first proposed in [20], Per phase φ ∈ Pn , let PL,n and QφL,n denote the active
[21], where multi-utility transmission systems were par- and reactive powers demanded by a wye-connected load at
titioned in autonomously managed areas. Augmented La- the bus n. Capacitor banks are usually mounted at some
grangian methods were employed to decompose the overall φ
nodes to provide reactive power support. Let yC,n denote the
OPF problem in per-area instances. A similar approach was susceptance of a capacitor connected at node n and phase φ.
followed by e.g., [22], [23] (see also references therein), where Finally, suppose that S DG units are located at a subset of
φ
standard Lagrangian approaches were utilized in conjunction nodes S ⊂ N , and let PG,s and QφG,s denote the active and
with Newton methods. Solving SDP in a distributed fashion reactive powers supplied by unit s ∈ S. For conventional DG
is challenging due to the couplings of local voltage-related units, such as diesel generators, the supplied powers can be
matrices enforced by the positive semidefinite constraint of controlled, and they will be variables of the OPF problem; on
φ
the global voltage matrix. Results related to positive semidef- the other hand, PG,s and QφG,s represent committed powers
inite matrix completion [24] were leveraged in [6], [19] to for renewable-based sources.
φ
develop a distributed OPF algorithms for balanced networks Given the demanded loads {PL,n , QφL,n }, the goal is to
via dual decomposition. Using the results of [24], and tapping select a feasible set of voltages {Vnφ }, currents {Inφ }, and
into the powerful alternating direction method of multipliers powers supplied by conventional DG units {PG,s φ
, QφG,s } so
(ADMM) [25, Sec. 3.4], a distributed optimization problem that the steady-state operation of the microgrid is optimal in
for unbalanced microgrids is formulated here, where each LC a well defined sense. To this end, one of the following two
solves an optimization sub-problem, and then exchanges sim- objectives is usually pursued.
ple messages with its neighboring LCs. Compared to the dual i) Minimization of power losses.
decomposition schemes of [6], [19], the proposed ADMM- P The active power flowing
on line (m, n) ∈ E is Pm→n := φ∈Pmn Vmφ (Im→n φ
)∗ , where
based approach offers a markedly improved convergence.
In the OPF context, augmented Lagrangian methods (re- 1 Notation: Upper (lower) boldface letters will be used for matrices (column
lated to ADMM) were first used in [20], [21] to develop a vectors); (·)T for transposition; (·)∗ complex-conjugate; and, (·)H complex-
decentralized optimization scheme for (balanced) transmission conjugate transposition;
√ ℜ{·} denotes the real part, and ℑ{·} the imaginary
networks. Off-the-shelf schemes were used to solve the OPF part; j := −1 the imaginary unit. Tr(·) the matrix trace; rank(·) the matrix
rank; | · | denotes the magnitude of a number or the cardinality of a set; and
sub-problem associated with each sub-network. More recently, k · kF stands for Frobenius norm. Given a vector v and a matrix V, [v]P
ADMM was advocated in [26] for state estimation, and in [27] denotes a |P|×1 sub-vector containing entries of v indexed by the set P, and
for distributed multi-period OPF in balanced systems. Here, [V]P1 ,P2 the |P1 |×|P2 | sub-matrix with row and column indexes described
by P1 and P2 . Finally, 0M ×N and 1M ×N denotes M × N matrices with
the approaches of [20], [21], [27] are considerably broadened all zeroes and ones, respectively.
by considering unbalanced distribution networks, and by em- 2 The symbols defined throughout the paper are recapitulated in Table I.
IEEE TRANSACTIONS ON SMART GRID 3

φ TABLE I
Im→n is the complex current from node m to node n on phase
N OMENCLATURE
φ ∈ Pmn . Thus, the cost to be minimized here is [6]
X
C1 (V) = (Pm→n + Pn→m ) (1) N Set collecting the nodes of the microgrid
(m,n)∈E E Set collecting the distribution lines of the microgrid
S Subset of nodes featuring DG
where V := {I0φ , {Vnφ }, {Inφ }, {PG,n
φ
, QφG,n }} collects all the Nn Set of neighboring nodes of n ∈ N
Pn Set of phases at node n
steady-state variables. Pmn Set of phases of line (m, n)
ii) Minimization of the cost of supplied power. Letting the cost Vnφ Complex line-to-ground voltage at phase φ of node n
of power drawn from the PCC be denoted by c0 > 0, and the φ
In Complex current injected at phase φ of node n
φ
one incurred by the use of DG unit s ∈ S as cs ≥ 0, one can Im→n Complex current on phase φ of line (m, n)
φ
minimize the cost of supplied power [5] PL,n (Qφ
L,n ) Active (reactive) load demanded at node n on phase φ
φ
X φ φ X X φ PG,n (Qφ
G,n ) Active (reactive) power supplied at node n on phase φ
C2 (V) := c0 V0 (I0 )∗ + cs PG,s . (2) φ
yC,n Susceptance of a capacitor at node n and phase φ
φ∈P0 s∈S φ∈Ps Pm→n Active power exiting node m on line (m, n)
Zmn Phase impedance matrix of line (m, n)
Notice that (1) and (2) are equivalent when c0 = 1 and cs = 1 (s)
Ymn Shunt admittance matrix of line (m, n)
for all s ∈ S. Vnmin , Vnmax Utilization and service voltage magnitude limits
Based on (1)–(2), the following OPF problem is considered: min (Qmin )
PG,n G,n Minimum active (reactive) power supplied at node n
max (Qmax )
PG,n Maximum active (reactive) power supplied at node n
G,n
(P1) min Cm (V) (3a) A(ℓ) Set of nodes forming the area ℓ of the microgrid
V
φ φ
Ā(ℓ) Extended area ℓ comprising the nodes in A(ℓ) and the
s.t. Vsφ (Isφ )∗ = PG,s − PL,s + j(QφG,s − QφL,s ), nodes of different areas connected to A(ℓ) by a line
N̄ (ℓ) Set of neighboring areas of A(ℓ)
∀ φ ∈ Ps , s ∈ S (3b)
φ
Vnφ (Inφ )∗ = −PL,n − jQφL,n + jyC,n
φ
|Vnφ |2 ,
∀ φ ∈ Pn , n ∈ N \S (3c)
X h 1 
Inφ = (s)
Ymn + Z−1
mn [vn ]Pmn X 1 
2 (s)
m∈Nm [Y]Pn ,Pn := Ỹmn + Ỹmn (4)
i 2
− Z−1
mn [vm ]Pmn ∀ φ ∈ Pn , n ∈ N (3d) m∈Nn
{φ}
Vnmin ≤ |Vnφ | ≤ Vnmax , ∀ φ ∈ Pn , n ∈ N (3e) where Ỹmn := Z−1 mn if Pn = Pmn , otherwise
min φ max [Ỹmn ]Pnm ,Pnm := Z−1 mn and [Ỹmn ]Pn \Pnm ,Pn \Pnm = 0
PG,s ≤ PG,s ≤ PG,s , ∀ φ ∈ Ps , s ∈ S (3f) (s)
(Ỹmn is formed likewise).
φ
Qmin max
G,s ≤ QG,s ≤ QG,s , ∀ φ ∈ Ps , s ∈ S (3g) The next step consists in expressing the active and reactive
where Vnmin and Vnmax in (3e) are given minimum and powers injected per node, active powers flowing on the lines,
maximum utilization and service voltages; Nn := {j|(n, j) ∈ and voltage magnitudes, as linear functions of the outer-
E}; (3f)–(3g) are box constraints for the power supplied by product matrix V := vvH . To this end, define the following
conventional DG units, and m ∈ {1, 2} depending on the admittance-related matrix per node n and phase φ
chosen cost.
Ynφ := ēφn (ēφn )T Y (5)
Similar to OPF variants for transmission networks and
balanced distribution networks, (P1) is a nonlinear nonconvex where ēφn := [0T|P0 | , . . . , 0T|Pn−1 | , eφ,T T T T
Pn , 0|Pn+1 | , . . . , 0|PN | ] ,
problem because of the load flow equations (3b)-(3c) as well φ |Pn |
as the constraints (3e). In the next section, an equivalent and {ePn }φ∈Pn denotes the P canonical basis of R ,
reformulation of (P1) will be introduced, and its solution will and let the |Pmn | × n∈N |P n | matrices A m→n

be pursued using the SDP relaxation technique. and Bm be defined as Am→n := [0|Pmn |×Pm−1 |Pi | ,
i=0
Z−1 P , −Z−1 P
mn , 0|Pmn |× n−1
i=m+1 |Pi |
mn , 0|Pmn |× N i=n+1 |Pi |
] and
III. SDP-BASED C ENTRALIZED S OLUTION Bm := [0|Pmn |×Pm−1 |Pi | , I|Pmn | , 0|Pmn |×PN ],
i=0 i=m+1 |Pi |
P respectively. Then, a linear model in V can be established
Consider the n∈N |Pn | × 1 complex vectors v :=
[v0T , v1T , . . . , vN ] and i := [iT0 , iT1 , . . . , iTN ]T , with vn and
T T using the following lemma (see also [28] and [12]).
in the |Pn |×1 complex vectors collecting voltages {Vnφ }φ∈Pn Lemma 1: For the Hermitian matrices
and currents {Inφ }φ∈Pn per node n ∈ N . Voltages and injected
1 φ
currents abide by Ohm’s lawP i = Yv, where PY is a symmetric ΦφP,n := (Y + (Ynφ )H ) (6a)
block matrix of dimensions n∈N |Pn | × n∈N |Pn |, whose 2 n
j
entries are given by: ΦφQ,n := (Ynφ − (Ynφ )H ) (6b)
i) matrix −Z−1 2
mn occupying the |Pmn | × |Pmn | off-diagonal
block corresponding to line (m, n) ∈ E; and, ΦφV,n := ēφn (ēφn )T (6c)
ii) the |Pn | × |Pn | diagonal block corresponding to node 1
AH H


Φm→n := m→n Bm + Bm Am→n (6d)
n ∈ N with Nm := {n|(m, n) ∈ E} 2
IEEE TRANSACTIONS ON SMART GRID 4

O(max{Nc , n∈N |Pn |}4

P pP
voltage magnitudes and active as well as reactive powers are n∈N |Pn | log(1/ǫ)) for gen-
linearly related with V as eral purpose SDP solvers, with Nc denoting the total number
of constraints and ǫ > 0 a given solution accuracy [17].
Tr(ΦφV,n V) = |Vnφ |2 (7a) Notice however that the sparsity of {ΦφP,n , ΦφQ,n , ΦφV,n } and
Tr(ΦφP,n V) = φ
PG,n φ
− PL,n (7b) the so-called chordal structure of a matrix can be exploited
to obtain substantial computational savings; see e.g., [29]. In
Tr(ΦφQ,n V) = QφG,n − QφL,n + φ
yC,n Tr(ΦφV,n V) (7c)
contrast, gradient descent-based solvers for nonconvex pro-
Tr(Φm→n V) = Pm→n (7d) grams, sequential quadratic programming, and particle swarm
φ optimization, do not guarantee global optimality of the ob-
with PG,n = QφG,n = 0 for n ∈ N \S, and yC,n
φ
= 0 if
tained solutions and are sensitive to initialization. Here, global
capacitor banks are not present at node n.
optimality translates to lower distribution losses and increased
Proof. See the Appendix. 
monetary savings compared to sub-optimal OPF solutions.
Using Lemma 1, problem (P1) is equivalently reformulated Since (P3) is a relaxed version of (P2), Vopt could have rank
as follows: greater than 1. In this case, rank reduction techniques may
be employed to find a feasible rank-1 approximation of Vopt
(P2) min C̃m (V) provided it exists. For instance, the randomization technique
V
s.t. φ φ
Tr(ΦP,n V) + PL,n = 0, ∀ φ, ∀ n ∈ N \S (8a) offers a viable way to obtain a rank-1 approximation with
quantifiable approximation error; see e.g., [17] and references
Tr(ΦQ,n V) + QL,n − yC,n Tr(ΦφV,n V) = 0,
φ φ φ
therein. Albeit feasible for (P2), the resultant solution is
∀ φ, ∀ n ∈ N \S (8b) generally suboptimum [17]. For balanced tree distribution net-
works, [13] and [6] established conditions under which a rank-
min
PG,s ≤ Tr(ΦφP,s V) + PL,s
φ max
≤ PG,s , ∀ φ, ∀ n ∈ S (8c)
1 solution is attainable provided the original OPF problem
min φ φ max
QG,s ≤ Tr(ΦQ,s V) + QL,s ≤ QG,s , ∀ φ, ∀ n ∈ S (8d) is feasible. Unfortunately, when the tree power network is
(Vnmin )2 ≤ Tr(ΦφV,n V) ≤ (Vnmax )2 , ∀ φ, ∀ n ∈ N (8e) unbalanced, the results of [13] and [6] no longer apply, as
explained in the ensuing Section III-B. However, an intuitive
V  0, [V]P0 ,P0 = v0 v0H (8f) argument will be provided in Section III-B to explain why a
rank(V) = 1 (8g) rank-1 solution is expected even in the unbalanced setup. But
first, SDP-consistent constraints on line flows are derived in
where the costs C̃1 (V) and C̃2 (V) are re-expressed as
Section III-A, and a remark is provided.
X
C̃1 (V) = (Tr(Φm→n V) + Tr(Φn→m V)) (9a) Remark 1. Step-down or in-line three- or single-phase trans-
(m,n)∈E former banks (if any) can be accommodated in the formulated
X X optimization problems by using their series component mod-
C̃2 (V) = cs Tr(ΦφP,s V) . (9b)
els [14, Ch. 8], [15]. If a delta connection is employed on one
s∈S∪{0} φ∈Ps
side of the transformer, a small “dummy” resistance should
Problem (P2) is still nonconvex because of the rank-1 be included between the primary and the secondary sides (one
constraint (8g). Nevertheless, the SDP relaxation technique, per phase) in order to ensure that the matrix Vopt obtained by
which amounts to dropping the rank constraint [17], can be solving (P3) has rank 1; see also [12]. 
leveraged to obtain the following convex problem:
A. Constraints on line flows
(P3) min C̃m (V) (10)
V Constraints on the power dispelled on the distribution lines,
s.t. (8a) − (8f) . or, on the line current magnitudes are generally adopted to
protect conductors from overheating (which may eventually
Clearly, if the optimal solution Vopt of (P3) has rank 1, then it
trigger an outage event). Using Lemma 1, it turns out that the
is a globally optimal solution also for the nonconvex problem
real power dissipated on a line (m, n) ∈ E can be limited by
(P2). Further, since (P1) and (P2) are equivalent, there exists
H simply adding the constraint Tr(Φm→n V) + Tr(Φn→m V) ≤
a vector vopt so that Vopt = vopt vopt , and the optimal costs
∆Pmn in (P3), for a given maximum power loss ∆Pmn .
of (P1) and (P2) coincide at the optimum. This is formally φ max max
Consider now the constraint |Imn | ≤ Imn , with Imn
summarized next. φ
a given upper bound on the magnitude of Imn . Aiming
Proposition 1: Let Vopt denote the optimal solution of the to an SDP-consistent reformulation of this constraint, let
SDP (P3), and assume that rank(V√ opt ) = 1. Then, the vector of imn := [{Imn φ
}]T denote the |Pmn | × 1 vector collecting
line-to-ground voltages vopt := λ1 u1 , where λ1 ∈ R+ is the the complex currents flowing through line (m, n) ∈ E,
unique non-zero eigenvalue of Vopt and u1 the corresponding and notice that imn is related to voltages vm and vn as
−1
eigenvector, is a globally optimal solution of (P1).  mn = Zmn ([vm ]Pmn − [vn ]Pmn ). Next, define the |Pmn | ×
iP
The upshot of the proposed formulation is that a globally n∈N |Pn | complex matrix

optimal solution of (P2) (and hence (P1)) can be obtained Bmn := [0|Pmn |×Pm−1 |Pn | , Žm
mn , . . .
n=0
via standard interior-point solvers in polynomial time. In n P
0|Pmn |×Pn−1 |Pn | , Žmn 0|Pmn |× N ] (11)
fact, the worst-case complexity of (P3) is on the order n=m+1 n=n+1 |Pn |
IEEE TRANSACTIONS ON SMART GRID 5

where Žm mn is a |Pmn | × |Pm | matrix with elements

−1
[Žm m
mn ]Pmn ,Pmn = Zmn and [Žmn ]Pmn ,Pm \Pmn = 0; like-
n
wise, Žmn has dimensions |Pmn | × |Pn |, and its entries are
[Žnmn ]Pmn ,Pmn = −Z−1 n
mn and [Žmn ]Pmn ,Pn \Pmn = 0. Thus,
an SDP-compliant re-formulation of the constraint on the
current magnitude is possible as follows.
Lemma 2: Consider the Hermitian matrix
ΦφI,mn := BH φ φ T
mn emn (emn ) Bmn (12)
(a) Nonconvex problem. (b) Relaxed problem.
where {eφmn }φ∈Pmn denotes the canonical basis of R|Pmn | .
φ max Fig. 1. Line flow region and its Pareto optimal points.
Then, constraint |Imn | ≤ Imn can be equivalently re-
expressed as
Tr{ΦφI,mn V} ≤ (Imn
max 2
) . (13)

Following similar steps, and using (3d), constraints on the
magnitude of the injected currents {Inφ } can be derived too.
In unbalanced microgrids, it is of prime interest to protect
from overheating also the neutral cable(s) of the distribution
(ϕ)
lines [14]. Towards this end, let Pmn denote the set of (a) Power injected at 3 nodes. (b) Injection at 2 and 3 (P1 = −5).
grounded neutral cables that are present on the line (m, n) ∈
(ϕ)
E. Further, let Tmn the |Pmn |× |Pmn | be the neutral transfor- Fig. 2. Feasible power injection region and its Pareto optimal points.
mation matrix, which is obtained from the primitive impedance
matrix of line (m, n) via Kron reduction [14, Sec. 4.1]. Thus,
(ϕ) (1) |P ϕ | that the Pareto regions of the feasible power injections at
the neutral currents imn := [Imn , . . . , Imnmn ]T are linearly
(ϕ) the nodes of the nonconvex OPF (the balanced counterpart
related to the line currents imn as imn = Tmn imn . It readily
of (P2)) and the relaxed SDP (the balanced counterpart of
follows from Lemma 2, that the magnitude of the current on
(P3)) amounts to showing that: i) the region of feasible
the neutral cables can be constrained in (P3) as
powers F := {(Pm→m , Pm→m )}(m,n)∈E |(Pm→m , Pm→m ) ∈
(ϕ) (ϕ),max 2 (ϕ)
Tr{ΦI,mn V} ≤ (Imn ) , ∀ ϕ ∈ Pmn (14) Fmn ∀ (m, n) ∈ E} is the Cartesian product of the regions
{Fmn }(m,n)∈E ; and, ii) the region of the injected powers is
(ϕ) (ϕ) (ϕ) (ϕ)
with ΦI,mn,t := BH H T
mn Tmn emn (emn ) Tmn Bmn , {emn } the given by an affine transformation of F . Specifically, the first
(ϕ) (ϕ),max
canonical basis of R|Pmn | , and Imn a cap on the magnitude property i) follows from the fact that power flows on different
(ϕ) lines are decoupled; that is, it is possible to modify the angle
of Imn .
φ
θmn of a line (m, n), while preserving the angle difference
φ
θkl of any other line (k, l) 6= (m, n).
B. The rank conundrum
Suppose now that the off-diagonal elements of Zmn are
Sufficient conditions under which a rank-1 solution is not zero; that is, Z−1 mn 6= (gmn − jbmn )I|Pnm | . The to-
always obtained provided the original OPF problem is fea- tal power flowing from node m to node n is given by
sible were established in [13] and [6] for balanced distri- H −1
Pm→n = ℜ{vm Zmn (vm − vn )}, and it is now a function
bution networks with a tree topology. Balanceness implies φ φ,ρ
of {θmn }φ∈Pmn , as well as of the angle differences {θm :=
−1
that equal single-phase loads are served, Zmn = (gmn − φ ρ φ,ρ φ ρ
θm −θm }φ,ρ∈Pm and {θn := θn −θn }φ,ρ∈Pn . Different from
(s) (s)
jbmn )I|Pnm | and Ymn = jbmn I|Pnm | for each line (m, n) ∈ the balanced case, the power flows {Pm→n } are no longer
(s)
E, where gmn , bmn , bmn > 0. To recapitulate the broad decoupled across lines. In fact, it is impossible to adjust the
φ φ,ρ φ,ρ
outline of the proofs in [13] and [6], assume for simplic- angles {θmn , θm , θn } to obtain a new flow on line (m, n),
φ
ity that the shunt admittances are all zero. Then, the total without affecting the angle differences {θlm , θlφ,ρ , θm
φ,ρ
} and
φ φ,ρ φ,ρ
power flowing from node m to n is given by Pm→n = {θnk , θn , θk } for one of the other lines (l, m) and (n, k)
3|Vmφ |2 + 3|Vmφ ||Vnφ |(bmn sin(θmn
φ φ
) − gmn cos(θmn )), with connected to the nodes m and n, respectively. Thus, the results
φ φ φ
θmn := θm − θn the angle difference between voltages of [13] and [6] no longer apply in the unbalanced setup.
Vmφ and Vnφ . Since the network is balanced, θmn φ
is the An analytical characterization of the flow region F in
same on each phase φ ∈ Pmn . Fixing the voltage mag- the unbalanced case is challenging because of the number
nitudes, the region of the feasible powers (Pm→n , Pn→m ), of voltage angles involved and the aforementioned coupling
which is denoted as Fmn , becomes an affine transforma- of the line power flows. Nevertheless, the following simple
tion of the unit circles. Then, if one minimizes a strictly examples illustrate why one should expect a rank 1 solution
increasing function of the powers (Pm→n , Pn→m ), it follows from the relaxed OPF even in an unbalanced setup.
that the Pareto front of Fmn and the one of the convex Consider a 2-node unbalanced network, and suppose that a
hull of Fmn coincide if − tan−1 (bnm /gnm ) < θnm φ
< two-phase line connects the two nodes. Let Z12 = [(0.0753 +
−1
tan (bnm /gnm ) [13]. Based on this observation, proving j0.1181), (0.0156 + j0.0502); (0.0156 + j0.0502), (0.0744 +
IEEE TRANSACTIONS ON SMART GRID 6

(s)
j0.1211)] (as in [30]) and Ymn = 0, and assume that |Vmφ | = IV. D ISTRIBUTED S OLUTION
1 for n = 1, 2 and φ = a, b. Fig. 1(a) depicts the feasible Albeit polynomial, the computational complexity incurred
region F12 of powers (P1→2 , P2→1 ) for θ1a = 0◦ , θ12 a
∈ by standard interior-point solvers for SDP does not scale well
[−180 , 180 ], θ1 ∈ [110 , 130 ], and θ2 ∈ [110 , 130◦].
◦ ◦ ab ◦ ◦ ab ◦
with the number of nodes N , and the number of constraints
Notice that F12 is given by the Minkowski sum of the per- Nc [17]. However, this lack of scalability is incurred also
φ φ
phase regions {(P1→2 , P2→1 )}. It can be seen that the feasible by alternative methods based on off-the-shelf solvers for
flow region F12 (the dark gray area) is a perturbed ellipsoid; nonlinear programs [10]. Furthermore, the communication
specifically, expanding the expression of Pm→n , it follows overhead required to collect data from all end users at the
that θ1ab and θ2ab entail a perturbation of the center and of MEM, and subsequently disseminate the OPF solution back
the axes of the ellipsoid that would be obtained if the line to the LCs, may lead to traffic congestions and substantial
was balanced. Next, consider minimizing a strictly increasing delays in the data delivery. Therefore, a distributed SDP solver,
function C(P1→2 , P1→2 ) over F12 . It follows that the set of with minimal computational and communication costs, is well
Pareto optimal points are the ones represented by the red dots motivated. A distributed approach is also desirable in order to
in Fig. 1(a). Consider now the convex hull of F12 , which address possible concerns regarding data privacy and integrity,
amounts to augmenting the dark gray area with the light gray and when the microgrid includes single- or multi-facilities that
one shown in Fig. 1(b). Clearly, points belonging to the light are managed independently from the rest of the network in
gray area lead to a solution of (P3) with rank higher than 1. order to pursue specific economic interests [18].
Notably, the Pareto points of F12 and the ones of the relaxed Consider partitioning the microgrid into L areas {A(l) ⊂
region in Fig. 1(b) coincide; since C(·) is strictly increasing, N }Ll=1 , and suppose that each area is controlled by an LC.
the solution of (P3) must be on the Pareto boundary; thus, In the distributed SDP solver to be derived, each LC will
the optimal solution of (P3) has rank 1, and it is an optimal solve an OPF problem of reduced dimension for its controlled
solution also for the nonconvex problem (P2). Granted that the area. Let Ā(l) be an “extended” area defined as Ā(l) :=
Pareto points in Fig.s 1(a) and (b) are the same, [13, Lemma 5] A(l) ∪ {n|(m, n) ∈ E, m ∈ A(l) , n ∈ A(i) , l 6= j}; that
can be used to show that the two Pareto regions coincide also is, Ā(l) collects also the nodes belonging to different areas
when constraints on the voltages are involved. that are connected to A(l) by a distribution line. Sets {Ā(l) }
can be interpreted as counterparts of the “regions” considered
The region of injected powers for a 3-node network is ex- in [20] in the context of transmission systems. Based on
amined next. Let Z12 = Z23 = [(0.0753+j0.1181), (0.0156+ the overlaps among {Ā(l) }L l=1 , define the set of neighboring
j0.0502); (0.0156 + j0.0502), (0.0744 + j0.1211)], and as- areas for the l-th one as N̄ (l) := {j|Ā(l) ∩ Ā(j) 6= 0}.
sume that |Vnφ | = 1 for all n = 1, 2, 3 and φ = a, b. Finally, let the vector v̄(l) stack the complex line-to-ground
Finally, let Pn := Pna + Pnb , with Pnφ the power injected voltages of the nodes in Ā(l) (that is, {vn }n∈Ā(l) ), and let
at node n = 1, 2, 3 and phase φ = a, b. The gray area Φφ,l φ,l φ,l l (l)
:= v̄(l) (v̄(l) )H denote
P,n , ΦQ,n , ΦV,n , Φm→n and V
depicted in Fig. 2(a) corresponds to the region of feasible
−1 −1 the sub-matrices of ΦφP,n , ΦφQ,n , ΦφV,n , Φm→n and V, respec-
power injections I for − tan−1 (ℜ{[Z12 ]1,1 }/ℑ{[Z12 ]1,1 }) <
a −1 −1 −1 tively, formed by extracting rows and columns corresponding
θnm < tan (ℜ{[Z12 ]1,1 }/ℑ{[Z12 ]1,1 }) [13], and for the
to nodes in Ā(l) . With these notational conventions, it is
line-line angles θ1ab , θ2ab , θ3ab ∈ [110◦, 130◦ ]; specifically,
possible to re-write the SDP (P3) as
tan−1 (ℜ{[Z−1 −1 ◦
12 ]1,1 }/ℑ{[Z12 ]1,1 }) ≈ 58 . The angle differ-
ence θ3 is confined in the set [110 , 130◦], as higher values
ab ◦ L
X
(l)
are not likely to happen in practice [30]. Clearly, the gray (P4) min C̃m (V(l) ) (15a)
V
region is nonconvex. If one considers minimizing a strictly l=1
(l)
increasing function of the injected powers {P1 , P2 , P3 } over s.t. V ∈ B (l) , l = 1, . . . L (15b)
I, it follows that the Pareto optimal points are the ones color V0 (15c)
coded red in Fig. 2(a). It can be noticed that the Pareto region (l) (l)
does not change if one takes the convex hull of I; therefore, where B denotes the set of sub-matrices V satisfying the
the solution of (P3) has rank 1, and it is an optimal solution following constraints per area A(l) :
for (P2). Next, fix the power injected (or absorbed) at node Tr(Φφ,l (l) φ
∀ φ, ∀ n ∈ A(l) ∩ S (16a)
P,n V ) + PL,n = 0,
3, and consider minimizing a strictly increasing function of
P1 and P2 . The two-dimensional region of feasible powers P1 Tr(Φφ,l (l) φ φ φ,l (l)
Q,n V ) + QL,n − yC,n Tr(ΦV,n V ) = 0,
and P2 is depicted in gray in Fig 2(b). Again, the Pareto front ∀ φ, ∀ n ∈ A(l) \(A(l) ∩ S) (16b)
of the gray region and the one of its convex hull coincide. min
≤ Tr(Φφ,l (l) φ max (l)
PG,s P,s V ) + PL,s ≤ PG,s , ∀ φ, ∀ n ∈ S (16c)
φ,l (l) φ
These examples suggest that the nonconvex OPF problem Qmin max
G,s ≤ Tr(ΦQ,s V ) + QL,s ≤ QG,s , ∀ φ, ∀ n ∈ S (16d)
(P2) and its relaxed counterpart (P3) share their optimal so- (Vnmin )2 ≤ Tr(Φφ,l (l) max 2
V,n V ) ≤ (Vn ) , ∀ φ, ∀ n ∈ A(l) (16e)
lution when − tan−1 (ℜ{[Z−1 −1 a
mn ]φ,φ }/ℑ{[Zmn ]φ,φ }) < θnm <
−1 −1 −1
tan (ℜ{[Zmn ]φ,φ }/ℑ{[Zmn ]φ,φ }), and the angle differences with the additional constraint [V(l) ]P0 ,P0 = v0 v0H for the area
(l)
between conductors θm φρ
are small enough. This further mo- that contains the PCC. Further, C̃m (V(l) ) represents the cost
(l)
tivates efforts toward analytical characterization of the power associated with the area A ; that is, it collects the terms
injection region in unbalanced distribution systems. in (9a) (if m = 1) or (9a) (if m = 2) that pertain to A(l) . For
IEEE TRANSACTIONS ON SMART GRID 7

example, if one wishes to minimize the distribution losses, the Proposition 2: Under (As1) and (As2), the graph GN is
corresponding cost becomes chordal, meaning that each of its cycles comprising four or
(l)
X more nodes has a chord. Furthermore, all its maximal cliques
C̃1 (V(l) ) = Tr(Φlm↔n V(l) )
correspond to the elements of {V(l) }L l=1 . 
(m,n)∈E|m,n∈A(l)
X 1 As established in [24], the PSD matrix V is “completable”
+ Tr(Φlm↔n V(l) ) (17) if and only if GN is chordal, and all its submatrices corre-
2
(m,n)∈E|m∈A(l) ,n∈Ā(l) \A(l) sponding to the maximal cliques of GN are PSD. Therefore,
constraining V to be PSD is tantamount to enforcing the
with Φlm↔n := Φlm→n + Φln→m . On the other hand, the
(l) constraint on all local matrices V(i)  0, ∀l = 1, . . . , L.
expression for C̃2 (V(l) ) is not unique, depending on the
Notice that (As1) requires the macro graph GA to be a tree,
specific network topology and operational setup. For instance,
while no conditions are imposed on the topology of the
if ones assumes that each area is formed by one lateral or
(l) microgrid. Thus, it may be possible to find network partitions
one sub-lateral [18], then C̃2 (V(l) ) accounts for the cost of with an associated chordal graph GN also in the case of
power flowing into the (sub-)lateral from the “backbone” of weakly-meshed microgrids.
the microgrid, and the cost of power generated within the Next, let Plj :=
S
n∈(Ā(l) ∩Ā(j) ) Pn collect the indexes
(sub-)lateral. Instead, if customers are allowed to use DG
(l) corresponding to the voltages {{Vnφ }φ∈Pn } of the nodes that
units only to satisfy their own needs, C̃2 (V(l) ) boils down
(l) (l) φ the extended areas Ā(l) and Ā(j) share. For example, if areas
to C̃2 (V(l) ) = c0 C̃1 (V(l) ) + s∈A(l) ∩S cs φ∈Ps PG,s
P P
. 1 and 2 share nodes n = 5 and n = 6, then P12 indexes the
Either way, the equivalent formulation (18) effectively ex- voltages {V5φ }φ∈P5 and {V6φ }φ∈P6 . Further, define as Vj
(l)
presses the cost as the superposition of local costs, and divides
the submatrix of V(l) collecting the rows and columns of V(l)
network constraints on a per-area basis. However, even with
corresponding to the voltages in Pl,j . With these definitions,
such a decomposition the main challenge lies in the PSD
and assuming that (As1) and (As2) hold, problem (P4) can be
constraint (15c) that couples local matrices {V(l) }. Indeed,
re-written in the following equivalent form:
if all submatrices {V(l) } were non-overlapping, the PSD
L
constraint on V would simplify to V(l)  0 per area l, and X
(l)
(P4) would be decomposable in L sub-problems. However, this (P5) min C̃m (V(l) ) (18a)
{V(l) }
l=1
equivalence fails to hold here, since submatrices share entries
of V. The idea is to identify valid network topologies (that is, s.t. V(l) ∈ B (l)
, l = 1, . . . L (18b)
valid partitions of the microgrid in smaller areas) for which (l) (j) (l)
Vj = Vl , j ∈ N̄ , l = 1, . . . , L (18c)
the PSD constraint decomposition is feasible. (l)
To this end, define the following two auxiliary graphs: V  0, l = 1, . . . L (18d)
i) a “macro” graph GA , where nodes represent the areas and where constraint (18c) enforces neighboring areas to consent
edges are defined by the neighborhood sets {N̄ (l) }L l=1 ; and, on the entries of V(l) and V(j) that they have in common.
ii) a “micro” graph GN induced by the sub-matrices Clearly, constraints (18c) couple the optimization problems
{V(l) }L
P
l=1 ; that is, a graph with n∈N |Pn | nodes (one per across areas. To enable a fully distributed solution, con-
phase and node), with an edge connecting the nodes represent- (l)
sider introducing the auxiliary variables {Wj }j∈N̄ (l) and
ing the voltages Vnφ and Vmθ if the entry of V corresponding to (l)
{Xj }j∈N̄ (l) per area. With these auxiliary variables, (P5) can
Vnφ (Vmθ )∗ is contained in one of the sub-matrices {V(l) }L l=1 . be equivalently re-stated as
Examples of macro graphs are provided in Figs. 3 and 4.
L
Based on these auxiliary graphs, results on completing X
(l)
partial Hermitian matrices will be leveraged to obtain PSD (P6) min C̃m (V(l) ) (19a)
{V(l) 0}
ones [24]. These results rely on the so-termed chordal property l=1

of the graph GN induced by {V(l) }L l=1 to establish the equiv-

s.t. V(l) ∈ B (l)
, l = 1, . . . L (19b)
alence between positive semidefiniteness of the overall matrix (l) (j)
ℜ{Vj } = Wl , j ∈ N̄ (l) , l = 1, . . . , L (19c)
V and that of all submatrices corresponding to the graph’s (l) (j) (l)
maximal cliques. Towards decomposing the PSD constraint ℑ{Vj } = Xl , j ∈ N̄ , l = 1, . . . , L (19d)
into local ones, the following assumptions are made, which (j) (l) (l)
Wl = Wj , j ∈ N̄ , l = 1, . . . , L (19e)
naturally suggest valid partitions of the microgrid: (j) (l)
Xl = Xj , j ∈ N̄ (l) , l = 1, . . . , L. (19f)
(As1) The graph GT A is a tree; and,
(As2) |Ā(l) \(Ā(l) Ā(i) )| > 0 for all i, l = 1, . . . , L, i 6= l; A similar approach was followed by [19], which utilized
that is, no nested extended areas are present. either primal or dual iterations to distribute the OPF in bal-
Condition (As1) is quite reasonable in tree distribution net- anced transmission networks. A distributed OPF for balanced
works; for example, an area can be formed by a pair of nodes distribution feeders was derived in [6], where the dual (sub-
that are connected by a distribution line [6], or by laterals )gradient ascent was used. Unfortunately, sub-gradient ascent
and sub-laterals [18]. (As2) is a technical condition ensuring methods do not always lead to a satisfactory solution; when
that the subgraph induced by V(l) is a maximal clique of the dual function is non-differentiable and the step size is
GN , which allows using the results of [24]. Based on these fixed, dual and primal iterates converge only on the average.
assumptions, the following can be readily proved. What is more, recovering the primal variables from the optimal
IEEE TRANSACTIONS ON SMART GRID 8

dual variables is not always guaranteed [31, Sec. 5.5.5]. Using Lemma 3, steps [S1]–[S3] can be simplified as
Besides dealing with unbalanced power networks, the novelty follows. Furthermore, convergence to the solution of the
here consists in solving (P6) distributedly by resorting to centralized SDP (P3) is established.
the ADMM [25, Sec. 3.4], a powerful scheme that has been (l) (l)
Proposition 3: If Γj (0) = Λj (0) = 0|Plj | , then [S1]–
successfully applied to distributed optimization and estimation
(l) (l) [S3] boil down to the following primal-dual updates:
in several contexts. To this end, let {Γj } and {Λj } be
the multipliers associated with (19c) and (19d), respectively, [S1′ ] Update V(l) (i + 1) per area l = 1, . . . , L as:
and consider the partial quadratically-augmented Lagrangian (P7(l) ) min L(l) (V(l) , {αj , βj }) (25a)
of (19) as V(l) 0,{αj ≥0,βj ≥0}
L n
(l)
L({V(l) }, {Wj }, {Xj }, {Γj }, {Λj }) =
(l) (l) (l)
X (l)
C̃m (V(l) ) s.t. V(l) ∈ B (l) (25b)
−αj rTj,ℜ
 
l=1
X h  (l) T   0, ∀j ∈ N̄ (l) (25c)
+
(l) (j)
Tr (Γj ) (ℜ{Vj } − Wl ) rj −I
−βj rTj,ℑ
 
j∈N̄ (l)
   0, ∀j ∈ N̄ (l) (25d)
(l) (l) (j)
+Tr (Λj )T (ℑ{Vj } − Xl ) rj,ℑ −I
κ (l) (j) κ (l) (j)
io
where the local Lagrangian (25a) is given by
+ kℜ{Vj } − Wl k2F + kℑ{Vj } − Xl k2F (20)
2 2 X κ
L(l) (V(l) , {αj , βj }) := C̃m
(l)
(V(l) ) +
(αj + βj )
2
j∈N̄ (l)
where κ ∈ R+ is a positive constant [25, Sec. 3.4]. Then, the X h  (l) (l)
 
(l) (l)
i
ADMM amounts to iteratively performing the following steps + Tr (Γj (i))T ℜ{Vj } + Tr (Λj (i))T ℑ{Vj }
(i denotes the iteration index): j∈N̄ (l)

[S1] Update primal variables: and the vectors rj,ℜ and rj,ℑ collect the real and imagi-
(l)
 of the entries of the matrix Vj −
(l) nary
{V (i + 1)} :=  parts, respectively,
1 (l) (j)
arg min L({V(l) }, {Wj (i)}, {Xj (i)},
(l) (l)
2 Vj (i) + Vl (i) .
{V(l) 0
(l) (l) [S2′ ] Update dual variables locally per area l = 1, . . . , L:
{Γj (i)}, {Λj (i)})
(l) κ
(l) (l) (j)

Γj (i + 1) = Γj (i) + ℜ{Vj (i + 1)} − ℜ{Vl (i + 1)
s.t. V(l) ∈ B (l) , l = 1, . . . , L. (21) 2
(26)
[S2] Update auxiliary variables: (l) (l) κ (l) (j)

Λj (i + 1) = Λj (i) + ℑ{Vj (i + 1)} − ℑ{Vl (i + 1)
(l) (l) 2
{Wj (i + 1), Xj (i + 1)} := (27)
(l) (l)
arg min L({V(l) (i + 1)}, {Wj }, {Xj }, Furthermore, for any κ > 0 the iterates {V(l) (i)},
(l) (l)
{Wj ,Xj }
(l) (l)
(l) (l)
{Γj (i), Λj (i)} produced by [S1′ ]–[S2′] are convergent, and
{Γj (i)}, {Λj (i)}) (l) (l)
limi→+∞ V(l) (i) = Vopt for all l = 1, . . . , L, with {Vopt }
(j) (l) (j) (l)
s.t. Wl = Wj , Xl = Xj , j ∈ N̄ (l) , l (22) sub-matrices of the optimal solution Vopt of (P3).
Proof. See the Appendix. 
[S3] Update dual variables:
(l) (l)
Γj (i + 1) = Γj (i) + κ(ℜ{V(l) (i + 1)} − Wl (i + 1)) (23)
(j) At each iteration, the LC of area l receives from the LCs
(j)
(l) (l) (l) (j) of its neighboring areas j ∈ N̄ (l) matrices Vl (i), and
Λj (i + 1) = Λj (i) + κ(ℑ{V (i + 1)} − Xl (i + 1)) (24) (l) (l)
updates the local multipliers {Γj (i), Λj (i)} via (26)–(27).
In step [S1], the per-area matrices {V(l) (i)} are obtained by These multipliers are locally stored at area l, and they are not
(l) (l)
minimizing (20), where variables {Wj (i + 1), Xj (i + 1)} exchanged among LCs (in contrast, multipliers are exchanged
and the multipliers are kept fixed to their previous iteration per iteration in [19]). Then, LC l updates V(l) (i + 1) by
(l)
values. Likewise, the auxiliary variables are updated in [S2] solving (P7(l) ), and transmits Vj (i + 1) to its neighboring
by fixing {V(l) (i + 1)} to their up-to-date values. Finally, the (l)
areas j ∈ N̄ .
dual variables are updated in [S3] via dual sub-gradient ascent.
Interestingly, the ADMM iterations [S1]–[S3] can be sim- V. N UMERICAL T ESTS
plified by exploiting the favorable decomposability of the
The SDP-based solver is tested here using the following two
Lagrangian. To this end, the following lemma is first needed.
networks, operating in a grid-connected mode:
(l)
Lemma 3: If the multipliers are initialized as Γj (0) = • the IEEE 37-node test feeder [30] shown in Fig. 3; and
(l)
Λj (0) = 0|Plj | , then for every pair of neighboring areas • the 10-node 3-phase network depicted in Fig. 4.
(l) (j)
l and j it holds that Γj (i) + Γl (i) = 0|Plj | for each The optimization package CVX3 , along with the interior-point
(l) (j)
i ≥ 1. Likewise, Λj (i) + Λl (i) = 0|Plj | for each i ≥ 1 if based solver SeDuMi [32] are employed to implement the
(l) (l) centralized and distributed solvers in MATLAB.
Λj (0) = Λj (0) = 0|Plj | .
Proof. See the Appendix.  3 [Online] Available: http://cvxr.com/cvx/
IEEE TRANSACTIONS ON SMART GRID 9

160

140

120

100

Ps [kW]
80

60
DG 1
DG 2
40 DG 3
DG 4
DG 5
20
DG 6
DG 7
0
0 0.25 0.5 0.75 1 1.25
c /c
s 0

Fig. 5. Power generated by the DG units [kW].

each phase. The voltage magnitude at the PCC is 4.16 kV. With
these two choices, the performance of the proposed approach
can be assessed for different network sizes, line characteristics,
loading, and different network partitions shown next.
The minimum and maximum utilization and service voltages
are set to Vnmin = 0.95 pu and Vnmax = 1.05 pu for all nodes.
Thus, voltage regulation is enforced without requiring changes
in the voltage regulator taps (as in e.g., [15]). Further, the
voltage at the PCC is set to v1 = [1∠0◦ , 1∠−120◦, 1∠120◦]T
pu. The average computational time required by SeDuMi to
Fig. 3. First Test: IEEE 37-node feeder.
solve the centralized problem (P3) was 9.0 sec and 0.3 sec
(machine with Intel Core i7-2600 CPU @ 3.40GHz), which
is significantly lower than the time required by commercial
solvers for non-linear programs (see e.g., [10]).
TABLE II
R ESULTS FOR THE IEEE 37- NODE FEEDER

cs /c0 Ploss [kW] C2 [$] P0 [MW] PG [MW] rank(Vopt )

0 39.30 57.9 1.4463 1.0500 1

0.25 39.30 68.4 1.4463 1.0500 1
Fig. 4. Second test: (a) 10-node microgrid; (b) corresponding graph GA . 0.50 39.30 78.9 1.4463 1.0500 1
0.75 38.97 89.3 1.4469 1.0491 1
1 36.60 99.7 1.5979 0.8957 1
1.25 57.22 102.6 2.3088 0.2054 1
The 4.8 kV network of Fig. 3 is an actual portion of
power a distribution network located in California; all the
demanded complex powers are “spot” loads, and the network TABLE III
loading is very unbalanced [30]. Compared to the original R ESULTS FOR THE 10- NODE NETWORK IN F IG . 4
scheme however, 7 DG units are placed at nodes S =
{10, 12, 16, 19, 26, 32, 36}. Specifically, single-phase conven- cs /c0 Ploss [kW] C2 [$] P0 [MW] PG [MW] rank(Vopt )
tional DG units supply a maximum real power of 50 kW per
0 18.38 59.9 1.4984 0.3000 1
phase, and are operated at a unit power factor (PF); that is, 0.25 18.38 62.9 1.4984 0.3000 1
Qmin max
G,s = QG,s = 0 for all s ∈ S. Line impedances and shunt 0.50 18.38 65.9 1.4984 0.3000 1
admittances are computed based on the dataset in [30]. Finally, 0.75 18.38 68.9 1.4984 0.3000 1
1 18.27 71.9 1.5457 0.2526 1
delta-wye conversions are performed whenever necessary. As 1.25 23.08 72.1 1.8031 0.0000 1
for the network of Fig. 4(a), the line admittances are all set
to Zmn = [0.0693 + j0.2036, 0.0312 + j0.1003, 0.0316 + Tables
P II and III list the real power drawn at the PCC
j0.0847; 0.0312 + j0.1003, 0.0675 + j0.2096, 0.0307 + P0 := φ∈P0 V0φ (I0φ )∗ , the total power generated by the DG
P P φ
j0.0770; 0.0316 + j0.0847, 0.0307 + j0.0770, 0.0683 + units PG := s φ PG,s , and the overall power losses Ploss
j0.2070] Ω (see [30]), which gives rise to an unbalanced and costs of supplied power, when the cost (2) is employed.
operation of the network. Shunt admittances are neglected. The costs of supplied power are set to c0 = 40 $/MW and
Single-phase conventional DG units are placed at nodes S = c1 = . . . = c7 ∈ {0, 10, 20, 30, 40, 50} $/MW. Notice that
{5, 7}, and they can supply a maximum real power of 50 kW minimizing (2) is tantamount to minimizing the power loss (1)
per phase, at unit PF. All the loads are “spot”, and the loading when c0 = 1 and cs = 1 for all s ∈ S. In fact, it can be clearly
is assumed balanced. Specifically, the active and reactive seen in Tables II and III that the power loss is minimized for
loads are set to {0, 50, 0, 130, 130, 110, 110, 0, 90, 90, } kW this choice of {cs }. Powers Ploss , P0 , and PG remain the same
and {0, 20, 0, 82, 82, 60, 60, 0, 30, 30, } kVAr, respectively, on when cs ≤ c0 ; however, when cs ≥ c0 (which holds during
IEEE TRANSACTIONS ON SMART GRID 10

1 1
the PCC from 1 to 1.02 pu. This further prompts an analytical
10 10
characterization of the feasibility region of (P1). Additional
tests were performed on the IEEE 13-node feeder [30], and
0
10 0
10 rank-1 matrices Vopt were again always obtained (results are
Sub−gradient,
s = 0.1/i
not reported here due to space limitation).
−1
10 −1
10
Sub−gradient,
s = 0.1 Convergence of the proposed distributed SDP solver is
showcased in Fig. 6, where the 37-node feeder is partitioned
ADMM, κ = 1
as shown in Fig. 3. This partition resembles the case where
ADMM, κ = 10
∆ Vi,j

i,j
−2 −2
laterals include multi-facilities that are managed independently
∆V
10 10

from the rest of the network [18]. The trajectories correspond-

(j) (l)
−3
10 −3
10
ing to ∆Vl,j (i) := kVl (i) − Vj (i)k1 /36 per iteration i,
are reported for different values of the ADMM parameter κ,
−4 −4
and are compared with the ones obtained by using the sub-
10 10
∆ V1,2 ∆ V1,2 gradient ascent-based distributed algorithm developed in [6] (s
∆V ∆ V1,3
∆V
1,2 denotes the step size, which is assumed to be either constant, or
1,3 ∆ V1,4

0 50 100 150 200 250 300 0 50 100 150 200 250 300
monotonically decreasing). It can be noticed that the proposed
Iteration index Iteration index
distributed solver exibit a considerably faster convergence
than the one based on the sub-gradient. Considering that
Fig. 6. Convergence of the ADMM, with feeder partitioned as in Fig. 3. the nominal voltage of the feeder is 4.8 kV, in less then
than 50 iterations the average gap between the entries of
(j) (l)
1 1
Vl (i) and Vj (i) is on the order of a few volts. Further,
10 10

Sub−grad, s = 0.01
Sub−grad, s = 0.01 notice that the convergence rate is approximately the same for
Sub−grad, s = 10
∆V1,2 (i), ∆V1,3 (i), and ∆V1,4 (i). The average computational
10
0
10
0
time required by SeDuMi to solve each sub-problem was
of 6, 3, 0.3, and 2.7 sec. Lower computational times can be
Sub−grad, s = 0.1

10
−1
Sub−grad, s = 10/i 10
−1
Sub−grad, s = 10
obtained by selecting areas of smaller size.
Sub−grad, s = 0.1
Sub−grad, s = 10/i Fig. 7 illustrates the convergence of the distributed SDP
solved when the considered 10-node network is partitioned as
1,2

∆ V2,3

−2 −2
10 10
shown in Fig 4. The ADMM-based method outperforms the
∆V

one based on the sub-gradient here too. In this case, the gaps
(j)
10
−3

ADMM, κ = 1
10
−3 between the elements of matrices {Vl (i)} rapidly vanish
ADMM, κ = 1

ADMM, κ = 10 ADMM, κ = 10
when κ = 100 and κ = 10 after approximately 100 iterations.
−4
ADMM, κ = 100 −4 ADMM, κ = 100
In this case, the average time required by SeDuMi to solve
10 10
each sub-problem was 0.1 sec.

0 50 100 150 200 250 300 0 50 100 150 200 250 300 VI. C ONCLUDING REMARKS
Iteration index Iteration index

The OPF problem was considered for microgrids oper-

ating in an unbalanced setup. Inspite of the inherent non-
Fig. 7. Convergence of the ADMM, network partitioned as shown in Fig. 4. convexity, the SDP relaxation technique was advocated to
obtain a convex problem. As corroborated by numerical tests,
the late-night hours) the DG units reduce the generated active the main contribution of the proposed approach consists in
powers; further, the power loss becomes significantly higher offering the potential to obtain the globally optimal solution
when DG units are not used at the maximum extent. Fig. 5 of the original nonconvex OPF. A distributed SDP solver was
depicts the active powers generated by DG units as a function also developed by resorting to the ADMM. The distributed
of cs /c0 , for the IEEE 37-node test feeder. It can be noticed algorithm ensures scalability with respect to the microgrid
that the DG units electrically close to the PCC are not utilized size, robustness to communication outages, and preserves data
when cs ≥ c0 ; on the other hand, DG 6 and 7 still operate at privacy and integrity.
more than 50% of their maximum capacity.
Interestingly, the rank of matrix Vopt was always 1. There- A PPENDIX
fore, the globally optimal solutions of (P2) (and hence of Proof of Lemma 1. Equality (7a) can be readily established
the nonconvex (P1)) were always attained. In other words, by noticing that |Vnφ |2 = vH ēφn (ēφn )T v = Tr(ēφn (ēφn )T V).
no lower power losses or costs of supplied power can be To prove (7b), notice first that the injected apparent power at
attained with alternative OPF solution approaches. The rank node n and phase φ is given by Vnφ (Inφ )∗ = (Vnφ∗ Inφ )∗ =
of matrix Vopt was greater than 1 for the IEEE 37-node test (vH ēφn (ēφn )T i)H . Next, using i = Yv, it follows that
feeder when cs /c0 > 1.75 (two non-zero eigenvalues, with (vH ēφn (ēφn )T i)H = (vH ēφn (ēφn )T Yv)H = (vH Ynφ v)H =
λ1 ∝ 102 and λ2 ∝ 10−2 ). Nevertheless, a rank-1 solution vH (Ynφ )H v, which can be equivalently rewritten as Tr(Ynφ V).
was readily obtained upon raising the voltage magnitude at Thus, the injected real and reactive powers can be obtained
IEEE TRANSACTIONS ON SMART GRID 11

by using, respectively, the real and imaginary parts of (Ynφ )H . [6] A. Y. Lam, B. Zhang, A. Dominguez-Garcia, and D. Tse, “Optimal
Finally, to prove (7d), define first the |Pmn | × 1 vector im→n distributed voltage regulation in power distribution networks,” 2012,
[Online] Available at http://arxiv.org/abs/1204.5226v1.
collecting the complex currents flowing from m to n on each [7] E. Sortomme and M. A. El-Sharkawi, “Optimal power flow for a system
−1
phase, and notice that im→n = Zmn (vm − vn ) = Am→n v of microgrids with controllable loads and battery storage,” in IEEE/PES
and vm = Bm v. Then, it follows that Pm→n = ℜ{iH m→n v} =
Power Systems Conference and Exposition, Seattle, WA, Mar. 2009.
[8] M. R. Irving and M. J. H. Sterling, “Efficient Newton-Raphson algorithm
ℜ{vH AH H
m→n Bm v} = ℜ{Tr(Am→n Bm V)}. for load-flow calculation in transmission and distribution networks,”
Proc. Inst. Elec. Eng. – Gener., Transm. Distrib., vol. 134, no. 5, pp.
Proof of Lemma 3 The proof is provided for the multipliers 325–330, Sept. 1987.
(l) (l)
Γj ; same steps can be followed for the dual variables Λj . [9] F.-C. Lu and Y.-Y. Hsu, “Fuzzy dynamic programming approach to
Suppose that i = 1, and notice that the optimiza- reactive power/voltage control in a distribution substation,” IEEE Trans.
Power Syst., vol. 12, no. 2, pp. 681–688, May 1997.
tion problem (22) decouples per pair of neighboring ar- [10] S. Paudyaly, C. A. Canizares, and K. Bhattacharya, “Three-phase dis-
eas; specifically, a number of sub-problems is to be tribution OPF in smart grids: Optimality versus computational burden,”
solved with respect to (wrt) the only pair of variables in 2nd IEEE PES Intl. Conf. and Exhibition on Innovative Smart Grid
(j) (l) Technologies, Manchester, UK, Dec. 2011.
(Wl , Wj ) for j ∈ N̄ (l) (and clearly l ∈ N̄ (j) ). Ne- [11] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming
glecting irrelevant terms, and setting the first-order deriva- for optimal power flow problems,” Int. J. of Electrical Power & Energy
(l) (l) (l) (l) Systems, vol. 30, no. 6–7, pp. 383–392, 2008.
tive of L({V(l) (1)}, {Wj }, {Xj }, {Γj (0)}, {Λj (0)}) [12] J. Lavaei and S. H. Low, “Zero duality gap in optimal power flow
(j) (l)
wrt (Wl , Wj ) to zero, it readily follows that the minimizer problem,” IEEE Trans. Power Syst., vol. 1, no. 1, pp. 92–107, 2012.
(l) [13] J. Lavaei, D. Tse, and B. Zhang, “Geometry of power flows in tree
Wj (1) is given by networks,” in IEEE Power Ener. Soc. Gen. Meet., San Diego, CA, Jul.
2012.
(l) 1  (l) (j)

[14] W. H. Kersting, Distribution System Modeling and Analysis. 2nd ed.,
Wj (1) = Vj (1) + Vl (1) . (28) Boca Raton, FL: CRC Press, 2007.
2
[15] S. Paudyal, C. A. Canizares, and K. Bhattacharya, “Optimal operation
(l) of distribution feeders in smart grids,” IEEE Trans. on Ind. Electron.,
Next, substituting (28) into [S3], and setting Γj (0)
 = vol. 58, no. 10, pp. 4495–4503, Oct. 2011.
(l) (l) κ (l) (j)
Γj (0) = 0 yields Γj (1) = 2 Vj (1) − Vl (1) and [16] S. Bruno, S. Lamonaca, G. Rotondo, U. Stecchi, and M. L. Scala,
  “Unbalanced three-phase optimal power flow for smart grids,” IEEE
(j) (j) (l)
Γl (1) = κ2 Vl (1) − Vj (1) . It then follows by induc- Trans. on Ind. Electron., vol. 58, no. 10, pp. 4504–4513, Oct. 2011.
(l) (j)
[17] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite
tion that Γj (i) = −Γl (i) also for all subsequent iterations. relaxation of quadratic optimization problems,” IEEE Sig. Proc. Mag.,
vol. 27, no. 3, pp. 20–34, May 2010.
Proof of Proposition 3 Substituting (28) into [S3], and [18] J. Guerrero, P. Loh, M. Chandorkar, and T. Lee, “Advanced control
(l) (l) architectures for intelligent microgrids - part I: Decentralized and
setting Γj (0) = Γj (0) = 0, one can readily obtain (26)–
hierarchical control,” IEEE Trans. on Ind. Electron., vol. 60, no. 4, pp.
(27). Consider now [S1′ ]. First, substitute (28) into (20), and 1254–1262, 2013.
discard irrelevant terms. Next, introduce an auxiliary variable [19] Y. Lam, B. Zhang, and D. Tse, “Distributed algorithms for op-
αj to upper bound the squared l2 norm of rj,ℜ , and add timal power flow problem,” Nov. 2011, [Online] Available at
http://arxiv.org/abs/1109.5229.
the constraint krj,ℜ k22 ≤ αj . Thus, (25d) follows by Schur’s [20] B. H. Kim and R. Baldick, “Coarse-grained distributed optimal power
complement. Indeed, it is certainly possible to introduce one flow,” IEEE Trans. Power Syst., vol. 12, no. 2, pp. 932–939, May 1997.
single variable α to upper bound the squared l2 norm of [21] R. Baldick, B. H. Kim, C. Chase, and Y. Luo, “A fast distributed
implementation of optimal power flow,” IEEE Trans. Power Syst.,
the vector [rT1,ℜ , rT1,ℑ , . . . , rT|N̄ (l) |,ℜ , rT|N̄ (l) |,ℑ ]T , and derive a vol. 14, no. 3, pp. 858–864, Aug. 1999.
linear matrix inequality similar to (25d) by using Schur’s [22] G. Hug-Glanzmann and G. Andersson, “Decentralized optimal power
complement. Finally, since the cost in (P6) is convex, and flow control for overlapping areas in power systems,” IEEE Trans. Power
Syst., vol. 24, no. 1, pp. 327–336, Feb. 2009.
the constraint set is bounded and convex, every limit point of [23] F. J. Nogales, F. J. Prieto, and A. J. Conejo, “A decomposition method-
{V(l) (i)} is an optimal solution to (P6), as established in [25, ology applied to the multi-area optimal power flow problem,” Ann. Oper.
Prop. 4.2]. Since (P6) and (P3) are equivalent under (As1) and Res., no. 120, pp. 99–116, 2003.
[24] R. Grone, C. R. Johnson, E. M. Sá, and H. Wolkowicz, “Positive
(As2) hold, every limit point of {V(l) (i)} is also an optimal definite completions of partial Hermitian matrices,” Linear Algebra and
solution to (P3). its Applications, vol. 58, pp. 109–124, 1984.
[25] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa-
tion: Numerical Methods. Englewood Cliffs, NJ: Prentice Hall, 1989.
R EFERENCES [26] V. Kekatos and G. B. Giannakis, “Distributed robust power system state
estimation,” IEEE Trans. Power Syst., vol. 28, Feb. 2013, see also Proc.
[1] N. Hatziargyriou, H. Asano, R. Iravani, and C. Marnay, “Microgrids: of Globecom Conf., Annaheim, CA, USA, Dec. 3-7, 2012.
An overview of ongoing research, development, and demonstration [27] M. Kraning, E. Chu, J. Lavaei, and S. Boyd, “Message passing
projects,” IEEE Power & Energy Mag., vol. 5, no. 4, pp. 78–94, July– for dynamic network energy management,” 2012, [Online]. Available:
Aug. 2007. http://arxiv.org/pdf/1204.1106.pdf.
[2] E. Tedeschi, P. Tenti, and P. Mattavelli, “Synergistic control and coop- [28] H. Zhu and G. B. Giannakis, “Estimating the state of AC power systems
erative operation of distributed harmonic and reactive compensators,” in using semidefinite programming,” in Proc. of North American Power
IEEE Power Elec. Spec. Conf., Rhodes, Greece, Jun. 2008. Symposium, Boston, MA, Aug. 2011.
[3] S. Bolognani and S. Zampieri, “A distributed control strategy for reactive [29] R. A. Jabr, “Exploiting sparsity in SDP relaxations of the OPF problem,”
power compensation in smart microgrids,” 2013, to appear. See also IEEE Trans. Power Syst., vol. 2, no. 27, pp. 1138–1139, May 2012.
http://arxiv.org/pdf/1106.5626. [30] W. H. Kersting, “Radial distribution test feeders,” in IEEE Power
[4] P. Carvalho, P. Correia, and L. Ferreira, “Distributed reactive power Engineering Society Winter Meeting, vol. 2, 2001, pp. 908–912.
generation control for voltage rise mitigation in distribution networks,” [31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
IEEE Trans. Power Syst., vol. 23, no. 2, pp. 766–772, May 2008. University Press, 2004.
[5] D. Forner, T. Erseghe, S. Tomasin, and P. Tenti, “On efficient use of [32] J. F. Sturm, “Using SeDuMi 1.02, a Matlab toolbox for optimization over
local sources in smart grids with power quality constraints,” in Proc. of symmetric cones,” Optimization Methods Software, vol. 11–12, no. 1,
IEEE Int. Conf. on Smart Grid Comm., Gaithersburg, MD, Oct. 2010. pp. 625–653, Aug. 1999.