238 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO.

3, SEPTEMBER 2015

A Distributed Approach for the Optimal

Power-Flow Problem Based on ADMM
and Sequential Convex Approximations
Sindri Magnússon, Pradeep Chathuranga Weeraddana, Member, IEEE, and Carlo Fischione, Member, IEEE

Abstract—The optimal power-flow (OPF) problem, which plays A. Previous Work

a central role in operating electrical networks, is considered. The
problem is nonconvex and is, in fact, NP hard. Therefore, design- The problem was originally presented by Carpentier [1], and
ing efficient algorithms of practical relevance is crucial, though has been extensively studied since then and become of great im-
their global optimality is not guaranteed. Existing semidefinite portance in efficient operation of power systems [2]. The OPF
programming relaxation-based approaches are restricted to OPF problem is nonconvex due to a quadratic relationship between
problems where zero duality holds. In this paper, an efficient
novel method to address the general nonconvex OPF problem the powers and the voltages and because of a lower bound on
is investigated. The proposed method is based on an alternating the voltage magnitudes. In fact, the problem is NP-hard, see [3].
direction method of multipliers combined with sequential convex Therefore, practical and general-purpose algorithms must rely
approximations. The global OPF problem is decomposed into on some approximations or heuristics. We refer the reader to [2]
smaller problems associated with each bus of the network, the and [4] for a contemporary survey of OPF.
solutions of which are coordinated via a light communication
protocol. Therefore, the proposed method is highly scalable. The It is well known that the OPF problem is equivalently
convergence properties of the proposed algorithm are mathemat- reformulated as a rank-constrained problem [5]. As a result,
ically substantiated. Finally, the proposed algorithm is evaluated classic convex approximation techniques are applied to handle
on a number of test examples, where the convergence properties nonconvexities of the rank constraint, which usually results
of the proposed algorithm are numerically substantiated, and the in a semidefinite program [6] (SDP). SDP relaxations to OPF
performance is compared with a global optimal method.
have gained a lot of attention recently, see [3], [7]–[9], and
Index Terms—Distributed optimization, optimal power flow references therein. The authors in [3] show that SDP relaxation
(OPF), smart grid. is equivalent to the dual problem of the original OPF. Moreover,
sufficient conditions for zero duality and mechanisms to recover
I. I NTRODUCTION the primal solution by the dual problem are given. Thus, [3]
identifies a class of OPF problems, where the global optimum
T HE optimal power-flow (OPF) problem in electrical net-
works optimally determines the amount of power to be
generated at each generator. Moreover, it decides how to dis-
is obtained efficiently by using convex tools. Some other classes
of OPF problems, for which zero duality holds, are investi-
patch the power such that a global network-wide objective gated in [7]–[9]. In particular, [7] derives zero duality results
criterion is optimized, while ensuring that the power demand for networks that have tree topology where oversatisfaction
of each consumer is met and that the related laws of physics of loads is allowed. On the other hand, [8] and [9] provide
are held. Traditionally, the OPF problem has only been solved graphically intuitive conditions for the zero duality gap for two-
in transmission networks. However, the extensive information bus networks, which are later generalized to tree topologies.
gathering of individual power consumption in the smart grid has The aforementioned references suggest that the applicability
made the problem relevant, not only in transmission networks, of SDP approaches is limited to special network classes. Of
but also in distribution networks which deliver electricity to course, SDP relaxations can always be used to compute a lower
end users. bound on the optimal value of the primal problem. However,
in practice, what is crucial is a network operating point. In
general, SDP relaxations fail to provide a network operating
point (i.e., a feasible point) due to nonzero duality gap [10].
Another drawback of SDP-based methods is that even when
zero duality holds, if the objective functions are nonquadratic,
Manuscript received January 18, 2014; revised August 14, 2014; accepted
December 19, 2014. Date of publication April 20, 2015; date of current version the dual machinery employed in constructing primal feasible
September 14, 2015. This work was supported in part by the Wireless@KTH solutions is not applied. The authors in [10] and [11] explore
project DevelopLTE4Grids and in part by EIT project LTE4SE. Recommended limitations of SDP approaches and give practical examples
by Associate Editor M. Chertkov.
The authors are with the Electrical Engineering School, Access Linnaeus where the sufficient conditions for zero duality do not hold.
Center, KTH Royal Institute of Technology, Stockholm 10 100 44, Sweden Centralized methods for OPF problem, of course, exhibit
(e-mail: sindrim@kth.se; chatw@kth.se; carlofi@kth.se). poor scalability. On the contrary, distributed and scalable OPF
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. solution methods are less investigated, though they are highly
Digital Object Identifier 10.1109/TCNS.2015.2399192 desirable in the context of rapidly growing real-world electrical

2325-5870 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 239

networks. Unlike centralized methods, distributed OPF solution neously general and scalable, are of crucial importance in the-
methods are also appealing in the context of privacy and se- ory as well as in practice and, therefore, deserve investigations.
curity, because they do not entail collecting possibly sensitive
problem data at a central node. In other words, when solving
B. Our Contributions
in a centralized manner the OPF problem in the smart grid,
the power companies must rely on private information, such as The main contributions of this paper are as follows.
the load profile of their costumers [12], [13], which might be 1) We develop a distributed algorithm for the general non-
of interest to a third party. For example, government agencies convex OPF problem. Our approach is not restricted to
might inquire the information to profile criminal activity, and any special classes of networks, where zero duality holds.
insurance companies might be interested in buying the infor- It also handles nonquadratic convex objective functions,
mation to determine if an individual is viable for insurance [14]. unlike the SDP-based distributed algorithms.
Therefore, the gathering of private information at a centralized 2) We capitalize on the alternating direction method of
node has raised serious concerns about personal privacy which, multipliers (ADMM) [29] to accomplish the distributed
in turn, discourages the use of centralized approaches. Inter- implementation (among electrical network buses) of the
estingly, the sparsity of most electrical networks brings out an proposed algorithm with a little coordination of the neigh-
appealing decomposition structure and, therefore, it is worth boring entities. Thus, the proposed algorithm is highly
investigating distributed methods for the OPF problem. scalable, which is favorable in practice.
Distributed methods for the OPF problem are first studied 3) In the case of subproblems, we capitalize on sequential
in [15]–[17], where the transmission network is divided into approximations, in order to gracefully manipulate the
regions and different decomposition methods, including the nonconvexity issues. The approach is adopted from an
auxiliary problem principle, predictor-corrector proximal mul- existing algorithm originally proposed in [30] in the
tiplier method, and alternating direction method, are explored context of the centralized OPF problem.
to solve the problem distributively among these regions. The 4) The convergent properties of the proposed algorithm are
formulation is restricted to 2-region network decompositions, mathematically and numerically substantiated.
and border variables cannot be shared among more than two 5) A number of numerical examples are provided to evaluate
regions. Another approach to decentralize the problem into the performance of the proposed algorithm.
regions is presented in [18]–[20]. The method is based on solv-
ing the Karush–Kuhn–Tucker (KKT) optimality conditions,
C. Organization and Notations
where a Newton procedure is adapted. The authors provide a
sufficient condition for convergent which can be interpreted as a This paper is organized as follows. Section II describes the
measurement of coupling between regions. However, when the system model and problem formulation. The solution method
condition is not satisfied, they rely on the generalized minimal is presented in Section III. In Section IV, we discuss some
residual method to find the Newton direction, which involves a fundamental properties of the algorithm. Numerical results are
lot of communications between entities. The methods presented provided in Section V. Finally, Section VI concludes this paper.
√
in [21] are limited to dc OPF. The imaginary unit is denoted by j, that is, j = −1.
More recent distributed algorithms are found in [22]–[28]. Boldface lowercase and uppercase letters represent vectors and
The decentralized methods in [22]–[24] capitalize on the SDP matrices, respectively, and calligraphic letters represent sets.
relaxation, which still has the drawbacks of being specific to The cardinality of A is denoted by |A|. len(x) denotes the
special classes of networks and a lack of flexibility with general length of x. The set of real and complex n-vectors are denoted
objective functions. Another relaxation method is presented in by IRn and Cn , respectively, and the set of real and complex
[25], where instead of the original nonconvex constraint sets, m × n matrices are denoted by IRm×n and Cm×n . We denote
the convex hull of those are used. However, the method can the real and imaginary parts of the complex number z ∈ C by
result in an infeasible point to the original unrelaxed problem, Re(z) and Im(z), respectively. The set of non-negative integers
entailing local methods to help construct good feasible points. is denoted by N, that is, N = {0, 1, . . .}. The superscript (·)T
Other recent works consider distributed methods for optimal stands for transpose. We use parentheses to construct column
reactive power flow in distribution networks [26], [27]. Both vectors from comma separated lists, for example, (a, b, c) =
papers first make approximations that yield a convex OPF [aT bT cT]T. We denote the diagonal block matrix with
problem and then distribute the computation by using dual de- A1 , · · · , AN on the diagonal by diag(A1 , · · · , AN ). The
composition [26] and ADMM [27]. The recent work in [28] em- Hadamard product of the matrices A and B is denoted by
ploys ADMM to the general nonconvex OPF problem to devise A ◦ B. We denote by x2 the 2 -norm of the vector x. We
a scalable algorithm. A major drawback of the method in [28] is denote the gradient of the function f in the point x by ∇x f .
that its convergence is very sensitive to the initialization. In fact,
the authors of [28] always initialize their algorithm with a point
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
which is close to the optimal solution. However, the optimal
solution is not known a priori, limiting the scope of the method. Consider an electrical network with N buses with N =
Almost all of the aforementioned methods can be classified {1, 2, . . . , N } denoting the set of buses and L ⊆ N × N rep-
as those which are general yet not scalable and those which are resenting the set of flow lines. We let ik = iRe Im
k + jik be the
Re Im
scalable yet not general. However, methods, which are simulta- current injection and vk = vk + jvk be the voltage at bus
240 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

k ∈ N . Let pD k +jqk ∈ C and pk +jqk ∈ C denote the complex

D G G
where the variables are pG , qG , p, q, iRe , iIm , vRe ,
power demand and the complex power generated by bus k ∈ N , vIm , and iRe ls , ils , pls , qls for (l, s) ∈ L. Here, constraint
Im

respectively. Thus, the complex power pk + jqk ∈ C injected to (2b) is from i = Yv, (2c) is derived from the conservation
bus k is given by pk + jqk = (pG k + jqk ) − (pk + jqk ). ls = Re(yls (vl − vs )),
G D D
of power flow holds, (2e) is from iRe
For notational compactness, we let p , q , p , qD , p,
G G D
(2f) is from the complex power being (v ◦ i∗ ), and iIm ls =
q, i, iRe , iIm , v, vRe , and vIm denote the vectors (pG k )k∈N , Im(yls (vl − vs )) with cls = (gls , −gls , bls , −bls ) and dls =
(qkG )k∈N , (pD k ) k∈N , (q D
k ) k∈N , (p k ) k∈N , (q k ) k∈N , (i k )k∈N , (bls , −bls , −gls , gls ), and (2f) is from pls = Re(vl i∗ls ) and
(iRe
k ) k∈N , (i Im
k ) k∈N , (v k ) k∈N , (v Re
k ) k∈N , and (v Im
k ) k∈N , re- qls = Im(vl i∗ls ). Note that (2b)–(2f) correspond to the con-
spectively. We denote by iRe ls + ji Im
ls ∈ C the complex current straints imposed by the laws of physics associated with the
and by pls + jqls ∈ C the complex power transferred from bus electrical network. In addition, (2g)–(2l) correspond to the
l to the rest of the network through the flow line (l, s) ∈ L. The constraints imposed by operational limitations, where the lower
admittance matrix Y ∈ CN ×N of the network is given by bound problem data (·)min and the upper bound problem data
⎧  (·) max determine the boundaries of the feasible regions of
⎨ yll + (l,t)∈L ylt , if l = s, power, current, as well as voltages in the network. Note that if a
Y = −yls , if (l, s) ∈ L, (1)
⎩ bus k is not a generator bus, then there is no power generation at
0, otherwise that bus and, thus, pG G
k + jqk = 0. Such situations can be easily
modeled by letting
where yls = gls + jbls ∈ C is the admittance in the flow line
(l, s) ∈ L, and yll = gll + jbll ∈ IC is the admittance to ground
at bus l. We let G ∈ IRN ×N and B ∈ IRN ×N denote the real pG,min
k = pG,max
k = qkG,min = qkG,max = 0, k ∈ N \G. (3)
and imaginary parts of Y, respectively. In particular, [G]ls =
gls and [B]ls = bls yield Y = G + jB. The constraints (2e), (2f), and (2l) are nonconvex, which,
in turn, makes problem (2) nonconvex. In fact, the problem is
NP-hard [3]. Thus, it hinders efficient algorithms from achiev-
A. Centralized Formulation ing optimality. However, in the sequel, we design an efficient
For fixed power demands, pD and qD , the goal of the OPF algorithm to address problem (2) in a decentralized manner.
problem is to find the optimal way to tune the variables pG ,
qG , p, q, iRe , iIm , vRe , and vIm , ensuring that the relation-
ships among the variables are held and system limitations are B. Distributed Formulation
respected. The objective function differs between applications. In this section, we derive an equivalent formulation of
In this paper, we consider the minimization of a convex cost problem (2), where all of the constraints, except for a single
function of real power generation. We denote by fkG , the cost of consistency constraint, are decoupled among the buses. In
generating power at bus k ∈ G, where G ⊆ N denotes the set particular, the resulting formulation is in the form of a general
of generator buses. The OPF problem can now be expressed as1 consensus problem [29, Sec. 7.2], where the fully decentralized
   implementation can be realized, without any coordination of a
min fkG pGk (2a) central authority. More generally, the proposed formulation can
k∈G be easily adapted to accomplish decoupling among subsets of
s.t. iRe + jiIm = GvRe − BvIm + j(BvRe + GvIm ) (2b) buses, each of which corresponds to buses located in a given
 G 
pk + jqk = pG k − pk + j q k − q k , k ∈ N ,
D D
(2c) area, for example, multiarea OPF [15].
  We start by identifying the coupling constraints of problem
iRe
ls +ji Im
ls = c T
ls +jd T
ls (v Re Re Im Im
l , v s , v l , v s ), (l, s)∈L
(2). From constraint (2b), note that the current injection of
(2d)
  each bus is affected by the voltages of its neighbors and by
p+jq = vRe ◦ iRe +vIm ◦ iIm +j vIm ◦ iRe −vRe ◦ iIm its own voltage. Therefore, constraint (2b) introduces coupling
(2e) between neighbors. To decouple constraint (2b), we let each
pls + jqls = vlRe iRe ls + v Im Im
l i ls + j(v Im Re
l i ls − v Re Im
l i ls ) bus maintain local copies of the neighbors’ voltages and then
(l, s) ∈ L (2f) enforce them to agree by introducing consistency constraints.
To formally express the idea from before, we first denote
pG,min ≤ pG G,max
k ≤ pk ,k ∈ N, (2g)
k
by Nk the set of bus k itself and its neighboring buses, thsat
qkG,min ≤ qkG ≤ qkG,max , k ∈ N , (2h) is, Nk = {k} ∪ {n|(k, n) ∈ L}. Copies of real and imaginary
ls ) + (ils ) ≤ (ils ) , (l, s) ∈ L,
(iRe 2 Im 2 max 2
(2i) parts of the voltages corresponding to buses in Nk are denoted
pls + qls ≤ (sls ) , (l, s) ∈ L,
2 2 max 2
(2j) by vkRe ∈ IR|Nk | and vkIm ∈ IR|Nk | , respectively. For notational
Re) Im)
|pls | ≤ pmax
ls , (l, s) ∈ L, (2k) convenience, we let (vk 1 = vkRe and (vk 1 = vkIm . We refer
 min 2  Re 2  Im 2 Re Im
to v and v as real and imaginary net variables, respec-
vk ≤ vk + vk ≤ (vkmax )2 , k ∈ N (2l)
tively. Note that the copies of either the net variable vkRe or vkIm
are shared among |Nk | entities, which we call the degree of net
1 Formulation (2) is equivalent to the OPF formulation in [3], and one can
variable vkRe or vkIm . The consistency constraints are given by
easily switch between the two formulation by using simple transformations.
We use formulation (2), because it is convenient, in terms of notations, when
describing the content in subsequent sections. vkRe = Ek vRe , vkIm = Ek vIm (4)
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 241

where Ek ∈ IR|Nk |×N is given by (6f), (6g), (6j), and by γ k (zk ) ≤ 0 the nonlinear convex in-
  equality constraints (6h) and (6i) as we will see next.2
1, if vkRe l is a local copy of vsRe Now we can express the distributed formulation of problem
(Ek )ls = (5)
0, otherwise. (2) as
Note that (4) ensures the agreement of the copies of the net   
min fkG pG
k (10a)
variables and that for any bus k, either vkRe or vkIm is local in
k∈G
the sense that they depend only on neighbors. 
The constraints (2b)–(2l) of problem (2) can be written s.t. zk = pG G Re Im Re Im
k , q k , pk , q k , ik , ik , x k , x k ,
by using local variables vkRe and vkIm . In particular, we can 
īRe Im
k , īk , p̄k , q̄k , k ∈ N, (10b)
equivalently list them as follows:
 T Re  (αk (zk ), λk (zk ), μk (zk )) = 0, k ∈ N, (10c)
k + jik = gk vk − bk vk + j bk vk + gk vk
iRe Im T Re T Im T Im
(6a)
 G  (β k (zk ), γ k (zk )) ≤ 0, k ∈ N, (10d)
pk + jqk = pk − pk + j qk − qk
G D D
(6b)
  (vkmin )2r ≤ (vkRe )2r + (vkIm )2r ≤ (vkmax )2r ,
īk +j īk = Ck vk +Dk vk +j Dk vk −Ck vk
Re Im Re Im Re Im
(6c)
 Im Re 
pk +jqk = (vk )1 ik +(vk )1 ik +j (vk )1 ik −(vk )1 iIm
Re Re Im Im Re
k r = 1, . . . , |Nk |, k ∈ N, (10e)
(6d) vkRe + jvkIm = Ek v Re
+ jEk v Im
, k∈N (10f)
 Im Re 
p̄k +j q̄k = (vk )1 īk +(vk )1 īk +j (vk )1 īk −(vk )1 īIm
Re Re Im Re Re
k
(6e) where the variables are pG G Re Im Re Im
k , q k , pk , q k , ik , ik , v k , v k ,
pkG,min
≤ pk ≤ pk
G G,max
, (6f) īk , īk , p̄k , q̄k , zk for k ∈ N and v and v . Note that
Re Im Re Im

(10f) establishes the consistency constraints [cf. (4)], which

qkG,min ≤ qkG ≤ qkG,max , (6g) affirms the consistency among neighbor voltages. The coupling
k )r + (īk )r ≤ (īk
(īRe )r , r = 1, . . . , |Nk | − 1
2 Im 2 max 2
(6h) in the original centralized formulation (2) has been subsumed
(p̄k )r + (q̄k )r ≤ (s̄k )r , r = 1, . . . , |Nk | − 1
2 2 max 2
(6i) in the consistency constraint (10f), which results in the form of
|(p̄k )r | ≤ (p̄max
k )r , r = 1, . . . , |Nk | − 1, (6j) general consensus problem [29, Sec. 7.2], where decomposition
(vk )r ≤ (vk )r +(vkIm )2r ≤ (vkmax )2r , r = 1, . . . |Nk |
min 2 Re 2
(6k) methods can be gracefully applied.

where k ∈ N , īRe Re Im Im
k = (ikl )l∈Nk \{k} , īk = (ikl )l∈Nk \{k} , p̄k =
(pkl )l∈Nk \{k} , and q̄k = (qkl )l∈Nk \{k} , with the order kept pre- III. D ISTRIBUTED S OLUTION M ETHOD
served as in (vkRe )1:|Nk | and (vkIm )1:|Nk | . In addition, gk (or In this section, we present our distributed algorithm to the
bk ) in constraint (6a) are obtained by first extracting the kth OPF problem (10). In particular, we use the ADMM as the basis
column of G (respectively, B) and then extracting the rows for our algorithm development, where we have fast conver-
corresponding to the buses in Nk , where the order of the gence properties, compared to the dual decomposition [29]. The
components are preserved as in vkRe and vkIm . In addition, Ck ∈ ADMM is also promising in the sense that it works on many
IR(|Nk |−1)×|Nk | and Dk ∈ IR(|Nk |−1)×|Nk | in constraint (6c) are nonconvex problems as a good heuristic [29, Sec. 9]. Once the
given by solution method is established, we investigate the properties in
⎛ (g ) −(g ) · · · 0 ⎞ Section IV.
k 2 k 2
⎜ (gk )3 0 ··· 0 ⎟
Ck = ⎜
⎝ .. .. .. .. ⎟
⎠ (7)
. . . . A. Outline of the Algorithm
(gk )|Nk |−1 0 ··· −(gk )|Nk | For notational simplicity, we let vk and Ēk denote
⎛ (bk )2 −(bk )2 ··· 0 ⎞ (vkRe , vkIm ) and diag(Ek , Ek ), respectively, for each k ∈ N .
⎜ (bk )3 0 ··· 0 ⎟ Moreover, we let v denote (vRe , vIm ). The ADMM essen-
Dk = ⎜
⎝ .. .. .. .. ⎟ . (8)
⎠ tially minimizes the augmented Lagrangian associated with the
. . . . problem in an iterative manner. Particularized to our problem
(bk )|Nk |−1 0 ··· −(bk )|Nk | (10), the partial augmented Lagrangian with respect to the
consistency constraints (10f) (i.e., vk = Ēk v) is given by
Moreover, vkmin , vkmax , īmaxk , s̄max
k , and p̄max
k of constraints
(6f)–(6k) are chosen in a straightforward manner [cf. (2g)–(2l)].    G  G
Lρ pG , (vk )k∈N , v, (yk )k∈N = f k pk
Finally, for notational convenience, associated with each bus
k∈G
k, we denote by  ρ 
  + ykT (vk − Ēk v) + vk − Ēk v22 (11)
zk = pG G Re Im Re Im Re Im
k , qk , pk , qk , ik , ik , vk , vk , īk , īk , p̄k , q̄k (9) k∈N
2

the local variables of bus k, by αk (zk ) = 0 the affine con-

straints (6a)–(6c), by λk (zk ) = 0 the nonlinear equality con- 2 The functions βk and γk depend on pG,min , pG,max , qkG,min , qkG,max ,
k k
straint (6d), by μk (zk ) = 0 the nonlinear equality constraint īmax , s̄max max
and p̄k
, which are intentionally omitted for clarity and space
k k
(6e), by β k (zk ) ≤ 0 the linear convex inequality constraints reasons.
242 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

where yk is the dual variable associated with (10f), and ρ is

called the penalty parameter. Together, with the separability of
(11) among k ∈ N , the steps of ADMM are formally expressed
below.

Algorithm 1: ADMM for distributed OPF (ADMM-DOPF)

(n)
1) Initialization: Set n = 0 and initialize yk and v(n) .
(n) Fig. 1. Feasible set of ((vkRe )r , (vkIm )r ) where A = (vkmax )r and
2) Private variable update: Set yk = yk and v = v(n) .
B = (vkmin )r . (a) Xrk ; (b) X̌rk ; (c) X̌rk .
Each bus k ∈ N updates xk locally, where we let
(n+1) (n+1)
(zk , uk ) be the primal and dual (possibly) opti-
of each bus k (Ek v − vk 2 ) is below a predefined threshold;
mal variables achieved for the following problem:
3) running the ADMM-DOPF algorithm until the objective
ρ value decrement between two successive iterations is below a
min fkG (pG k ) + yk (vk − Ēk v)+ ||vk − Ēk v||2
T 2
(12a)
2 predefined threshold. In the sequel, we discuss in detail the
G G Re Im Re Im Re Im
s.t. zk = (pk , qk , pk , qk , ik , ik , vk , vk īk , īk , p̄k , q̄k ) algorithm steps (12)–(14).
(12b)
αk (zk ) = 0 (12c)
λk (zk ) = 0 (12d) B. Subproblems: Private Variable Update
μk (zk ) = 0 (12e) In this section, problem (12) is considered. Since Problem
β k (zk ) ≤ 0 (12f) (12) is NP-hard, only exponentially complex global methods
γ k (zk ) ≤ 0 (12g) can guarantee its optimality. We capitalize on sequential convex
(vkmin )2r ≤ (vkRe )2r + (vkIm )2r ≤ (vkmax )2r approximations [31] to design an algorithm, which is efficient
r = 1, . . . , |Nk | (12h) compared to global methods. Similar techniques are used in
[30] for centralized OPF, which we use as the basis for design-
where the variables are pG G Re Im Re Im
k , q k , pk , q k , ik , ik , v k , v k ,
ing our subproblem algorithm.
Re Im
īk , īk , p̄k , q̄k , and zk . We denote by vk
(n+1)
, the part We start by noting that constraints (12b), (12c), (12f), and
(n+1) Re Im (12g) are convex as opposed to constraints (12h), (12d), and
of zk corresponding to (vk , vk ).
(12e), which are clearly nonconvex. The idea is to approximate
3) Net variable update: We let v(n+1) be the solution to the
the nonconvex constraints.
problem
In the case of (12h), we note that for any r ∈ {1, . . . , |Nk |},
   ρ 2
(n+1)  (n+1)  the values of (vkRe )l and (vkIm )l represent a donut, see Fig. 1(a).
min ykT vk − Ēk v + vk − Ēk v (13)
2 2 In other words, the 2-D set
k∈N
  
where the variable is v. Xrk = (vkRe )r , (vkIm )r ∈ IR2 (vkmin )2r ≤ (vkRe )2r + (vkIm )2r

4) Dual variable update: Each bus k ∈ N updates its dual ≤ (vkmax )2r
variable yk as
  is a donut, which is clearly nonconvex.
(n+1) (n) (n+1)
yk = yk + ρ vk − Ēk v(n+1) . (14) We approximate the nonconvex set Xrk by considering a
convex subset of Xrk instead, which we denote by X̌rk , see
5) Stopping criterion: Set n := n+1. If the stopping criterion Fig. 1(b). To do this, we simply consider the hyperplane tangent
is not met, go to step 2, otherwise STOP and return (z(n) , to the inner circle of the donut at the point C in Fig. 1(b).
(n) (n) (n)
v(n), u(n), y(n)) = ((zk )k∈N , v(n), (uk )k∈N , (yk )k∈N ). Specifically, given a point ((v̌kRe )r , (v̌kIm )r ) ∈ Xrk , X̌rk is the
intersection of Xrk and the halfspace
 Re       
The first step initializes the net and dual variables. In the (vk )r , (vkIm )r ∈ IR2  ar vkRe r +br vkIm r ≥ cr (15)
second step, each bus solves a nonconvex optimization problem
in order to update its private variable (see Section III-B). where
In the third step, the net variable is updated by solving the 
  2
unconstrained quadratic optimization problem (13), which has a 
  vkmin r
ar = sign v̌k r 
Re
    2
closed-form solution. The net variable update can be performed 1 + v̌kIm r / v̌kRe r
in a distributed fashion with a light communication protocol   
(see Section III-D). The fourth step is the dual variable update, v̌kIm r  2
br = a r  Re  , cr = vkmin r
which can be performed locally on each bus (see Section III- v̌k r
D). The fifth step is the stopping criterion. Natural stopping
criterions include: 1) running the ADMM-DOPF algorithm for if (v̌kRe )r = 0 and
a fixed number of iterations; 2) running the ADMM-DOPF al-    
gorithm until the decrement between the local and net variables ar = 0, br = sign v̌kIm r , cr = vkmin r
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 243

if (v̌kRe )r = 0. In the case of nonlinear nonconvex constraints optimization problem and the third step is the stopping crite-
(12d) and (12e), we capitalized on the well-known Taylor’s rion. A natural stopping criterion is to run the algorithm until
ẑk the decrement between two successive iterations is below a
approximation. Specifically, given a point ẑk , we denote by λ̂k (m+1) (m)
the first-order Taylor’s approximation of λk at ẑk . Similarly, we certain predefined threshold, that is, zk − zk  <  for
denote by μ̂ẑkk , the first-order Taylor’s approximation of μk at a given  > 0. However, since zk only depends on vk , the
ẑk . The approximation is refined in an iterative manner until a component related to (vkRe , vkIm ), we use
 
stopping criterion is satisfied.  (m+1) (m) 
ẑk vk − vk  < sub (17)
It is worth noting that to construct the functions μ̂ẑkk and λ̂k , 2

one only needs the values of v̂k , where v̂k is the component of (m) (m+1) (m)
where vk and vk are the components of zk and
ẑk corresponding to (vkRe , vkIm ).3 (m+1)
By using the constraint approximations discussed before, zk , respectively, corresponding to the variable vk and
we design a subroutine to perform step 2 of the ADMM- sub > 0 is a given threshold. Furthermore, we do not need
DOPF algorithm. The outline of this successive approximation to reach the minimum accuracy in every ADMM iteration, but
algorithm is given as follows. only as the ADMM method progresses. Therefore, it might be
practical to set an upper bound on the number of iterations,
that is
Algorithm 2: Subroutine for step 2 of the ADMM-DOPF
m ≥ max_iter (18)
1) Initialize: Given v and yk from ADMM-DOPF nth it-
eration. Set (vRe , vIm ) = v. For all r ∈ {1, . . . , |Nk |}, for some max_iter ∈ N.
set ((v̌kRe )r , (v̌kIm )r ) = ((Ek vRe )r , (Ek vIm )r ) and con-
struct X̌rk . Let m = 1 and initialize ẑk .
2) Solve the approximated subproblem C. On the Use of Quadratic Programming (QP) Solvers
  ρ Problem (16) can be efficiently solved by using general
min fkG pG k + yk (vk − Ēk v) + vk − Ēk v2
T 2
2 interior-point algorithms for convex problems. However, even
(16a) higher efficiencies are achieved if problem (16) can be handled
 G G by specific interior-point algorithms. For example, if the objec-
s.t. zk = pk , qk , pk , qk , iRe Im Re Im
k , ik , v k , v k , tive function (16a) is quadratic, sophisticated QP solvers can be
 easily employed. See the Appendix for details.
īRe Im
k , īk , p̄k , q̄k (16b)
αk (zk ) = 0 (16c)
D. Net Variables and Dual Variable Updates
ẑk
λ̂k (zk ) = 0 (16d) Note that the net variable v(n+1) is the unique solution of the
μ̂ẑkk (zk ) =0 (16e) unconstrained convex quadratic optimization problem (13), and
is given by
β k (zk ) ≤ 0 (16f)  −1
   
γ k (zk ) ≤ 0 (16g) (n+1) T T (n+1) 1 (n)
v = Ēk Ēk Ēk vk + yk
 Re   Im   ρ
vk r , vk r ∈ X̌rk , r = 1, . . . , |Nk | (16h) k∈N k∈N

(19)
where the variables are pG G Re Im Re
k , q k , pk , q k , ik , ik , v k ,  −1
vkIm , īRe , ī Im
, p̄ , q̄ , and z . The solution corresponding   (n+1)
k k k k k
(m)
= ĒT
k Ēk ĒT
k vk (20)
to the variable zk is denoted by zk and all of the dual k∈N k∈N
optimal variables are denoted by uk .
(m) ⎛ −1  −1 ⎞
3) Stopping criterion: If stopping criterion is not met, set  
(m)
= diag ⎝ ET
k Ek , ET
k Ek
⎠
ẑk = zk , m := m + 1 and go to step 2. Otherwise, k∈N k∈N
(m) (m)
STOP and return (zk , uk ).  
 Re(n+1)
 Im(n+1)
× ET
k vk , ET
k vk (21)
k∈N k∈N
The initialization in the first step is done by setting v̂k =   
1 1 Re(n+1)
(v̌kRe , v̌kIm ), where v̂k is the component of ẑk corresponding = diag ,..., ET
k vk ,
to the variable (vkRe , vkIm ). The rest of the vector ẑk is then |N1 | |NN |
k∈N
 
initialized according to (6a)–(6e), which have a unique solution
vRe(n+1)
when v̂k is given. The second step involves solving a convex   
1 1 Im(n+1)
diag ,..., ET
k vk (22)
|N1 | |NN |
k∈N
3 Thisfollows directly from the definition of the first-order Taylor approxi-  
mation and (6a) and (6c). vIm(n+1)
244 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

where (19) follows trivially from the differentiation of the objec-

tive function of problem (13), (20) follows byinvoking the op-
timality conditions for problem (13), that is, k∈N Ēk Tyk = 0,
 follows from Ēk = diag(Ek , Ek ), and (22) follows from
(21)
k∈N Ek TEk = diag(|N1 |, . . . , |NN |). From (22), it is not
difficult to see that any net variable component update is equiv-
alently obtained by averaging its copies maintained among the
neighbor nodes. Such averaging can be accomplished by using
fully distributed algorithms, such as gossiping [32]. Therefore,
step 3 of the ADMM-DOPF algorithm can be carried out in a
fully distributed manner.
The dual variable update (14) can be carried out in a fully
distributed manner, where every bus increments the current
dual variables by a (scaled) discrepancy between current net
variables and its own copies of those net variables.

IV. P ROPERTIES OF THE D ISTRIBUTED Fig. 2. Graphical illustration of Algorithm 2. (a) Scenario 1, improper approx-
S OLUTION M ETHOD imation of set Z, see (24), makes the approximated problem (26) infeasible.
(b) Scenario 2, improper choice of the approximation point x̂ = x(0) makes
Recall that the original problem (2) or, equivalently, problem the approximated problem (26) infeasible. (c) Scenario 3, the sequence of
(10) is nonconvex and NP-hard. Therefore, ADMM-based ap- approximations eventually converges to a desired point A. (d) Scenario 4, the
proaches are not guaranteed to converge [29, Sec. 9], though algorithm jumps between points A = (xa , g(xa )) and B = (xb , g(xb )).
general convergence results are available for the convex case
[29, Sec. 3.2]. Nevertheless, in the sequel, we highlight some where the variables are p ∈ IR and x ∈ IR. Recall that
of the convergence properties of our proposed ADMM-DOPF Algorithm 2 approximates nonconvex functions in constraints
algorithm. In particular, we first illustrate, by using an example, (12d) and (12e) of problem (12) by using their first-order
the possible scenarios that can be encountered by Algorithm 2, Taylor’s approximations [see (16d) and (16e)] and the noncon-
that is, step 2 of the ADMM-DOPF algorithm. Then, we capi- vex constraint (12h) by using a convex constraint [see (16h)].
talized on one of the scenarios, which is empirically observed Particularized to the smaller dimensional problem (25), the
to be the most dominant, in order to characterize the solutions approximations pointed above are equivalent to replacing g
of the ADMM-DOPF algorithm. by its first-order Taylor’s approximation ĝ and to approximate
Z by some convex set Ž, where Ž ⊆ Z. The result is the
A. Graphical illustration of Algorithm 2 approximated subproblem given by

We start by focussing on step 2, the main ingredient of the minimize f (p, x)

ADMM-DOPF algorithm. To get insights into the subroutine
(i.e., Algorithm 2) performed at step 2, we first rely on a subject to p = ĝ(x) = g(x̂) + g (x̂)(x − x̂)
simple graphical interpretation. Here, instead of problem (12),
we consider a small dimensional problem to demonstrate some (p, x) ∈ Ž (26)
essential ingredients of the analysis. In particular, we con-
sider the convex objective function f (p, x) in place of (12a). where the variables are p ∈ IR and x ∈ IR, and x̂ represents the
Moreover, instead of the nonconvex constraints (12d) and (12e) point at which the first-order Taylor’s approximation is made.
[cf. (6d) and (6e)], we consider the constraint Let us next examine the behavior of Algorithm 2 by consider-
ing, instead of problem (16), the representative smaller dimen-
p = g(x) (23) sional problem (26). Recall that the key idea of Algorithm 2
is to iteratively refine the first-order Taylor’s approximations
where g is a nonconvex function, which resembles right-hand ẑk
λ̂k (zk ) and μ̂ẑkk (zk ) [see steps 2 and 3 of Algorithm 2], until
side of (6d) and (6e). Finally, instead of the remaining con-
a stopping criterion is satisfied. This behavior is analogously
straints (12b), (12c), (12f), (12g), and (12h) of problem (12),
understood from problem (26), by iteratively refining the first-
we consider the constraint
order Taylor’s approximation ĝ of g.
(p, x) ∈ Z (24) Fig. 2 illustrates the sequential refinement of ĝ, where the
shaded area represents the set Z, the rectangular box represents
where Z is not a convex set [cf. (12h)]. Thus, the smaller dimen- the convex set Ž, the solid curve represents the function g, the
sional problem, which resembles subproblem (12) is given by dotted lines represent the sequential approximations ĝ, and the
thick solid curves represent the contours of f . Note that there
minimize f (p, x) are several interesting scenarios, which deserve attention to
subject to p = g(x) built intuitively the behavior of Algorithm 2 [see Fig. 2(a)–(d)].
(p, x) ∈ Z (25) Fig. 2(a) shows the first scenario, where an improper
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 245

approximation of set Z makes the approximated problem (26) To quantify formally the optimality properties of ADMM-
infeasible, irrespective of the choice of x̂. In contrast, Fig. 2(b) DOPF, we rely on the following definition:
depicts a scenario, where an improper choice of the approxima- Definition 1 ((δ, )-KKT Optimality): Consider the possibly
tion point x̂ makes the approximated problem (26) infeasible. nonconvex problem of the form
Fig. 2(c) shows a sequence of approximations, which eventually
converge to the optimal point “A.” Finally, Fig. 2(d) shows a minimize f0 (x)
scenario, where a sequence of approximations switch between
subject to fi (x) ≤ 0, i = 1, . . . , q
two points “A” and “B,” that is, there is no convergence. Any
other scenario can be constructed by combining cases from hi (x) = 0, i = 1, . . . , p
Fig. 2(a)–(d).
ri (x) = 0, i = 1, . . . , s (27)
Analogously, the discussion from before suggests that the
approximation points (ẑk )k∈N [cf. x̂] were used when con- where f0 : IRn → IR is the objective function, fi : IRn →
ẑk
structing λ̂k (zk ) and μ̂ẑkk (zk ) [cf. ĝ(x)] and the approxima- IR, i = 1, . . . , q are the associated inequality constraint func-
tions used in the set (X̌rk )r=1,...,|Nk | [cf. Ž], can heavily influ- tions, hi : IRn → IR, i = 1, . . . , p and ri : IRn → IR, i =
ence the performance of Algorithm 2. Therefore, especially if 1, . . . , s are the equality constraint functions, and x ∈ IRn
scenarios 1 and 2 depicted in Fig. 2(a) and (b) occur, during is the optimization variable. Moreover, let λi denote the
the algorithm iterations, they have to be avoided by changing dual variable associated with constraint fi (x) ≤ 0, and νi
the initializations. However, extensive numerical experiments and ωi denote the dual variables associated with constraint
show that there are specific choices of ẑk , and X̌rk can make hi (x) = 0 and ri (x) = 0, respectively. Then, an arbitrary point
Algorithm 2 often converge to a point as depicted in Fig. 2(c) (x , λ1 , . . . , λq , ν1 , . . . , νp , ω1 , . . . , ωp ) is called (δ, )-KKT
and barely encounters the scenarios depicted in Fig. 2(a), (b), optimal, if
and (d). See Section V-A for details.
fi (x ) ≤ 0, i = 1, . . . , q (28)

B. Optimality Properties of the Algorithm 2 Solution hi (x ) = 0, i = 1, . . . , q (29)


s
Results obtained in this section are based on the empirical ob- (1/s) ri (x )22 = δ (30)
servations (see Section V) that scenario 3 depicted in Fig. 2(c) i=1
is more dominant compared to others. In particular, we make
the following assumptions. λi ≥ 0, i = 1, . . . , q (31)
Assumption 1: For any k ∈ N , there exists (zk , uk ), to λi fi (x ) = 0, i = 1, . . . , q (32)
which Algorithm 2 can converge. Specifically, there exists
(m) (m)
(zk , uk ), where limm→∞ (zk , uk ) = (zk , uk ) for all k ∈
N . In addition, for all k ∈ N , the components (vkRe , vkIm ) of 
 
q
zk , strictly satisfy the constraint (16h). 
(1/n) ∇x f0 (x ) +

λi ∇x fi (x )
Under Assumption 1, the following assertion can be made: 
i=1
Proposition 1: Suppose Assumption 1 holds. Then, the 2
output (zk , uk ) of Algorithm 2 satisfies Karush-Kuhn-Tucker 
p 
s 

+ νi ∇x hi (x ) + ωi ∇x ri (x ) = . (33)
(KKT) conditions for problem (12). 
i=1 i=1
Proof: See Appendix II-A.  2

Combined with our empirical observations that Algorithm 2 Note that (28)–(33) are closely related to the well-
almost always converges to a point as depicted in Fig. 2(c) (i.e., known KKT optimality criterions, see [6, Sec. 5.5.3]. It
Assumption 1 holds), Proposition 1 claims that the point satis- suggests that the smaller δ and  are, better the point
fies the first-order necessary conditions for local optimality. (x , λ1 , . . . , λq , ν1 , . . . , νp , ω1 , . . . , ωp ) to its local optimality.
We use Definition 1 to formally analyze the optimality proper-
ties of ADMM-DOPF as discussed in the sequel.
C. Optimality Properties of the ADMM-DOPF Solution
Recall that we have used z = (zk )k∈N to denote the vector
As we already pointed out, there is no guarantee that the of all local primal variables in (9), v = (vRe , vIm ) to denote
eventual output (z, v, u, y) of ADMM-DOPF is optimal, or the vector of all net variables, u = (uk )k∈N to denote the dual
even feasible to the original problem (10), because the problem variables associated with constraints (10b)–(10e), and, finally,
is NP-hard. However, Proposition 1, asserts that the eventual y to denote the dual variables associated with constraint (10f).
output (zk , uk ) of Algorithm 2 is a KKT point for problem (12) Let us assume that at the termination of ADMM-DOPF, the
solved at step 2 of ADMM-DOPF. One can easily relate this output corresponding to v and y is v and y , respectively. The
result to characterize the properties of (z, u) of the ADMM- output of ADMM-DOPF corresponding to z and u are simply
DOPF output, as we will see later. However, the properties of the output of Algorithm 2 given by z = (zk )k∈N and u =
the remaining output (v, y) have yet to be investigated. In this (uk )k∈N . However, unlike in convex problems, in the case of
section, combined with the results of Proposition 1, we analyze problem (10), one cannot take for granted that the consistency
the optimality properties of the ADMM-DOPF output. constraint (10f) is satisfied (cf. [29, Sec. 3.2.1]). In particular,
246 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

(n) TABLE I
vk − Ēk v 22 → 0 does not necessarily hold when n → ∞, S PECIFICATIONS OF THE T EST P ROBLEMS . T HE F IRST C OLUMN
where k ∈ N and n is the ADMM-DOPF iteration index. How- I NDICATES THE N UMBER OF B USES . T HE S ECOND C OLUMN G IVES THE
ever, an appropriate choice of the penalty parameter ρ in the R EFERENCE TO THE O RIGINAL P ROBLEMS . T HE T HIRD C OLUMN S HOW
H OW W E M ODIFY THE O RIGINAL P ROBLEM (p̄D D
k AND q̄k I NDICATE THE
ADMM-DOPF algorithm usually allows finding outputs, where O RIGINAL P ROBLEM DATA A SSOCIATED W ITH THE P OWER D EMANDS ).
the consistency constraints are almost satisfied with a small T HE F OURTH AND F IFTH C OLUMNS S PECIFY THE N UMBER OF
error floor, which is negligible in real practical implementations G ENERATORS AND L OADS , R ESPECTIVELY. T HE S IXTH C OLUMN
S PECIFIES THE T YPE OF F LOW L INE L IMIT (FL) U SED , I F A NY, T HAT I S ,
as we will see empirically in Section V. For latter use, let us W HICH C ONSTRAINTS (2 I ), (2 J ), AND (2 K ) A RE I NCLUDED
quantify this error floor from δ k , i.e.,

δ k = vk − Ēk v , k ∈ N . (34)

Now we can formally establish the optimality properties of

ADMM-DOPF as follows.
Proposition 2: Given Assumption 1 holds, the output
(z , v , u , y ) at the termination ofADMM-DOPF is
(a−1 δ̄, b−1 ρ2 δ̄)-KKT optimal, where δ̄ = k∈N |δ k 22 , ρ is the
penalty parameter used in the ADMM-DOPF iterations, and
a = len((δ k )k∈N ), b = len(z , v ) are normalization factors.
Proof: See Appendix II-B.  TABLE II
F REQUENCY OF THE T ERMINATION OF Algorithm 2
We note that deriving an analytical expression of δ by using F ROM THE S TOPPING C RITERIA (18)
problem (10) data is very difficult. However, we can numeri-
cally compute δ and  given in Proposition 1 as

δ = a−1 δ̄,  = b−1 ρ2 δ̄. (35)

Extensive numerical experiences show that we usually have

very small values for δ. For example, for all considered sim-
ulations with ρ = 106 [see Section V], we have δ on the order
of 10−12 (or smaller) and  on the order of 10−1 (or smaller)
after 5000 ADMM-DOPF iterations. The convex problem (16) is solved with the convex solution
method presented in Section III-C, together with the built-
V. N UMERICAL R ESULTS in Matlab QP solver quadprog. As a stopping criterion for
Algorithm 2, we use sub = 10−10 and max_iter = 20, un-
In this section, we present numerical experiments to illus- less stated otherwise. For the ADMM method, we use vRe =
trate the proposed algorithm. We compare our algorithm with (1, · · · , 1), vIm = (0, · · · , 0), and y = (0, · · · , 0) as an initial
the branch-and-bound algorithm [11], centralized OPF solver point.
provided by Matpower [33], and the SDP relaxation from [3].
In order to study the convergence properties of the algorithm,
we evaluate it on four examples that have a (nonzero) duality A. Properties of Algorithm 2
gap, see Table I, rows 1-4. These four examples come from [10] In this subsection, we investigate the convergence properties
and [11], and are obtained by making a small modification to of Algorithm 2. In particular, we relate the convergence behav-
standard test examples, see Table I, column 3. It is worth noting ior of Algorithm 2 to the analysis in Sections IV-A and IV-B,
that the methods based on the SDP relaxation do not apply here where four scenarios, or possible outcomes, of Algorithm 2
due to the nonzero duality gap [10]. To study the scalability (Fig. 2) were identified.
properties of the proposed algorithm, we also evaluate it on two During the numerical evaluations, Algorithm 2 was executed
larger examples, see Table I, rows 5-6. The exact specifications 11 060 000 times. Scenarios 1 and 2 [Fig. 2(a) and (b), respec-
of the considered examples are found in Table I and references tively] are where the approximated subproblem is infeasible
therein. The objective functions in all examples are quadratic. and occurred only 6 times (≈ 0.00008%). These occurrences
The units of real power, reactive power, apparent power, volt- occurred at one of the buses in the 300-bus example, during
age magnitude, and the objective function values are megawatts ADMM iterations 137–143. This suggests even if a particular
(MW), MVAr, megavolt-amperes (MVA), per unit (p.u.),4 bus fails to converge with Algorithm 2 in consecutive ADMM
and U.S.$/h, respectively. In all six problems, the average iterations, the bus can recover to find a solution in latter ADMM
power demand of the loads is in the range 10–100 MW and iterations.
1–10 MVAr, for the real and reactive powers, respectively. To numerically study the occurrences of Scenarios 3 and 4,
The simulations were executed in a sequential computational we run Algorithm 2 with 1) max_iter = 20; 2) max_iter =
environment, using Matlab version 8.1.0.604 (R2013a) [35]. 1000; and 3) max_iter = 25 000. In all considered cases, we
use sub = 10−10 [compare with (17)]. Table II summarizes
4 The voltages base is 400 kV. how frequently the stopping criteria (18) of Algorithm 2 is met.
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 247

Fig. 3. CDF displaying DFk,n of (36) for every subproblem k and every
ADMM iteration n for each of the four examples.

Note that the entries of Table II suggest an upper bound

on the frequencies of Scenario 4. Therefore, Table II shows
that the frequencies of Scenario 4 decrease (or unchanged) as
max_iter is increased. However, the effects are marginal as
indicated in the table, especially for max_iter ≥20. On the
other hand, recall that sub = 10−10 , that is, the decrement
of voltages between two successive iterations is below 10−10
[compare with (17)]. Such an infinitesimal accuracy in the
stopping criteria (17) suggests the algorithm’s convergence, see
Scenario 3, Fig. 2(c). For example, consider the case N = 3
and max_iter = 20 in Table II. From the results, the frequency
of termination of Algorithm 2 from the stopping criteria (17)
(i.e., Scenario 3) becomes 99.995%. Thus, from Proposition 1,
it follows that when max_iter = 20, 99.995% of the cases
Algorithm 2 converge to a point satisfying the KKT conditions.
It is worth noting that for all considered cases, the convergence
of the algorithm is in the range 99.99%–100%. The results also Fig. 4. Histograms displaying DFk,n in (36) for every subproblem k and
every ADMM iteration n for the considered test networks: (a) three-bus
suggest that the convergence properties of Algorithm 2 can be networks, (b) nine-bus network, (c) 14-bus network, (d) 30-bus network,
improved (see the case N = 300) or remain intact (see the cases (e) 118-bus network, and (f) 300-bus network.
N = 3, 9, 14, 30, and 118) at the expense of the increase in
max_iter. As a consequence of this promising behavior of Algorithm 2,
Note that the voltages returned by Algorithm 2 are always we will proceed under Assumption 1.
feasible to problem (12), that is, it satisfies (12h). However, the
resulting power injections might be infeasible [compared with
(12c)–(12g)]. To measure the feasibility of the returned power B. Connection to Proposition 2
injections of Algorithm 2, we define the following metric called In this section, we relate the numerical evaluations to
the degree of feasibility (DF): Proposition 2. In particular, we inspect the behavior of δ
 
 (n) (n)  [compared with (30) and (35)], and  [compared with (33) and
DFk,n = min pk + jqk − (p + jq) (36) (35)] with respect to ρ, which are defined in Section IV-C.
p+jq∈Sk
The unit of measurements for δ is p.u.2 and the unit of 
where k and n indicate the bus and ADMM iteration, respec- can be interpreted as the square of the decrease/increase in
(n) (n) U.S.$/hour with respect to a small perturbation in the variable
tively, pk + jqk is the returned power injection, and
 G,min ! z = (zk )k∈N .
p − pD ≤ Re(z) ≤ pG,max − pD Fig. 5 depicts δ at every 500 ADMM iterations, for ρ =

Sk = z ∈ C G,min
k k k k (37)
qk − qkD ≤ Im(z) ≤ qkG,max − qkD 106 , · · · , 1013 . In the 30-bus example, the results are almost
identical for ρ = 109 , 1010 , · · · , 1013 and, accordingly, we only
The unit of measurement for DFk,n is MVA. In order to include the results for ρ = 106 , 107 , 108 , 1013 . Since δ mea-
provide a statistical description of DFk,n for every execution of sures the inconsistency between the subproblems, the point
Algorithm 2, we consider an empirical cumulative distribution returned by ADMM-DOPF can only be considered feasible
function (CDF) (Fig. 3) and a histogram (Fig. 4), for each when δ has reached acceptable accuracy, that is, δ < γ for some
example separately. These results suggest that Algorithm 2 γ > 0. We do not consider any particular threshold γ, since we
returns a feasible solution with high accuracy in all cases, where are only interested in observing the convergence behavior. In
the worst case accuracy is 2.5 × 10−10 . this aspect, the result shows a promising behavior, as δ has a
248 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

Fig. 5. δ versus the number of ADMM iterations: (a) three-bus network, Fig. 7. δ versus the objective function value: (a) three-bus network; (b) nine-
(b) nine-bus network, (c) 14-bus network, and (d) 30-bus network. bus network, (c) 14-bus network, and (d) 30- bus network.

to the KKT optimal point. Therefore, the decreasing trend in ,

which is observed from the results, is desired. In the case of the
3-, 9-, and 14-bus examples,  reaches values between 10−2 and
10−11 in almost every case. However, in the 30-bus example,
only when ρ = 106 does epsilon reach a value below 10−2 .

C. Convergence and Scalability Properties

By convention, the objective value of problem (10) is ∞ if
the problem is infeasible and is given by (10a) if it is feasi-
ble. Therefore, when computing the objective function of the
problem, one has to verify whether the constraints (10b)–(10f)
are feasible or not. Based on Fig. 4, the feasibility of the
−10
subproblem variables, including pG k , is on the order of 10 in
the worst case for every ADMM iteration [compare with (36)].
In other words, (pG k )k∈G returned by Algorithm 2 in every
ADMM iteration is feasible (with very high precision). There-
fore, our proposed Algorithm 1, which includes Algorithm 2 as
a subroutine, ensures the feasibility of constraints (10b)–(10f)
Fig. 6.  versus the number of ADMM iterations: (a) three-bus network, (with very high precision). However, the feasibility of the re-
(b) nine-bus network, (c) 14-bus network, and (d) 30-bus network. maining constraint (10f) has to be verified in order to compute a
sensible operating point. In the sequel, we numerically analyze
decreasing trend in all cases. Furthermore, for the 3-, 9-, and the feasibility of the constraint (10f) together with the objective
14-bus examples, δ converges to a fixed error floor for the larger value computed by using (10a).
values of the penalty parameter ρ. In particular, as ρ increases, Fig. 7 shows the objective value versus δ, at every 100
δ converges to a point closer to zero, which suggests a negative or 200’ ADMM iterations. In the case of the 3- and 9-bus
relationship between δ and ρ. Therefore, this indicates that examples, we compare the objective value with the branch-
increasing the penalty parameter enforces higher accuracy of and-bound algorithm from [11] where the relative tolerance, the
consistency among the subproblems. On the contrary to the 3-, difference between the best upper and lower bounds relative to
9-, and 14-bus examples, δ decreases more slowly when the the best upper bound, is 0.001. The upper bound is obtained
penalty parameter increases in the case of the 30-bus example. from Matpower, and the lower bound is obtained by using the
However, in the case of the 30-bus example, δ is still decreasing Matlab toolbox YALMIP [36] and the solver SEDUMI [37] to
after the last iteration considered when ρ = 109 , · · · , 1013 . solve the dual SDP relaxation. In the case of the 14- and 30-bus
Fig. 6 depicts  at every 500th ADMM iterations, for different examples, the branch-and-bound algorithm failed due to mem-
ρ’s. In contrast to δ, the decreasing trend in  is not necessary ory errors. In all cases, we compare our results with the SDP re-
to obtain a feasible solution to the problem. However, under laxation from [3]. The results show that the algorithm converges
Assumption 1, as δ and  go to zero, the algorithm converges to some objective value in relatively few iterations, which can
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 249

Fig. 8. Relative objective function: (a) three-bus example and (b) nine-bus
example.

even be optimal with an appropriate choice of ρ. For example,

for the considered cases, ρ = 106 yields almost optimal objec-
tive values. Moreover, the desired consistency metric δ is driven
toward zero as the number of ADMM iterations increases.
Fig. 8 depicts the relative objective function (|f − f  |/f  )
for the 3- and 9-bus examples. The results are consistent with
those presented in Fig. 7. For example, in the case of ρ = 106 , Fig. 9. Tp , |f − f  |/f  , δ, and  as a function of ADMM iterations. (a) Tp ;
the relative objective function value is on the order of 10−6 . The (b) |f − f  |/f  ; (c) δ; (d) .
results suggest that a proper choice of ρ is beneficial to achieve
a good network operating point. Fig. 9(c) and (d) depicts  and δ as a function of ADMM
To study the scalability properties of the proposed algorithm, iterations. The results show that irrespective of the number of
we compute the CPU time, relative objective value, δ, and  of buses, the metrics δ and  decrease as desired. Results further
the algorithm for all of the considered examples. In the case of suggest that those values are driven toward small values as
3-, 9-, 14-, and 30-bus examples, we choose ρ = 106 and in the ADMM iterations increase.
case of 118- and 300-bus examples, we chose ρ = 107 . Table III shows the running time and the objective value
Fig. 9(a) shows the parallel running times Tp versus ADMM obtained by different approaches. As benchmarks, we consider
iterations. In particular, we define Tp = Ts /|N |, where Ts is the centralized algorithms, SDP relaxation [3], branch and
the sequential CPU time. The behavior of the plots in the case bound [11], and Matpower [33]. Table III shows that our
of 9-, 14-, 30-, 118-, and 300-bus examples are very similar. proposed method yields network operating points, which are
In other words, the parallel running time Tp is independent of almost optimal, where the discrepancy with respect to the
the number of buses, indicating the promising scalability of optimal is on the order of 0.1% (respectively, 1%) or less with
the proposed algorithm. Note that the 3-bus example has to 10 000 (respectively, 3000) ADMM iterations. Note that the
handle more variables per subproblem, compared with the other running time of ADMM-DOPF is insensitive to the network
examples, see column 6 of Table I. This is clearly reflected size, see Fig. 9(a) for more details. However, even in small
in the plot of the three-bus example, as an increase of the networks (e.g., the case with N = 14, N = 30), the running
associated parallel running time Tp . time of the branch-and-bound algorithm can explode. This is
Fig. 9(b) depicts the relative objective function |f − f  |/f  . expected because the worst case complexity of the branch-and-
In the case of 3-, 9-, 118-, and 300-bus examples, the global bound algorithm grows exponentially with the problem size.
optimum f  is found by a branch-and-bound algorithm. How- Results further suggest that the running time of the centralized
ever, in the case of 14- and 30-bus examples, the branch-and- algorithm SDP relaxation increases as the network size grows,
bound algorithm failed and, therefore, the best known objective unlike the proposed ADMM-DOPF. Note that the running
value found by Matpower was considered as f  .5 The results time of ADMM-DOPF is large compared to the centralized
show that for large and relatively large test examples (e.g., N = Matpower. However, those values can be further reduced if
30, 118, 300), the relative objective function value is on the ADMM-DOPF is deployed in a parallel computation environ-
order of 10−3 and is not affected by network size. A similar ment, where every subproblem is handled at a dedicated set of
independence of the performance is observed for very small test resources, including processors and memory among others. In
examples as well (e.g., N = 3, 9, 14) with relative objective addition to the centralized benchmarks, we also consider the
function values on the order of 10−6 . The reduction of the decentralized one proposed in [28], which employs ADMM for
relative objective function values of smaller networks compared general nonconvex OPF. However, the results of [28] are not
with larger network examples are intuitively expected due to documented in Table III, because for all considered examples,
substantial size differences of those networks. the algorithm therein did not converge. This agrees with the
numerical results of [28], where the authors mentioned that the
5 Since the 118- and 300-bus examples have zero duality gap, the branch-
convergence of their algorithm is more sensitive to the initial
and-bound algorithm worked efficiently. However, in the case of 14- and 30-bus
examples, where there is nonzero duality gap, the branch-and-bound algorithm point in the case of meshed networks [28, p. 5]. We note that
failed. their method converges if it is initialized close to the optimal
250 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

TABLE III
C OMPARISON OF THE O UTPUT OF ADMM-DOPF W HEN AT ADMM I TERATION n = 3000 AND n = 10 000 W ITH THE SDP R ELAXATION [3],
B RANCH AND B OUND [11], AND M ATPOWER [33]. Ts , Tp , O BJ , AND # I TER I NDICATE THE S EQUENTIAL AND PARALLEL RUNNING
T IMES IN S ECONDS , THE O BJECTIVE VALUE IN U.S.$/h, AND THE N UMBER OF I TERATIONS , R ESPECTIVELY

solution. However, in practice, such an initialization point is by constructing a series of sets of the form Y̌rk and affine
unknown, thus limiting dramatically the applicability of the functions of the form γ̌k that approximate the feasible set
method in [28]. specified by (16g) and (16h) in an increasing precision. The
Finally, from all of our numerical experiments discussed QP-based algorithm to solve problem (16) can be summarized
before, we note that the power losses in the flow lines are as follows.
typically on the order of 4.4% (or less) of the total power flow
in the line. Since the losses are not negligible, approximations,
such as the linearization of power-flow equations, can be less Algorithm 3: QP to solve Problem (16)
applicable to compute better network operating points. 1) Initialize: Given the initial approximated set Y̌rk and
affine function γ̌k . Let m̄ = 1.
2) Solve the QP
VI. C ONCLUSION
ρ
k )+yk (vk − Êk v) + vk −Êk v2 (38a)
fkG (pG T 2
We proposed a distributed algorithm for the OPF, by de- min
2
composing the OPF problem among the buses that compose
the electrical network. A light communication protocol among s.t. zk = (pG G Re Im Re Im
k , q k , pk , q k , ik , ik , v k , v k , (38b)
neighboring buses is needed during the algorithm, resulting in îRe Im
k , îk , p̂k , q̂k ) (38c)
high scalability properties. The subproblems related to each bus ẑk
capitalize on sequential convex approximations to gracefully (αk (zk ), λ̂k (zk ), μ̂ẑkk (zk )) =0 (38d)
manipulate the nonconvexity of the problem. We showed the
(β k (zk ), γ̌ k (zk )) ≤ 0 (38e)
convergence of subproblem solutions to the first-order nec-
essary condition for local optimality, under mild conditions. ((vkRe )r , (vkIm )r ) ∈ Y̌rk , r = 1, . . . , |Nk | (38f)
Furthermore, by using the local optimality results associated
with the subproblems, we quantified the optimality of the where the variables are pG G Re Im Re
k , q k , pk , q k , ik , ik , v k ,
Im Re Im ¯
overall algorithm. We evaluated the proposed algorithm on vk , īk , īk , bf pk , q̄k , and zk . The solution correspond-
a number of test examples to demonstrate its convergence ing to the variables zk , ((vkRe )r , (vkIm )r ) are denoted
(m̄) (m̄)
properties and to compare it with the global optimal method. by zk , vr , respectively, and all of the dual optimal
In all considered cases, the proposed algorithm achieved close (m̄)
variables are denoted by uk .
to optimal solutions. Moreover, the proposed algorithm showed (m̄)
3) Stopping criterion: If γ k (zk ) ≤ 0 and vr ∈ X̌rk for
(m̄)
appealing scalability properties when tested on larger examples. (m̄) (m̄)
all r ∈ {1, . . . , |Nk |}, STOP and return (zk , uk ).
Otherwise, increase the precession of set Y̌r and func- k

A PPENDIX I tion γ̌k by adding a hyperplane and an affine function,

O N THE U SE OF Q UADRATIC P ROGRAMMING QP S OLVERS respectively, set m̄ := m̄ + 1 and go to step 2.
Note that not all of the constraints of problem (16) are
affine (or linear). In particular, constraints (16g) and (16h) are
not affine. Therefore, QP solvers are not directly applied to The set Y̌rk is initialized in the first step by approximating the
solve the problem. However, if constraints (16g) and (16h) are exterior boundary of the donut Xrk [Fig. 1(a)] by an equilateral
approximated by using affine constraints, then QP is readily octagon as shown in Fig. 1(c), and γ̌k is initialized corre-
applied to the modified problem. spondingly. The second step simply involves solving a QP pro-
Let us start by considering the feasible regions defined by gramming problem. The algorithm terminates in the third step
(m̄)
(16h), which accounts for X̌rk , r ∈ {1, . . . , |Nk |}, see Fig. 1(b). if γ k (z(m̄) ) ≤ 0 and (vk )r ∈ X̌rk for all r ∈ {1, . . . , |Nk |}.
(m̄)
Next, we approximate the nonlinear boundary of X̌rk by affine However, if (vk )r ∈ Y̌rk \ X̌rk , we increase the precession of
functions as depicted in Fig. 1(c). We denote by Y̌rk the ap- Y̌r by adding a hyperplane on the exterior boundary of the
k
(m̄)
proximated polyhedral set. We can apply similar ideas to ap- donut Xrk , so that (vk )r ∈ Y̌rk . In particular, we set Y̌rk =
proximate the feasible regions specified by (16g) [cf. (6h)–(6i)], Y̌rk ∩ W where W is the halfspace
where we use γ̌k to denote the resulting affine function. Finally,   
the idea is to find the desired optimal solution of problem (16) W = ((vkRe )r , (vkIm )r ) ∈ IR2  αr (vkRe )r +βr (vkIm )r ≤γr (39)
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 251

where where ∇ ¯ is used to represent the component-wise differentia-

" tion of associated functions.
(m̄) (vkmax )2r
αr=sign(Re((vk )r )) Now we can easily conclude that zk is primal feasible for
(m̄) (m̄)
1+((Im((vk )r )/(Re((vk )r ))2 problem (12). This follows from (44), the fact that constraints
 (m̄)

(Im((vk )r ) (12b), (12c), (12f), and (12g) are identical to (16b), (16c), (16f),
βr=ar (m̄)
, γr = (vkmax )2r and (16g), respectively, and that X̌rk ⊆ Xrk .
(Re((vk )r )
Dual feasibility of uk , associated with constraints (16f) and
if Re(xr ) = 0 and (16g), affirms the dual feasibility of uk associated with identical
(m̄) constraints (12f) and (12g). In the case of constraint (16h), the
αr = 0, βr = sign(Re((vk )r )), γr = (vkmax )r . recall from (15) that X̌rk is characterized by ((vkRe )r , (vkIm )r ) ∈
γ̌k can be treated identically. IR2 such that

A PPENDIX II cr ≤ ar (vkRe )r + br (vkIm )r and (vkRe )2r + (vkIm )2r ≤ (vkmax )2r .
P ROOFS (44)
A. Proof of Proposition 1
Thus, dual feasibility of uk components associated with
Proof: Obviously, problem (16) is convex and in any iter- the first (respectively, second) constraint above ensures the
(m) (m)
ation m of Algorithm 2, (zk , uk ) [so is (zk , uk )] are primal dual feasibility of the same uk components associated with
and dual optimal, with zero duality gap. Thus, (zk , uk ) satisfies (vkmin )2r ≤ (vkRe )2r + (vkIm )2r (respectively, (vkRe )2r + (vkIm )2r ≤
KKT conditions for problem (16) [6, Sec. 5.5.3]. However, (vkmax )2r ) of (12h). Thus, we conclude uk is dual feasible for
in order to show that (zk , uk ) satisfies KKT conditions for problem (12).
problem (12), we need to show: 1) zk is primal feasible; 2) uk From (42), the fact that constraints (12b), (12c), (12f), and
is dual feasible; 3) zk and uk satisfy complementary slackness (12g) are identical to (16b), (16c), (16f), and (16g), respectively,
conditions; and 4) derivative of the Lagrangian of problem (12) and that the components (vkRe , vkIm ) of zk , strictly satisfy the
vanishes with zk and uk [6, Sec. 5.5.3]. constraint (16h) [see Assumption 1], it follows that zk and uk
We start by noting that the original functions definitions satisfy complementary slackness conditions for problem (12).
λk (zk ) and μk (zk ) [see (12d) and (12e)] are characterized by In addition, Assumption 1, together with the complementary
using the basic form slackness condition, ensures that the components of uk associ-
h(p, x1 , x2 , y1 , y2 ) = p − x1 y1 − x2 y2 (40) ated with constraints (16h) are identically zero.
  Finally, recall that (zk , uk ) are optimal primal and dual
z
variables for problem (16). Therefore, the derivative of the
where p ∈ IR represents power, x1 , x2 ∈ IR represent cur- Lagrangian L̂k (zk , uk ) associated with problem (16) vanishes
rents, y1 , y2 ∈ IR represent voltages, and we have denoted at (zk , uk ), that is, ∇zk L̂k (zk , uk ) = 0. This result, combined
(p, x1 , x2 , y1 , y2 ) compactly by z. Let ĥẑ denote the first-order with (43), and the fact that constraints (12b), (12c), (12f), and
Taylor’s approximation of h at ẑ. That is, ĥẑ characterizes the (12g) are identical to (16b), (16c), (16f), and (16g), respectively,
basic form of the first-order Taylor’s approximation of function and the fact that the components of uk associated with con-
ẑk straints (16h) are identically zero, affirms that the derivative
definitions λ̂k (zk ) and μ̂ẑkk (zk ), see (16d) and (16e). There-
fore, without loss of generality, we make our assertions based of the Lagrangian Lk (zk , uk ) associated with problem (12)
on h and ĥẑ , together with the assumption limm→∞ z(m) = z , vanishes at zk and uk , i.e.,
(m)
where z plays the role of zk and z(m) plays the role of zk .
Let us next summarize some intermediate results, which will ∇zk Lk (zk , uk ) = 0 (45)
be useful later.
Lemma 1: Given the function h on the form (40), and which concludes the proof. 
(m)
limm→∞ z(m) = z , we have 1) limm→∞ ĥz (z ) = h(z )
(m)
and 2) limm→∞ ∇z ĥz (z ) = ∇z h(z ).
B. Proof of Proposition 2
Proof: See Appendix B-C. 
From Lemma 1 above, we conclude that Proof: Given that Assumption 1 holds, Proposition 1
 
asserts that all constraints, but (10f) of problem (10) are
h(z ) = ĥz (z ) and ∇z h(z ) = ∇z ĥz (z ). (41) primal feasible. Combined with (34), it trivially
 follows that
δ = a−1 δ̄, where a = len((δ k )k∈N ) and δ̄ = k∈N δ k 22 [cf.
By relating the result (41) to our original problems (12) and (30)]. To show that  = b−1 ρ2 δ̄ [cf. (33)], let us consider the
(16), we can deduce that Lagrangian L(z, v, u, y) associated with problem (12). Note
z z that L(z, v, u, y) is related to the Lagrangian Lk (zk , uk |yk ) of
λk (zk ) = λ̂kk (zk ), μk (zk ) = μ̂kk (zk ) (42)
problem (12) as
and  
  L(z, v, u, y) = Lk (zk , uk |yk ) − (ρ/2)xk − Êk v22 .
∇ ¯ z λ̂zkk (zk ), ∇
¯ z λk (zk )=∇ ¯ z μ̂zk (zk )
¯ z μk (zk )=∇ (43)
k k k k k k∈N
252 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 2, NO. 3, SEPTEMBER 2015

⎡ ⎤
∇z1 L1 (z1 , u1 |yk ) − ρz̄1 ⎡ ⎤
⎢ .. ⎥ −ρz̄1
⎢ ⎥
⎥ ⎢ .. ⎥
. .
⎢
∇z,v L(z , v , u , y ) = ⎢ ∇z LN (z , u |y ) − ρz̄N ⎥ = ⎢ ⎥ (46)
⎢ N  N N k  ⎥ ⎣ −ρz̄N ⎦
⎣  ⎦
∇v −ykT Êk v 0
k∈N
⎛ ⎞
⎜ ⎟
z̄k = ⎜
⎝0, 0, 0, 0, 0, 0, (xk − Ek v ),
Re Re
(xIm − Ek vkIm ), 0, 0, 0, 0⎟
⎠ (47)
 k

x
k
−Ēk v =δ k

Note the notation used when passing the parameters to Lk , ACKNOWLEDGMENT

where we have highlighted the dependence of Lk on yk [cf.
The authors would like to thank A. Forsgren for valuable
(12a)]. Let us now inspect the derivative of the Lagrangian
comments.
L(z, v, u, y), evaluated at (z , v , u , y ). In particular, we
have (46), shown at the top of the page, where z̄k is given in
(47), shown at the top of the page [compared with (34)]. Here, R EFERENCES
the first equality follows from the standard derivation combined [1] J. Carpentier, “Contribution à l’étude du dispatching économique,” Bull.
with Proposition 1, the second equality follows from (45), and Soc. Francaise Elect., vol. 3, no. 8, pp. 431–447, 1962.
by [2] S. Frank, I. Steponavice, and S. Rebennack, “Optimal power flow:
 invoking T 
the optimality conditions for problem (13), that is,
A bibliographic survey I,” Energy Syst., vol. 3, no. 3, pp. 221–258,
k∈N Ēk yk = 0. From (46) and (47), we conclude that  = Sep. 2012.
b−1 ρ2 δ̄ [cf. (33)], where b = len(z , v ). Finally, conditions [3] J. Lavaei and S. H. Low, “Zero duality gap in optimal power flow prob-
(28), (29), (31), and (32), associated with problem (10) follow lem,” IEEE Trans. Power Syst., vol. 27, no. 1, pp. 92–107, Feb. 2012.
[4] S. Frank, I. Steponavice, and S. Rebennack, “Optimal power flow:
from straightforward arguments, which concludes the proof.  A bibliographic survey II,” Energy Syst., vol. 3, no. 3, pp. 259–289,
Sep. 2012.
[5] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming for
optimal power flow problems,” Int. J. Elect. Power Energy Syst., vol. 30,
C. Proof of Lemma 1 no. 6–7, pp. 383–392, Jul. 2008.
[6] S. Boyd and L. Vandenberghe, Convex Optimization. New York, USA:
Let H denote the Hessian of function h. Note that H is Cambridge University Press, 2004.
a matrix with constant entries and, thus, does not depend on [7] S. Bose, D. Gayme, S. H. Low, and K. M. Chandy, “Optimal power flow
z. From the definition of the Taylor series expansion at zm , over tree networks,” presented at the 49th Annu. Allerton Conf. Commun.,
Control, Comput., Monticello, IL, USA, 2011.
we have [8] B. Zhang and D. Tse, “Geometry of injection regions of power networks,”
IEEE Trans. Power Syst., vol. 28, no. 2, pp. 788–797, May 2013.
(m) [9] J. Lavaei, D. Tse, and B. Zhang, “Geometry of power flows and optimiza-
h(z) − ĥz (z) = (1/2)(z(m) − z)T H(z(m) − z). (48) tion in distribution networks,” IEEE Trans. Power Syst., vol. 29, no. 2,
pp. 572–583, Mar. 2014.
[10] B. C. Lesieutre, D. K. Molzahn, A. R. Borden, and C. L. DeMarco,
Moreover, the differentiation of (48) yields “Examining the limits of the application of semidefinite programming
to power flow problems,” in Proc. 49th Annu. Allerton Conf. Commun.,
(m) Control, Comput., 2011, pp. 1492–1499.
∇h(z) − ∇ĥz (z) = H(z(m) − z). (49) [11] A. Gopalakrishnan, A. U. Raghunathan, D. Nikovski, and L. T. Biegler,
“Global optimization of optimal power flow using a branch and bound al-
gorithm,” in Proc. 50th Annu. Allerton Conf. Commun., Control, Comput.,
To show case 1 of the proposition, we consider the following 2012, pp. 609–616.
relations: [12] E. L. Quinn, “Privacy and the new energy infrastructure,” Center for
Energy and Environmental Security (CEES), University of Colorado,
(m) CO 80309, USA, Tech. Rep. 09-001, Feb. 2009. [Online]. Avail-
h(z ) − ĥz (z ) = (1/2)(z(m) − z )T H(z(m) − z ) (50) able: http://ssrn.com/abstract=1370731 or http://dx.doi.org/10.2139/ssrn.
1370731
(m) [13] A. Cavoukian, J. Polonetsky, and C. Wolf, “Smartprivacy for the smart
(1/2)λmin (H)z(m) − z 22 ≤ h(z ) − ĥz (z ) grid: Embedding privacy into the design of electricity conservation,”
Identity Inf. Soc., vol. 3, no. 2, pp. 275–294, 2010.
≤ (1/2)λmax (H)z(m) − z 22 (51) [14] S. Salinas, M. Li, and P. Li, “Privacy-preserving energy theft detection in
smart grids: A P2P computing approach,” IEEE J. Sel. Areas Commun.,
vol. 31, no. 9, pp. 257–267, Sep. 2013.
where (50) follows from (48) with z = z , and (51) follows from [15] B. H. Kim and R. Baldick, “Coarse-grained distributed optimal power
(50) and basics of linear algebra. By letting m → ∞ in (51), we flow,” IEEE Trans. Power Syst., vol. 12, no. 2, pp. 932–939, May 1997.
(m) (m) [16] R. Baldick, B. H. Kim, C. Chase, and Y. Luo, “A fast distributed imple-
conclude limm→∞ ĥz (z ) = h(z ), since limm→∞ zk = zk. mentation of optimal power flow,” IEEE Trans. Power Syst., vol. 14, no. 3,
Similarly, by using (49) and that H(z −z ) ≤ λmax (HT H)
(m)  pp. 858–864, Jul. 1999.
[17] B. H. Kim and R. Baldick, “A comparison of distributed optimal power
z(m) − z 22 and H(z(m) −z ) ≥ λmin (HT H)z(m) −z 22 , flow algorithms,” IEEE Trans. Power Syst., vol. 15, no. 2, pp. 599–604,
(m)
we conclude limm→∞ ∇ĥz (z ) = ∇h(z ). May 2000.
MAGNÚSSON et al.: DISTRIBUTED APPROACH FOR THE OPF PROBLEM 253

[18] A. J. Conejo, F. J. Nogales, and F. J. Prieto, “A decomposition procedure Sindri Magnússon received the B.Sc. degree in math-
based on approximate Newton directions,” Math. Program., vol. 93, no. 3, ematics from the University of Iceland, Reykjavík,
pp. 495–515, 2002. Iceland, in 2011 and the M.Sc. degree in mathe-
[19] F. J. Nogales, F. J. Prieto, and A. J. Conejo, “A decomposition method- matics from KTH Royal Institute of Technology,
ology applied to the multi-area optimal power flow problem,” Ann. Oper. Stockholm, Sweden, in 2013, where he is currently
Res., vol. 120, no. 1–4, pp. 99–116, 2003. pursuing the Ph.D. degree in automatic control at
[20] G. Hug-Glanzmann and G. Andersson, “Decentralized optimal power the School of Electrical Engineering and ACCESS
flow control for overlapping areas in power systems,” IEEE Trans. Power Linnaeus Center.
Syst., vol. 24, no. 1, pp. 327–336, Feb. 2009. His research interests include distributed opti-
[21] A. G. Bakirtzis and P. N. Biskas, “A decentralized solution to the DC-OPF mization, in theory and applications.
of interconnected power systems,” IEEE Trans. Power Syst., vol. 18, no. 3,
pp. 1007–1013, Aug. 2003.
[22] A. Y. S. Lam, B. Zhang, and D. Tse, “Distributed algorithms for optimal
power flow problem,” in 2012 IEEE 51st Annual Conference on Decision Pradeep Chathuranga Weeraddana (S’08–M’11)
and Control (CDC), Dec. 2012, pp. 430–437. received the M.Eng. degree in telecommunication
[23] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power from the School of Engineering and Technology,
flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, Asian Institute of Technology, Khlong Luang, Thai-
pp. 1464–1475, Sep. 2013. land, in 2007 and the Ph.D. degree from the Univer-
[24] B. Zhang, A. Y. S. Lam, A. Dominguez-Garcia, and D. Tse, “Optimal sity of Oulu, Oulu, Finland, in 2011.
distributed voltage regulation in power distribution networks,” ArXiv Currently, he is a Postdoctoral Researcher with
e-prints, Apr. 2012. [Online]. Available: http://adsabs.harvard.edu/abs/ the Automatic Control Lab, Electrical Engineering
2012arXiv1204.5226Z Department and ACCESS Linnaeus Center, KTH
[25] M. Kraning, E. Chu, J. Lavaei, and S. Boyd, “Dynamic network energy Royal Institute of Technology, Stockholm, Sweden.
management via proximal message passing,” Found. Trends Optimiz., His research interests include the application of opti-
vol. 1, no. 2, pp. 73–126, 2014. mization techniques in various application domains, such as signal processing,
[26] S. Bolognani, G. Cavraro, R. Carli, and S. Zampieri, “A distributed wireless communications, smart grids, privacy, and security.
feedback control strategy for optimal reactive power flow with voltage
constraints,” ArXiv e-prints, Mar. 2013.
[27] P. Šul, S. Backhaus, and M. Chertkov, “Optimal distributed control of
reactive power via the alternating direction method of multipliers,” IEEE Carlo Fischione (M’05) received the Dr.Eng. degree
Transactions on Energy Conversion, vol. 29, no. 4, pp. 968–977, Dec. in electronic engineering (Hons.) and the Ph.D. de-
2014. [Online]. Available: http://dx.doi.org/10.1109/TEC.2014.2363196. gree in electrical and information engineering from
[28] A. X. Sun, D. T. Phan, and S. Ghosh, “Fully decentralized ac optimal the University of L’Aquila, L’Aquila, Italy, in 2001
power flow algorithms,” in Proc. IEEE Power Energy Soc. Gen. Meeting, and 2005, respectively.
Jul. 2013, pp. 1–5. Currently, he is a tenured Associate Professor at
[29] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed KTH Royal Institute of Technology, Electrical En-
optimization and statistical learning via the alternating direction method gineering and ACCESS Linnaeus Center, Automatic
of multipliers,” Found. Trends Mach. Learning, vol. 3, no. 1, pp. 1–122, Control Lab, Stockholm, Sweden. He held research
Jan. 2011. positions at the University of California at Berkeley,
[30] R. O’Neil, A. Castillo, and M. B. Cain, “The IV formulation and Berkeley, CA, USA, (Visiting Scholar from 2004
linear approximation of the AC optimal power flow problem,” to 2005 and Research Associate from 2007 to 2008) and Royal Institute
2012. [Online]. Available: http://www.ferc.gov/industries/electric/ of Technology, Stockholm (Research Associate from 2005 to 2007). He has
indus-act/market-planning/opf-papers/acopf-2-iv-linearization.pdf co-authored more than 100 publications, including books, book chapters,
[31] S. Boyd, “Sequential convex programming,” University lecture, international journals and conferences, and an international patent. He has
2013. [Online]. Available: http://web.stanford.edu/class/ee364b/lectures/ also offered his advice as a Consultant to numerous technology companies,
seq_slides.pdf such as Berkeley Wireless Sensor Network Lab, Ericsson Research, Synopsys,
[32] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, “Randomized gossip and United Technology Research Center. He is co-funder and CTO of the
algorithms,” IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2508–2530, sensor networks start-up companies Lokkupp and MIND. His research interests
Jun. 2006. include optimization and parallel computation with applications to wireless-
[33] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, “Matpower: sensor networks, networked control systems, and wireless networks.
Steady-state operations, planning, analysis tools for power systems re- Dr. Fischione received a number of awards, including the best paper award
search and education,” IEEE Trans. Power Syst., vol. 26, no. 1, pp. 12–19, from the IEEE T RANSACTIONS ON I NDUSTRIAL I NFORMATICS of 2007, the
Feb. 2011. best paper awards at the IEEE International Conference on Mobile Ad-hoc and
[34] R. D. Christie, Power systems test case archive. Seattle, Washington, Sensor System 2005 and 2009 (IEEE MASS 2005 and IEEE MASS 2009); the
USA, Aug. 1999. [Online]. Available: http://www.ee.washington.edu/ Best Paper Award of the IEEE Sweden VT-COM-IT Chapter of 2014; the Best
research/pstca/. Business Idea award from VentureCup East Sweden in 2010; the Ferdinando
[35] MATLAB, ver. 8.1.0.604 (R2013a), The MathWorks Inc., Natick, MA, Filauro award from the University of L’Aquila, Italy, in 2003; the Higher
USA, 2013. Education award from Abruzzo Region Government, Italy, in 2004; and the
[36] J. Löfberg, “Yalmip: A toolbox for modeling and optimization in Junior Research award from Swedish Research Council, 2007; and the Silver
MATLAB,” presented at the CACSD Conf., Taipei, Taiwan, 2004. Ear of Wheat award in history from the Municipality of Tornimparte, Italy,
[Online]. Available: http://users.isy.liu.se/johanl/yalmip 2012. He has chaired or served as a technical member of program committees of
[37] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization several international conferences and is serving as referee for technical journals.
over symmetric cones,” Optimiz. Meth. Softw., vol. 11–12, pp. 625–653, He is Ordinary Member of the academy of history Deputazione Abruzzese di
1999. [Online]. Available: http://plato.asu.edu/ftp/usrguide.pdf Storia Patria.