666 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO.

1, JANUARY 2022

An Exact Sequential Linear Programming Algorithm

for the Optimal Power Flow Problem
Sleiman Mhanna , Member, IEEE, and Pierluigi Mancarella , Senior Member, IEEE

Abstract—Despite major advancements in nonlinear program- L Set of all branches ij where i is the “from” bus.
ming (NLP) and convex relaxations, most system operators around Lt Set of all branches ji where j is the “to” bus.
the world still predominantly use some form of linear program-
ming (LP) approximation of the AC power flow equations. This B. Parameters and Input Data
is largely due to LP technology’s superior reliability and com-
putational efficiency, especially in real-time market applications, bsh
i Shunt susceptance (pu) at bus i.
security-constrained applications, and extensions involving integer bch
ij Charging susceptance (pu) in the π-model of line ij.
variables, in addition to its ability to readily generate locational c0,gi Constant coefficient ($/h) term of generator g’s cost
marginal prices (LMP) for market applications. In the aim of lever- function.
aging the advantages of LP while retaining the accuracy of NLP c1,gi Coefficient ($/MWh) of the linear term of generator
interior-point methods (IPMs), this paper proposes a sequential
linear programming (SLP) approach consisting of a sequence of g’s cost function.
carefully constructed supporting hyperplanes and halfspaces. The c2,gi Coefficient ($/MWh2 ) of the quadratic term of gen-
algorithm is numerically demonstrated to converge on 138 test cases erator g’s cost function.
with up the 3375 buses to feasible high-quality solutions (i) without gish Shunt conductance (pu) at bus i.
AC feasibility restoration (i.e., using LP solvers exclusively), (ii) in k Iteration number.
computation times generally within the same order of magnitude as
those from a state-of-the-art NLP solver, and (iii) with robustness pdi /qid Active/Reactive power demand (pu) at bus i.
against the choice of starting point. In particular, the (relative) sij Apparent power rating (pu) of branch ij.
optimality gaps and the mean constraint violations are on average θij /θij Upper/Lower limit of the difference of voltage angles
around 10−3 % and 10−7 , respectively, under a single parameter of buses i and j.
setting for all the 138 test cases. To the best of our knowledge, the ρij Penalty parameter used to guide convergence.
proposed SLP approach is the first to use LP exclusively to reach
feasible and high-quality solutions to the nonconvex AC OPF in a t̃ij Complex tap ratio of a phase-shifting transformer
shift
reliable way, which paves the way for system and market operators (t̃ij = τij eiθi ).
to keep using their LP solvers but now with the ability to accurately ỹij Series admittance (pu) in the π-model of line ij.
capture transmission losses, price reactive power (Q-LMP), and ζ Percentage apparent power loading of a branch.
obtain more accurate LMP.
Index Terms—Sequential linear programming, optimal power C. Operators
flow, supporting hyperplanes, locational marginal pricing, reactive •∗ Conjugate operator.
power pricing. domf Domain of function f .
epif Epigraph of function f .
NOTATION ∇f Gradient of function f .
/{•} Imaginary/Real value operator.
A. Sets •/• Minimum/Maximum magnitude operator.
B Set of all buses in the power network. |•| Magnitude operator/Cardinality of a set.
Bi Set of buses connected to bus i. •T Transpose operator.
G Set of all generators gi such that i is the bus and g is
D. Variables
the generator connected to it.
Gi Set of all generators connected to bus i. pgi /qgi Active/Reactive power (pu) generation of generator
g at bus i.
Manuscript received January 31, 2021; revised May 17, 2021 and June 29,
pij /qij Active/Reactive power (pu) flow on branch ij.
2021; accepted July 11, 2021. Date of publication July 14, 2021; date of current rij Non-negative slack variable used to (i) guide conver-
version December 23, 2021. Paper no. TPWRS-00174-2021. (Corresponding gence and (ii) circumvent infeasbility.
author: Sleiman Mhanna.)
The authors are with the Electrical and Electronic Engineering Depart-
θi Voltage angle (rad) at bus i.
ment, University of Melbourne Melbourne School of Engineering, Mel- vi Voltage magnitude (pu) at bus i.
bourne, Victoria 3010, Australia (e-mail: [email protected];
[email protected]). E. Acronyms
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TPWRS.2021.3097066. AC Alternating current.
Digital Object Identifier 10.1109/TPWRS.2021.3097066 AEMO Australian Energy Market Operator.

0885-8950 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 667

DC Direct current. mixed-integer LP (MILP) solvers is particularly advantageous

IPM Interior-point method. in computationally demanding areas such as real-time markets,
LMP Locational marginal prices/pricing. security-constrained OPF, day-ahead security-constrained unit
LP Linear programming. commitment, and transmission network expansion planning to
MILP Mixed-integer linear programming. name a few [16], e) In complex engineering systems such as
NLP Nonlinear programming. the power network, where controls are constantly changing, the
NZEM New Zealand’s Electricity Market. concept of optimality (local or global) may be elusive, if not
OPF Optimal power flow. completely irrelevant, compared to other more basic require-
QCP Quadratically constrained programming. ments [14]. In other words, the industry traded exactness for
Q-LMP Locational marginal reactive power prices/pricing. reliability, robustness, and computational efficiency.
RTO Regional Transmission Organizations. Because the DC OPF inherently ignores losses, a common
SDP Semidefinite programming. technique in academic literature is to incorporate losses by
SLP Sequential linear programming. approximating the cosine terms in the original polar-form AC
SOCP Second-order cone programming. power flow constraints by their Maclaurin series expansion and
then linearizing [17] or convexifing [18] the resulting noncon-
vex quadratic constraints. On the other hand, system operators
I. INTRODUCTION around the world adopt different approaches to capture active
HE classical optimal power flow (OPF) problem consists power (MW) losses. The Australian Energy Market Operator
T of finding the least-cost dispatch of active and reactive
power from generators to satisfy the loads at all buses in a
(AEMO) and New Zealand’s Electricity Market (NZEM) model
the losses as marginal loss factors in the form of quadratic
way that is governed by physical laws, such as Ohm’s law functions of real power, which are then approximated by piece-
and Kirchhoff’s law, as well as other technical restrictions such wise linear segments [19]. Some system operators and Regional
as thermal limit constraints. The main complexity of the OPF Transmission Organizations (RTO) in the USA use loss factors
problem is directly attributed to the definition of complex power together with loss distribution factors to represent the sensitivity
in which the product of the complex voltage and the conjugate of the losses with respect to power injections at each bus [20].
of the complex current gives rise to the nonlinear and nonconvex System operators in China distribute an estimate of the losses to
alternating current (AC) power flow constraints. Consequently, the loads [20].
finding a globally optimal solution to this nonconvex nonlinear Unfortunately, even after modeling the losses, the DC OPF
programming (NLP) problem is proven to be non-deterministic can result in a poor approximation of the true solution and the
polynomial-time hard (NP-hard) [1], [2]. Since its inception by associated LMP [16]. In fact, it is mathematically proven in [21]
Carpentier in 1962 [3], the OPF problem has garnered rigorous that solutions from the DC OPF are never AC feasible. In an
research attention that gave rise to a rich body of books and attempt to restore AC feasibility, some system operators in the
literature on solution approaches that can be broadly classified USA resort to subsequent corrective measures in the form of
into (i) exact methods, (ii) approximations, and (iii) convex a quasi-optimization process that iterates between a DC OPF
relaxations. with loss factors and an AC power flow [22]. Furthermore,
Exact methods solve the AC OPF directly to find a feasible, because the DC OPF ignores reactive power, system operators
globally or locally optimal solution. The set of exact meth- and RTO is the USA do not price reactive power, which leads to
ods guaranteeing locally optimal solutions includes primal-dual uplift, whereby out-of-market costs are incurred by out-of-merit
interior-point methods (IPMs) with a trust-region and/or filter resources to relieve network constraints and provide reactive
line search [4]–[8], sequential quadratic programming [9], and power services [23].
sequential linear programming (SLP) [10]. The computationally Another research strand revolves around convex relaxations
more demanding global solution methods include single-tree which generally aim to find the convex hull of the nonconvex
methods such as spatial branch and bound [11], and multi-tree feasible region in the hope that the solution of the resulting
methods such as outer approximation [12]. State-of-the-art IPMs convex problem is exact, in which case the global optimum of an
are computationally efficient and have been shown to solve NP-hard problem can be computed in polynomial time. Convex
large-scale instances (more than 50 000 buses) in minutes [13]. relaxations trace back to [24], which numerically showed that
Nonetheless, despite the tractability of today’s IPMs, most a second-order cone programming (SOCP) relaxation of the
system and market operators around the world still favor LP alternative-form OPF [25] can be exact on radial networks, and
approximations such as the direct current (DC) OPF for the to [26], which numerically showed that a semidefinite program-
following reasons: a) IPMs are generally sensitive to the choice ming (SDP) relaxation can be exact on meshed networks. Since
of starting point [14] and to the choice of formulation [15] then, it has been well understood that convex relaxations are only
(especially in cases involving piecewise linear generator cost exact under specific conditions that are unfortunately difficult
functions), which means that the computation time can go to predict in practice and can only be truly verified after the
from seconds to minutes, b) LP solvers are reliable and always problem is solved [27], [28]. This has therefore inspired a myriad
converge to a unique solution, c) Locational Marginal Prices of solution approaches on tightening techniques towards recov-
(LMP), which are the linchpin of electricity markets, are easier ering feasible solutions to the original AC OPF with guaranteed
to obtain using LP, d) The computational superiority of LP and global optimality in the best case, and computing tighter lower

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
668 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

bounds in the worst case (see [29] for a comprehensive review). order of magnitude as a state-of-the-art NLP IPM solver, without
Alas, despite the reliability of convex programming technology, intermediate AC feasibility restoration and with robustness of
convex relaxations still have not found their way into real-world the choice of starting point, under a single parameter setting for
industrial applications as they generally lack the computational all the test cases. In particular, the (relative) optimality gaps and
efficiency of LP-based approximations. the mean constraint violations are on average around 10−3 %
In light of the above, one way to capture both losses and and 10−7 , respectively. Furthermore, the method is not only
reactive power while still using LP technology is by using a evaluated on meshed network instances from MATPOWER [8]
carefully designed SLP approach. SLP, which was originally (as is the case in [10], [33], and [32]), but also on the more
introduced in [30], [31] as the cutting plane method, consists of challenging ones from PGLib-OPF [34], as well as on radial
solving the original NLP problem by solving a series of LP prob- distribution network instances, for a total of 138 test cases.
lems generated by approximating all the nonlinear constraints by Finally, because it uses only LP solvers to directly solve the AC
their first-order Taylor series expansion. Recently, [10] proposed OPF to feasibility, the proposed SLP method can also be viewed
an SLP approach using the IV formulation and numerically as a promising alternative to the widely used quasi-optimization
demonstrated its potential in solving OPF instances with up to process that alternates between DC OPF with loss factors and
3375 buses to high-quality solutions. The method controls the AC feasibility, which can lead to uplift, costlier operation, and
step-size by introducing a penalty function and corresponding inaccurate LMP.
slack variables. However, rigorous numerical analysis in [10] In a nutshell, this paper advances the state of the art in the
revealed the sensitivity of the method to the choice of starting following ways:
point and the authors indeed acknowledge this being one of the r It introduces a novel SLP algorithm for the nonconvex
pitfalls of the approach. Another major limitation of applying AC OPF problem consisting of a sequence of carefully
an SLP scheme to the IV formulation is the eventuality of constructed supporting hyperplanes and halfspaces. The
dealing with nonconvex inequality constraints (the nonconvex exactness, computational efficiency, and robustness against
quadratic voltage bounds) whose linearization cuts the interior the choice of starting point are demonstrated in an extensive
of the feasible region, which increases the risk of infeasible LP numerical analysis on 138 test cases with both radial and
problems and suboptimal solutions. In an attempt to circumvent meshed topologies with up to 3375 buses.
the shortcomings of [10], the work in [32] uses a trust-region r In addition to LMP, the method can readily generate accu-
based SLP approach on the polar-form OPF but now with an rate Q-LMP using LP solvers exclusively, which lays the
AC feasibility restoration phase consisting of a Newton-Raphson foundation for reactive power pricing.
scheme to bring the iterates close to the boundary of the feasible The paper is organized as follows. Section II describes the
region. However, the work in [32] requires problem-specific alternative form of the OPF problem as well as the SOCP relax-
parameter tuning and the numerical evaluation shows that the ation. Section III introduces three novel SLP approaches in Sec-
method converges to poor-quality solutions with a marginal tions III-A, III-B, and III-C for the SOCP relaxation of the OPF,
improvement in computation time. The approach in [33] also the OPF in radial networks, and the OPF in general (meshed)
uses a second-order AC feasibility restoration phase but now networks, respectively. Section IV numerically evaluates the
with a combination of careful mathematical transformations proposed algorithms for general networks on 138 instances with
and first-order approximations of the bilinear voltage terms up to 3375 buses. The paper concludes in Section V.
instead of a step-size control. A related work by the same
authors proposes an improved DC OPF model that captures both
reactive power and voltages (and therefore losses) in the aim II. OPTIMAL POWER FLOW
of bestowing reactive power pricing to the incumbent AC-DC Consider a power network N = (B, L ∪ Lt ), where B is the
quasi-optimization processes adopted by most system operators set of buses and L ∪ Lt is the set of edges such that L is the set of
and RTO in the USA [20]. all branches ij where i is the “from” bus, and Lt is the set of all
Against this background, this work attempts to further narrow branches ji where j is the “to” bus. Examples of power system
the gap between LP approximations and IPMs by proposing an branches include transmission lines and phase-shifting trans-
exact SLP method that leverages the structure of the alternative formers. There are two types of power network topologies, radial
form [25] of the OPF problem by dynamically introducing affine and meshed. Medium and low-voltage distribution networks are
and linear cuts as a sequence of carefully constructed supporting generally radial, whereas high-voltage transmission networks
hyperplanes and an increasing number supporting halfspaces. are almost exclusively meshed. The set of generating units
The success of this particular construction hinges on a single is denoted by the two-dimensional tuple set G = {(g, i)|g ∈
parameter, which is automatically tuned. This is in contrast to the Gi , i ∈ B}, where Gi is the set of all generators connected to bus
trust-region-like approaches which may require introducing a lot i. In the classical AC OPF, the demand at corresponding buses
more slack variables and parameters that require cumbersome is assumed to be static. The complex voltage ṽi (pu) at bus i can
and sometimes problem-specific tuning. Another distinctive fea- be expressed as ṽi = vi eiθ√i
= vi ∠θi = vi cos(θi ) + ivi sin(θi )
ture of the proposed method is that it does not require an AC in polar form, where i = −1. In power system analysis, the
feasibility restoration phase. voltage magnitude is usually nondimensionalized against a base
The method is numerically demonstrated to converge to feasi- voltage level and is expressed in per-unit (pu). For instance, if
ble high-quality solutions in computation times within the same the base voltage is 100 kV, then the pu equivalent of 110 kV

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 669

would be 1.1 (pu). In practice, bus voltage magnitudes vi at the contains the nonconvex atan2 function. In meshed networks,
transmission system level are desired to be close to the base (2k) is necessary to ensure that angle differences sum to zero
voltage to ensure system security and stability. Transmission over every cycle in the network. In radial networks, which are
lines and phase-shifting transformers are represented by their acyclic, constraint (2k) can be dropped from (2) and the angles
π-model equivalents, in which the admittance is defined by can be recovered by solving a system of linear equations of the
ỹij = gij + ibij , where gij and bij are the conductance (pu) form (2k) from the optimal wij r
and wij
i
. Specifically, the OPF
and susceptance (pu), respectively. Additionally, the charging in balanced 3-phase radial distribution networks can be written
susceptance in the π-model of branch ij is denoted by bch
ij (pu). as
By defining 
minimize
p ,q ,w ,
fgi (pgi ) (3a)
gi gi i
wi = |ṽi |2 = vi2 , i ∈ B, (1a) r
wij i
,wij ,pij ,qij gi∈G
 
wij
r
=  ṽi ṽj∗ = vi vj cos (θi − θj ) , ij ∈ L, (1b) subject to (2b)–(2d), (2f)–(2j), (2l). (3b)
 
wij
i
=  ṽi ṽj∗ = vi vj sin (θi − θj ) , ij ∈ L, (1c) The OPF formulation in (2) is equivalent to the polar, rectangu-
lar, and IV OPF formulations (see [35] for a review).
the alternative formulation [25] of the OPF problem can be The alternative formulation of the OPF problem possesses
written as an appealing property. In particular, because voltages are non-

minimize fgi (pgi ) (2a) negative (i.e., wi ≥ 0 ∀i ∈ B), relaxing (2j) into an inequality
pgi ,qgi ,wi ,θi , constraint makes it convex. In fact, this property was the intuition
r i gi∈G
wij ,wij ,pij ,qij
behind the well-known SOCP relaxation in [24], which further
subject to ignores (2k). As such, the SOCP relaxation of the OPF can be
written as
pgi ≤ pgi ≤ pgi , gi ∈ G (2b) 
minimize
p ,q ,w ,
fgi (pgi ) (4a)
q gi ≤ qgi ≤ q gi , gi ∈ G (2c) r
wij
gi
i
,wij
gi i
,pij ,qij gi∈G

v 2i ≤ wi ≤ v 2i , i∈B (2d) subject to (2b)–(2d), (2f)–(2i), (2l) (4b)

θij ≤ θi − θj ≤ θij , ij ∈ L (2e) wi wj ≥ (wij ) + (wij
r 2
) , ij ∈ L
i 2
(4c)
    r   r
pgi − pdi = pij + gish wi , i∈B (2f) tan θij wij ≤ wij
i
≤ tan θij wij . ij ∈ L (4d)
g∈Gi j∈Bi
  Problem (4) can be written in the general form
qgi − qid = qij − bsh
i wi , i∈B (2g)
minimize f0 (x) (5a)
g∈Gi j∈Bi x

pij = gij
c
wi − gij wij
r
+ bij wij
i
, ij ∈ L ∪ Lt (2h) subject to fi (x) ≤ 0, i = 1, . . . , m (5b)

qij = bcij wi − bij wij

r
− gij wij
i
, ij ∈ L ∪ Lt (2i) aTi x ≤ bi , i = 1, . . . , p (5c)

wi wj = (wij
r 2
) + (wij
i 2
) , ij ∈ L (2j) where x ∈ Rn , f0 , f1 , . . . , fm : Rn → R, a1 , . . . , ap ∈ Rn , and
b1 , . . . , bp ∈ R. Function f0 (x) represents the cost functions in
θi − θj = atan2(wij
i
, wij
r
), ij ∈ L (2k) (4a), whereas (5b) represent the convex nonlinear constraints
(4c) and (2l). The linear inequality constraints in (5c) represent
p2ij + qij
2
≤ s2ij , ij ∈ L ∪ Lt (2l)
(2b)–(2d) and (2f)–(2i).3
∗ ∗ In light of the above, this work capitalizes on this specific
where gij c
= {(ỹij −0.5ibch
ij )/|t̃ij | }, bij = {(ỹij −0.5ibij )
2 c ch
2 ∗ ∗ c ∗
/|t̃ij | }, gij = {ỹij /t̃ij }, bij = {ỹij /t̃ij }, gji = {ỹji − property of the alternative formulation to design an SLP ap-
∗ ∗ ∗ proach incorporating a sequence of carefully constructed sup-
0.5ibch ji }, bji = {ỹji − 0.5ibji }, gji = {ỹji /t̃ji }, bji =
c ch
shift porting hyperplanes in conjunction with an increasing number
∗
{ỹji /t̃∗ji }, and t̃ij = τij eiθij is the complex tap ratio of of halfspaces bounded by the supporting hyperplanes of a non-
a phase-shifting transformer.1 The objective function in (2a) convex set of the form
is generally a quadratic cost function of the form fgi (pgi ) =
 
c2,gi (pgi )2 + c1,gi (pgi ) + c0,gi . The relationship between wi , Wij = wi , wj , wij
r
, wij
i
∈ R4 |
wij
r
and wij i
in (1) is captured by (2j), which is a rotated  r 2  
second-order cone (SOC) constraint first introduced in [25].2 wi = (wij ) + (wij ) /wj , wi , wj ≥ 0 , ij ∈ L.
i 2
(6)
The nonconvexities in (2) stem from equality constraint (2j), A similar treatment cannot be applied to (2k) as this constraint
which describes the boundary of a rotated SOC, and (2k) which does not become convex when the equality is relaxed into an

1 For shift = 0.
a transmission line τij = 1 and θij 3 Recall that an affine equality constraint of the form aT x = b can be
i i
i = −w i and w r = w r .
that wji rewritten as {aT T
i x ≤ bi } ∩ {−ai x ≤ −bi }.
2 Note
ij ji ij

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
670 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

⎡ ⎤T
inequality. However, (2k) is quasilinear in θi − θj and wiji
/wijr 4
, r,(k)
2wij /wj
(k)
and therefore a simple first-order approximation is enough to ⎢ ⎥ ⎡ r r,(k)
⎤
 ⎢ ⎥ wij − wij
i,(k) (k)
 ⎢ 2w /w
approximate it to a very high accuracy in an iterative LP-based ij
2  j
2 ⎥
⎥ ⎢ i,(k) ⎥
algorithm, especially when used in tandem with the above con- = fij w(k) + ⎢
⎢ wij
r,(k)
+ wij
i,(k)
⎥ ⎣ wij − wij ⎦ .
i
⎢ ⎥ (k)
struction for the set Wij . This elaborate construction, which is ⎣  2 ⎦ wj − wj
(k)
detailed in the next section, obviates the need for a step-size − wj
control in the classical sense (see [10] and [32]).
Another advantage of the alternative formulation (Problem 2) It is easy to show that the hyperplane {(w, wi )|wi −
of the AC OPF problem is that it enjoys linear branch current fij
a
(w, w(k) ) = 0} supports the set epifij at the boundary point
limit constraints. The derivation of those linear constraints can be (w , fij (w(k) )), since for any
(k)

found in Appendix A. Some system operators deal directly with  T

branch current limits, as opposed to branch MVA limits, since ∇fij w(k)
the current magnitude is directly related to conductor tempera- (w, wi ) ∈ epifij ⇒
−1
ture [36]. Although, those linear branch current limit constraints    
can be directly handled by the proposed SLP approach without w w(k)
extra treatment, the branch MVA limits are used (and then × −   ≤ 0.
wi fij w(k)
linearized) instead for generality and comparison purposes.
Said differently, the set {(w, wi )|wi ≥ fij a
(w, w(k) )} is a
III. SEQUENTIAL LINEAR PROGRAMMING supporting halfspace to the set {(w, wi )|wi ≥ ((wij ) +
r 2

This section introduces three novel SLP approaches in Sec- (wij ) )/wj , wj > 0} at the boundary point (w , fij (w )).
i 2 (k) (k)
tions III-A, III-B, and III-C for the SOCP relaxation of the The second step consists of approximating the convex
OPF (Problem 4), the nonconvex AC OPF in radial networks quadratic thermal limit constraint in (2l). Instead of directly ap-
(Problem 3), and the nonconvex AC OPF in meshed networks proximating (2l) by a first-order Taylor series expansion around
(Problem 2), respectively. Instead of directly introducing the (k) (k)
(pij , qij ), the approximation is applied to the projection of
SLP algorithm for the full nonconvex AC OPF problem in (2), (k) (k)
point (pij , qij ) on the (convex) set defined by (2l). In more
which may be overwhelming at first glance, the below specific
sequence of introductions aims at easing the exposition by detail, let C = {y ∈ R2 |y22 ≤ s2ij }, where y = [pij , qij ]T . The
walking the reader through the evolution of the SLP algorithm projection of y (k) on C, denoted by PC (y (k) ), can now be
from one that solves the convex SOCP relaxation of the OPF obtained by solving the following problem
problem in (4) to one that solves the original nonconvex AC    2 
 
OPF problem in (2). PC y (k) = argmin y − y (k)  |y ∈ C ,
y 2

A. SLP for the SOCP Relaxation of the OPF which admits the analytical solution
Before introducing an SLP approach for Problem 4, constraint ⎡ ⎤T
(4c) is rewritten as   ⎢ (k) (k)
sij pij sij qij ⎥
(wij
r 2
) + (wij
i 2
) PC y (k) = ⎢
⎣ 2   2 ,  2 
⎥
2 ⎦ .
wi ≥ , ij ∈ L (7) (k)
pij
(k)
+ qij
(k)
pij
(k)
+ qij
wj
which is also a convex set as it can be interpreted as the epi- After factorizing and rearranging the terms, the first-order Taylor
graph of a convex quadratic-over-linear function fij : R3 → R series approximation of gij (y) = p2ij + qij
2
− s2ij at the bound-
defined as ary point PC (y (k) ) can be written as
epifij = {(w, wi ) |w ∈ domfij , fij (w) ≤ wi } ,    
gij
a
y, y (k) = −2s2ij + 2y T PC y (k) .
where fij (w) = ((wij ) + (wij
r 2
) )/wj , domfij = R2 × R++ ,
i 2

and w = [wij r
, wij
i
, wj ]T . This form is more conducive to con- The set {y|gija
(y, y (k) ) ≤ 0} is a supporting halfspace to set C
vergence as now the linearization is in the reduced space of 3 at the boundary point PC (y (k) ).
variables instead of 4 (when using (4c) directly). The first-order The last step is to substitute the quadratic terms in fgi (pgi )
Taylor series approximation of fij (w) at a point w(k) can be by corresponding variables and rotated SOC constraints which
written as can then be tightly approximated by a lifted polyhedron as
   T   in [37]. This construction requires far less inequality constraints
fij
a
(w, w(k) ) = fij w(k) + ∇fij w(k) w − w(k) than standard piecewise linear (PWL) approximations for the
same accuracy. Specifically, |G| variables pg and associated
√
4 This r ∈ [v v cos θ , v v ] and w i ∈ [v v sin θ ,
is because wij
linear constraints of the form pg = c2,gi pgi , for all gi ∈ G
i j ij i j ij i j ij
i , w r ) lies in a quasilinear
v i v j sin θ ij ], and therefore the range of atan2(wij
are introduced along with N = |G|/2 + |G|/2 − |G|/2
ij
region. variables αn and constraints of the form αn ≥ p22n−1 + p22n for

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 671

requires no other parameters to ensure convergence. In fact,

Algorithm 1: SLP for the SOCP Relaxation of the OPF.
since the SOCP relaxation of the OPF is a convex problem,
(0) (0)
1: Initialization: Choose starting point (vi , θi ) for all the SLP in Algorithm 1 with an arbitrary starting point that lies
i ∈ B. Set k := 1, , th ∈ [10−10 , 10−3 ], k = 50, in the domain of all the constraint functions is guaranteed to
ζ ∈ (0, 1), and Tij := {∅} for all ij ∈ L ∪ Lt . converge to an optimal solution of Problem 4 in a finite number
of iterations (if Problem 4 is feasible and a solution exists).
2: while min(F ) < − and max(G) > th and k ≤ k Theorem 1: Let the feasible set of Problem 5 be defined
3:
(k) r,(k) i,(k) (k) (k)
Obtain (wi , wij , wij , pij , qij ) by solving as G = {x|fi (x) ≤ 0, i = 1, . . . , m, aTi x ≤ bi , i = 1, . . . , p},
which is nonempty and compact. Problem 5 then consists of

N  finding a vector x such that f0 (x ) = min{f0 (x)|x ∈ G}. If
minimize αn + c1,gi (pgi ) + c0,gi (9a)
pgi ,pg ,αn ,qgi ,wi , x(k) ∈ S (k) is such that
r ,w i ,p ,q
wij ij ij ij
n=1 gi∈G  
f0 (x(k) ) = min f0 (x)|x ∈ S (k) , (10)
subject to (2b)–(2d), (2f)–(2i), (4d), (8) (9b)
a
fij (w, w(κ) ) ≤ wi , ij ∈ L, κ = 0, . . . , k − 1, (9c) where S (0) = {x|aTi x ≤ bi , i = 1, . . . , p} is compact and

a
gij (y, y (κ) ) ≤ 0, ij ∈ L ∪ Lt , κ ∈ Tij , (9d) S (k) = S (k−1) ∩ x|fi (x(k) ) + ∇fi (x(k) )T (x − x(k ) ≤ 0,
4: for ij ∈ L ∪ Lt do i = 1, . . . , m} ,
(k) (k)
5: if ((pij )2 + (qij )2 ) > (ζsij )2 then
6: Tij := Tij ∪ {k}. then the sequence x(k) contains a Cauchy subsequence that
7: end if converges to a point x ∈ G with f0 (x(0) ) ≤ f0 (x(1) ) ≤ · · · ≤
8: end for f0 (x(k−1) ) ≤ f0 (x(k) ) ≤ f0 (x ) ≤ f0 (x) for all x ∈ G.
9: k := k + 1. The proof can be found in Appendix B. Since Algorithm 1
10: end while terminates when min(F ) ≥ − and max(G) ≤ th , x is called
an −effective solution. Finally, although the proof of Theorem 1
depended on it, not all the supporting halfspaces in (9c) are nec-
which the lifted polyhedral construction is denoted as essary for convergence. Previous inactive supporting halfspaces
  can be carefully removed in exchange for a reduction in the
PkR p22n−1 + p22n ≤ αn , n ∈ {1, . . . , N } . (8) overall computation time.
A polyhedral outer approximation of Problem 4 can now Although quadratically constrained programming (QCP)
be iteratively constructed by dynamically adding supporting solvers such as Gurobi [38] and CPLEX [39] can efficiently
halfspaces to the LP approximation of Problem 4 at iteration handle large-scale problems of the form (4), the Barrier method
k − 1. The SLP method for solving the SOCP relaxation of the is widely known to be numerically more stable on large-scale LP
OPF in Problem 4 is described in Algorithm 1, where than on large-scale QCP problems [37], [38]. More importantly,
   in applications requiring binary or integer decision variables,
2  2  
another advantage of using Algorithm 1 is that a crossover
(k) r,(k) i,(k) (k)
F = wi − wij + wij /wj ,
ij∈L
strategy can be used to recover an LP basis that can be used
in the branch-and-cut algorithms at the core of MILP solvers (as
and these use the primal or dual simplex which allow for efficient
  
 2  2 re-optimization at each node of the branch-and-bound tree),
(k) (k)
G= pij + qij − s2ij . which are more efficient than both the branch-and-cut with the
ij∈L∪Lt linearized outer approximation and the branch-and-bound with
At every iteration k, Algorithm 1 solves an LP approximation IPM (usually Barrier) techniques used in mixed-integer QCP
of Problem 4 in the form of Problem 9 (line 3), which features (MIQCP) solvers.
at most |L| + |L ∪ Lt | additional linear constraints compared to Under certain conditions, the SOCP relaxation can be exact
the problem at k − 1. These supporting halfspaces, delineated by in radial networks [27], [28], i.e., constrains (7) are all active at
(9c) and (9d), are dynamically added to Problem 9 at k. Set Tij the optimum, in which case Algorithm 1 can be directly used in
registers all previous iteration numbers that qualify as halfspaces a sequential MILP framework when binary or integer variables
for (2l). The condition on line 5 is satisfied when constraint are present. Moreover, Algorithm 1 can easily be adapted to the
(2l) is (1 − ζ 2 )s2ij away from being violated, in which case the SOC relaxation of the DistFlow model [40] for radial networks.
In the next section, Algorithm 1 is extended to a general setting
supporting halfspace {y|gij a
(y, y (k) ) ≤ 0} is added to Problem
where the SOCP relaxation may not be exact in 3-phase balanced
9 at k. In other words, the condition on line 5 is true when the
radial distribution networks.
apparent power flow on branch ij is greater than ζsij , where ζ is a
parameter that captures a predetermined percentage loading of a
B. Proposed SLP Algorithm for the OPF in Radial Networks
branch. The algorithm terminates when the maximum violations
of constraints (7) and (2l) do not exceed the small tolerances In general, constraints (7) may not all be active at the op-
 and th . Aside from these tolerances and ζ, Algorithm 1 timum; thereby making the relaxation inexact and therefore

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
672 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

is violated at iteration k, in which case the supporting halfspace

Algorithm 2: SLP for the OPF in radial networks.
{(w, wi )|wi − fij a
(w, w(k−1) ) ≥ 0} is added to Problem 11 at
(0) (0)
1: Initialization: Choose starting point (vi , θi ) for all k + 1.
i ∈ B. Set k := 1, , th ∈ [10−10 , 10−3 ], k = 50, In contrast to Algorithm 1, there is no theoretical guaran-
(0)
ζ ∈ (0, 1), ρ > ρij > 0, γ > 0, Iij := {∅} for all tee that Algorithm 2 will converge to a feasible solution of
ij ∈ L, and Tij := {∅} for all ij ∈ L ∪ Lt . Problem 3, let alone an optimal one. In fact, [41] proves that
an SLP algorithm in conjunction with a trust-region and an
2: while max(F ) >  and max(G) > th and k ≤ k
(k) r,(k) i,(k) (k) (k) NLP filter applied to a general NLP with only inequality con-
3: Obtain (wi , wij , wij , pij , qij ) by solving straints is guaranteed to converge to a feasible solution (global
 (k−1) convergence5 ), from an arbitrary starting point that lies in the
minimize (9a) +
p ,p ,α ,q ,w ,
ρij rij (11a)
gi g n gi i non-empty bounded region of the subset of the linear inequality
r
wij i
,wij ,pij ,qij ,rij ij∈L
constraints. However, the proof is predicated on the assumption
subject to (2b)–(2d), (2f)–(2i), (4d), (8) (11b) that a Newton-like (second-order) feasibility restoration phase
  is used if an LP subproblem becomes infeasible, which means
fij
a
w, w(k−1) + rij = wi , ij ∈ L, (11c) that LP solvers cannot be used exclusively. A geometric inter-
  pretation of Algorithm 2 on a small nonconvex two-dimensional
fij
a
w, w(κ) ≤ wi , ij ∈ L, κ ∈ Iij , (11d) problem can be found in Appendix C. The next section extends
  Algorithm 2 to meshed networks.
gij
a
y, y (κ) ≤ 0, ij ∈ L ∪ Lt , κ ∈ Tij , (11e)
C. Proposed SLP Algorithm for the OPF in Meshed Networks
rij ≥ 0, ij ∈ L, (11f) Problem 3, and therefore Algorithm 2, is generally inexact on
4: for ij ∈ L meshed networks as it lacks the treatment of the cycle constraints
5: if Fij ∈/ [−, ] captured by (2k). The first-order Taylor series approximation of a
6: Iij := Iij ∪ {k − 1}. function of the form hij (wij
i
, wij
r
) = atan2(wiji
, wij
r
) at a point
i,(k) r,(k)
7: end if (wij , wij ) can be written as
8: if rij ≥     
(k) (k−1) i,(k) r,(k) i,(k) r,(k)
9: ρij := min(ρ, γρij ). haij wij
i
, wij
r
, wij , wij = atan2 wij , wij
10: end if
⎡ r,(k)
⎤T
11: end for −wij
12: for ij ∈ L ∪ Lt ⎢   2 ⎥  
⎢ r,(k) 2 i,(k) ⎥
13:
(k) (k)
if ((pij )2 + (qij )2 ) > (ζsij )2 ⎢ wij + wij ⎥ wij
i
− w
i,(k)
+⎢
⎢
⎥ ij
⎥ wr − wr,(k) . (12)
14: Tij := Tij ∪ {k}. ⎢
i,(k)
wij ⎥ ij ij
15: end if ⎣ 2  2 ⎦
r,(k) i,(k)
16: end for wij + wij
17: k := k + 1.
18: end while By defining
   
(k) (k) i,(k) r,(k)
H = θi − θj − atan2 wij , wij , (13)
ij∈L

infeasible in practice. In this case, Algorithm 1 is extended the SLP algorithm is described in Algorithm 3, which uses the
to Algorithm 2, which now includes supporting hyperplanes same slack variables rij in the Taylor series approximation of
in conjunction with supporting halfspaces of a set of the form (2k) to circumvent infeasible LP problems that may arise if (12)
(6) to recover an −effective solution to Problem 3. However, is used directly instead of (14c) and (14d).
instead of directly using supporting hyperplanes of the form The weight parameter is automatically tuned exactly as de-
{(w, wi )|wi = fija
(w, w(k) )} (to the nonconvex set of the form scribed in Algorithm 2, and numerical experiments have shown
(6)), which can lead to infeasible LP problems, Algorithm 2 (0)
that setting ρij to a fixed high value, say six orders of magnitude
introduces non-negative slack variables rij whose purpose is larger than max(c2,gi , c1,gi ) would result in fast convergence
twofold. The first is to prevent infeasible LP problems, espe- (0)
to a feasible but poor-quality solution. Instead, setting ρij to
cially during the first few iterations of the algorithm when it is
around one order of magnitude larger than max(c2,gi , c1,gi )
initialized from a poor-quality starting point. The second is to
and adaptively increasing it as described on lines 8 to 10 in
guide convergence by adaptively tuning the weight parameter
(k−1) Algorithm 2 was numerically found to strike a good tradeoff
ρij . More specifically, when the condition on line 8 is true, between computation time and solution quality.
(k−1)
ρij is increased by a factor of γ, as long as it remains
smaller than a predefined upper limit ρ. Set Iij registers all 5 Global convergence of an algorithm entails convergence to a local optimum
previous iteration numbers that qualify as supporting halfspaces from any starting point. Global convergence on general NLP problems should
for (6). The condition on line 5 is satisfied when constraint (6) not be confused with convergence to a global optimum.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 673

TABLE I
Algorithm 3: SLP for the OPF in Meshed Networks. EXACTNESS AND COMPUTATIONAL PERFORMANCE OF ALGORITHM 2
1: Initialization: Same as in Algorithm 2. COMPARED TO KNITRO ON 33-NODE, 69-NODE, AND 119-NODE IEEE
DISTRIBUTION SYSTEMS FOR THREE DIFFERENT STARTING POINTS
2: while max(F ∪ H) >  and max(G) > th and
k≤k
(k) r,(k) i,(k) (k) (k) (k)
3: Obtain (wi , wij , wij , pij , qij , θi ), by
solving
 (k−1)
minimize (9a) + ρij rij (14a)
pgi ,pg ,αn ,qgi ,wi ,θi
r ,w i ,p ,q ,r
wij ij∈L
ij ij ij ij

subject to (11b)–(11f) (14b)

 
i,(k−1) r,(k−1)
haij i
wij r
, wij , wij , wij ≤ rij , ij ∈ L, (14c)
 
i,(k−1) r,(k−1)
haij i
wij r
, wij , wij , wij ≥ −rij , ij ∈ L, (14d)

4: Lines 4 to 16 in Algorithm 2.
5: k := k + 1.
6: end while (0)
10 max((max(c2,gi , c1,gi ))gi∈G ), γ = 5, and ρ = γ 4 ρij .
Finally, let the relative optimality gap be defined as
Aside from the tolerances and ζ, the convergence of Al-   
† † †
Gap = PIPM − PSLP /PIPM × 100, (15)
gorithm 3 hinges on a single parameter ρij , which is then
automatically tuned by introducing parameters γ and ρ. More †
importantly, an AC feasibility restoration phase is not needed where PIPM is the locally optimal solution obtained by KNITRO
†
in Algorithm 3 owing to the iteratively refined polyhedral outer and PSLP is the -effective solution obtained by Algorithm 2 or
approximation of (i) the nonconvex set of the form (6) by the Algorithm 3.
supporting halfspaces in (11d) in conjunction with the support-
ing hyperplanes in (11c), and (ii) the convex set of the form A. Exactness
{(pij , qi )|p2ij + qij
2
≤ s2ij , ij ∈ L ∪ Lt } by the supporting half-
1) Radial Networks: The radial test cases consist of the
spaces in (11e). Extensive numerical evaluation of Algorithms 2
33-node [44], 69-node [45], and 119-node [46] IEEE distri-
and 3 is conducted in the next section to demonstrate their
bution systems with the tie lines removed to ensure a radial
exactness, computational efficiency, and robustness against the
topology.7 The exactness and computational performance of
choice of starting point.
Algorithm 2 for three different starting points is shown in Table I.
More specifically, let FS1 designate a flat start consisting of
IV. NUMERICAL EVALUATION (0) (0)
(vi = 1, θi = 0), FS2 designate a flat start consisting of
In this experimental setup, Julia v1.4.0 [42] is used as a (0) (0)
(vi = v i , θi = 0), and FS3 designate a flat start consisting
programming language along with JuMP v0.21.1 [43] as a (0) (0)
frontend modeling language for the optimization problems. The of (vi = v i , θi = 0), for all i ∈ B. In Table I, columns 2
original NLP problem in (2) is solved using KNITRO v12.2.2 [9] and 3 show the objective function value (2a) at the optimum
with default settings. The LP problems in Algorithm 1 and and the associated computation time of KNITRO, respectively.
Algorithm 3 are solved using Gurobi v9.0.2 [38] with the Barrier The computation times and the associated number of iterations
method (“Method = 2”) and the following parameters: of Algorithm 2 are listed in columns 4 and 5, respectively.
r “Crossover = 0,” which disables crossover,6 Column 6 shows the mean violations of constraints (2j), and the
r “ObjScale = −1,” which uses the reciprocal of the maxi- last column shows the relative optimality gap of Algorithm 2
mum coefficient as scaling for the objective function, defined in (15). It is evident from Table I that Algorithm 2
r “ScaleFlag = −1,” which implements model scaling converges on all the three test cases from all three starting point
for improving the numerical properties of the constraint strategies to high-quality solutions in computation times within
matrix, the same order of magnitude as KNITRO’s. In particular, the
r “Presolve = 2,” which implements an aggressive presolve mean violations of constraints (2j) and the (relative) optimality
strategy in an effort to tighten the model. gaps are on average around 10−8 and 10−4 %, respectively. In
For all the test cases, Algorithms 2 and 3 are fact, Algorithm 2 converges from any starting point such that
(0)
initialized with  = 10−5 , th = 10−3 , ζ = 0.9, ρij =
(0) v i ≤ vi ≤ v i for all i ∈ B. The robustness of Algorithm 2

6 The crossover strategy transforms the interior solution returned by the Barrier
method into a basic solution to be used at the root node of an MILP, and is 7 It is worth noting that keeping the tie lines results in weakly meshed
therefore not needed for the continuous models in this work. topologies which necessitate the use of Algorithm 3 instead of Algorithm 2.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
674 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

against the choice of starting point will be analyzed in more TABLE II

EXACTNESS AND COMPUTATIONAL PERFORMANCE OF ALGORITHM 3
detail in Section IV-B.8 COMPARED TO KNITRO ON ALL THE MATPOWER AND PGLIB-OPF
2) Meshed Networks: Algorithm 3 is implemented on a wide TEST CASES WITH SIZES RANGING FROM 2383 TO 3375 BUSES
range of meshed network test cases, namely, 27 instances FOR STARTING POINT FS1

available with MATPOWER [8], and 108 of the more diffi-

cult instances from PGLib-OPF [34], for a total of 135 test
cases with up the 3375 buses. The PGLib-OPF test cases are
designed specifically to incorporate key network parameters
such as line thermal limit and small angle difference, which
are critical in assessing the robustness of OPF optimization
algorithms. In particular, active power increase (API) cases
are designed to emulate heavily loaded systems (with binding
thermal limit constraints) whereas the small angle difference
(SAD) cases are designed to emulate small angle differences
in some practical systems.9 The exactness and computational
performance of Algorithm 3 are demonstrated on four different
choices of starting points, namely three flat starts and a DC OPF
warm start. The three flat starts are the same ones defined in
the previous section, i.e., FS1, FS2, and FS3. The exactness
and computational performance of Algorithm 3 compared to
KNITRO’s on MATPOWER and PGLib-OPF test cases with
sizes ranging from 2383 to 3375 buses for FS1 and a DC OPF
warm start are shown in Table II and Table III, respectively.
The DC OPF is infeasible on most PGLib-OPF SAD test cases
(also evidenced in [34]), which is why Table III only shows the
results for the corresponding MATPOWER, PGLib-OPF TYP,
and PGLib-OPF API test cases. For lack of space, not all results
for the 135 test cases are shown here. Results for the remaining
test cases under FS1 and the DC OPF warm start as well as those
under FS2 and FS3 can be found in [47]. Numerical evaluation
of Algorithm 1 under FS1 can also be found in [47]. In Tables II
and III, columns 2 and 3 show the objective function value (2a) at
the optimum and the associated computation time of KNITRO,
respectively. The computation times and the associated number
of iterations of Algorithm 3 are listed in columns 4 and 5,
respectively. Column 6 shows the mean violations of constraints
(2j) and (2k), respectively, and the last column shows the relative
optimality gap defined in (15).
As evidenced in [47] and Tables II and III, Algorithm 3
converges on all the 135 test cases from all four starting point
strategies to high-quality solutions in computation times gen-
erally within the same order of magnitude as KNITRO’s. In
particular, the mean violations of constraints (2j) and (2k) and
the (relative) optimality gaps are on average around 10−7 and
10−3 %, respectively. In fact, Algorithm 3 converges for any
(0) (0)
starting point such that v i ≤ vi ≤ v i for all i ∈ B and θi such
(0)
that θij ≤ θij ≤ θij for all ij ∈ L, both of which are trivial
to ensure in practice. The robustness of Algorithm 3 against
the choice of starting point will be further analyzed in the next
section. More interestingly, most of the cases that take more than

8 Please note that Algorithm 3 is for general networks and can thus be applied
to any grid topology, including radial and weakly meshed topologies. However,
since radial networks are acyclic, the hyperplanes (12) approximating the cycle
10 iterations to converge under the generic parameter setting
constraints are redundant and keeping them may add unnecessary computational described above can converge faster under a different value
(0)
burden. It is for this reason that Algorithm 2 is used instead of Algorithm 3 for of ρij . For instance, MATPOWER’s Polish test case 2383wp
radial networks.
(0)
9 The angle difference in the PGLib-OPF SAD are ±10◦ . with ρij = 100 max(c2,gi , c1,gi ) converges in 6 iterations and

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 675

TABLE III
EXACTNESS AND COMPUTATIONAL PERFORMANCE OF ALGORITHM 3
COMPARED TO KNITRO ON MATPOWER, PGLIB-OPF TYP,10
AND PGLIB-OPF API TEST CASES WITH SIZES RANGING FROM 2383 TO 3375
BUSES FOR A DC OPF WARM START

Fig. 1. Average number of iterations to convergence of Algorithm 2 on the

33-node, 69-node, and 119-node IEEE distribution systems. The error bars
show the maximum and minimum number of iterations to convergence for each
test case.

Fig. 2. Average number of iterations to convergence of Algorithm 3 on the 27

meshed cases from MATPOWER [8]. The error bars show the maximum and
minimum number of iterations to convergence for each test case.

B. Robustness
9 seconds and 8 iterations and 12 seconds with the DC Warm This section further analyses the robustness of Algorithms 2
start and FS1, respectively. A similar performance enhancement and 3 beyond the three flat starts identified in the previous sec-
is witnessed on the PGLib-OPF TYP and PGLib-OPF SAD ver- tion. Towards this aim, Algorithms 2 and 3 are initialized from
sions of the 2383wp test case. Another example is PGLib-OPF (0)
100 different starting points with bus voltages vi drawn from a
SAD’s case2316_sdet which converges in 10 iterations and 18 (0)
seconds under FS1. In practice, a power system operator can uniform distribution in the interval (v i , v i ), i.e., vi ∼ U (v i , v i )
(0)
(0)
tailor ρij and γ specifically to the power system at hand for an and θi = 0 for all i ∈ B. For each test case, Algorithms 2
ideal performance. Furthermore, as mentioned earlier, not all the and 3 converge from all the 100 different starting points to
supporting half spaces in (11d) are necessary for convergence, the same objective values obtained from KNITRO [9], and
which means that previous inactive ones can be removed in favor the average number of iterations to convergence is shown in
of faster convergence. Fig. 1 and Fig. 2 for the three radial test cases and the 27
meshed cases from MATPOWER [8], respectively. The error
bars in Fig. 1 and Fig. 2 show the maximum and minimum
number of iterations to convergence for each test case. The
robustness is corroborated by the following two observations: (i)
10 TYP is short for Typical Operating Conditions in [34]. for each test case, the two algorithms converge from all the 100

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
676 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

TABLE IV
ALGORITHM 3’S LMP, Q-LMP, NODE VOLTAGES, AND BRANCH POWER FLOWS
COMPARED TO THOSE FROM KNITRO [9]

Fig. 3. LMP and Q-LMP of KNITRO [9] and Algorithm 3 on the IEEE 57-bus
test system [8].

V. CONCLUSION
This paper presented an SLP approach that exploits the struc-
ture of the alternative-form OPF by carefully constructing iter-
atively refined polyhedral outer approximations and supporting
hyperplanes. Rigorous numerical evaluation on a wide range
of test cases with up to 3375 buses shows that the method
different starting points to the same objective values obtained consistently converges to feasible high-quality solutions with
from KNITRO [9], and (ii) the average number of iterations to robustness against the choice of starting point. While not claim-
convergence for each test case is very close, if not equal, to the ing it is computationally superior to existing NLP solvers, the
one obtained from starting point FS1 where all the bus voltages proposed method represents a competitive LP-based alternative,
(0)
vi are set to 1 pu. A similar observation is witnessed on the especially in settings where only LP solvers can be used or are
remaining test cases from PGLib-OPF [34]. Those remaining practically deployed, e.g., market applications with shadow pric-
results are omitted here for ease of exposition but can be found ing of active and reactive power as well as congestion and losses.
in [47]. Another important benefit of the proposed SLP approach over
NLP-based approaches is its potential straightforward extension
to applications involving discrete decision variables, whereby
powerful state-of-the-art MILP solvers can be used.
C. LMP and Q-LMP
The exactness of the solutions from Algorithm 3 also trans- APPENDIX A
lates to accurate LMP and Q-LMP as shown in Table IV. Table IV BRANCH CURRENT LIMIT CONSTRAINTS
also shows the maximum errors in nodal voltages and branch
In the polar-form OPF, the branch current magnitude is ex-
power flows. Table IV affirms that Algorithm 3 consistently
pressed as
produces accurate LMP and Q-LMP as well as accurate nodal
voltages and branch power flows. This property entails that 
market operators already using LP solvers can not only gen- iij = a1 vi2 + a2 vj2 − 2vi vj (a3 cos θij − a4 sin θij ), (A.1a)
erate more accurate LMP by using Algorithm 3 compared to 
the classical DC OPF, but can now offer a new avenue for iji = a2 vi2 + a1 vj2 − 2vi vj (a3 cos θij + a4 sin θij ), (A.1b)
pricing reactive power through Q-LMP, without changing their ∗ ∗
∗ 2 ∗
LP solvers. Detailed LMP and Q-LMP profiles for the IEEE where a1 = (g̃ij ) , a2 = (g̃ij
c 2
) , a3 = {g̃ijc
}{g̃ij }+
c∗ ∗ c∗ ∗ c∗ ∗
57-bus test system obtained from KNITRO and Algorithm 3 are {g̃ij }{g̃ij }, and a4 = {g̃ij }{g̃ij } − {g̃ij }{g̃ij }.
shown in Fig. 3. The limit constraint on the branch current magnitude can now

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 677

be written as
iij ≤ I ij , ij ∈ L ∪ Lt , (A.2)
where I ij = I ji (pu) is the current rating of branch ij.
Constraints (A.1) are nonconvex and, as shown in [36], lin-
earizing them is a daunting task. However, in the alternative-
form OPF, those constraints can be replaced by linear equiva-
lents. In more detail, by defining ij := i2ij , and after substituting
(1) in (A.1) and squaring both sides of the resulting equation,
the branch current magnitude and associated limit become
ij = a1 wi + a2 wj − 2a3 wij
r
+ 2a4 wij
i
, ij ∈ L, (A.3a)
ji = a2 wi + a1 wj − 2a3 wij
r
− 2a4 wij
i
, ij ∈ L, (A.3b)
2
ij ≤ I ij . ij ∈ L ∪ Lt (A.3c)
The original current magnitude iij can be straightforwardly
obtained from the solution by taking the square root of ij in
a computationally cheap post-processing step.

APPENDIX B
PROOF OF THEOREM 1
If x(k) minimizes f0 (x) on S (k) then it must satisfy the
inequalities
fi (x(κ) ) + ∇fi (x(κ) )T (x(k) − x(κ) ) ≤ 0,
κ = 0, . . . , k − 1, i = 1, . . . , m. (B.1)
Additionally, if x(k) contains a subsequence that converges to
a point in G then every fi (x(k) ), i = 1, . . . , m must contain
a subsequence that converges to a value that is less than or
equal to zero. If there exists a finite constant K such that
max((∇fi (x(k) ))i=1,...,m ) ≤ K for all x ∈ S (0) , and an r >
0 independent of k, such that
   
  Fig. 4. Geometric interpretation of the SLP algorithm at k = 3 and k = 7.
r ≤ fi x(κ) ≤ ∇fi (x(κ) )T (x(κ) − x(k) ) ≤ K x(κ) − x(k)  ,

κ = 0, . . . , k − 1, i = 1, . . . , m,
whose optimal solution is x∗ = [x∗1 , x∗2 ]T = [0, 1]T . The novel
then one can find a subsequence of iteration numbers where
  SLP algorithm in this work solves the problem
 (u) 
x − x(v)  ≤ r/K, ∀u < v, minimize f (x) + ρr (C.2a)
x,r
so that x (k)
does not contain a Cauchy sequence. However,    T  
because S (0) is compact, no such r exists, and x(k) therefore subject to g x(k) + ∇g x(k) x − x(k) = −r, (C.2b)
contains a subsequence that converges to a point x† ∈ S (0) .    T  
Finally, it follows from (B.1) that the corresponding subsequence g x(κ) + ∇g x(κ) x − x(κ) ≤ 0,
of every fi (x(k) ), i = 1, . . . , m converges to a value that is less
than or equal to zero, which entails that x† ∈ G and x = x† .  κ = 0, . . . , k − 1 (C.2c)
−2 ≤ x1 , x2 ≤ 2, (C.2d)
APPENDIX C
SLP EXAMPLE r ≥ 0, (C.2e)
Consider the nonconvex problem at each iteration k until convergence, measured by |g(x(k) )| ≤ .
minimize f (x) = x1 − x2 (C.1a) A geometric interpretation of the SLP algorithm with ρ = 10 and
x from starting point x(0) = [−2, 2]T is shown in Fig. 4(a) and
subject to g(x) = 3x21 − 2x1 x2 + x22 − 1 = 0, (C.1b) Fig. 4(b) at k = 3 and k = 7, respectively. In Fig. 4, supporting
hyperplanes (C.2b) are shown in red and supporting halfspaces
−2 ≤ x1 , x2 ≤ 2, (C.1c) (C.2c) are shown in green. The contour lines of f (x) − f (x(k) )

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
678 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 37, NO. 1, JANUARY 2022

Fig. 5. Average number of iterations to convergence of the proposed SLP Fig. 7. Average number of iterations to convergence of the proposed SLP
algorithm initialized from 1000 different starting points for ρ = 1, 10, 100. The algorithm with the projection of point x(k) on the set defined by g(x) = 0.
error bars show the maximum and minimum number of iterations to convergence. The algorithm is again initialized from 1000 different starting points for
ρ = 1, 10, 100. The error bars show the maximum and minimum number of
iterations to convergence.

More interestingly, the SLP algorithm applied to Problem C.1

converges much faster if the projection of point x(k) on the set
defined by g(x) = 0 is used instead of the first-order Taylor
series approximation of g(x) directly at x(k) (See Section III-A
for more details on the projection). Fig. 6 shows that the modified
SLP algorithm is already close to the optimal solution after
only 3 iterations. In fact, as shown in Fig. 7, the proposed SLP
algorithm with the projection of point x(k) on the set defined
by g(x) = 0 is on average 40% faster compared to directly
applying the first-order Taylor series approximation of g(x) at
x(k) (9 iterations on average for the modified SLP algorithm
compared to 15 for the one with the first-order Taylor series
approximation of g(x) directly at x(k) ). Finally, it is also evident
Fig. 6. Geometric interpretation of the SLP algorithm with the projection of form Fig. 4 and Fig. 6 that not all the supporting halfspaces
point x(k) on the set defined by g(x) = 0 at k = 3. are necessary for convergence. Some of the older inactive half-
spaces can be removed for an improvement in computational
and the nonconvex constraint g(x) = 0 are shown in black and speed.
blue, respectively. The simultaneous inclusion of the supporting
hyperplanes (C.2b) and the supporting halfspaces (C.2c) is REFERENCES
essential for convergence. Without the hyperplanes in (C.2b) [1] K. Lehmann, A. Grastien, and P. Van Hentenryck, “AC-Feasibility on
the algorithm converges to the solution of a convex relaxation tree networks is NP-Hard,” IEEE Trans. Power Syst., vol. 31, no. 1,
pp. 798–801, Jan 2016.
of Problem C.1, i.e., one with g(x) ≤ 0 instead of g(x) = 0. [2] D. Bienstock and A. Verma, “Strong NP-hardness of AC power flows
Moreover, without the halfspaces in (C.2c) the algorithm fails feasibility,” Operations Res. Lett., vol. 47, no. 6, pp. 494–501, Dec. 2019.
to converge altogether. Furthermore, without the positive slack [3] J. Carpentier, “Contribution to the Economic Dispatch Problem,” Bull.
Soc. Franc. Elect., vol. 8, no. 3, pp. 431–447, 1962.
variable r and the second term in (C.2a), some starting points, [4] J. A. Momoh, M. E. El-Hawary, and R. Adapa, “A review of selected
such as x(0) = [0, 0]T in this case, can lead to an infeasible optimal power flow literature to 1993 Part II: Newton, linear programming
LP Problem C.2. The robustness against the choice of start- and interior point methods,” IEEE Trans. Power Syst., vol. 14, no. 1,
pp. 105–111, Feb. 1999.
ing point bestowed by ρr and (C.2e) is shown in Fig. 5 for [5] J. A. Momoh, M. E. El-Hawary, and R. Adapa, “A review of selected
 = 10−4 , three choices of ρ, and 1000 starting points drawn optimal power flow literature to 1993 Part I: Nonlinear and quadratic
from a uniform distribution in the interval (−10, 10) for both programming approaches,” IEEE Trans. Power Syst., vol. 14, no. 1,
(0) (0) (0) (0) pp. 96–104, Feb. 1999.
x1 and x2 , i.e., x1 , x2 ∼ U (−10, 10). The error plot in [6] S. Frank, I. Steponavice, and S. Rebennack, “Optimal power flow: A
Fig. 5 demonstrates a small variation in the number of iterations bibliographic survey I formulations and deterministic methods,” Energy
Syst., vol. 3, no. 3, pp. 221–258, 2012.
to convergence for ρ = 1, 10, 100 over all the 1000 starting [7] F. Capitanescu, “Critical review of recent advances and further devel-
points, thus corroborating the robustness of the proposed SLP opments needed in AC optimal power flow,” Elect. Power Syst. Resric,
algorithm on Problem C.1. vol. 136, pp. 57–68, 2016.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.
MHANNA AND MANCARELLA: EXACT SEQUENTIAL LINEAR PROGRAMMING ALGORITHM FOR THE OPTIMAL POWER FLOW PROBLEM 679

[8] R. Zimmerman, C. Murillo-Sánchez, and R. Thomas, “MATPOWER: AC optimal power flow problems,” Electric Power Syst. Res., vol. 165,
Steady-State operations, planning, and analysis tools for power systems pp. 134–143, 2018.
research and education,” IEEE Trans. Power Syst., vol. 26, no. 1, pp. 12–19, [33] Z. Yang, H. Zhong, Q. Xia, A. Bose, and C. Kang, “Optimal power flow
Feb. 2011. based on successive linear approximation of power flow equations,” IET
[9] R. H. Byrd, J. Nocedal, and R. A. Waltz, “Knitro: An integrated package for Gener., Transmiss. Distrib., vol. 10, no. 14, pp. 3654–3662, 2016.
nonlinear optimization,” in Large Scale Nonlinear Optimization. Berlin, [34] S. Babaeinejadsarookolaee et al., “The power grid library for benchmark-
Germany: Springer Verlag, 2006, pp. 35–59. ing AC optimal power flow algorithms,” 2019, arXiv:1908.02788.
[10] A. Castillo, P. Lipka, J.-P. Watson, S. S. Oren, and R. P. O’Neill, “A [35] R. P. O’Neill, A. Castillo, and M. B. Cain, “The IV formulation and linear
successive linear programming approach to solving the IV-ACOPF,” IEEE approximations of the AC optimal power flow problem,” Federal Energy
Trans. Power Syst., vol. 31, no. 4, pp. 2752–2763, Jul. 2016. Regulatory Commission Staff Tech. Paper, pp. 1–18, Dec. 2012.
[11] A. Gopalakrishnan, A. U. Raghunathan, D. Nikovski, and L. T. Biegler, [36] D. Shchetinin, T. T. De Rubira, and G. Hug, “On the construction of linear
“Global optimization of optimal power flow using a branch & bound al- approximations of line flow constraints for AC optimal power flow,” IEEE
gorithm,” in Proc. 50th Annu. Allerton Conf. Commun., Control, Comput., Trans. Power Syst., vol. 34, no. 2, pp. 1182–1192, Mar. 2019.
2012, pp. 609–616. [37] S. Mhanna, G. Verbič, and A. C. Chapman, “Tight LP approximations
[12] J. Liu, M. Bynum, A. Castillo, J. P. Watson, and C. D. Laird, “A mul- for the optimal power flow problem,” in Proc. 19th Power Syst. Comput.
titree approach for global solution of ACOPF problems using piece- Conf., 2016, pp. 1–7.
wise outer approximations,” Comput. Chem. Eng., vol. 114, pp. 145–157, [38] “Gurobi optimizer reference manual, version 9.1,” 2019. [Online]. Avail-
2018. able: www.gurobi.com
[13] J. Kardos, D. Kourounis, O. Schenk, and R. Zimmerman, “Complete [39] IBM, “IBM ILOG CPLEX ver. 12.10 User’s Manual. IBM ILOG CPLEX
results for a numerical evaluation of interior point solvers for large-scale Division, Incline Village, NV, 2019,” https://www.ibm.com/products.
optimal power flow problems,” 2018, arXiv:1807.03964. [40] M. E. Baran and F. F. Wu, “Network reconfiguration in distribution systems
[14] B. Stott and O. Alsaç, “Optimal power flow-basic requirements for real-life for loss reduction and load balancing,” IEEE Trans. Power Del., vol. 9,
problems and their solutions (White paper),” in Proc. SEPOPE XII Symp., no. 4, pp. 1401–1407, Apr. 1989.
Rio de Janeiro, Brazil, 2012, pp. 1866–1876. [41] R. Fletcher, S. Leyffer, and P. L. Toint, “On the global convergence of an
[15] C. Coffrin, B. Knueven, J. Holzer, and M. Vuffray, “The impacts of convex SLP-Filter algorithm,” Numerical Analysis Report NA/183, Univ. Dundee,
piecewise linear cost formulations on AC optimal power flow,” Electric Aug. 1998.
Power Systems Research, vol. 199, p. 107191, 2021. [42] J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, “Julia: A fresh
[16] B. Stott, J. Jardim, and O. Alsaç, “DC power flow revisited,” IEEE Trans. approach to numerical computing,” SIAM Rev., vol. 59, no. 1, pp. 65–98,
Power Syst., vol. 24, no. 3, pp. 1290–1300, Aug. 2009. 2017.
[17] T. N. Dos Santos and A. L. Diniz, “A dynamic piecewise linear model [43] I. Dunning, J. Huchette, and M. Lubin, “Jump: A modeling language for
for DC transmission losses in optimal scheduling problems,” IEEE Trans. mathematical optimization,” SIAM Rev., vol. 59, no. 2, pp. 295–320, 2017.
Power Syst., vol. 26, no. 2, pp. 508–519, May 2011. [44] M. E. Baran and F. F. Wu, “Optimal capacitor placement on radial dis-
[18] H. Zhong, Q. Xia, Y. Wang, and C. Kang, “Dynamic economic dispatch tribution systems,” IEEE Trans. Power Del., vol. 4, no. 1, pp. 725–734,
considering transmission losses using quadratically constrained quadratic Jan. 1989.
program method,” IEEE Trans. Power Syst., vol. 28, no. 3, pp. 2232–2241, [45] J. S. Savier and D. Debapriya, “Impact of network reconfiguration on
Aug. 2013. distribution systems performance,” IEEE Power Del., vol. 22, no. 4,
[19] B. F. Hobbs, E. Bartholomew, Y. Chen, G. Drayton, and W. Lise, pp. 2473–2480, 2007.
“Improved transmission representations in oligopolistic market models: [46] D. Zhang, Z. Fu, and L. Zhang, “An improved TS algorithm for loss-
Quadratic losses, phase shifters, and DC lines,” IEEE Trans. Power Syst., minimum reconfiguration in large-scale distribution systems,” Electric
vol. 23, no. 3, pp. 1018–1029, Aug. 2008. Power Syst. Res., vol. 77, no. 5/6, pp. 685–694, 2007.
[20] Z. Yang, H. Zhong, A. Bose, T. Zheng, Q. Xia, and C. Kang, “A linearized [47] “SLP results,” Accessed: Aug. 2021. [Online]. Available: https:
OPF model with reactive power and voltage magnitude: A pathway to //drive.google.com/drive/folders/1qU81wlSvGu9QMknOuLtFV_
improve the MW-only DC OPF,” IEEE Trans. Power Syst., vol. 33, no. 2, gUfMUkaWPb
pp. 1734–1745, Mar. 2018.
[21] K. Baker, “Solutions of DC OPF are never AC feasible,” Proc. 12th ACM
Int. Conf. Future Energy Systems, ser. e-Energy ’21. New York, NY, USA: Sleiman Mhanna (Member, IEEE) received the
Association for Computing Machinery, 2021, p. 264268. B.Eng. degree (with high distinction) in electrical
[22] R. O’Neill, T. Dautel, and E. Krall, “Recent ISO software enhancements engineering from Notre Dame University, Zouk Mos-
and future software and modeling plans,” Federal Energy Regulatory beh, Lebanon, in 2010, the M.Eng. degree in electrical
Commission Staff Tech. Paper, Nov. 2011. engineering from the American University of Beirut,
[23] W. Sauer, “Uplift in RTO and ISO markets,” Federal Energy Regulatory Beirut, Lebanon, in 2012, and the Ph.D. degree from
Commission, Washington, DC, USA, Tech. Rep., 2014. the School of Electrical and Information Engineering,
[24] R. A. Jabr, “Radial distribution load flow using conic programming,” IEEE Centre for Future Energy Networks, The University
Trans. Power Syst., vol. 21, no. 3, pp. 1458–1459, Aug. 2006. of Sydney, Sydney, NSW, Australia, in 2016. He is
[25] A. G. Expósito and E. R. Ramos, “Reliable load flow technique for currently a Research Fellow with the Department of
radial distribution networks,” IEEE Trans. Power Syst., vol. 14, no. 3, Electrical and Electronic Engineering, The University
pp. 1063–1069, Aug. 1999. of Melbourne, Melbourne, VIC, Australia. His research interests include compu-
[26] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming for tational methods for integrated multi-energy systems, decomposition methods,
optimal power flow problems,” Int. J. Elect. Power Energy Syst., vol. 30, and demand response.
no. 6/7, pp. 383–392, 2008.
[27] S. H. Low, “Convex relaxation of optimal power flow-Part II: Exactness,”
IEEE Control Netw. Syst., vol. 1, no. 2, pp. 177–189, Jun. 2014.
[28] S. Huang, Q. Wu, J. Wang, and H. Zhao, “A sufficient condition on convex Pierluigi Mancarella (Senior Member, IEEE) is a
relaxation of AC optimal power flow in distribution networks,” IEEE Trans. Chair Professor of electrical power systems with the
Power Syst., vol. 32, no. 2, pp. 1359–1368, Mar. 2017. University of Melbourne, Melbourne, VIC, Australia,
[29] F. Zohrizadeh, C. Josz, M. Jin, R. Madani, J. Lavaei, and S. Sojoudi, “A and a Professor of smart energy systems with The
survey on conic relaxations of optimal power flow problem,” Eur. J. Oper. University of Manchester, Manchester, U.K. His re-
Res., vol. 287, no. 2, pp. 391–409, 2019. search interests include techno-economic modeling
[30] E. W. Cheney and A. A. Goldstein, “Newton’s method for convex program- of integrated multienergy systems, security, reliability
ming and Tchebycheff approximation,” Numerische Mathematik, vol. 1, and resilience of future networks, and energy infras-
no. 1, pp. 253–268, 1959. tructure planning under uncertainty. He is the Editor
[31] J. E. Kelley, Jr., “The cutting-plane method for solving convex programs,” of the IEEE TRANSACTIONS ON POWER SYSTEMS,
J. Soc. Ind. Appl. Math., vol. 8, no. 4, pp. 703–712, 1960. IEEE TRANSACTIONS ON SMART GRID, and IEEE
[32] L. P. Sampath, B. V. Patil, H. B. Gooi, J. M. Maciejowski, and K. V. SYSTEMS JOURNAL, and an IEEE Power and Energy Society Distinguished
Ling, “A trust-region based sequential linear programming approach for Lecturer.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2022 at 14:08:02 UTC from IEEE Xplore. Restrictions apply.