SDP_PF: Implementation of Applications of a

Semidefinite Programming Relaxation of the Power

Flow Equations
Version 1.0
Daniel K. Molzahn∗
December 17, 2014

Abstract
This report documents the implementation of applications for a semidefinite pro-
gramming relaxation of the power flow equations. The code integrates with the
software package Matpower. When a rank condition is satisfied, the semidefinite
relaxation yields the global solution to the optimal power flow (OPF) problem in
polynomial time. Through the use of matrix decomposition techniques, practically
sized systems are computationally tractable. Additional functionality includes imple-
mentation of a sufficient condition for global optimality of an OPF solution, sufficient
conditions for power flow insolvability including the ability to respect generator re-
active power limits, and calculation of voltage stability margins.

∗
University of Michigan Department of Electrical Engineering and Computer Science.
molzahn@umich.edu

1
Publications
The function sdpopf_solver implements SDPOPF, an optimal power flow solver based on
semidefinite programming. A detailed description of the theory incorporated in this code
is available in reference [1] and the references therein. Please cite reference [1] in work that
uses SDPOPF.
D.K. Molzahn, J.T. Holzer, B.C. Lesieutre, and C.L. DeMarco, “Implemen-
tation of a Large-Scale Optimal Power Flow Solver Based on Semidefinite
[1]
Programming,” IEEE Transactions on Power Systems, vol. 28, no. 4, pp. 3987-
3998, November 2013.
Also included with this code is the function testGlobalOpt, which implements a suffi-
cient condition for global optimality of a given OPF solution. Please cite reference [2] in
work that uses testGlobalOpt.
D.K. Molzahn, B.C. Lesieutre, and C.L. DeMarco, “A Sufficient Condition for
Global Optimality of Solutions to the Optimal Power Flow Problem,” IEEE
[2]
Transactions on Power Systems (Letters), vol. 29, no. 2, pp. 978-979, March
2014.
The function insolvablepf implements a sufficient condition for the insolvability of a
specified set of power flow equations. Please cite reference [3] in work that uses insolvablepf.
D.K. Molzahn, B.C. Lesieutre, and C.L. DeMarco, “A Sufficient Condition
for Power Flow Insolvability with Applications to Voltage Stability Margins,”
[3]
IEEE Transactions on Power Systems, vol. 28, no. 3, pp. 2592-2601, Au-
gust 2013.
The function insolvablepf does not incorporate generator reactive power limits. A
mixed-integer semidefinite programming formulation implemented by the function
insolvablepf_limitQ incorporates generator reactive power limits. Please cite reference [4]
in work that uses insolvablepf_limitQ.
An alternative approach to proving power flow insolvability uses the Positivstellensatz
theorem from the field of real algebraic geometry and sum of squares programming to
generate infeasibility certificates. The function insolvablepfsos implements the proposed
method for the power flow equations without considering reactive power limited generators.
The function insolvablepfsos_limitQ implements the proposed method for the power flow
equations with consideration of upper limits on generators’ reactive power outputs. Please
cite reference [4] in work that uses insolvablepfsos or insolvablepfsos_limitQ.
D.K. Molzahn, V. Dawar, B.C. Lesieutre, and C.L. DeMarco, “Sufficient Con-
ditions for Power Flow Insolvability Considering Reactive Power Limited Gen-
[4] erators with Applications to Voltage Stability Margins,”Bulk Power System
Dynamics and Control - IX. Optimization, Security and Control of the Emerg-
ing Power Grid, 2013 IREP Symposium, 25-30 August 2013.

2
1 Introduction
The optimal power flow (OPF) problem seeks decision variable values that yield an opti-
mal operating point for an electric power system in terms of a specified objective function,
subject to both network equality constraints (i.e., the power flow equations) and engineer-
ing inequality constraints (e.g., limits on voltage magnitudes, active and reactive power
generations, and flows on transmission lines and transformers).
The OPF problem is nonconvex due to the nonlinear power flow equations [5], may have
local solutions [6], and is, in general, NP-hard [7]. Nonconvexity of the OPF problem has
made solution techniques an ongoing research topic. Many OPF solution techniques have
been proposed, including successive quadratic programs, Lagrangian relaxation, genetic
algorithms, particle swarm optimization, and interior point methods (for relevant survey
papers, see [8–12]).
Recently, significant research attention has focused on a semidefinite programming re-
laxation of the optimal power flow problem [7, 13]. Semidefinite programming is a type
of convex optimization that minimizes a linear objective function over the intersection of
a cone of positive semidefinite matrices (i.e., symmetric matrices constrained to have all
non-negative eigenvalues) and an affine plane. Semidefinite programming has been suc-
cessful in solving or approximating the solutions of many practical problems that are oth-
erwise computationally challenging, including NP-hard optimization problems. Overviews
of semidefinite programming theory and practice are available in references [14–16].
By appropriate selection of Ai matrices, the power flow equations in rectangular voltage
coordinates can be written in the form xT Ai x = ci , where x is a vector of orthogonal voltage
components.
h iT
x = Vd1 Vd2 . . . Vdn Vq1 Vq2 . . . Vqn (1)
By defining the matrix W = xxT , the power flow equations can then be rewritten as the
combination of the linear equations trace (Ai W) = ci and the condition rank (W) = 1.
The nonconvexity in this formulation of the power flow equations is entirely due to the
rank condition. The semidefinite relaxation of the power flow equations does not enforce
the rank condition; rather, the constraint W  0, where the symbol  indicates that the
corresponding matrix is positive semidefinite, is used to define a convex feasible space. If
the solution to the resulting semidefinite relaxation satisfies the rank condition (i.e., the
semidefinite relaxation is “exact”), a voltage profile satisfying the power flow equations can
be obtained from an eigenvector associated with a non-zero eigenvalue of the solution’s W
matrix [7]. The voltage profile obtained from a semidefinite programming relaxation of the
optimal power flow is guaranteed to be the globally optimal solution if the rank condition
is satisfied. No prior OPF solution method offers a guarantee of finding the global solution;
semidefinite programming approaches thus have a substantial advantage over traditional
solution techniques.
Existing literature [7] claims that the rank condition is satisfied for most practical
power system models, including the IEEE test systems [17]. However, the rank condition

3
is not always satisfied, which means that semidefinite relaxations do not give physically
meaningful solutions for all realistic power system models [18, 19]. Recent research has
investigated the conditions under which the rank condition is satisfied; to date, sufficient
conditions for rank condition satisfaction include requirements on power injection and
voltage magnitude limits and either radial networks (typical of distribution system models)
or or unrealistically dense placement of controllable phase shifting transformers [20–23].
In addition to theoretical concerns regarding satisfaction of the rank condition, practical
computational issues are also of interest. Semidefinite programming relaxations of the OPF
problem constrain a 2n × 2n symmetric matrix to be positive semidefinite, where n is the
number buses in the system. The semidefinite program size thus grows as the square of
the number of buses, which makes solution of OPF problems by semidefinite programming
computationally intractable for large systems. Recent work using matrix completion [24–
26] reduces the computational burden inherent in solving large systems by taking advantage
of the sparse matrix structure created by realistic power system models. Sojoudi and
Lavaei [22], Bai and Wei [27], and Jabr [28] present formulations that decompose the single
large 2n×2n positive semidefinite matrix constraint into positive semidefinite constraints on
many smaller matrices. If the matrices from these decompositions satisfy a rank condition,
the 2n × 2n matrix also satisfies the rank condition and the globally optimal solution can
be obtained.
This code incorporates several computational and modeling advances in applying the
semidefinite relaxation to practical power systems. Modeling advances include allowing
multiple generators at the same bus and incorporating flow limits on parallel lines. Compu-
tational advances include enhancements to the existing matrix decomposition algorithms.
A first enhancement is a heuristic algorithm for combining some of the many small matrices
resulting from the decomposition. Since linking constraints are required between terms of
the decomposed matrices that refer to the same term in the 2n × 2n matrix, it is not al-
ways advantageous to create the smallest possible matrices. Combining matrices eliminates
some of these linking constraints, which results in a factor of two to three improvement in
computational speed over existing matrix decomposition algorithms [1].
A second enhancement presented is a technique for recovering the optimal voltage profile
from the decomposed matrices. None of the existing literature describes a method for
actually obtaining the optimal voltage profile from a solution to a decomposed formulation.
A final enhancement modifies the maximal clique decomposition as formulated by
Jabr [28] to allow for application to general power systems. This formulation uses a
Cholesky factorization of the absolute value of the imaginary part of the bus admittance
matrix to form a chordal extension of the network. However, this matrix may not be
positive definite (for instance, in networks with sufficiently large shunt capacitances), thus
preventing formation of a Cholesky factorization. This code uses an alternative matrix that
is always positive definite and gives an equivalent chordal extension in order to extend this
decomposition for general networks.
Additionally, this code includes the function testGlobalOpt, which uses the Karush–
Kuhn–Tucker (KKT) conditions of the semidefinite relaxation of the OPF problem to
create a sufficient condition for global optimality of an OPF solution. If an OPF solution

4
from any solver (e.g., the interior point solvers in Matpower [12]) satisfies the sufficient
condition, the solution is guaranteed to be globally optimal.
Finally, this code includes several functions that implement sufficient conditions for
power flow insolvability. The function insolvablepf gives a sufficient condition for in-
solvability of the power flow equations (neglecting generator reactive power limits). This
function uses the semidefinite programming relaxation of the power flow equations to min-
imize the slack bus voltage with proportional changes in PV bus voltages. If the minimum
achievable slack bus voltage in the semidefinite relaxation is greater than the specified slack
bus voltage, the power flow equations are insolvable. Conversely, if the minimum slack bus
voltage is less than or equal to the specified slack bus voltage, solvability of the power flow
equations is indeterminate. This function also provides a controlled voltage margin and a
power injection margin to the power flow solvability boundary. See [3] for further details.
The function insolvablepf_limitQ gives a sufficient condition for insolvability of the
power flow equations considering generator reactive power limits. This function uses a
mixed-integer semidefinite programming formulation to maximize the power injections in
a uniform, constant power factor profile. If the factor by which the power injections can be
changed is less than one, no power flow solution exists that satisfies the generator reactive
power limits. This function also gives a power injection margin as a measure of the distance
to the power flow solvability boundary considering generator reactive power limits. See [4]
for further details.
Using the fact that the power flow equations are polynomials in the rectangular voltage
components Vd and Vq , the function insolvablepfsos applies the Positivstellensatz theorem
from the field of real algebraic geometry and sum of squares programming to generate
infeasibility certificates. An infeasibility certificate proves that the power flow equations
are insolvable. Using a similar approach, the function insolvablepfsos_limitQ considers
upper limits on reactive power generation through a polynomial formulation of these limits.
See [4] for further details.

2 Dependencies
This code is implemented as an add-on to the power systems software package Matpower1
[12]. Matpower is a package of MATLAB M-files that solve the power flow and optimal
power flow problems.
Other required software packages include the optimization modeling tool YALMIP2 [29]
and a semidefinite programming solver compatible with YALMIP, such as SeDuMi3 [30]
or SDPT34 [31]. (Any YALMIP compatible solver should work, but this code was tested
most thoroughly with SeDuMi version 1.3 and SDPT3 version 4.0.)

1
Matpower is available at http://www.pserc.cornell.edu/matpower/
2
YALMIP is available at http://users.isy.liu.se/johanl/yalmip/
3
SeDuMi is available at http://sedumi.ie.lehigh.edu/
4
SDPT3 is available at http://www.math.nus.edu.sg/~mattohkc/sdpt3.html

5
3 Installation
1. Install YALMIP.

2. Install a semidefinite programming solver compatible with YALMIP, such as SeDuMi

or SDPT3. Successful installation of YALMIP and a semidefinite programming solver
can be tested using the command yalmiptest; the solver is correctly installed if the
“SDP” test indicates a correct solution.

3. Install Matpower. Ensure that the folder containing the Matpower files is on
the MATLAB path.

4. Unzip the files in sdp_pf.zip into a folder on the MATLAB path.

5. Solve an optimal power flow problem using the standard Matpower commands
with the option opf.ac.solver set to 'SDPOPF' or the new runsdpopf command, e.g.,
runsdpopf('case9').

4 Outputs
The optimal power flow solver sdpopf_solver outputs three variables: results, success,
and raw. The output results is a Matpower case structure (mpc) with the usual baseMVA,
bus, branch, gen, and gencost fields. There are three additional fields in this structure:

• f : The final objective function value. This is a lower bound on the the system oper-
ating cost and equals the minimum operating cost if the rank condition is satisfied.

• mineigratio : A measure of the satisfaction of the rank condition.

The rank condition is satisfied for the dual problem if the dimension of the nullspaces
of all decomposed A matrices are less than or equal to two. If the optimal voltage
profile is obtained from the dual problem, mineigratio is the minimum ratio between
the third smallest and second smallest eigenvalues (i.e., between an eigenvalue that
should not be zero and one that should be zero when the rank condition is satisfied)
among all matrices corresponding to the maximal cliques.
The rank condition is satisfied for the primal problem if the ranks of all decomposed
W matrices are less than or equal to two. If the optimal voltage profile is obtained
from the primal problem, mineigratio is the minimum ratio between the second
largest and third largest eigenvalues (i.e., between an eigenvalue that should not be
zero and one that should be zero when the rank condition is satisfied) among all
matrices corresponding to the maximal cliques.

• zero_eval : A measure of solution consistency in recovering the optimal voltage

profile when matrix decomposition is used.

6
 The optimal voltage profile is recovered from the semidefinite programming solution
using a vector in the nullspace of a matrix (see [1] for more details). This matrix
has a non-empty nullspace (i.e., it has a zero eigenvalue) for semidefinite program
solutions that satisfy the rank condition. The output zero_eval is the smallest
magnitude eigenvalue of this matrix. If zero_eval is non-zero, the optimal voltage
profile can not be constructed in a consistent manner from the decomposed matrices.

The output success is a binary variable indicating whether SDPOPF found a solution
that satisfies the rank condition. The rank condition is considered satisfied (and thus
success set to 1) when mineigratio is greater than the option sdp_pf.mineigratio_tol
and the magnitude of zero_eval is less than the option sdp_pf.zeroeval_tol.
The output raw contains the variables returned by the SDP solver in the following fields.

• xr : Final value of optimization variables

– A : Cell array of matrix variables for the dual problem

– W : Cell array of matrix variables for the primal problem

• pimul : Constraint Lagrange multipliers on

– lambdaeq_sdp : Active power equalities

– psiU_sdp : Generator active power upper inequalities
– psiL_sdp : Generator active power lower inequalities
– gammaeq_sdp : Reactive power equalities
– gammaU_sdp : Reactive power upper inequalities
– gammaL_sdp : Reactive power lower inequalities
– muU_sdp : Square of voltage magnitude upper inequalities
– muL_sdp : Square of voltage magnitude lower inequalities

• info : solver specific termination code

7
5 Solver Options
SDPOPF has several options that control the semidefinite programming relaxation of the
optimal power flow problem.

Option Default Description

yalmip.opts empty Specifies a struct of YALMIP solver options
passed to yalmip_options to override defaults,
applied after overrides from yalmip.opt_fname.
The YALMIP options struct specifies the de-
sired solver and solver parameters. For example,
the following code sets the solver to SeDuMi
with a tolerance parameter eps of 1 × 10−9 .

mpopts = mpoption;
mpopts.yalmip.opts = sdpsettings;
mpopts.yalmip.opts.solver = ’sedumi’;
mpopts.yalmip.opts.sedumi.eps = 0;

The following code sets the solver to SDPT3

with tolerance parameters equal to 1 × 10−9 .

mpopts = mpoption;
mpopts.yalmip.opts = sdpsettings;
mpopts.yalmip.opts.solver = ’sdpt3’;
mpopts.yalmip.opts.sdpt3.gaptol = 1e-9;
mpopts.yalmip.opts.sdpt3.inftol = 1e-9;
mpopts.yalmip.opts.sdpt3.steptol = 1e-9;

See help YALMIP_OPTIONS for further details.

yalmip.opt_fname '' Specifies the name of a user-supplied function passed
as the FNAME argument to yalmip_options to override
the defaults. See help YALMIP_OPTIONS for further de-
tails.
sdp_pf.max_number_of_cliques 0.1 Maximum number of maximal cliques (stopping crite-
rion for maximal clique combination).

• 0 : No maximum

• (0, 1) : Specified proportion of the original num-

ber of maximal cliques

• ≥ 1 : Specified number of maximal cliques

8
sdp_pf.eps_r 1 × 10−4 Minimum branch series resistance. The semidefinite
relaxation requires a non-zero series resistance on each
branch. This parameter specifies a floor on this resis-
tance.
sdp_pf.recover_voltage 4 Specifies the method for recovering the voltage from
the semidefinite program solution.

• 1 : From the dual problem (A matrices)

• 2 : From the primal problem (W matrices)

• 3 : Whichever of the primal or dual problem

has the smallest difference between the speci-
fied power injections at PQ buses and the PQ
injections calculated from the resulting voltage
profile, measured with a 2-norm.

• 4 : Whichever of the primal or dual problem has

the largest minimum eigenvalue ratio, indicating
the closest to satisfying the rank condition.

sdp_pf.recover_injections 2 Specifies the method for recovering the active and re-
active power injections and the line flows from the
semidefinite program solution.

• 1 : Calculate from the voltage profile.

• 2 : Recover from the primal problem (W ma-

trices).

sdp_pf.min_Pgen_diff 1 Large power system models with tight inequality con-

straints may cause numerical problems in the semidef-
inite program solver. To avoid these problems, specify
a minimum tolerance in MW for active power gener-
ation inequality constraints. If the difference between
a generator’s upper and lower active power generation
bounds is smaller than this tolerance, the active power
generation is fixed to the midpoint of the generation
range.
sdp_pf.min_Qgen_diff 1 See sdp_pf.min_Pgen_diff; this setting controls a sim-
ilar minimum generation tolerance in MVAr for reac-
tive power inequality constraints.
sdp_pf.max_line_limit 9900 Line flow limits greater than this value are considered
unlimited.

9
 sdp_pf.max_gen_limit 99998 Generators with reactive power generation limits
greater than this value are considered unlimited.
sdp_pf.ndisplay 1 When verbose is greater than 1, diagnostic and tim-
ing information will display as SDPOPF executes.
sdp_pf.ndisplay controls how frequently this diagnos-
tic information is displayed.
sdp_pf.choldense 10 Controls the density parameter of the minimum degree
ordering function amd in the Cholesky decomposition
used to create a chordal extension of the power sys-
tem graph. See the MATLAB function amd for further
details.
sdp_pf.cholaggressive 1 Controls the aggressive parameter of the minimum de-
gree ordering function amd in the Cholesky decompo-
sition used to create a chordal extension of the power
system graph. See the MATLAB function amd for fur-
ther details.
sdp_pf.bind_lagrange 1 × 10−3 A binding constraint is required to extract the opti-
mal voltage profile from the solution to the semidefi-
nite programming relaxation. Binding constraints are
identified by non-zero corresponding Lagrange multi-
pliers. A Lagrange multiplier is considered non-zero,
and therefore binding, if its magnitude is greater than
sdp_pf.bind_lagrange.
sdp_pf.zeroeval_tol 1 × 10−5 Maximum magnitude of zero_eval to consider SD-
POPF successful. Also used in testGlobalOpt for sat-
isfaction tolerances for the complementarity and posi-
tive semidefinite feasibility conditions for global opti-
mality of a candidate OPF solution.
sdp_pf.mineigratio_tol 1 × 105 Minimum value of mineigratio to consider SDPOPF
successful.

6 Functions
Several new functions were required to implement SDPOPF.

• [RESULTS,SUCCESS,RAW] = SDPOPF_SOLVER(OM, MPOPT)

This is the main SDPOPF function. It implements a semidefinite programming
relaxation of the power flow equations with matrix decomposition techniques for
exploiting power system sparsity [1].

• [RESULTS, SUCCESS] = RUNSDPOPF(CASEDATA, MPOPT, FNAME, SOLVEDCASE)

This is a function for calling SDPOPF in Matpower.

10
• [SDPOPT] = YALMIP_OPTIONS(OVERRIDES, FNAME)
[SDPOPT] = YALMIP_OPTIONS(OVERRIDES, MPOPT)
This function is used to set the YALMIP solver options (i.e., a SDPSETTINGS variable).
Overrides can be provided directly as a struct in the first argument, as a user-supplied
function name FNAME in the second argument, or as a struct or function name in
options yalmip.opts or yalmip.opt_fname, respectively, of the Matpower options
struct MPOPT. See help YALMIP_OPTIONS for further details.
• [MAXCLIQUES,E] = COMBINEMAXCLIQUES(MAXCLIQUES, E, MAXNUMBEROFCLIQUES, NDISPLAY)
This function combines the maximal cliques to reduce computation time. This func-
tion uses the heuristic that an additional variable in a matrix is equivalent to an
additional linking constraint between overlapping maximal cliques. See [1] for more
details.
• [QUANTITY] = RECOVERFROMW(SDPMAT, WREF_DD, WREF_QQ, WREF_DQ, MATIDX_DD, MATIDX_QQ,
MATIDX_DQ, W, SDPVAR, MAXCLIQUE)
This function calculates trace(sdpmat*W) for the decomposed matrices. This function
is used for recovering power injections and line flows from the primal problem when
the option sdp_pf.recover_injections is set to 2.
• [SDPVEC] = MAT2VEC(SDPMAT, WREF_DD, WREF_QQ, WREF_DQ, MATIDX_DD,
MATIDX_QQ, MATIDX_DQ)
This function is used in the formation of the matrix completion decomposition for the
semidefinite programming relaxation of the optimal power flow. It converts a 2n × 2n
symmetric matrix sdpmat into a list form. For each nonzero element of sdpmat, the
list form in sdpvec gives the appropriate decomposed matrix, the location in the
matrix, and the value for that element.
• [A] = ADDTOA(SDPMAT, WREF_DD, WREF_QQ, WREF_DQ, MATIDX_DD, MATIDX_QQ,
MATIDX_DQ, A, SDPVAR, MAXCLIQUE)
This function adds a matrix multiplied by a SDP variable to the A matrices in the
dual problem.
• [E] = PRIM(AADJ)
This is an implementation of Prim’s algorithm for calculating a minimum spanning
tree from a graph adjacency matrix. This implementation can incorporate negative
edge weights. A maximal spanning tree can be created by specifying the negative of
the graph adjacency matrix.
• [MAXCLIQUE,ISCHORDAL] = MAXCARDSEARCH(AADJ)
This function determines the maximal cliques for a chordal graph described by the
adjacency matrix Aadj. It also tests the graph for chordality. The algorithms for
determining the maximal cliques and testing for chordality are described in [32].

11
 • [COST] = COMBINECOST(MAXCLIQUES, MAXCLIQUEIDX)
This function calculates the cost of combining two maximal cliques in terms of the
number of scalar variables and linking constraints that will be required after com-
bining the maximal cliques specified in maxcliquesidx. This function implements
the clique combination heuristic described in [1]. A negative cost indicates that
the heuristic predicts decreased computational costs after combining the specified
maximal cliques.

• [AINC] = MAKEINCIDENCE(BUS, BRANCH)

This function builds the bus incidence matrix. This matrix has size nline by nbus,
with each row having two nonzero elements: +1 in the entry for the “from” bus of
the corresponding line and −1 in the entry for the “to” bus of the corresponding line.

• [YK, YK_, MK, YLINEFT, YLINETF, Y_LINEFT, Y_LINETF, YL, YL_]

= MAKESDPMAT(MPC)
This creates functions that return the matrices used in the semidefinite programming
relaxation of the optimal power flow problem.

• [MPC] = CASE3SC
OPF data for 3-bus system used in [18]. The semidefinite relaxation of the OPF
problem successfully solves case3sc with a value of 60 MVA for the line-flow limit
on the line from bus 3 to bus 2. The semidefinite relaxation fails to give a physically
meaningful solution to case3sc with a value of 50 MVA for the line-flow limit on this
line. See [18] for further details.

• [MPC] = CASE2LOCAL
OPF data for 2-bus system used in [19]. The semidefinite relaxation fails to give a
physically meaningful solution to case2local for values of the upper voltage magni-
tude limit at bus two that are between 0.985 and 1.04. See [19] for further details.

7 Additional Functionality
Three functions provide additional tools that are based on the semidefinite programming
relaxation of the power flow equations.

• [GLOBALOPT, COMP, APSD] = TESTGLOBALOPT(MPC, MPOPT)

The function testGlobalOpt uses the KKT conditions of the semidefinite relaxation
of the OPF problem to create a sufficient condition for global optimality of a specified
OPF solution. The sufficient condition requires satisfaction of both a complemen-
tarity condition (trace (AW) = 0) and a feasibility condition (A  0). Failure to
satisfy these conditions may result when the semidefinite relaxation does not satisfy
the rank condition (i.e., rank (W) > 2) [18, 19], in which case the solution may still

12
 be globally optimal. Alternatively, failure to satisfy the sufficient condition for global
optimality may indicate that a better solution exists. See [2] for details.
Specify an OPF solution in the input mpc that was solved with the associated op-
tions variable mpopt (i.e., mpc = runopf(mpc, mpopt)). The function testGlobalOpt
outputs information on satisfaction of the complementarity and feasibility conditions
in comp and Apsd, respectively. The output comp is the value of trace (AW). The
output Apsd is a flag indicating if the matrix A + I has a Cholesky factorization,
where  is specified by sdp_pf.zeroeval_tol, which is both necessary and sufficient
for positive definiteness. If comp is equal to zero (within tolerances specified by the
option sqrt(sdp_pf.zeroeval_tol) and Apsd is equal to one, the output globalopt
is equal to one; otherwise, globalopt equals zero.
Note that tight solution tolerances are typically required for good numeric results.
Additionally, enforcing small minimum branch resistances may result in satisfaction of
the complementarity and feasibility conditions for problems that would not otherwise
satisfy these conditions.

• [INSOLVABLE, VSLACK_MIN, SIGMA, ETA, MINEIGRATIO]

= INSOLVABLEPF(MPC, MPOPT)
The function insolvablepf evaluates a sufficient condition for insolvability of the
power flow equations (neglecting generator reactive power limits). A semidefinite
programming relaxation of the power flow equations minimizes the slack bus voltage
(with proportional changes in voltage at PV buses). If the minimum achievable slack
bus voltage is greater than the specified slack bus voltage, no power flow solutions
exist. The converse does not necessarily hold; a power flow solution may not exist
even if the minimum slack bus voltage is less than or equal to the specified slack bus
voltage.
As a by-product of the optimization problem used to evaluate the insolvability condi-
tion, insolvablepf yields two voltage stability margins to the power flow solvability
boundary. The output sigma is a margin in terms of controlled voltages which gives
the factor by which the slack bus voltage can be changed (with proportional changes
in PV bus voltages) while maintaining the possibility of power flow solution existence.
The output eta is a margin in terms of power injections which gives the factor by
which the power injections can be uniformly changed at constant power factor while
maintaining the possibility of power flow solution existence.
Due to incorporation of the matrix completion decomposition described in [1], this
function is suitable for large systems.
See [3] for further details.

• [INSOLVABLE, ETA, MINEIGRATIO] = INSOLVABLEPF_LIMITQ(MPC, MPOPT)

The function insolvablepf_limitQ evaluates a sufficient condition for power flow
insolvability considering generator reactive power limits. Generator reactive power

13
 limits are incorporated using a mixed-integer semidefinite programming relaxation
of the power flow equations. Using the semidefinite programming relaxation of the
power flow equations, this function maximizes the power injections in a uniform,
constant power factor injection profile. The result of this optimization problem, eta,
is the largest factor by which the power injections can be changed in the direction of
this profile while maintaining the possibility of power flow solution existence. Thus,
eta is a margin of the distance to the power flow solvability boundary, considering
generator reactive power limits. See [4] for further details.
Note that due to the computational limitations of YALMIP’s branch-and-bound
solver, this function is not suitable for large systems. This function has been success-
fully run on test systems with sizes up to the 57-bus system; 57 buses is therefore
chosen as a limit on system size. This limit can be modified using the variable
maxSystemSize.

• [INSOLVABLE] = INSOLVABLEPFSOS(MPC,MPOPT)
The function insolvablepfsos implements a sufficient condition for power flow in-
solvability using sum of squares programming to generate an infeasibility certificate
for the power flow equations. An infeasibility certificate proves insolvability of the
power flow equations. Note that absence of an infeasibility certificate does not neces-
sarily mean that a solution does exist (that is, insolvable = 0 does not imply solution
existence). This implementation uses constant (degree zero) polynomials. See [4] for
more details.

• [INSOLVABLE] = INSOLVABLEPFSOS_LIMITQ(MPC,MPOPT)
The function insolvablepfsos_limitQ implements a sufficient condition for power
flow insolvability considering gnenerators with upper limits on reactive power output.
Sum of squares programming is used to generate an infeasibility certificate for the
power flow equations. An infeasibility certificate proves insolvability of the power
flow equations. Note that absence of an infeasibility certificate does not necessarily
mean that a solution does exist (that is, insolvable = 0 does not imply solution
existence). This implementation uses constant (degree zero) polynomials. See [4] for
more details.

8 Limitations and Future Work

SDPOPF has several important differences from other OPF solvers that deserve special
attention. First, although line flow constraints specified in terms of current flow are the-
oretically possible, they are not yet implemented in this code (i.e., Matpower option
opf.flow_lim is restricted to 'S' and 'P' for apparent power and active power line flow
limits, respectively). The ability to specify flow limits based on line currents may be added
in a future release of SDP_PF.

14
 Second, since non-convex objective functions would eliminate the convexity of the
semidefinite relaxation, only convex objective functions are permissible. Quadratic and
piecewise-linear objective functions of active power generation are currently implemented in
SDPOPF. (Note that the current implementation does not allow buses with both piecewise-
linear generator cost functions and quadratic generator cost functions.) More flexible ob-
jective function specifications, such functions of reactive power generation, are theoretically
possible, but not planned for implementation at this time.
Third, DC lines are not explicitly implemented in SDPOPF. It is theoretically possible
to model DC lines in the semidefinite relaxation as linked positive and negative power
injections at the DC line terminal buses, but this functionality is not implemented nor
planned for implementation at this time. SDPOPF will throw an error for input files with
dcline fields.
Fourth, limits on branch angle differences are not implemented in SDPOPF. Future
work includes incorporating angle difference constraints using the method described in [33].
Fifth, SDPOPF does not handle user-defined cost functions or constraints defined in
Matpower. That is, some of the Matpower functionalities detailed in Section 5.3 (dis-
patchable loads, generator capability curves, and branch angle difference limits) and in
Chapter 6 (customized callbacks, reserve requirements, interface flow limits, DC transmis-
sion lines, etc.) of the Matpower manual [34] are not implemented in SDPOPF. It may be
possible to add some of these functionalities through modifications to the sdpopf_solver.m
file.
A noteworthy difference in behavior between the current version of SDPOPF and other
Matpower solvers is in the handling of dispatchable loads. In other solvers, dispatchable
loads are specified as generators with negative generation limits. An additional constraint
enforces constant power factor at the dispatchable load. In SDPOPF, dispatchable loads
can still be specified with negative generation limits, but the constant power factor con-
straint is not enforced; the dispatchable load can demand any reactive power within the
specified reactive power limits. This behavior may be changed in a future version of SDP_PF
to be consistent with other solvers.
Sixth, note that recovering the optimal voltage profile requires a binding constraint.
SDPOPF can currently recover solutions with binding voltage magnitude or line-flow con-
straints, which is sufficient for the majority of practical OPF problems. However, it is
possible to construct cases with solutions that have neither binding voltage magnitude nor
binding line-flow constraints. Since there is at least one binding constraint at the optimal
solution for an OPF problem, it is possible, in principle, to recover the optimal voltage
profile for any OPF problem. These cases are expected to be rare.
Finally, note that even with the computational advances from reference [1] the semidef-
inite program solver typically has significantly higher processing and memory requirements
than other OPF solvers, particularly for large system models. With Matpower option
verbose set to 2, SDPOPF will display diagnostic information that is useful for gauging
the progress of SDPOPF for large system models.

15
9 Conclusion
This document has introduced a package of code implementing applications of a semidef-
inite programming relaxation of the power flow equations. When a rank condition is
satisfied, the semidefinite programming relaxation obtains the globally optimal solution
to an OPF problem. This code is an implementation of the research published in [1],
incorporating both the modeling and computational advances contained therein. Model-
ing advances include the capability to limit flows on parallel branches and allow multiple
generators at the same bus. Computational advances in matrix decomposition algorithms
extend the applicability of these algorithms and significantly reduce computation time.
Additional functionality includes testGlobalOpt, an implementation of a sufficient con-
dition for global optimality of an OPF solution, insolvablepf, an implementation of a
sufficient condition for power flow insolvability, insolvablepf_limitQ, an implementation
of a sufficient condition for power flow insolvability considering generator reactive power
limits, insolvablepfsos, an implementation of method for generating infeasibility certifi-
cates for the power flow equations, and insolvablepfsos_limitQ, an implementation of a
method for generating infeasibility certificates for the power flow equations with consider-
ation of upper limits on reactive power generation.

Acknowledgements
The author would like to thank Dr. Bernard Lesieutre, Dr. Christopher DeMarco, Jesse
Holzer, Alex Borden, and Vikas Dawar at the University of Wisconsin-Madison and Ray
Zimmerman at Cornell University for their insights and assistance preparing and testing
this code.

References
[1] D. Molzahn, J. Holzer, B. Lesieutre, and C. DeMarco, “Implementation of a Large-
Scale Optimal Power Flow Solver Based on Semidefinite Programming,” IEEE Trans-
actions on Power Systems, vol. 28, no. 4, pp. 3987–3998, 2013.

[2] D. Molzahn, B. Lesieutre, and C. DeMarco, “A Sufficient Condition for Global Op-
timality of Solutions to the Optimal Power Flow Problem,” IEEE Transactions on
Power Systems, vol. 29, pp. 978–979, March 2014.

[3] D. K. Molzahn, B. C. Lesieutre, and C. L. DeMarco, “A Sufficient Condition for Power

Flow Insolvability with Applications to Voltage Stability Margins,” IEEE Transactions
on Power Systems, vol. 28, no. 3, pp. 2592–2601, 2013.

[4] D. K. Molzahn, V. Dawar, B. C. Lesieutre, and C. L. DeMarco, “Sufficient Conditions

for Power Flow Insolvability Considering Reactive Power Limited Generators with
Applications to Voltage Stability Margins,” in Bulk Power System Dynamics and

16
 Control - IX. Optimization, Security and Control of the Emerging Power Grid, 2013
IREP Symposium, August 25-30 2013.

[5] B. Lesieutre and I. Hiskens, “Convexity of the Set of Feasible Injections and Revenue
Adequacy in FTR Markets,” IEEE Transactions on Power Systems, vol. 20, pp. 1790–
1798, November 2005.

[6] W. Bukhsh, A. Grothey, K. McKinnon, and P. Trodden, “Local Solutions of the

Optimal Power Flow Problem,” vol. 28, no. 4, pp. 4780–4788, 2013.

[7] J. Lavaei and S. Low, “Zero Duality Gap in Optimal Power Flow Problem,” IEEE
Transactions on Power Systems, vol. 27, pp. 92–107, February 2012.

[8] M. Huneault and F. Galiana, “A Survey of the Optimal Power Flow Literature,” IEEE
Transactions on Power Systems, vol. 6, pp. 762–770, May 1991.

[9] J. Momoh, R. Adapa, and M. El-Hawary, “A Review of Selected Optimal Power Flow
Literature to 1993. I. Nonlinear and Quadratic Programming Approaches,” IEEE
Transactions on Power Systems, vol. 14, pp. 96–104, February 1999.

[10] J. Momoh, M. El-Hawary, and R. Adapa, “A Review of Selected Optimal Power Flow
Literature to 1993. II. Newton, Linear Programming and Interior Point Methods,”
IEEE Transactions on Power Systems, vol. 14, pp. 105–111, February 1999.

[11] Z. Qiu, G. Deconinck, and R. Belmans, “A Literature Survey of Optimal Power Flow
Problems in the Electricity Market Context,” in Power Systems Conference and Ex-
position, 2009. PSCE ’09. IEEE/PES, pp. 1–6, March 2009.

[12] R. Zimmerman, C. Murillo-Sánchez, and R. Thomas, “MATPOWER: Steady-State

Operations, Planning, and Analysis Tools for Power Systems Research and Educa-
tion,” IEEE Transactions on Power Systems, no. 99, pp. 1–8, 2011.

[13] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite Programming for Optimal
Power Flow Problems,” International Journal of Electrical Power & Energy Systems,
vol. 30, no. 6-7, pp. 383–392, 2008.

[14] A. Nemirovski, “Lectures on Modern Convex Optimization,” 2012.

[15] L. Vandenberghe and S. Boyd, “Semidefinite Programming,” SIAM Review, pp. 49–95,
March 1996.

[16] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Univ Pr, 2004.

[17] “Power Systems Test Case Archive. University of Washington Department of Electri-
cal Engineering [Online] Available at: http://www.ee.washington.edu/research/
pstca/,”

17
[18] B. C. Lesieutre, D. K. Molzahn, A. R. Borden, and C. L. DeMarco, “Examining the
Limits of the Application of Semidefinite Programming to Power Flow Problems,” in
49th Annual Allerton Conference on Communication, Control, and Computing, 2011,
September 28-30 2011.
[19] W. Bukhsh, A. Grothey, K. McKinnon, and P. Trodden, “Local Solutions of Op-
timal Power Flow,” Tech. Rep. ERGO-11-017, University of Edinburgh School of
Mathematics, Edinburgh Research Group in Optimization, 2011. [Online]. Available:
http://www.maths.ed.ac.uk/ERGO/pubs/ERGO-11-017.html.
[20] B. Zhang and D. Tse, “Geometry of Feasible Injection Region of Power Networks,” in
49th Annual Allerton Conference on Communication, Control, and Computing, 2011,
September 28-30 2011.
[21] S. Bose, D. Gayme, S. Low, and K. Chandy, “Optimal Power Flow Over Tree Net-
works,” in 49th Annual Allerton Conference on Communication, Control, and Com-
puting, 2011, September 28-30 2011.
[22] S. Sojoudi and J. Lavaei, “Physics of Power Networks Makes Hard Optimization Prob-
lems Easy To Solve,” in To appear in 2012 IEEE Power & Energy Society General
Meeting, July 22-27 2012.
[23] J. Lavaei, D. Tse, and B. Zhang, “Geometry of Power Flows in Tree Networks,” in To
appear in 2012 IEEE Power & Energy Society General Meeting, July 22-27 2012.
[24] M. Fukuda, M. Kojima, K. Murota, K. Nakata, et al., “Exploiting Sparsity in Semidef-
inite Programming via Matrix Completion I: General Framework,” SIAM Journal on
Optimization, vol. 11, no. 3, pp. 647–674, 2001.
[25] K. Nakata, K. Fujisawa, M. Fukuda, M. Kojima, and K. Murota, “Exploiting Sparsity
in Semidefinite Programming via Matrix Completion II: Implementation and Numer-
ical Results,” Mathematical Programming, vol. 95, no. 2, pp. 303–327, 2003.
[26] S. Kim, M. Kojima, M. Mevissen, and M. Yamashita, “Exploiting Sparsity in Lin-
ear and Nonlinear Matrix Inequalities via Positive Semidefinite Matrix Completion,”
Mathematical Programming, vol. 129, no. 1, pp. 33–68, 2011.
[27] X. Bai and H. Wei, “A Semidefinite Programming Method with Graph Partitioning
Technique for Optimal Power Flow Problems,” International Journal of Electrical
Power & Energy Systems, 2011.
[28] R. Jabr, “Exploiting Sparsity in SDP Relaxations of the OPF Problem,” IEEE Trans-
actions on Power Systems, vol. 27, pp. 1138–1139, May 2012.
[29] J. Lofberg, “YALMIP: A Toolbox for Modeling and Optimization in MATLAB,” in
IEEE International Symposium on Computer Aided Control Systems Design, 2004,
pp. 284–289, IEEE, 2004.

18
[30] J. Sturm, “Using SeDuMi 1.02, A MATLAB Toolbox for Optimization Over Symmet-
ric Cones,” Optimization Methods and Software, vol. 11, no. 1, pp. 625–653, 1999.

[31] R. Tütüncü, K. Toh, and M. Todd, “Solving Semidefinite-Quadratic-Linear Programs

using SDPT3,” Mathematical Programming, vol. 95, no. 2, pp. 189–217, 2003.

[32] R. Tarjan and M. Yannakakis, “Simple Linear-Time Algorithms to Test Chordality

of Graphs, Test Acyclicity of Hypergraphs, and Selectively Reduce Acyclic Hyper-
graphs,” SIAM Journal on Computing, vol. 13, p. 566, 1984.

[33] R. Madani, S. Sojoudi, and J. Lavaei, “Convex Relaxation for Optimal Power Flow
Problem: Mesh Networks,” IEEE Transactions on Power Systems.

[34] R. D. Zimmerman and C. E. Murillo-Sánchez, “MATPOWER 5.0 User’s Manual,”

December 17, 2014.

19