Exact Convex Relaxation of Optimal Power Flow in Radial Networks

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TAC.2014.2332712, IEEE Transactions on Automatic Control
Limited circulation. For review only
1
Exact Convex Relaxation of Optimal Power Flow in

Radial Networks
Lingwen Gan, Na Li, Ufuk Topcu, and Steven H. Low
Abstract—The optimal power flow (OPF) problem determines may not be feasible (may not satisfy the nonlinear power flow
a network operating point that minimizes a certain objective such equations). In this case an operator typically tightens some
as generation cost or power loss. It is nonconvex. We prove that constraints in DC OPF and solves again. This may not only
a global optimum of OPF can be obtained by solving a second-
order cone program, under a mild condition after shrinking reduce efficiency but also relies on heuristics that are hard to
the OPF feasible set slightly, for radial power networks. The scale to larger systems or faster control in the future. Finally,
condition can be checked a priori, and holds for the IEEE 13, DC approximation is unsuitable for distribution systems where
34, 37, 123-bus networks and two real-world networks. loss is much higher than in transmission systems, voltages
can fluctuate significantly, and reactive powers are used to
stabilize voltages [17]. See [18] for a more accurate power
I. I NTRODUCTION
flow linearization that addresses these shortcomings of the DC
The optimal power flow (OPF) problem determines a net- approximation.
work operating point that minimizes a certain objective such as Many nonlinear algorithms that seek a local optimum of
generation cost or power loss. It has been one of the fundamen- the OPF problem have also been developed to avoid these
tal problems in power system operation since 1962. As dis- shortcomings. Representative algorithms include successive
tributed generation (e.g., photovoltaic panels) and controllable linear/quadratic programming [19], trust-region based methods
loads (e.g., electric vehicles) proliferate, OPF problems for [20], [21], Lagrangian Newton method [22], and interior-point
distribution networks become increasingly important. To use methods [23]–[25]. Some of them, especially those based on
controllable loads to integrate volatile renewable generation, Newton-Ralphson, are quite successful empirically. However,
solving the OPF problem in real-time will be inevitable. Power when they converge, these algorithms converge to a local
distribution networks are usually radial (have a tree topology). minimum without assurance on the suboptimality gap.
The OPF problem is difficult because power flow is gov- In this paper we focus on the convexification approach (see
erned by nonlinear Kirchhoff’s laws. There are three ways [26], [27] for a tutorial). Solving OPF through semidefinite
to deal with this challenge: 1) approximate the power flow relaxation is first proposed in [28] as a second-order cone pro-
equations; 2) look for a local optimum of the OPF problem; gram (SOCP) for radial networks and in [29] as a semidefinite
and 3) convexify the constraints imposed by the Kirchhoff’s program (SDP) for general networks in a bus injection model.
laws. After a brief discussion of the first two approaches, we It is first proposed in [30], [31] as an SOCP for radial networks
will focus on the third. See extensive surveys in e.g. [1]–[12]. in the branch flow model of [32], [33]. While these convex
Power flow equations can be approximated by linear equa- relaxations have been illustrated numerically in [28] and [29],
tions known as the DC power flow equations [13]–[15] if whether or when they are exact is first studied in [34] (i.e.,
1) power losses on the lines are small; 2) voltages are close when an optimal solution of the original OPF problem can be
to their nominal values; 3) voltage angle differences between recovered from every optimal solution of an SDP relaxation).
adjacent buses are small. With the DC approximation, the OPF Exploiting graph sparsity to simplify the SDP relaxation of
problem, called DC OPF, reduces to a linear program. For OPF is first proposed in [35], [36] and analyzed in [37]. These
transmission networks, these three assumptions are reasonable relaxations are equivalent for radial networks in the sense that
and DC OPF is widely used in practice. It, however, has there is a bijective map between their feasible sets [38]. The
three limitations. First, it is not applicable for applications SOCP relaxation, however, has a much lower computational
such as power routing (e.g. [16]) and volt/var control (e.g. complexity. We will hence focus on the SOCP relaxation in
[17]) since it assumes fixed voltage magnitudes and ignores this paper.
reactive powers. Second, a solution of the DC approximation Solving OPF through convex relaxation offers several ad-
vantages. It provides the ability to check if a solution is
This work was supported by NSF NetSE grant CNS 0911041, ARPA-E
grant DE-AR0000226, Southern California Edison, National Science Council globally optimal. If it is not, the solution provides a lower
of Taiwan, R.O.C, grant NSC 103-3113-P-008-001, Los Alamos National Lab bound on the minimum cost and hence a bound on how far any
through a DoE grant, Resnick Institute, and AFOSR award number FA9550- feasible solution is from optimality. Unlike approximations, if
12-1-0302.
Lingwen Gan, and Steven H. Low are with the Engineering and Applied a relaxation is infeasible, it is a certificate that the original
Science Division, California Institute of Technology, Pasadena, CA 91125 OPF is infeasible.
USA (e-mail: [email protected]). Na Li is with the Department of Electrical Convex relaxations may not be exact [39]–[41]. For radial
Engineering, Harvard University, Boston, MA 02138 USA. Ufuk Topcu is
with the Department of Electrical and Systems Engineering, University of networks, three types of sufficient conditions have been devel-
Pennsylvania, Philadelphia, PA 19104 USA. oped in the literature that guarantee their exactness. They are
Preprint submitted to IEEE Transactions on Automatic Control. Received: June 11, 2014 14:31:13 PST
0018-9286 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
2
not necessary in general and have implications on allowable For each bus i ∈ N , let vi denote the square of the magni-
power injections, voltage magnitudes, or voltage angles: tude of its complex voltage, e.g., if the voltage is 1.05∠120◦
A Power injections: These conditions require that not both per unit, then vi = 1.052 . The substation voltage v0 is fixed
constraints on real and reactive power injections be bind- and given. Let si = pi + iqi denote the power injection of bus
ing at both ends of a line [30], [31], [42]–[44]. i where pi and qi denote the real and reactive power injections
B Voltage angles: These conditions require that the voltage respectively. Let Pi denote the unique path from bus i to bus
angles across each line be sufficiently close [45]. This is 0. Since the network is radial, the path Pi is well-defined. For
needed also for stability reasons. each line (i, j) ∈ E, let zij = rij + ixij denote its impedance.
C Voltages magnitudes: These conditions require that the Let ìj denote the square of the magnitude of the complex
upper bounds on voltage magnitudes not be binding [46], current from bus i to bus j, e.g., if the current is 0.5∠10◦ ,
[47]. They can be enforced through affine constraints on then ìj = 0.52 . Let Sij = Pij + iQij denote the sending-end
power injections. This paper generalizes these results. power flow from bus i to bus j where Pij and Qij denote
the real and reactive power flow respectively. Some of the
Summary of contributions: The goal of this paper is to
notations are summarized in Fig. 1. We use a letter without
show that after modifying the OPF problem for radial networks
subscripts to denote a vector of the corresponding quantities,
slightly, the corresponding SOCP relaxation is exact under
e.g., v = (vi )i∈N + , ` = (ìj )(i,j)∈E . Note that subscript 0 is
a mild condition that can be checked a priori. In particular,
not included in nodal quantities such as v and s. For a complex
contributions of this paper are threefold.
number a ∈ C, let ā denote the conjugate of a.
First, we prove in Theorem 1 that if voltage upper bounds do
not bind at optimality, then the SOCP relaxation is exact under Pi
a mild condition. The condition can be checked a priori and
holds for the IEEE 13, 34, 37, 123-bus networks and two real- Bus 0 Bus j Bus i
world networks. The condition has a physical interpretation vj zij vi
that all upstream reverse power flows increase if the power Sij , ìj
loss on a line is reduced. si
Second, in Section IV we modify the OPF problem by
limiting power injections to a region where voltage upper Fig. 1. Some of the notations.
bounds do not bind so that the SOCP relaxation is exact
Given the network (N , E), the impedance z, and the substa-
under the aforementioned condition. We illustrate that this only
tion voltage v0 , the other variables (s, S, v, `, s0 ) are described
eliminates power injections from the original feasible set that
by the branch flow model for radial networks [32], [33]:
are close to voltage upper bounds. Examples exist where the X
SOCP relaxation is not exact without this modification. Sij = si + (Shi − zhi `hi ), ∀(i, j) ∈ E; (1a)
Third, we prove in Theorem 4 that the result in this paper h: h→i
X
unifies and generalizes the results in [46], [47]. 0 = s0 + (Sh0 − zh0 `h0 ); (1b)
The rest of this paper is organized as follows. The OPF h: h→0
problem and the SOCP relaxation are introduced in Section II, vi − vj = 2Re(z̄ij Sij ) − |zij |2 ìj , ∀(i, j) ∈ E; (1c)
and a sufficient condition for exactness is provided in Section
|Sij |2
III. The condition consists of two parts, C1 and C2. Since ìj = , ∀(i, j) ∈ E (1d)
C2 cannot be checked a priori, we propose in Section IV a vi
modified OPF problem that always satisfies C2 and therefore B. The OPF problem
its SOCP relaxation is exact under C1. We compare C1 with
We consider the following controllable devices in a distri-
prior works in Section V and show that C1 holds with large
bution network: distributed generators, inverters, controllable
margin for a number of test networks in Section VI.
loads such as electric vehicles and smart appliances, and
shunt capacitors. For application examples, in volt/var control,
II. T HE OPTIMAL POWER FLOW PROBLEM reactive power injection of inverters and shunt capacitors
are controlled to regulate voltages; in demand response, real
A. Power flow model
power consumption of controllable loads is reduced or shifted.
A distribution network is composed of buses and lines Mathematically, power injection s is the control variable,
connecting these buses, and is usually radial. The root of the after specifying which the other variables (S, v, `, s0 ) are
network is a substation bus that connects to the transmission determined by the power flow laws in (1).
network. It has a fixed voltage and redistributes the bulk power The power injection si of a bus i ∈ N + is constrained to
it receives from the transmission network to other buses. Index be in a pre-specified set Si , i.e.,
the substation bus by 0 and the other buses by 1, . . . , n. Let
si ∈ Si , i ∈ N +. (2)
N := {0, . . . , n} denote the collection of all buses and define
N + := N \{0}. Each line connects an ordered pair (i, j) The set Si for some controllable devices are:
of buses where bus j lies on the unique path from bus i to • If si represents a shunt capacitor with nameplate capacity
bus 0. Let E denote the collection of all lines, and abbreviate q i , then Si = {s ∈ C | Re(s) = 0, Im(s) = 0 or q i }.
(i, j) ∈ E by i → j whenever convenient. Note that Si is nonconvex and disconnected in this case.
3
• If si represents a solar panel with generation capacity cone programming (SOCP) relaxation [31]:
pi , that is connected to the grid through an inverter with X
SOCP: min fi (Re(si ))
nameplate capacity si , then Si = {s ∈ C | 0 ≤ Re(s) ≤
i∈N
pi , |s| ≤ si }.
• If si represents a controllable load with constant power
over s, S, v, `, s0
factor η, whose real power consumption can vary contin- s.t. (5a) − (5c), (5e) − (5f);
uously
n from −pi to −pi (here pi ≤ pi ≤ 0), then Si = o |Sij |2
p ` ij ≥ , ∀(i, j) ∈ E. (6)
s ∈ C | pi ≤ Re(s) ≤ pi , Im(s) = 1 − η 2 Re(s)/η . vi
Note that si can represent the aggregate power injection of Note that SOCP is not necessarily convex, since we allow fi
multiple such devices with an appropriate Si , and that the set to be nonconvex and Si to be nonconvex. Nonetheless, we call
Si is not necessarily convex or connected. it SOCP for brevity.
An important goal of control is to regulate the voltages to If an optimal SOCP solution w = (s, S, v, `, s0 ) is feasible
lie within pre-specified lower and upper bounds v i and v i , i.e., for OPF, i.e., w satisfies (5d), then w is a global optimum of
OPF. This motivates the following definition.
v i ≤ vi ≤ v i , i ∈ N +. (3)
Definition 1. SOCP is exact if every of its optimal solutions
For example, if voltages must not deviate by more than 5% satisfies (5d).
from their nominal values, then 0.952 ≤ vi ≤ 1.052 per unit.
We consider the control objective III. A SUFFICIENT CONDITION
X We now provide a sufficient condition that ensures SOCP
C(s, s0 ) = fi (Re(si )) (4)
i∈N
is exact. It motivates a modified OPF problem in Section IV.
where fi : R → R denotes the generation cost at bus i for

A. Statement of the condition
i ∈ N . If fi (x) = x for i ∈ N , then C is the total power loss
on the network. We start with introducing the notations that will be used in
The OPF problem seeks to minimize the generation cost the statement of the condition. One can ignore the ` terms in
(4), subject to power flow constraints (1), power injection (1a) and (1c) to obtain the Linear DistFlow Model [32], [33]:
X
constraints (2), and voltage constraints (3): Sij = si + Shi , ∀(i, j) ∈ E;
X h: h→i
OPF: min fi (Re(si ))
vi − vj = 2Re(z̄ij Sij ), ∀(i, j) ∈ E.
i∈N
over s, S, v, `, s0 Let (Ŝ, v̂) denote the solution of the Linear DistFlow model,
X then
s.t. Sij = si + (Shi − zhi `hi ), ∀(i, j) ∈ E; X
h: h→i Ŝij (s) = sh , ∀(i, j) ∈ E;
(5a) h: i∈Ph
X X
0 = s0 + (Sh0 − zh0 `h0 ); (5b) v̂i (s) := v0 + 2 Re z̄jk Ŝjk (s) , ∀i ∈ N
h: h→0 (j,k)∈Pi
vi − vj = 2Re(z̄ij Sij ) − |zij |2 ìj , ∀(i, j) ∈ E;
(5c) as in Fig. 2. Physically, Ŝij (s) denotes the sum of power
2 injections sh towards bus 0 that go through line (i, j). Note
|Sij |
ìj = , ∀(i, j) ∈ E; (5d) that (Ŝ(s), v̂(s)) is affine in s, and equals (S, v) if and only if
vi line loss zij ìj is 0 for (i, j) ∈ E. For two complex numbers
si ∈ Si , i ∈ N + ; (5e)
v i ≤ vi ≤ v i , i ∈ N +. (5f)
v0
The following assumptions are made throughout this paper. 0 k
v̂i
2Re(z̄jk Ŝjk (s))
A1 The network (N , E) is a tree. Distribution networks are j i {h : i 2 Ph }
Ŝij (s)
usually radial.
A2 The substation voltage v0 is fixed and given. In practice, Ŝij = sum of s in shaded region
v̂i = v0 + sum of terms over dashed path
v0 can be modified several times a day, and therefore can
be considered as a given constant at the timescale of OPF. Fig. 2. Illustration of Ŝij and v̂i . The shaded region is downstream of bus
A3 Line resistances and reactances are strictly positive, i.e., i, and contains the buses {h : i ∈ Ph }. Quantity Ŝij (s) is defined to be the
rij > 0 and xij > 0 for (i, j) ∈ E. This holds in practice sum of bus injections in the shaded region. The dashed lines constitute the
path Pi from bus i to bus 0. Quantity v̂i (s) is defined as v0 plus the terms
because lines are passive (consume power) and inductive.
2Re(z̄jk Ŝjk (s)) over the dashed path.
A4 Voltage lower bounds are strictly positive, i.e., v i > 0 for
i ∈ N + . In practice, v i is slightly below 1 per unit. a, b ∈ C, let a ≤ b denote Re(a) ≤ Re(b) and Im(a) ≤ Im(b).
The equality constraint (5d) is nonconvex, and one can relax For two vectors a, b of the same dimension, let a ≤ b denote
it to inequality constraints to obtain the following second-order componentwise inequality. Define <, >, and ≥ similarly.
4
Lemma 1. If (s, S, v, `, s0 ) satisfies (1a)–(1c) and ` ≥ 0 denote the power injection region where v̂(s) is upper bounded
componentwise, then S ≤ Ŝ(s) and v ≤ v̂(s). by v. Since v ≤ v̂(s) (Lemma 1), the set Svolt is a power
injection region where voltage upper bounds do not bind.
Lemma 1 implies that v̂(s) and Ŝ(s) provide upper bounds
The following theorem provides a sufficient condition that
on v and S. It is proved in Appendix A. Let P̂ (s) and Q̂(s)
guarantees the exactness of SOCP.
denote the real and imaginary parts of Ŝ(s) respectively. Then
X Theorem 1. Assume that f0 is strictly increasing, and that
P̂ij (s = p + iq) = P̂ij (p) = ph , (i, j) ∈ E; there exist pi and q i such that Si ⊆ {s ∈ C | Re(s) ≤
h: i∈Ph
X pi , Im(s) ≤ q i } for i ∈ N + . Then SOCP is exact if the
Q̂ij (s = p + iq) = Q̂ij (q) = qh , (i, j) ∈ E. following conditions hold:
h: i∈Ph C1 Als Als+1 · · · Alt−1 ult > 0 for any l ∈ L and any s, t
Assume that there exist pi and q i such that such that 1 ≤ s ≤ t ≤ nl ;
C2 every optimal SOCP solution w = (s, S, v, `, s0 ) satisfies
Si ⊆ {s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } s ∈ Svolt .
for i ∈ N + as in Fig. 3, i.e., Re(si ) and Im(si ) are upper Theorem 1 implies that if C2 holds, i.e., optimal power
bounded by pi and q i respectively. Define a+ := max{a, 0} injections lie in the region Svolt where voltage upper bounds
do not bind, then SOCP is exact under C1. C2 depends on
Im SOCP solutions and cannot be checked a priori. This drawback
motivates us to modify OPF such that C2 always holds and
qi therefore the corresponding SOCP is exact under C1, as will
be discussed in Section IV.
Si
We illustrate the proof idea of Theorem 1 via a 3-bus linear
network in Fig. 5. The proof for general radial networks is
0 pi Re
0 1 2
S0, 1 S10 S21
A1 u2
Fig. 3. We assume that Si lies in the left bottom corner of (pi , q i ), but do s1 s2
not assume that Si is convex or connected. S0, 1 = s0
for a ∈ R. Let I := diag(1, 1) denote the 2×2 identity matrix, Fig. 5. A 3-bus linear network.
and define
provided in Appendix B. Assume C1 and C2 hold. If SOCP
rij 2 rij +
is not exact, then there exists an optimal SOCP solution
uij := , Aij := I − P̂ij (p) Q̂+
ij (q)
xij v i xij w = (s, S, v, `, s0 ) that violates (5d). We will construct
another feasible point w0 = (s0 , S 0 , v 0 , `0 , s00 ) of SOCP that has
for (i, j) ∈ E. For each i ∈ N + , (i, j1 ) ∈ E and (i, j2 ) ∈ E
a smaller objective value than w, contradicting the optimality
implies j1 = j2 , and therefore we can abbreviate uij and Aij
of w and implying SOCP is exact.
by ui and Ai respectively without ambiguity.
There are two ways (5d) gets violated: 1) (5d) is violated on
Further, let L := {l ∈ N | @k ∈ N such that k → l}
line (1, 0); or 2) (5d) is satisfied on line (1, 0) but violated on
denote the collection of leaf buses in the network. For a leaf
line (2, 1). To illustrate the proof idea, we focus on the second
bus l ∈ L, let nl + 1 denote the number of buses on path Pl ,
case, i.e., the case where `10 = |S10 |2 /v1 and `21 > |S21 |2 /v2 .
and suppose
In this case, the construction of w0 is
Pl = {lnl → lnl −1 → . . . → l1 → l0 }
Initialization: s0 = s, S21
0
= S21 ;
with lnl = l and l0 = 0 as in Fig. 4. Let Forward sweep: `021 = |S21
0 2
| /v2 ,
0 0
S10 = S21 − z21 `021 + s01 ;
L `010 = |S10
0 2
| /v1 ,
0 0
l0 u l1
l1 l nl 1
S0,−1 = S10 − z10 `010 ;
Al 1 u l2
Al 2 l 2
Backward sweep: v10 = v0 + 2Re(z̄10 S10
0
) − |z10 |2 `010 ;
l nl v20 = v10 + 2Re(z̄21 S21
0
) − |z21 |2 `021
l0 = 0, lnl = l where S0,−1

0
= −s00 . The construction consists of three steps:
S1 In the initialization step, s0 and S21
0
are initialized as the
Fig. 4. The shaded region denotes the collection L of leaf buses, and the corresponding values in w.
path Pl of a leaf bus l ∈ L is illustrated by a dashed line.
S2 In the forward sweep step, `0k,k−1 and Sk−1,k−2 0
are
recursively constructed for k = 2, 1 by alternatively
Svolt := {s ∈ Cn | v̂i (s) ≤ v i for i ∈ N + } applying (5d) (with v 0 replaced by v) and (5a)/(5b).
5
This recursive construction updates `0 and S 0 alternatively For practical parameter ranges of (r, x, p, q, v), line resis-
along the path P2 from bus 2 to bus 0, and is therefore tance and reactance rij , xij 1 per unit for (i, j) ∈ E,
called a forward sweep. line flows P̂ij (p), Q̂ij (q) are on the order of 1 per unit for
S3 In the backward sweep step, vk0 is recursively constructed (i, j) ∈ E, and voltage lower bound v i ≈ 1 per unit for
for k = 1, 2 by applying (5c). This recursive construction i ∈ N + . Hence, Ai is close to I for i ∈ N + , and therefore
updates v 0 along the negative direction of P2 from bus 0 C1 is likely to hold. As will be seen in Section VI, C1 holds
to bus 2, and is therefore called a backward sweep. for several test networks, including those with big (p, q) (high
One can show that w0 is feasible for SOCP and has a smaller penetration of distributed generation).
objective value than w. This contradicts the optimality of w, C1 has a physical interpretation. Recall that Sk,k−1 denotes
and therefore SOCP is exact. the reverse power flow on line (k, k − 1) for k = 1, . . . , n
and introduce S0,−1 := −s0 for convenience. If the power
Remark 1. Theorem 1 still holds if there is an additional loss on a line is reduced, it is natural that all upstream reverse
power injection constraint s ∈ S in OPF, where S can be an power flows will increase. More specifically, the power loss
arbitrary set. This is because we set s0 = s in the construction on line (t, t − 1) where t ∈ {1, 2, . . . , n} is reduced if the
of w0 , and therefore s ∈ S implies s0 ∈ S. Hence, an current `t,t−1 is reduced by −d`t,t−1 > 0. When power loss
additional constraint s ∈ S does not affect the fact that w0 is gets smaller, reverse power flow Ss−1,s−2 is likely to increase,
feasible for SOCP and has a smaller objective value than w. i.e., dSs−1,s−2 > 0, for s = 1, 2, . . . , t.
Let dSs−1,s−2 = dPs−1,s−2 + idQs−1,s−2 > 0 for s =
B. Interpretation of C1 1, . . . , t. It can be verified that (dPt−1,t−2 dQt−1,t−2 )T =
We illustrate C1 through a linear network as in Fig. 6. The −ut d`t,t−1 , and one can compute from (1) the Jacobian matrix
collection of leaf buses is a singleton L = {n}, and the path ∂P ∂P
!
k−1,k−2 k−1,k−2
from the only leaf bus n to bus 0 is Pn = {n → n − 1 → Ak := ∂Pk,k−1
∂Qk−1,k−2
∂Qk,k−1
∂Qk−1,k−2
· · · → 1 → 0}. Then, C1 takes the form ∂Pk,k−1 ∂Qk,k−1

As As+1 · · · At−1 ut > 0, 1 ≤ s ≤ t ≤ n. 2 rk,k−1
= I− (Pk,k−1 Qk,k−1 )
vk xk,k−1
That is, given any network segment (s − 1, t) where 1 ≤
s ≤ t ≤ n, the multiplication As As+1 · · · At−1 of A over the for k = 1, . . . , n. Therefore
segment (s − 1, t − 1) times ut is strictly positive. (dPs−1,s−2 dQs−1,s−2 )T = −As As+1 · · · At−1 ut d`t,t−1
network segment (s 1, t) for s = 1, . . . , t. Then, dSs−1,s−2 > 0 implies

0 s 1 s t 2 t 1 t n As As+1 · · · At−1 ut > 0 (7)
As At 1 ut for s = 1, 2, . . . , t. Note that Ak is obtained by replacing
✓ ◆
dPs 1,s 2 (P, Q, v) in Ak by (P̂ + (p), Q̂+ (q), v) (so that Ak only de-
= As · · · At 2 ut d`t,t 1
dQs 1,s 2 pends on SOCP parameters), and then (7) becomes C1.
Fig. 6. In the above linear network, L = {n} and Pn = {n → n − 1 →
· · · → 1 → 0}. C1 requires that given any highlighted segment (s − 1, t) IV. A MODIFIED OPF PROBLEM
where 1 ≤ s ≤ t ≤ n, the multiplication of A over (s − 1, t − 1) times ut
is strictly positive (componentwise). The condition C2 in Theorem 1 depends on SOCP solutions
and cannot be checked a priori. It can however be enforced
C1 only depends on SOCP parameters (r, x, p, q, v). It can by the additional constraint
be checked a priori and efficiently since A and u are simple
s ∈ Svolt (8)
functions of (r, x, p, q, v) that can be computed in O(n) time
and there are no more than n(n + 1)/2 inequalities in C1. on OPF. Condition (8) is equivalent to n affine constraints on
Proposition 1. If (p, q) ≤ (p , q ) and C1 holds for 0 0 s, v̂i (s) ≤ v i for i ∈ N + . Since vi ≤ v̂i (s) (Lemma 1), the
(r, x, p0 , q 0 , v), then C1 also holds for (r, x, p, q, v). constraints vi ≤ v i in (5f) become redundant after imposing
(8). To summarize, the modified OPF problem is
Proposition 1 implies that if C1 holds a set of power X
injections, then C1 also holds for smaller power injections. OPF-m: min fi (Re(si ))
It is proved in Appendix C. i∈N
over s, S, v, `, s0
Proposition 2. If (p, q) ≤ 0, then C1 holds.
s.t. (5a) − (5e);
Proposition 2 implies that if every bus only consumes real v i ≤ vi , v̂i (s) ≤ v i , i ∈ N +. (9)
and reactive power, then C1 holds. This is because when
(p, q) ≤ 0, the quantities P̂ij (p) ≤ 0, Q̂ij (q) ≤ 0 for A modification to OPF is necessary to ensure an exact
(i, j) ∈ E. It follows that Ai = I for i ∈ N + . Hence, SOCP, since otherwise examples exist where SOCP is not
Als · · · Alt−1 ult = ult > 0 for any l ∈ L and any s, t such exact. Remarkably, the feasible sets of OPF-m and OPF are
that 1 ≤ s ≤ t ≤ nl . similar since v̂i (s) is close to vi in practice [17], [32], [33].
6
One can relax (5d) to (6) to obtain the corresponding SOCP 3) rij /xij ≥ rjk /xjk whenever (i, j), (j, k) ∈ E; and
relaxation for OPF-m: P̂ij (p) ≤ 0, v i − 2xij Q̂+ 0
ij (q) > 0 for all (i, j) ∈ E .
X 4) rij /xij ≤ rjk /xjk whenever (i, j), (j, k) ∈ E; and
SOCP-m: min fi (Re(si ))
Q̂ (q) ≤ 0, v i − 2rij P̂ij+ (p) > 0 for all (i, j) ∈ E 0 .
i∈N  ij Y X 
over s, S, v, `, s0 ckl − dkl
 (k,l)∈Pj (k,l)∈Pj 
s.t. (5a) − (5c), (6), (5e), (9). 5)  X Y  rij > 0 for all (i, j) ∈ E
− ekl fkl  xij
Note again that SOCP-m is not necessarily convex, since we (k,l)∈Pj (k,l)∈Pj
allow fi and Si to be nonconvex. where ckl := 1 − 2rkl P̂kl +
(p̄)/v k , dkl := 2rkl Q̂+
kl (q̄)/v k ,
Since OPF-m is obtained by imposing additional constraint ekl := 2xkl P̂kl (p̄)/v k , and fkl := 1 − 2xkl Q̂+
+
(q̄)/v
kl k.
(8) on OPF, it follows immediately from Remark 1 that SOCP-
m relaxation is exact under C1—a mild condition that can be The results in [46], [47] say that, if there are no voltage
checked a priori. upper bounds, i.e., v = ∞, then SOCP is exact if any one of
1)–5) holds. Since C2 holds automatically when v = ∞ and
Theorem 2. Assume that f0 is strictly increasing, and that C1 holds if any one of 1)–5) holds (Theorem 4), the results
there exist pi and q i such that Si ⊆ {s ∈ C | Re(s) ≤ in [46], [47] follow from Theorem 1. Besides, the following
pi , Im(s) ≤ q i } for i ∈ N + . Then SOCP-m is exact if C1 corollary follows immediately from Theorems 2 and 4.
holds.
Corollary 2. Assume that f0 is strictly increasing, and that
The next result implies that SOCP (SOCP-m) has at most there exist pi and q i such that Si ⊆ {s ∈ C | Re(s) ≤
one optimal solution if it is convex and exact. The theorem is pi , Im(s) ≤ q i } for i ∈ N + . Then SOCP-m is exact if any
proved in Appendix D. one of 1)–5) holds.
Theorem 3. If fi is convex for i ∈ N , Si is convex for i ∈
N + , and SOCP (SOCP-m) is exact, then SOCP (SOCP-m) VI. C ASE S TUDIES
has at most one optimal solution. In this section, we use six test networks to demonstrate that
The proof of Theorem 3 also implies that the feasible set of 1) SOCP is simpler computationally than SDP.
OPF (OPF-m) is hollow, as stated in the following corollary. 2) C1 holds. We define C1 margin that quantifies how well
C1 is satisfied, and show that the margin is big.
˜ ṽ0 ) and x̂ = (ŝ, Ŝ, v̂, `,
Corollary 1. Let x̃ = (s̃, S̃, ṽ, `, ˆ v̂0 ) be
3) The feasible sets of OPF and OPF-m are similar. We
two distinct feasible points of OPF (OPF-m), then any convex define modification gap that quantifies the difference
combination of x̃ and x̂ cannot be feasible for OPF (OPF- between the feasible sets of OPF and OPF-m, and show
m), i.e., the point x = θx̃ + (1 − θ)x̂ is infeasible for OPF that this gap is small.
(OPF-m) for any θ ∈ (0, 1).
The proof of Corollary 1 is similar to that of Theorem 3 A. Test networks
and omitted for brevity.
The test networks include IEEE 13, 34, 37, 123-bus net-
works [48] and two real-world networks [30], [49] in the
V. C ONNECTION WITH PRIOR RESULTS
service territory of Southern California Edison (SCE), a utility
Theorem 1 unifies and generalizes the results in [46], [47] company in California, USA [50].
due to Theorem 4 proved in Appendix E. Theorem 4 below The IEEE networks are unbalanced three-phase radial net-
says that C1 holds if at least one of the followings hold: 1) works with some devices (regulators, circuit switches, trans-
Every bus only consumes real and reactive power; 2) lines formers, and distributed loads) not modeled in (1). Therefore
share the same resistance to reactance ratio; 3) The buses we modify the IEEE networks as follows.
only consume real power and the resistance to reactance ratio
1) Assume that each bus has three phases and split its load
increases as lines branch out from the substation; 4) The buses
uniformly among the three phases.
only consume reactive power and the resistance to reactance
2) Assume that the three phases are decoupled so that the
ratio decreases as lines branch out from the substation; 5)
network becomes three identical single phase networks.
upper bounds P̂ + (p̄), Q̂+ (q̄) on reverse power flows are
3) Model closed circuit switches as shorted lines and ignore
sufficiently small. Let
open circuit switches. Model regulators as multiplying
E 0 := {(i, j) ∈ E | i ∈
/ L} the voltages by fixed constants (set to 1.08 in the sim-
ulations). Model transformers as lines with appropriate
denote the set of all non-leaf lines.
impedances. Model the distributed load on a line as two
Theorem 4. Assume that there exist pi and q i such that Si ⊆ identical spot loads located at two ends of the line.
{s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1 The SCE networks, a 47-bus network and a 56-bus network,
holds if any one of the following statements is true: are shown in Fig. 7 with parameters given in Tables I and II.
1) Ŝij (p + iq) ≤ 0 for all (i, j) ∈ E 0 . These networks have increasing penetration of distributed
2) rij /xij is identical for all (i, j) ∈ E; and generation (DG) as listed in Table III. While the IEEE
v i − 2rij P̂ij+ (p) − 2xij Q̂+ 0
ij (q) > 0 for all (i, j) ∈ E . networks do not have any DG, the SCE 47-bus network
7
Fig. 7. Topologies of the SCE 47-bus and 56-bus networks [30], [49].
TABLE I
L INE IMPEDANCES , PEAK SPOT LOAD , AND NAMEPLATE RATINGS OF CAPACITORS AND PV GENERATORS OF THE 47- BUS NETWORK .
Network Data
Line Data Line Data Line Data Load Data Load Data PV Generators
From To R X From To R X From To R X Bus Peak Bus Peak Bus Nameplate
Bus Bus (Ω) (Ω) Bus Bus (Ω) (Ω) Bus Bus (Ω) (Ω) No MVA No MVA No Capacity
1 2 0.259 0.808 8 41 0.107 0.031 21 22 0.198 0.046 1 30 34 0.2
2 13 0 0 8 35 0.076 0.015 22 23 0 0 11 0.67 36 0.27 13 1.5MW
2 3 0.031 0.092 8 9 0.031 0.031 27 31 0.046 0.015 12 0.45 38 0.45 17 0.4MW
3 4 0.046 0.092 9 10 0.015 0.015 27 28 0.107 0.031 14 0.89 39 1.34 19 1.5 MW
3 14 0.092 0.031 9 42 0.153 0.046 28 29 0.107 0.031 16 0.07 40 0.13 23 1 MW
3 15 0.214 0.046 10 11 0.107 0.076 29 30 0.061 0.015 18 0.67 41 0.67 24 2 MW
4 20 0.336 0.061 10 46 0.229 0.122 32 33 0.046 0.015 21 0.45 42 0.13
4 5 0.107 0.183 11 47 0.031 0.015 33 34 0.031 0.010 22 2.23 44 0.45 Shunt Capacitors
5 26 0.061 0.015 11 12 0.076 0.046 35 36 0.076 0.015 25 0.45 45 0.2 Bus Nameplate
5 6 0.015 0.031 15 18 0.046 0.015 35 37 0.076 0.046 26 0.2 46 0.45 No. Capacity
6 27 0.168 0.061 15 16 0.107 0.015 35 38 0.107 0.015 28 0.13
6 7 0.031 0.046 16 17 0 0 42 43 0.061 0.015 29 0.13 Base Voltage (kV) = 12.35 1 6000 kVAR
7 32 0.076 0.015 18 19 0 0 43 44 0.061 0.015 30 0.2 Base kVA = 1000 3 1200 kVAR
7 8 0.015 0.015 20 21 0.122 0.092 43 45 0.061 0.015 31 0.07 Substation Voltage = 12.35 37 1800 kVAR
8 40 0.046 0.015 20 25 0.214 0.046 32 0.13 47 1800 kVAR
8 39 0.244 0.046 21 24 0 0 33 0.27
has 56.6% DG penetration (6.4MW nameplate DG capac- 2) The power injection constraints are as follows. For each
ity against 11.3MVA peak spot load), and the SCE 56-bus bus i ∈ N + , there may be multiple devices including
network has 130.4% DG penetration (5MW nameplate DG loads, capacitors, and PV panels. Assume that there is a
capacity against 3.835MVA peak spot load). total of Di such devices and label them by 1, 2, . . . , Di .
Let si,d denote the power injection of device d for d =
B. SOCP is more efficient to compute than SDP 1, 2, . . . , Di . If device d is a load with given real and
We compare the computation times of SOCP and SDP for reactive power consumptions p and q, then we impose
the test networks, and summarize the results in Table III. All
si,d = −p − iq. (10)
simulations in this paper use matlab 7.9.0.529 (64-bit) with
toolbox cvx 1.21 on Mac OS X 10.7.5 with 2.66GHz Intel If device d is a load with given peak apparent power
Core 2 Due CPU and 4GB 1067MHz DDR3 memory. speak , then we impose
We use the following OPF setup throughout the simulations.
1) The objective is to minimize power loss in the network. si,d = −speak exp(jθ) (11)
8
TABLE II
L INE IMPEDANCES , PEAK SPOT LOAD , AND NAMEPLATE RATINGS OF CAPACITORS AND PV GENERATORS OF THE 56- BUS NETWORK .
Network Data
Line Data Line Data Line Data Load Data Load Data Load Data
From To R X From To R X From To R X Bus Peak Bus Peak Bus Peak
Bus. Bus. (Ω) (Ω) Bus. Bus. (Ω) (Ω) Bus. Bus. (Ω) (Ω) No. MVA No. MVA No. MVA
1 2 0.160 0.388 20 21 0.251 0.096 39 40 2.349 0.964 3 0.057 29 0.044 52 0.315
2 3 0.824 0.315 21 22 1.818 0.695 34 41 0.115 0.278 5 0.121 31 0.053 54 0.061
2 4 0.144 0.349 20 23 0.225 0.542 41 42 0.159 0.384 6 0.049 32 0.223 55 0.055
4 5 1.026 0.421 23 24 0.127 0.028 42 43 0.934 0.383 7 0.053 33 0.123 56 0.130
4 6 0.741 0.466 23 25 0.284 0.687 42 44 0.506 0.163 8 0.047 34 0.067 Shunt Cap
4 7 0.528 0.468 25 26 0.171 0.414 42 45 0.095 0.195 9 0.068 35 0.094 Bus Mvar
7 8 0.358 0.314 26 27 0.414 0.386 42 46 1.915 0.769 10 0.048 36 0.097 19 0.6
8 9 2.032 0.798 27 28 0.210 0.196 41 47 0.157 0.379 11 0.067 37 0.281 21 0.6
8 10 0.502 0.441 28 29 0.395 0.369 47 48 1.641 0.670 12 0.094 38 0.117 30 0.6
10 11 0.372 0.327 29 30 0.248 0.232 47 49 0.081 0.196 14 0.057 39 0.131 53 0.6
11 12 1.431 0.999 30 31 0.279 0.260 49 50 1.727 0.709 16 0.053 40 0.030 Photovoltaic
11 13 0.429 0.377 26 32 0.205 0.495 49 51 0.112 0.270 17 0.057 41 0.046 Bus Capacity
13 14 0.671 0.257 32 33 0.263 0.073 51 52 0.674 0.275 18 0.112 42 0.054
13 15 0.457 0.401 32 34 0.071 0.171 51 53 0.070 0.170 19 0.087 43 0.083 45 5MW
15 16 1.008 0.385 34 35 0.625 0.273 53 54 2.041 0.780 22 0.063 44 0.057
15 17 0.153 0.134 34 36 0.510 0.209 53 55 0.813 0.334 24 0.135 46 0.134 Vbase = 12kV
17 18 0.971 0.722 36 37 2.018 0.829 53 56 0.141 0.340 25 0.100 47 0.045 Sbase = 1MVA
18 19 1.885 0.721 34 38 1.062 0.406 27 0.048 48 0.196 Zbase = 144Ω
4 20 0.138 0.334 38 39 0.610 0.238 28 0.038 50 0.045
TABLE III
DG PENETRATION , C1 MARGINS , MODIFICATION GAPS , AND COMPUTATION TIMES FOR DIFFERENT TEST NETWORKS .
DG penetration numerical precision SOCP time SDP time C1 margin estimated modification gap
IEEE 13-bus 0% 10−10 0.5162s 0.3842s 27.6762 0.0362
IEEE 34-bus 0% 10−10 0.5772s 0.5157s 20.8747 0.0232
IEEE 37-bus 0% 10−9 0.5663s 1.6790s +∞ 0.0002
IEEE 123-bus 0% 10−8 2.9731s 32.6526s 52.9636 0.0157
SCE 47-bus 56.6% 10−8 0.7265s 2.5932s 2.5416 0.0082
SCE 56-bus 130.4% 10−9 1.0599s 6.0573s 1.2972 0.0053
where θ = cos−1 (0.9), i.e, power injection si,d is con- 35

SDP
12
sidered to be a constant, obtained by assuming a power 30 SOCP 10
SDP time / SOCP time

computation time (s)
factor of 0.9 at peak apparent power. If device d is a 25

8
capacitor with nameplate q, then we impose 20
6
15
Re(si,d ) = 0 and 0 ≤ Im(si,d ) ≤ q. (12) 10
4
5 2
If device d is a PV panel with nameplate s, then we
0 0
impose 20 40 60 80
number of buses
100 120 20 40 60 80
number of buses
100 120
Re(si,d ) ≥ 0 and |si,d | ≤ s. (13)

Fig. 8. Comparison of the computation times for SOCP and SDP.
The power injection at bus i is
Di
X
si = si,d dramatically. Hence SOCP is much more efficient than SDP
d=1 for medium to large networks.
where si,d satisfies one of (10)–(13). SOCP and SDP can only be solved to certain numerical
3) The voltage regulation constraint is considered to be precisions. The best numerical precision we obtain without
0.92 ≤ vi ≤ 1.12 for i ∈ N + . Note that we choose a applying pre-conditioning techniques are listed in Table III.
small voltage lower bound 0.9 so that OPF is feasible
for all test networks. We choose a big voltage upper
bound 1.1 such that Condition C2 holds and therefore
SDP/SOCP is exact under C1. C. C1 holds with a large margin
The computation times of SDP and SOCP for different test
networks are summarized in Fig. 8. The number of buses In this section, we show that C1 holds with a large margin
determines the number of constraints and variables in the for all test networks. Noting that C1 becomes more difficult
optimization, and therefore reflects the problem size. Network to hold as (p, q) increases (Proposition 1), one can increase p,
topology also affects the computation time. As the number q until C1 fails. More specifically, let pfix
i and qi denote the
fix
of buses increases, the computation time of SOCP scales up fixed real and reactive loads at bus i ∈ N , let PVi and Capi
+
much more slowly than that of SDP and their ratio increases denote the nameplate capacities of the photovoltaic panels and
9
the shunt capacitors at bus i ∈ N + , and define stricter voltage upper bound constraint:
X
pi (η) := pfix
i + η · PVi , i ∈ N + , η ≥ 0; OPF-ε: min fi (Re(si ))
q i (η) := qifix + η · (PVi + Capi ), i ∈ N + , η ≥ 0. i∈N
over s, S, v, `, s0
When η = 0, one has (p(η), q(η)) ≤ 0 and therefore C1 holds
s.t. (5a) − (5e);
according to Proposition 2. According to Proposition 1, there
exists a unique η ∗ ∈ R+ ∪ {+∞} such that v i ≤ vi ≤ v i − ε, i ∈ N +.
η < η ∗ ⇒ C1 holds for (r, x, p(η), q(η), v); (14a) The feasible set FOPF-ε of OPF-ε is contained in FOPF . Hence,
∗
η > η ⇒ C1 does not hold for (r, x, p(η), q(η), v). (14b) for every (s, S, `, v, s0 ) ∈ FOPF-ε ⊆ FOPF , one has
Definition 2. C1 margin is defined as the unique η ∗ ≥ 0 that v̂i (s) ≤ vi + kv̂(s) − vk∞ ≤ v i − ε + ε = v i , i ∈ N+
satisfies (14).
Physically, η ∗ is the multiple by which one can scale up by (15). It follows that FOPF-ε ⊆ FOPF-m and therefore
distributed generation (PVs) and shunt capacitors before C1
fails to hold. Noting that p = p(1) and q = q(1), C1 holds FOPF-ε ⊆ FOPF-m ⊆ FOPF
for (r, x, p, q, v) if and only if η ∗ > 1 (ignore the corner case
where η ∗ = 1). The larger η ∗ is, the “more easily” C1 holds. as illustrated in Fig. 9.
The C1 margins of different test networks are summarized If ε is small, then FOPF-m is similar to FOPF . Any point w
in Table III. The minimum C1 margin is 1.30, meaning that that is feasible for OPF but infeasible for OPF-m is close to
one can scale up distributed generation and shunt capacitors the voltage upper bound since vi > v i − ε for some i ∈ N + .
by 1.30 times before C1 fails to hold. C1 margin of the IEEE Such points are perhaps undesirable for robust operation.
37-bus network is +∞, and this is because there is neither Definition 3. The value ε defined in (15) is called the
distributed generation nor shunt capacitors in the network. modification gap.
The C1 margin is above 20 for all IEEE networks, but much
smaller for SCE networks. This is because SCE networks have We demonstrate that the modification gap ε is small for all
big p and q (due to big PVi and Capi ) that make C1 more test networks through Monte-Carlo simulations. Note that ε is
difficult to hold. However, note that the SCE 56-bus network difficult to compute since the objective function in (15) is not
already has a DG penetration of over 130%, and that one can concave and the constraints in (15) are not convex. We choose
still scale up its DG by a factor of 1.30 times before C1 breaks 1000 samples of s, calculate the corresponding (S, v, `, s0 ) by
down. This highlights that C1 is a mild condition. solving the power flow equations (1a)–(1d) (using the forward
backward sweep algorithm [51]) for each s, and compute
ε(s) := kv̂(s) − vk∞ if (s, S, v, `, s0 ) ∈ FOPF . We use the
D. The feasible sets of OPF and OPF-m are similar maximum ε(s) over the 1000 samples as an estimate for ε. The
In this section, we show that OPF-m eliminates some estimated modification gap εset we obtained for different test
feasible points of OPF that are close to the voltage upper networks are listed in Table III. For example, εset = 0.0362 for
bounds for all test networks. To present the result, let FOPF the IEEE 13-bus network, in which case the voltage constraints
denote the feasible set of OPF, let k·k∞ denote the `∞ norm,1 are 0.81 ≤ vi ≤ 1.21 for OPF and 0.81 ≤ vi ≤ 1.1738 for
and let OPF-ε (assuming ε = εset ).
ε := max kv̂(s) − vk∞ s.t. (s, S, v, `, s0 ) ∈ FOPF (15)

VII. C ONCLUSION
denote the maximum deviation of v from its linear approxi-
mation v̂(s) over all OPF feasible points (s, S, v, `, s0 ).
We have proved that SOCP is exact if conditions C1 and
C2 hold. C1 can be checked a priori, and has the physical
w
interpretation that upstream power flows should increase if the
power loss on a line is reduced. C2 requires that optimal power
OPF-‐ε OPF-m OPF
injections lie in a region Svolt where voltage upper bounds do
not bind. We have proposed a modified OPF problem that
includes the additional constraint that power injections lie in
Fig. 9. Feasible sets of OPF-ε, OPF-m, and OPF. The point w is feasible for Svolt , such that the corresponding SOCP relaxation is exact
OPF but not for OPF-m.
under C1. We have also proved that SOCP has at most one
The value ε quantifies the difference between the feasible optimal solution if it it convex and exact. These results unify
sets of OPF and OPF-m. Consider the OPF problem with a and generalize our prior works [46], [47]. Empirical studies
show that C1 holds with large margin and that the feasible
1 The `
∞ norm of a vector x = (x1 , . . . , xn ) ∈ R is defined as kxk∞ :=
n sets of OPF and OPF-m are close, for the IEEE 13, 34, 37,
max{|x1 |, . . . , |xn |}. 123-bus networks and two real-world networks.
10
R EFERENCES [28] R. Jabr, “Radial distribution load flow using conic programming,” IEEE
Transactions on Power Systems, vol. 21, no. 3, pp. 1458–1459, 2006.
[1] J. A. Momoh, Electric Power System Applications of Optimization, ser. [29] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming
Power Engineering, H. L. Willis, Ed. Markel Dekker Inc., 2001. for optimal power flow problems,” International Journal of Electrical
[2] M. Huneault and F. D. Galiana, “A survey of the optimal power flow Power and Energy Systems, vol. 30, no. 6-7, pp. 383–392, 2008.
literature,” IEEE Transactions on Power Systems, vol. 6, no. 2, pp. 762– [30] M. Farivar, C. R. Clarke, S. H. Low, and K. M. Chandy, “Inverter VAR
770, 1991. control for distribution systems with renewables,” in IEEE SmartGrid-
[3] J. A. Momoh, M. E. El-Hawary, and R. Adapa, “A review of selected op- Comm, 2011, pp. 457–462.
timal power flow literature to 1993. Part I: Nonlinear and quadratic pro- [31] M. Farivar and S. H. Low, “Branch flow model: relaxations and
gramming approaches,” IEEE Transactions on Power Systems, vol. 14, convexification (parts I, II),” IEEE Transactions on Power Systems,
no. 1, pp. 96–104, 1999. vol. 28, no. 3, pp. 2554–2572, 2013.
[4] ——, “A review of selected optimal power flow literature to 1993. [32] M. E. Baran and F. F. Wu, “Optimal Capacitor Placement on radial
Part II: Newton, linear programming and interior point methods,” IEEE distribution systems,” IEEE Transactions on Power Delivery, vol. 4,
Transactions on Power Systems, vol. 14, no. 1, pp. 105 – 111, 1999. no. 1, pp. 725–734, 1989.
[5] K. S. Pandya and S. K. Joshi, “A survey of optimal power flow methods,” [33] ——, “Optimal sizing of capacitors placed on a radial distribution
J. of Theoretical and Applied Information Technology, vol. 4, no. 5, pp. system,” IEEE Transactions on Power Delivery, vol. 4, no. 1, pp. 735–
450–458, 2008. 743, 1989.
[6] S. Frank, I. Steponavice, and S. Rebennack, “Optimal power flow: a [34] J. Lavaei and S. H. Low, “Zero duality gap in optimal power flow
bibliographic survey, I: formulations and deterministic methods,” Energy problem,” IEEE Transactions on Power Systems, vol. 27, no. 1, pp.
Systems, vol. 3, no. 3, pp. 221–258, 2012. 92–107, 2012.
[7] ——, “Optimal power flow: a bibliographic survey, II: nondeterministic [35] X. Bai and H. Wei, “A semidefinite programming method with graph
and hybrid methods,” Energy Systems, vol. 3, no. 3, pp. 259–289, 2013. partitioning technique for optimal power flow problems,” International
[8] M. B. Cain, R. P. O’Neill, and A. Castillo, “History of optimal power Journal on Electrical Power and Energy Systems, vol. 33, no. 7, pp.
flow and formulations (OPF Paper 1),” US FERC, Tech. Rep., 2012. 1309–1314, 2011.
[9] R. P. O’Neill, A. Castillo, and M. B. Cain, “The IV formulation and [36] R. A. Jabr, “Exploiting sparsity in sdp relaxations of the opf problem,”
linear approximations of the AC optimal power flow problem (OPF IEEE Transactions on Power Systems, vol. 27, no. 2, pp. 1138–1139,
Paper 2),” US FERC, Tech. Rep., 2012. 2012.
[10] ——, “The computational testing of AC optimal power flow using the [37] D. Molzahn, J. Holzer, B. Lesieutre, and C. DeMarco, “Implementation
current voltage formulations (OPF Paper 3),” US FERC, Tech. Rep., of a large-scale optimal power flow solver based on semidefinite
2012. programming,” IEEE Transactions on Power Systems, vol. 28, no. 4,
[11] A. Castillo and R. P. O’Neill, “Survey of approaches to solving the pp. 3987–3998, 2013.
ACOPF (OPF Paper 4),” US FERC, Tech. Rep., 2013. [38] S. Bose, S. Low, and K. Chandy, “Equivalence of branch flow and bus
[12] ——, “Computational performance of solution techniques applied to the injection models,” in Allerton, 2012.
ACOPF (OPF Paper 5),” US FERC, Tech. Rep., 2013. [39] B. Lesieutre, D. Molzahn, A. Borden, and C. L. DeMarco, “Examining
[13] B. Stott and O. Alsac, “Fast decoupled load flow,” IEEE Transactions on the limits of the application of semidefinite programming to power flow
Power Apparatus and Systems, vol. PAS-93, no. 3, pp. 859–869, 1974. problems,” in Allerton, 2011, pp. 1492–1499.
[14] O. Alsac, J. Bright, M. Prais, and B. Stott, “Further developments in [40] W. A. Bukhsh, A. Grothey, K. McKinnon, and P. Trodden, “Local
lp-based optimal power flow,” IEEE Transactions on Power Systems, solutions of optimal power flow,” IEEE Transactions on Power Systems,
vol. 5, no. 3, pp. 697–711, 1990. vol. 28, no. 4, pp. 4780–4788, 2013.
[15] B. Stott, J. Jardim, and O. Alsac, “Dc power flow revisited,” IEEE [41] D. K. Molzahn, B. C. Lesieutre, and C. L. DeMarco, “Investigation
Transactions on Power Systems, vol. 24, no. 3, pp. 1290–1300, 2009. of non-zero duality gap solutions to a semidefinite relaxation of the
[16] Y. Xiao, Y. Song, and Y. Sun, “Power flow control approach to power optimal power flow problem,” in Hawaii International Conference on
systems with embedded FACTS devices,” IEEE Transactions on Power System Sciences, 2014.
Systems, vol. 17, no. 4, pp. 943–950, 2002. [42] S. Bose, D. F. Gayme, S. Low, and K. M. Chandy, “Optimal power flow
[17] K. Turitsyn, P. Sulc, S. Backhaus, and M. Chertkov, “Local control over tree networks,” in Allerton, 2011, pp. 1342–1348.
of reactive power by distributed photovoltaic generators,” in IEEE [43] B. Zhang and D. Tse, “Geometry of injection regions of power net-
SmartGridComm, 2010, pp. 79–84. works,” IEEE Transactions on Power Systems, vol. 28, no. 2, pp. 788–
[18] C. Coffrin and P. Van Hentenryck, “A linear-programming approxima- 797, 2013.
tion of ac power flows,” arXiv:1206.3614, 2012. [44] S. Sojoudi and J. Lavaei, “Physics of power networks makes hard
optimization problems easy to solve,” in IEEE Power and Energy Society
[19] G. C. Contaxis, C. Delkis, and G. Korres, “Decoupled optimal power
General Meeting, 2012, pp. 1–8.
flow using linear or quadratic programming,” IEEE Transactions on
Power Systems, vol. 1, no. 2, pp. 1–7, 1986. [45] J. Lavaei, D. Tse, and B. Zhang, “Geometry of power flows in tree
networks,” in IEEE PES General Meeting, 2012, pp. 1–8.
[20] W. Min and L. Shengsong, “A trust region interior point algorithm
[46] L. Gan, N. Li, U. Topcu, and S. H. Low, “On the exactness of convex
for optimal power flow problems,” International Journal on Electrical
relaxation for optimal power flow in tree networks,” in IEEE Conference
Power and Energy Systems, vol. 27, no. 4, pp. 293–300, 2005.
on Decision and Control, 2012, pp. 465–471.
[21] A. A. Sousa and G. L. Torres, “Robust optimal power flow solution
[47] ——, “Optimal power flow in distribution networks,” in IEEE Confer-
using trust region and interior methods,” IEEE Transactions on Power
ence on Decision and Control, 2013.
Systems, vol. 26, no. 2, pp. 487–499, 2011.
[48] “IEEE distribution test feeders,” online at available at http://ewh.ieee.
[22] E. C. Baptista, E. A. Belati, and G. R. M. da Costa, “Logarithmic barrier-
org/soc/pes/dsacom/testfeeders/.
augmented lagrangian function to the optimal power flow problem,”
[49] M. Farivar, R. Neal, C. Clarke, and S. Low, “Optimal inverter var
International Journal on Electrical Power and Energy Systems, vol. 27,
control in distribution systems with high pv penetration,” in PES General
no. 7, pp. 528–532, 2005.
Meeting, 2012, pp. 1–7.
[23] G. L. Torres and V. H. Quintana, “An interior-point method for non-
[50] “Southern california edison,” online at http://www.sce.com/.
linear optimal power flow using voltage rectangular coordinates,” IEEE
[51] W. H. Kersting, Distribution System Modeling and Analysis. CRC
Transactions on Power Systems, vol. 13, no. 4, pp. 1211–1218, 1998.
Press, 2006.
[24] R. A. Jabr, “A primal-dual interior-point method to solve the optimal
power flow dispatching problem,” Optimization and Engineering, vol. 4,
no. 4, pp. 309–336, 2003.
[25] F. Capitanescu, M. Glavic, D. Ernst, and L. Wehenkel, “Interior-point
A PPENDIX A
based algorithms for the solution of optimal power flow problems,” P ROOF OF L EMMA 1
Electric Power System Research, vol. 77, no. 5-6, pp. 508–517, 2007.
[26] S. H. Low, “Convex relaxation of optimal power flow, I: formulations Let (s, S, v, `, s0 ) satisfy (1a)–(1c) and ` ≥ 0 component-
and relaxations,” IEEE Transactions on Control of Network Systems, wise. It follows from (1a) that
vol. 1, no. 1, pp. 15–27, 2014. X X
[27] ——, “Convex relaxation of optimal power flow, II: exactness,” IEEE Sij = si + (Shi − zhi `hi ) ≤ si + Shi
Trans. on Control of Network Systems, vol. 1, no. 2, June 2014.
h: h→i h: h→i
11
for (i, j) ∈ E. On the other hand, Ŝij (s) is the solution of Algorithm 1 Construct a feasible point
X Input: an optimal SOCP solution w = (s, S, v, `, s0 ) that
Ŝij = si + Ŝhi , (i, j) ∈ E. violates (5d), a leaf bus l ∈ L with 1 ≤ m ≤ nl such
h: h→i
that (16) holds (assume lk = k for k = 0, . . . , m without
By induction from the leaf lines, one can show that loss of generality).
Sij ≤ Ŝij (s), (i, j) ∈ E. Output: w0 = (s0 , S 0 , v 0 , `0 , s00 ).
1: Initialization.
It follows from (1c) that (Construct s0 , `0 outside Pm , and S 0 outside Pm−1 .)
vi − vj = 2Re(z̄ij Sij ) − |zij |2 ìj keep s:
s0 ← s;
≤ 2Re(z̄ij Sij )
keep ` outside path Pm :
≤ 2Re(z̄ij Ŝij (s))
for (i, j) ∈ E. Sum up the inequalities over Pi to obtain `0ij ← ìj , (i, j) ∈
/ Pm ;
X keep S outside path Pm−1 :
vi − v0 ≤ 2 Re(z̄jk Ŝjk (s)),
0
(j,k)∈Pi Sij ← Sij , (i, j) ∈
/ Pm−1 ;
i.e., vi ≤ v̂i (s), for i ∈ N . 2: Forward sweep.
(Construct `0 on Pm , S 0 on Pm−1 , and s00 .)
A PPENDIX B for k = m, m − 1, . . . , 1 do
P ROOF OF T HEOREM 1 0
|Sk,k−1 |2
The proof idea of Theorem 1 has been illustrated via a `0k,k−1 ← ;
3-bus linear network in Section III-A. Now we present the vk
proof of Theorem 1 for general radial networks. Assume that
0
Sk−1,k−2 ← sk−1 1k6=1
X
f0 is strictly increasing, and that C1 and C2 hold. If SOCP + 0
(Sj,k−1 − zj,k−1 `0j,k−1 );
is not exact, then there exists an optimal SOCP solution j: j→k−1
w = (s, S, v, `, s0 ) that violates (5d). We will construct
end for
another feasible point w0 = (s0 , S 0 , v 0 , `0 , s00 ) of SOCP that
s00 ← −S0,−1
0
;
has a smaller objective value than w. This contradicts the
optimality of w, and therefore SOCP is exact. 3: Backward sweep.
(Construct v 0 .)
Construction of w0
v00 ← v0 , Nvisit = {0};
The construction of w0 is as follows. Since w violates (5d),
there exists a leaf bus l ∈ L with m ∈ {1, . . . , nl } such that while Nvisit 6= N do
w satisfies (5d) on (l1 , l0 ), . . . , (lm−1 , lm−2 ) and violates (5d) find i ∈
/ Nvisit and j ∈ Nvisit such that i → j;
on (lm , lm−1 ). Without loss of generality, assume lk = k for vi0 ← vj0 + 2Re(z̄ij Sij
0
) − |zij |2 `0ij ;
k = 0, . . . , m as in Fig. 10. Then
Nvisit ← Nvisit ∪ {i};
|Sm,m−1 |2
`m,m−1 > , (16a) end while
vm
|Sk,k−1 |2
`k,k−1 = , k = 1, . . . , m − 1. (16b)
vk
values in w. Since s0 = s, the point w0 satisfies (5e).
Furthermore, since `0ij = ìj for (i, j) ∈ / Pm and Sij 0
=
Sij for (i, j) ∈
/ Pm−1 , the point w also satisfies (5a) for
0
/ Pm−1 .
(i, j) ∈
0 1 m 1 m l
S2 In the forward sweep step, `0k,k−1 and Sk−1,k−2 0
are
recursively constructed for k = m, . . . , 1 by alternatively
applying (5d) (with v 0 replaced by v) and (5a)/(5b).
Hence, w0 satisfies (5a) for (i, j) ∈ Pm−1 and (5b).
S3 In the backward sweep step, vi0 is recursively constructed
from bus 0 to leaf buses by applying (5c) consecutively.
Fig. 10. Bus l is a leaf bus, with lk = k for k = 0, . . . , m. Equality (5d) is Hence, the point w0 satisfies (5c).
satisfied on [0, m − 1], but violated on [m − 1, m].
The point w0 satisfies another important property given below.
One can then construct w = (s , S , v , `
0 0
as in Algo-
0 0 0
, s00 )
Lemma 2. The point w0 satisfies `0ij ≥ |Sij
0 2
| /vi for (i, j) ∈ E.
rithm 1. The construction consists of three steps:
S1 In the initialization step, s0 , `0 outside path Pm , and S 0 Proof. When (i, j) ∈
/ Pm , it follows from Step S1 that `0ij =
outside path Pm−1 are initialized as the corresponding 2 0 2
ìj ≥ |Sij | /vi = |Sij | /vi . When (i, j) ∈ Pm , it follows
12
from Step S2 that `0ij = |Sij

0 2
| /vi . This completes the proof For any k ∈ {1, . . . , m − 1}, one has
of Lemma 2.
∆Sk−1,k−2 = ∆Sk,k−1 − zk,k−1 ∆`k,k−1
Lemma 2 implies that if v ≥ v, then w satisfies (6).
0 0 0
|Sk,k−1 |2 − |Sk,k−1 |2
= ∆Sk,k−1 − zk,k−1 ,
vk
Feasibility and Superiority of w0 which is equivalent to

We will show that w0 is feasible for SOCP and has a smaller ∆Pk−1,k−2
= Bk
∆Pk,k−1
objective value than w. This result follows from Claims 1 and ∆Qk−1,k−2 ∆Qk,k−1
2. where

Claim 1. If C1 holds, then 0
Sk,k−1 > Sk,k−1 for k = 2 rk,k−1 h Pk,k−1 +Pk,k−1
0
Qk,k−1 +Q0k,k−1
i
Bk = I − .
0, . . . , m − 1 and v 0 ≥ v. vk xk,k−1 2 2
Claim 1 is proved later in this appendix. Here we illustrate Hence, one has

with Fig. 11 that Sk,k−1
0
> Sk,k−1 for k = 0, . . . , m − 1 ∆Pk−1,k−2
= −Bk Bk+1 · · · Bm−1 um ∆`m,m−1
seems natural to hold. Note that Sm,m−1
0
= Sm,m−1 and ∆Qk−1,k−2
for k = 1, . . . , m. To prove ∆Sk,k−1 > 0 for k =
Sm 1,m 2 Sm,m 1 0, . . . , m − 1, it suffices to show that Bk · · · Bm−1 um > 0
for k = 1, . . . , m.
0 m 1 m C1 implies that As · · · At−1 ut > 0 when 1 ≤ s ≤ t ≤ m.
One also has Bk − Ak = uk bTk where
Fig. 11. Illustration of Sk,k−1
0 > Sk,k−1 for k = 0, . . . , m − 1.  + 
0
2P̂k,k−1 (p) Pk,k−1 +Pk,k−1
v − v
that `0m,m−1 = |Sm,m−1
0
|2 /vm = |Sm,m−1 |2 /vm < `m,m−1 . bk =  2Q̂+ k (q) Q k
0
≥0
k,k−1 k,k−1 +Qk,k−1
Define ∆w = (∆s, ∆S, ∆v, ∆`, ∆s0 ) = w0 − w, then v − vk
k
∆`m,m−1 < 0 and therefore for k = 1, . . . , m − 1. To show that Bk · · · Bm−1 um > 0 for
∆Sm−1,m−2 = ∆Sm,m−1 − zm,m−1 ∆`m,m−1 k = 1, . . . , m, we prove the following lemma.
= − zm,m−1 ∆`m,m−1 > 0. (17) Lemma 3. Given m ≥ 1 and d ≥ 1. Let
A1 , . . . , Am−1 , A1 , . . . , Am−1 ∈ Rd×d and u1 , . . . , um ∈ Rd
Intuitively, after increasing Sm−1,m−2 , upstream reverse satisfy
power flow Sk,k−1 is likely to increase for k = 0, . . . , m − 2. • As · · · At−1 ut > 0 when 1 ≤ s ≤ t ≤ m;
C1 is a condition that ensures Sk,k−1 to increase for k = d
• there exists bk ∈ R that satisfies bk ≥ 0 and Ak − Ak =
0, . . . , m − 1. T
uk bk , for k = 1, . . . , m − 1.
Claim 2. If C2 holds, then v 0 ≤ v. Then
As · · · At−1 ut > 0 (18)
Proof. If C2 holds, then it follows from Lemma 1 that v 0 ≤
v̂(s0 ) = v̂(s) ≤ v. when 1 ≤ s ≤ t ≤ m.
It follows from Claims 1 and 2 that v ≤ v ≤ v 0 ≤ v, and Proof. We prove that (18) holds when 1 ≤ t ≤ s ≤ m by
therefore w0 satisfies (5f). Besides, it follows from Lemma mathematical induction on t − s.
2 that `0ij ≥ |Sij 0 2
| /vi0 for (i, j) ∈ E, i.e., w0
0 2
| /vi ≥ |Sij i) When t − s = 0, one has As · · · At−1 ut = ut =
satisfies (6). Hence, w is feasible for SOCP. Furthermore, w0
0 As · · · At−1 ut > 0.
has a smaller objective value than w because ii) Assume that (18) holds when t − s = 0, 1, . . . , K (0 ≤
X X K ≤ m − 2). When t − s = K + 1, one has
fi (Re(s0i )) − fi (Re(si ))
As · · · Ak Ak+1 · · · At−1 ut
i∈N i∈N
= 0
f0 (−Re(S0,−1 )) − f0 (−Re(S0,−1 )) < 0. = As · · · Ak−1 Ak Ak+1 · · · At−1 ut
+ As · · · Ak−1 (Ak − Ak )Ak+1 · · · At−1 ut
This contradicts with the optimality of w, and therefore SOCP
= As · · · Ak−1 Ak · · · At−1 ut
is exact. To complete the proof, we are left to prove Claim 1.
+ As · · · Ak−1 uk bTk Ak+1 · · · At−1 ut
= As · · · Ak−1 Ak · · · At−1 ut
Proof of Claim 1
+ bTk Ak+1 · · · At−1 ut As · · · Ak−1 uk
Assume C1 holds. First show that ∆Sk,k−1 > 0 for k =
0, . . . , m−1. Recall that S = P +iQ and that ui = (rij xij )T . for k = s, . . . , t−1. Since bk ≥ 0 and Ak+1 · · · At−1 ut >
It follows from (17) that 0, the term bTk Ak+1 · · · At−1 ut ≥ 0. According to induc-
tion hypothesis, As · · · Ak−1 uk > 0. Hence,
∆Pm−1,m−2
= −um ∆`m,m−1 > 0. As · · · Ak Ak+1 · · · At−1 ut ≥ As · · · Ak−1 Ak · · · At−1 ut
∆Qm−1,m−2
13
for k = s, . . . , t − 1. By substituting k = t − 1, . . . , s in Since SOCP is exact, ṽi `˜ij = |S̃ij |2 and v̂i `îj = |Ŝij |2 for
turn, one obtains (i, j) ∈ E. Define w := (w̃ + ŵ)/2. Since SOCP is convex,
w also solves SOCP. Hence, vi ìj = |Sij |2 for (i, j) ∈ E.
As · · · At−1 ut ≥ As · · · At−2 At−1 ut
Substitute vi = (ṽi + v̂i )/2, ìj = (`˜ij + `îj )/2, and Sij =
≥ ··· (S̃ij + Ŝij )/2 to obtain
≥ As · · · At−1 ut > 0, H
Ŝij S̃ij H
+ S̃ij Ŝij = v̂i `˜ij + ṽi `îj
i.e., (18) holds when t − s = K + 1.
for (i, j) ∈ E where the superscript H stands for hermitian
According to i) and ii), (18) holds when t − s = 0, . . . , m − 1.
transpose. The right hand side
This completes the proof of Lemma 3.
|S̃ij |2 |Ŝij |2
Lemma 3 implies that Bs · · · Bt−1 ut > 0 when 1 ≤ s ≤ v̂i `˜ij + ṽi `îj = v̂i + ṽi ≥ 2|S̃ij ||Ŝij |,
t ≤ m. In particular, Bk · · · Bm−1 um > 0 for k = 1, . . . , m, ṽi v̂i
and therefore ∆Sk,k−1 > 0 for k = 0, . . . , m − 1. and the equality is attained if and only if |S̃ij |/ṽi = |Ŝij |/v̂i .
Next show that v 0 ≥ v. Noting that ∆Sij = 0 for (i, j) ∈ / The left hand side
Pm−1 and ∆ìj = 0 for (i, j) ∈ / Pm , it follows from (5c) that H H
Ŝij S̃ij + S̃ij Ŝij ≤ 2|S̃ij ||Ŝij |,
2
∆vi − ∆vj = 2Re(z̄ij ∆Sij ) − |zij | ∆ìj = 0
and the equality is attained if and only if ∠Ŝij = ∠S̃ij . Hence,
for (i, j) ∈
/ Pm . When (i, j) ∈ Pm , one has (i, j) = (k, k − 1) S̃ij /ṽi = Ŝij /v̂i for (i, j) ∈ E.
for some k ∈ {1, . . . , m}, and therefore Introduce v̂0 := ṽ0 := v0 and define ηi := v̂i /ṽi for i ∈ N ,
∆vi − ∆vj = 2Re(z̄k,k−1 ∆Sk,k−1 ) − |zk,k−1 |2 ∆`k,k−1 then η0 = 1 and Ŝij = ηi S̃ij for (i, j) ∈ E. Hence,
≥ Re(z̄k,k−1 ∆Sk,k−1 ) − |zk,k−1 |2 ∆`k,k−1 |Ŝij |2 |ηi S̃ij |2 |S̃ij |2
`îj = = = ηi = ηi `˜ij
= Re(z̄k,k−1 (∆Sk,k−1 − zk,k−1 ∆`k,k−1 )) v̂i ηi ṽi ṽi
= Re(z̄k,k−1 ∆Sk−1,k−2 ) > 0. and therefore
Hence, ∆vi ≥ ∆vj whenever (i, j) ∈ E. Add the inequalities v̂j H
v̂i − 2Re(zij Ŝij ) + |zij |2 `îj
ηj = = H S̃ ) + |z |2 `˜
= ηi
over path Pi to obtain ∆vi ≥ ∆v0 = 0 for i ∈ N + , i.e., ṽj ṽi − 2Re(zij ij ij ij
v 0 ≥ v. This completes the proof of Claim 1.
for (i, j) ∈ E. Since the network (N , E) is connected, ηi =
η0 = 1 for i ∈ N . This implies ŵ = w̃ and completes the
A PPENDIX C
proof of Theorem 3.
P ROOF OF P ROPOSITION 1
Let A and A0 denote the matrices with respect to (p, q) and A PPENDIX E
(p , q 0 ) respectively, i.e., let
0
P ROOF OF T HEOREM 4
2
Theorem 4 follows from Claims 3–7.
A0i = I − ui P̂ij+ (p0 ) Q̂+ 0
ij (q ) , (i, j) ∈ E;
vi
2 Claim 3. Assume that there exist pi and q i such that Si ⊆
Ai = I − ui P̂ij+ (p) Q̂+ ij (q) , (i, j) ∈ E. {s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1
vi
holds if Ŝij (p + iq) ≤ 0 for all (i, j) ∈ E 0 .
When (p, q) ≤ (p0 , q 0 ), one has Alk − A0lk = ulk bTlk where
" # Proof. If Ŝij (p + iq) ≤ 0 for all (i, j) ∈ E 0 , then Alk = I
2 P̂l+k lk−1 (p0 ) − P̂l+k lk−1 (p) for all l ∈ L and all k ∈ {1 . . . , nl − 1}. It follows that
blk = ≥0 Als · · · Alt−1 ult = ult > 0 for all l ∈ L and all s, t such that
v lk Q̂+ 0 +
lk lk−1 (q ) − Q̂lk lk−1 (q)
1 ≤ s ≤ t ≤ nl , i.e., C1 holds.
for any l ∈ L and any k ∈ {1 . . . , nl }.
If A0ls · · · A0lt−1 ult > 0 for any l ∈ L and any s, t such Claim 4. Assume that there exist pi and q i such that Si ⊆
that 1 ≤ s ≤ t ≤ nl , then it follows from Lemma 3 that {s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1
Als · · · Alt−1 ult > 0 for any l ∈ L any s, t such that 1 ≤ s ≤ holds if 1) rij /xij is identical for all (i, j) ∈ E; and 2) v i −
t ≤ nl . This completes the proof of Proposition 1. 2rij P̂ij+ (p) − 2xij Q̂+ 0
ij (q) > 0 for all (i, j) ∈ E .
Proof. Assume the conditions in Claim 4 hold. Fix an arbitrary

A PPENDIX D l ∈ L, and assume lk = k for k = 0, . . . , nl without loss
P ROOF OF T HEOREM 3 of generality. Fix an arbitrary t ∈ {1, . . . , nl }, and define
In this appendix, we prove that SOCP has at most a unique (αs βs )T := As · · · At−1 ut for s = 1, . . . , t. Then it suffices
solution under the conditions in Theorem 3. The proof for to prove that αs > 0 and βs > 0 for s = 1, . . . , t. In particular,
SOCP-m is similar and omitted for brevity. we prove
Assume that fi is convex for i ∈ N , Si is convex for
αs > 0, βs > 0, αs /βs = r10 /x10 (19)
i ∈ N + , SOCP is exact, and SOCP has at least one solution.
˜ s̃0 ) and ŵ = (ŝ, Ŝ, v̂, `,
Let w̃ = (s̃, S̃, ṽ, `, ˆ ŝ0 ) denote two inductively for s = t, t − 1, . . . , 1. Define η := r10 /x10 and
arbitrary SOCP solutions. It suffices to show that w̃ = ŵ. note that rij /xij = η for all (i, j) ∈ E.
14
i) When s = t, one has αs = rt,t−1 , βs = xt,t−1 , and and it follows that

αs /βs = η. Therefore (19) holds.
αk−1 2 r αk
ii) Assume that (19) holds for s = k (2 ≤ k ≤ t), then = I− P Q
βk−1 v x βk
T T k−1
αk βk = c η 1 αk 2 r
= − Qβk .
βk v k−1 x
for some c ∈ {c ∈ R | c > 0}. Abbreviate rk−1,k−2 by Hence,
+
r, xk−1,k−2 by x, P̂k−1,k−2 (p̄) by P , and Q̂+
k−1,k−2 (q̄)
2xQ 1
by Q for convenience. Then βk−1 = βk − βk = v − 2xQ βk > 0.
v k−1 v k−1 k−1
v k−1 − 2rP − 2xQ > 0 Then,
and it follows that 2rQ
αk−1 = αk − βk
v k−1
αk−1 2 r αk
= I− P Q 2rQ
βk−1 v k−1 x βk ≥ η− βk
v k−1
2 η η
= I− x P Q c 2xQ
v k−1 1 1 ≥ η 1− βk
v k−1
η 2 η η
=c − c P Q x = ηβk−1 > 0.
1 v k−1 1 1

αk 2 αk r The second inequality is due to r/x ≤ η. Hence, (20)
= − P Q
holds for s = k − 1.
βk v k−1 βk x

2 αk According to i) and ii), (20) holds for s = t, t − 1, . . . , 1. This
= 1− (rP + xQ) completes the proof of Claim 5.
v k−1 βk

1 αk Claim 6. Assume that there exist pi and q i such that Si ⊆
= v − 2rP − 2xQ >0
v k−1 k−1 βk {s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1
holds if 1) rij /xij ≤ rjk /xjk whenever (i, j), (j, k) ∈ E; and
and αk−1 /βk−1 = αk /βk = η. Hence, (19) holds for 2) Q̂ij (q) ≤ 0, v i − 2rij P̂ij+ (p) > 0 for all (i, j) ∈ E 0 .
s = k − 1.
Proof. The proof of Claim 6 is similar to that of Claim 5 and
According to i) and ii), (19) holds for s = t, t − 1 . . . , 1. This
omitted for brevity.
completes the proof of Claim 4.
Claim 7. Assume that there exist pi and q i such that Si ⊆
Claim 5. Assume that there exist pi and q i such that Si ⊆ {s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1
{s ∈ C | Re(s) ≤ pi , Im(s) ≤ q i } for i ∈ N + . Then C1 holds if
holds if 1) rij /xij ≥ rjk /xjk whenever (i, j), (j, k) ∈ E; and  Y X 
2) P̂ij (p) ≤ 0, v i − 2xij Q̂+ 0 ckl − dkl
ij (q) > 0 for all (i, j) ∈ E .  (k,l)∈Pj
 X (k,l)∈Pj   rij > 0, (i, j) ∈ E (21)
Y
Proof. Assume the conditions in Claim 5 hold. Fix an arbitrary − ekl fkl  xij
l ∈ L, and assume lk = k for k = 0, . . . , nl without loss (k,l)∈Pj (k,l)∈Pj
of generality. Fix an arbitrary t ∈ {1, . . . , nl }, and define
+
(αs βs )T := As · · · At−1 ut for s = 1, . . . , t. Then it suffices where ckl := 1−2rkl P̂kl (p̄)/v k , dkl := 2rkl Q̂+kl (q̄)/v k , ekl :=
to prove that αs > 0 and βs > 0 for s = 1, . . . , t. In particular, 2xkl P̂kl (p̄)/v k , and fkl := 1 − 2xkl Q̂+
+
kl (q̄)/v k.
we prove
The following lemma is used in the proof of Claim 7.
αs > 0, βs > 0, αs /βs ≥ rt,t−1 /xt,t−1 (20) Lemma 4. Given i ≥ 1; c, d, e, f ∈ Ri such that 0 < c ≤ 1,
d ≥ 0, e ≥ 0, and 0 < f ≤ 1 componentwise; and u ∈ R2
inductively for s = t, t − 1, . . . , 1. Define η := rt,t−1 /xt,t−1 that satisfies u > 0. If
and note that rs,s−1 /xs,s−1 ≤ η for s = 1, 2, . . . , t.  i 
Y X i
i) When s = t, one has αs = rt,t−1 , βs = xt,t−1 , and cj − dj 

αs /βs = η. Therefore (20) holds.  j=1 
 j=1  u > 0, (22)
ii) Assume that (20) holds for s = k (2 ≤ k ≤ t), then  X i Yi 
 
− ej fj
αk ≥ ηβk > 0. j=1 j=1
+
Abbreviate rk−1,k−2 by r, xk−1,k−2 by x, P̂k−1,k−2 (p̄) then
cj −dj ci −di
+
by P , and Q̂k−1,k−2 (q̄) by Q for convenience. Then ··· u>0 (23)
−ej fj −ei fi
P = 0, v k − 2xQ > 0 for j = 1, . . . , i.
15
Proof. Lemma 4 can be proved by mathematical induction on Apply the induction hypothesis to obtain
i. 0 0
0 c2 −d02 cK −d0K
i) When i = 1, Lemma 4 is trivial. v2 := ··· u > 0,
−e02 f20 −e0K fK0
ii) Assume that Lemma 4 holds for i = K (K ≥ 1). When 0
i = K + 1, if c1 −d01 c0K −d0K
v10 := ··· u > 0.
 i  −e01 f10
−eK 0
fK0
Y X i
 cj − dj  It follows that
 j=1 
 j=1  u > 0, c1 −d1 cK+1 −dK+1
 X i Yi  ··· u
  −e1 f1 −eK+1 fK+1
− ej fj
j=1 j=1 c1 −d1 c2 −d2 0
= v2
−e1 f1 −e2 f2
one can prove that (23) holds for j = 1, . . . , K + 1 as
follows. c1 c2 + d1 e2 −c1 d2 − d1 f2 0
= v2
First prove that (23) holds for j = 2, . . . , K + 1. The −e1 c2 − f1 e2 f1 f2 + e1 d2

idea is to construct some c0 , d0 , e0 , f 0 ∈ RK and apply c1 c2 −d2 − d1 0
≥ v2
the induction hypothesis. The construction is −e1 − e2 f1 f2
0
c0 = (c2 , c3 , . . . , cK+1 ), c1 −d01 0
= v2
−e01 f10
d0 = (d2 , d3 , . . . , dK+1 ),
= v10 > 0,
e0 = (e2 , e3 , . . . , eK+1 ),
f 0 = (f2 , f3 , . . . , fK+1 ). i.e., (23) holds for j = 1.
To this end, we have proved that (23) holds for j =
Clearly, c0 , d0 , e0 , f 0 satisfies 0 < c0 ≤ 1, d0 ≥ 0, e0 ≥ 0, 1, . . . , K + 1, i.e., Lemma 4 also holds for i = K + 1.
0 < f 0 ≤ 1 componentwise and
 K   K+1  According to i) and ii), Lemma 4 holds for i ≥ 1.
Y XK Y K+1
X

0
cj − 0
dj   cj − dj  Proof of Claim 7. Fix an arbitrary l ∈ L, and assume lk = k
 j=1    for k = 0, . . . , nl without loss of generality. Fix an arbitrary
 j=1  u =  j=2 j=2 u
 X K YK   K+1
X K+1
Y  t ∈ {1, . . . , nl }, then it suffices to prove that As · · · At−1 ut >
   
− e0j fj0 − ej fj 0 for s = 1, . . . , t. Denote rk := rk,k−1 and Sk := Sk,k−1 for
j=1 j=1 j=2 j=2 k = 1, . . . , t for brevity.
 
K+1
Y K+1
X Substitute (i, j) = (k, k − 1) in (21) to obtain
 cj − dj  k−1 ! 
 j=1  Y k−1
X 2rs Q̂+
2rs P̂s+
≥ 
 K+1
j=1
K+1
u
  1− − s

 X Y   s=1

vs vs  rk

− ej fj 
s=1 ! >0
k−1
X + k−1
Y +  xk
j=1 j=1  2xs P̂s 2xs Q̂s 
− 1−
> 0. s=1
vs s=1
vs
(24)
Apply the induction hypothesis to obtain that
0 0 for k = 1, . . . , t. Hence,
cj −d0j cK −d0K k−1
! k−1
··· u>0 Y 2rs P̂s+ X 2rs Q̂+ (q)
−e0j fj0 −e0K 0
fK 1− rk > s
xk ≥ 0
v v
for j = 1, . . . , K, i.e., (23) holds for j = 2, . . . , K + 1. s=1 s s=1 s
Next prove that (23) holds for j = 1. The idea is still to for k = 1, . . . , t. It follows that 1 − 2rk P̂k+ /v k > 0 for k =
construct some c0 , d0 , e0 , f 0 ∈ RK and apply the induction 1, . . . , t−1. Similarly, 1−2xk Q̂+ k /v k > 0 for k = 1, . . . , t−1.
hypothesis. The construction is Then, substitute k = t in (24) and apply Lemma 4 to obtain
c0 = (c1 c2 , c3 , . . . , cK+1 ),  
2rs P̂s+ 2rs Q̂+
s
d0 = (d1 + d2 , d3 , . . . , dK+1 ), 1 − v −
vs 
 s 
e0 = (e1 + e2 , e3 , . . . , eK+1 ),  2xs P̂ + +···
s 2xs Q̂s
− 1−
f 0 = (f1 f2 , f3 , . . . , fK+1 ). vs vs
 +

Clearly, c0 , d0 , e0 , f 0 satisfies 0 < c0 ≤ 1, d0 ≥ 0, e0 ≥ 0, 2rt−1 P̂t−1 (p) 2rt−1 Q̂+ t−1 (q)
1 − − 
0 < f 0 ≤ 1 componentwise and  v t−1 v t−1  rt
>0
 + 
 K K   K+1 K+1   2xt−1 P̂t−1 (p) 2xt−1 Q̂+ (p)  xt
Y X Y X − 1− t−1
0 0
 cj − dj   cj − dj  v t−1 v t−1
 j=1   j=1 
 j=1 u =  j=1  u > 0. for s = 1, . . . , t, i.e., As · · · At−1 ut > 0 for s = 1, . . . , t. This
 X K YK   K+1 K+1 
 0 0   X Y  completes the proof of Claim 7.
− ej fj − ej fj
j=1 j=1 j=1 j=1
16
Lingwen Gan received BE in electronics engineer-

ing from Tsinghua in 2010 and MS in electrical
engineering from Caltech in 2012. He is currently
a PhD candidate in electrical engineering at Caltech
working with Professor Steven H. Low. His research
interests are on distributed load control, optimal
power flow, and renewable energy.
Na Li (M’13) received the B.S. degree in mathemat-

ics from Zhejiang University, Hangzhou, China, in
2007, and the Ph.D. degree in control and dynamical
systems from the California Institute of Technology,
Pasadena, in 2013. She is currently a Postdoctoral
Associate of the Laboratory for Information and
Decision Systems, Massachusetts Institute of Tech-
nology, Cambridge. Her research is on power and
energy networks, systems biology and physiology,
optimization, game theory, decentralized control and
dynamical systems. Dr. Li was a Best Student Paper
Award Finalist in the IEEE Conference on Decision and Control, 2011.
Ufuk Topcu (M’08) received the Ph.D. degree from

the University of California, Berkeley, USA, in 2008.
He is a Research Assistant Professor at the Univer-
sity of Pennsylvania, Philadelphia, PA, USA. His re-
search is on the analysis, design, and verification of
networked, information-based systems with projects
in autonomy, advanced air vehicle architectures, and
energy networks. He was a Postdoctoral Scholar
at the California Institute of Technology, Pasadena,
CA, USA, between 2008 and 2012.
Steven H. Low (F’08) is a Professor of the Depart-

ment of Computing & Mathematical Sciences and
the Department of Electrical Engineering at Caltech.
Before that, he was with AT&T Bell Laboratories,
Murray Hill, NJ, and the University of Melbourne,
Australia. He was a co-recipient of IEEE best pa-
per awards, the R&D 100 Award, and an Okawa
Foundation Research Grant. He is on the Technical
Advisory Board of Southern California Edison and
was a member of the Networking and Information
Technology Technical Advisory Group for the US
President’s Council of Advisors on Science and Technology (PCAST) in
2006. He is a Senior Editor of the IEEE Journal on Selected Areas in
Communications, the IEEE Transactions on Control of Network Systems,
and the IEEE Transactions on Network Science & Engineering, and is on
the editorial boards of NOW Foundations and Trends in Networking, and in
Electric Energy Systems. He received his BS from Cornell and PhD from
Berkeley, both in EE.

Exact Convex Relaxation of Optimal Power Flow in Radial Networks

Uploaded by

Copyright:

Available Formats

Exact Convex Relaxation of Optimal Power Flow in Radial Networks

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exact Convex Relaxation of Optimal Power Flow in Radial Networks

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Exact Convex Relaxation of Optimal Power Flow in

where fi : R → R denotes the generation cost at bus i for

l0 = 0, lnl = l where S0,−1

network segment (s 1, t) for s = 1, . . . , t. Then, dSs−1,s−2 > 0 implies

where θ = cos−1 (0.9), i.e, power injection si,d is con- 35

sidered to be a constant, obtained by assuming a power 30 SOCP 10

SDP time / SOCP time

factor of 0.9 at peak apparent power. If device d is a 25

Re(si,d ) ≥ 0 and |si,d | ≤ s. (13)

ε := max kv̂(s) − vk∞ s.t. (s, S, v, `, s0 ) ∈ FOPF (15)

from Step S2 that `0ij = |Sij

Proof. Assume the conditions in Claim 4 hold. Fix an arbitrary

i) When s = t, one has αs = rt,t−1 , βs = xt,t−1 , and and it follows that

Lingwen Gan received BE in electronics engineer-

Na Li (M’13) received the B.S. degree in mathemat-

Ufuk Topcu (M’08) received the Ph.D. degree from

Steven H. Low (F’08) is a Professor of the Depart-

You might also like