4128

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

Computation of Maximum Loading Points

via the Factored Load Flow
Catalina Gmez-Quiles, Member, IEEE, Antonio Gmez-Expsito, Fellow, IEEE, and Walter Vargas
AbstractThis paper presents a fast and straightforward algorithm to obtain the maximum loading points of power systems by
simply performing bisection searches between feasible and infeasible load ow cases. The proposed method exploits the ability and
robustness of the factored load ow solution procedure to converge
to complex solutions well beyond the maximum loadability points.
Test cases on a diversity of benchmark cases show that the new
method is computationally attractive when compared with existing
continuation power ow methods.
Index TermsLoad ow, maximum loading point, solution infeasibility.

I. INTRODUCTION

ETERMINING the steady-state maximum loading point

of a power system has uninterruptedly been a topic of
research since the seventies [1][4]. In fact, the rst attempts to
address this issue for the simplest single-line case can be traced
back to the twenties [5][7].
It was in the eighties and nineties when, in addition to several theoretical contributions dealing with bifurcations and multiplicity of solutions in simplied lossless systems [8][10], the
rst practical procedures arose aimed at estimating steady-state
stability margins [11][13] and determining maximum loading
points [14][20].
Since then, a plethora of algorithms have been published
which cannot be properly summarized here due to space limitations. Fortunately, this has been done in several technical
reports [21], book chapters [22], [23] and books [24], [25],
where a taxonomy of methods, along with their advantages
and limitations, are provided. Following [22], [23], existing
methods for the determination of maximum loading (point of
collapse) can be broadly classied as follows:
Direct methods, in which a nonlinear equation system
mathematically characterizing the point of collapse (PoC)
is formulated and solved [16]. Even though this can be
done by resorting to ad hoc procedures [17], [18], the most
systematic way of addressing this issue is by explicitly
posing a nonlinear programming problem [19], [26].
Homotopy (or continuation power ow) methods, in which
a set of intermediate points (nose curve), corresponding to
Manuscript received July 13, 2015; revised October 13, 2015; accepted
November 23, 2015. Date of publication December 25, 2015; date of current
version August 17, 2016. This work was supported by the Spanish government
(Ministry of Economy and Competitiveness) under grant ENE2013-48428-C2.
Paper no. TPWRS-01010-2015.
The authors are with the Department of Electrical Engineering, University of
Seville, Seville, Spain.
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TPWRS.2015.2505185

increasingly higher loading levels, are sequentially computed [14], [15], [20], [27][29]. Each point in the curve
involves two phases, named as predictor and corrector
steps. The main advantage of these techniques is that, as a
byproduct, the so-called nose curves are obtained.
In [30] it has been shown that both continuation and optimization methods are actually equivalent.
A related problem to which some attention has also been paid
is that of restoring solvability in the presence of infeasible cases
for which no real solution exists [31][33]. The alternative PoC
method proposed in [34] approaches the bifurcation point from
the infeasible region, for which a globally convergent load ow
procedure is resorted to [35].
Recently, a factored approach for solving the load ow
problem has been presented [36], which directly stems, as a
particular case, from the equality-constrained WLS problem
[37], [38]. The main idea behind the factored load ow (FLF)
procedure, which can be applied as well to a much broader class
of nonlinear equations [39], is to unfold the original nonlinear
system into a sequence of simpler problems. At each iteration,
this involves the solution of two linear systems with a trivial
nonlinear mapping embedded in between.
Apart from its enhanced convergence rate when compared
to that of the conventional Newton-Raphson (NR) method, specially in cases near the maximum loadability limit, a particularly
relevant feature of the FLF procedure, preliminarily explored
in [40], is its ability to reliably converge to complex solutions
when solving infeasible cases, provided the at start prole is
suitably modied. The possibility of solving infeasible cases
by explicitly resorting to state variables in the complex domain
can be traced back to [41], where the multiplicity of solutions
issue was addressed. To the authors' knowledge, however, such
a promising idea has not been so far sufciently exploited.
This paper takes advantage of the robustness of the FLF
method, capable of indistinctly solving feasible or infeasible
cases, to develop a simple and competitive procedure for
determining the maximum loading points of power systems, by
simply performing a bisection-like search alternatively solving
feasible and infeasible cases.
The structure of this paper is as follows: Section II reviews the
recently introduced FLF formulation. Then, Section III presents
the bisection-based procedure to determine maximum loading
points. Finally, the proposed strategy is tested on several benchmark systems and compared with several PoC and continuation
power ow methods.
II. REVIEW OF THE FACTORED LOAD FLOW
For a network with buses, let the power ow equations be
written in compact form as follows:
(1)

0885-8950 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/
redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

GMEZ-QUILES et al.: COMPUTATION OF MAXIMUM LOADING POINTS VIA THE FACTORED LOAD FLOW

4129

where,

(12)
(13)

is the vector of
specied quantities,
namely net active power injection at all buses,
but the slack, net reactive power injection at
load (or PQ) buses and voltage magnitude at
generation (or PV) buses:

and
where and are rectangular matrices of sizes
, respectively, and vector
represents a set of elementary
nonlinear functions, which can be trivially inverted:

(2)

(14)

is the state vector, composed of two unknowns

per bus (voltage magnitude and phase angle, in
polar form), except for the slack bus, which is
taken as phase reference
:

The above augmented system can also be expressed in more

compact form, by simply eliminating vector :

(3)
From an initial guess, , the NR method proceeds by successively solving the following system in unknowns:
(4)
,
is the Jacobian of
, comwhere
puted at the current point , and
is the power mismatch
or residual vector.
In polar form, the set of generation voltage magnitudes can
be trivially removed from both and , leading to a reduced
system in
unknowns, but in the sequel they will be kept
in the model for completeness and convenience.
The FLF solution resorts to slightly modied specied and
state vectors, and , and requires the introduction of two auxiliary vectors and as follows [36]:
In vector , specied generator voltages are replaced by
their squared values:
(5)
.
where
In the state vector, voltage magnitudes are replaced for
convenience by their log counterparts:
(6)
where
Intermediate vector :

(15)
(16)
Notice that (15) is a linear underdetermined system whereas
(16) is an overdetermined one. Among the innite solutions to
(15) only that exactly satisfying (16) constitutes a solution to
the original nonlinear system (1).
Obviously, when the remaining auxiliary vector is eliminated, the original system is obtained:
(17)
It is worth noting that the NR scheme can be applied as well
to the above factorized expression, leading to,
(18)

where is the Jacobian of

, whose inverse is trivially obtained [36].
The two-stage FLF procedure, arising from the above unfolded representation of nonlinear systems, can then be formally
stated as follows [36], [39]:
. Initialize iteration counter,
, set
.
:
1.1. Compute vector

by solving the linear system,

(7)
where, for each branch connecting buses and , the following pair of variables is adopted [42]:
(8)
(9)
variables ( being
Vector then comprises
the number of branches).
Intermediate vector , composed also of variables,
where,

(10)

. For an initial guess,

(19)
1.2 Obtain

from,
(20)

1.3 Perform the nonlinear transformation

the inverted Jacobian,
.

and obtain

:
2.1 Obtain

by solving,
(21)

The factored solution method unfolds the original system (1)

into the following simpler problems [39]:
(11)

. Check for reactive power

2.2 Update
limits and switch bus types if needed.
2.3 Convergence check: if
else, return to step 1.

, then stop;

4130

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

Fig. 1. Two-bus illustrative system and associated data.

Fig. 3. Region of attraction of the FLF's method for a two-bus system, indicating number of iterations needed to obtain the solution within a tolerance of
.

Fig. 2. Region of attraction of the Newton method for a two-bus system, indicating number of iterations needed to obtain the solution within a tolerance of
.

Steps 1.1 and 1.2 provide a new vector that, satisfying (15),
minimizes
(least-distance problem). Note that the
computational effort involved in step 1.1 is actually very small
if the Cholesky factors of matrix
are available from previous iterations and the mismatch vector
was saved in
step 2.3 (only when bus-type switchings take place has matrix
to be updated). Tests performed elsewhere [36] show that the
computational effort of step 1 is well offset by step 2 being much
less costly than a regular NR step. Overall, each iteration of the
above two-step factored procedure is on average faster than that
of NR for large realistic problems.
III. BEHAVIOR

OF THE FLF
LOADING

BEYOND
POINT

THE

MAXIMUM

It was shown in [39] that, near the solution, the factored approach has a similar or even better convergence rate (quadratic)
as that of NR method, while its basin of attraction around the
solution point is much wider. This means that the FLF method
can converge from starting points which are much further away
from the solution.
Consider for instance the two-bus system used in [23] (pp.
8182) to illustrate the region of attraction of Newton's method
for a normal operating point. For the impedance and loading
data provided in Fig. 1 the solution point is
and
.
The region of attraction corresponding to the Newton's
method is represented in Fig. 2, which is in agreement with
that provided in [23], where iteration numbers refer to a convergence tolerance of
. Note that the number of iterations
quickly increases for increasing voltage magnitudes and that
very large phase angles are not tolerated. The external region
labelled NC represents starting points for which either convergence is not achieved or another solution point is reached.

The counterpart for the FLF method is shown in Fig. 3,

where it is clearly seen that convergence to the solution point is
achieved, usually in much less iterations, starting from almost
any point in the plane, except for very low voltage magnitudes
and very high phase angles. Actually the thin region at the
bottom labelled NC denotes the region of attraction of the
alternative solution point in the lower part of the nose curve,
rather than non-convergent cases.
This enhanced capability to converge from starting points
which are far away from the solution also applies to realistic
systems, as shown in [36], [39].
The most outstanding feature of the FLF, though, is its ability
to converge to a complex solution when no real satisfying (1)
exists. As discussed in [40], by simply adding a small imaginary
component to the initial value of voltage magnitudes, the capability of the FLF to reach complex solutions, when real solutions
do not exist, is signicantly enhanced. This way, the FLF can
solve both feasible and infeasible cases in a seamless fashion
by slightly modifying the customary at start.
Just as an example, in addition to those reported in [40], let
us consider the IEEE 30-bus system. Table I shows the number
of iterations for increasing loading scenarios (real powers at
all buses and reactive powers at load buses are multiplied by a
scalar ), when reactive power limits of generators are ignored.
Flat start is used for the NR method while
is used to
initialize the voltage magnitudes of PQ buses in the FLF procedure. A column showing the largest imaginary component of the
computed voltage magnitudes is also added for convenience.
As can be seen, the point of collapse for the 30-bus system lies
somewhere between
and
. The last
point for which the NR method converges is with
,
whereas the FLF converges for much larger values of , well in
the infeasible region. For practical purposes, though, choosing
as the bifurcation point would be sufciently accurate, since the imaginary component for
is already
larger than
.
Needless to say, if the Jacobian matrix
is
exactly singular (bifurcation point), then both methods will
break down. But, as claimed elsewhere [23], when working
with oating point arithmetic it is extremely unlikely or simply

GMEZ-QUILES et al.: COMPUTATION OF MAXIMUM LOADING POINTS VIA THE FACTORED LOAD FLOW

4131

TABLE I
NUMBER OF ITERATIONS TAKEN BY THE FLF AND NR METHODS ON THE
30-BUS SYSTEM FOR INCREASING VALUES OF THE LOAD SCALING FACTOR
WITHIN A CONVERGENCE THRESHOLD OF

Fig. 4. Evolution of

impossible that such a critical point is exactly or mathematically hit, particularly in realistic systems. Stated in other terms,
computationally speaking there are no singular matrices, but
very ill-conditioned ones. For the 30-bus system, the condition
number of the Jacobian is around
for all tested cases
between
and , which is absolutely tolerable for modern
computers.
IV. MAXIMUM LOADING POINT DETERMINATION
BY MEANS OF THE FLF
Given the base-case vector of specied quantities, (net injected powers and generator voltages), the maximum loading
point of a system can be found by letting the loads (and generators) follow an assumed trajectory, , until no real solution exists for (1). Throughout this process, generator voltages are kept
constant unless reactive power limits are hit, in which case the
involved buses are switched to PQ type. In the simplest form, the
loading level can be expressed in terms of a scalar , as follows:

obtain

. Choose a value for

.
.

. For

run the FLF to

. Run the FLF to obtain

. If
, then set
,
return to step 2; else continue (bisection phase).
. Run the FLF to obtain

. If
else,

and

, then STOP; else continue.

with
. Starting from the base case (
or
), the loading level is increased until a Hopf or limit-induced
bifurcation is reached for
, representing the maximum
(critical) loadability under the mentioned assumptions.
The procedure to determine the maximum loading point via
the FLF procedure comprises two phases. The rst phase starts

increases for the 30-bus system.

with
, which is assumed without loss of generality to be
feasible. Then, successive points are computed for increasing
values of until the FLF algorithm provides a solution with
non-negligible imaginary components, indicating that an infeasible point has been reached. This infeasible point, along with
the latest feasible one, dene an interval which is successively
bisected or halved by means of the binary search strategy, until
its width is sufciently small. For the sake of computational
saving, each execution of the FLF starts with the solution of the
closest feasible solution rather than from at start (in all cases a
certain imaginary component should be added to speed up convergence in case an infeasible case is being solved, which is
unknown in advance).
The proposed procedure can be formally stated as follows:

. If
where vectors
and
represent the growth rate of net
active power for all buses but the slack and net reactive power
for load buses, respectively. In absence of better hypotheses, it
is customarily assumed that both load and generation increase
in proportion to the base case quantities, leading to,

as

and

, then
and
. Return to step 4.

.
;

The mean value of the interval,

, tends to the critical
loading point, . The stopping decision is governed by the
width of the interval
, which can be set between
and
depending on the required accuracy. Regarding ,
it should be of the same order as the convergence threshold
adopted in step 2.3 of the FLF algorithm.
The initial value of
should be carefully chosen. If it is too
small, then a large number of FLF executions will be required
during the rst phase, before the second (bisection) phase can
begin. If it is too large, then too many steps may be needed
during the bisection phase.

4132

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

Fig. 5. Sequence of points computed by the FLF-based proposed method for the two-bus system. Real (imaginary) values are shown on the left (right) axis.

Regarding the bisection strategy, simply taking the average of

the current interval, as done by step 5, has shown in practice to
be effective enough, but more elaborated interpolation strategies
are possible taking into account the values of
and
. These will be explored in a future work.
Note that the point of collapse obtained at the end will be
conditioned by the set of reactive power limits which are hit
throughout the search process.
Consider again the 30-bus system (the reader is referred to
Table I). Starting with
and
the points
,
2, 2.5 and 3 will be successively explored. Then, the bisection
procedure will start in the interval [2.5 3] and will come up
with a critical value which is close to
. Fig. 4 graphically shows for this example the evolution of
as increases from 1 to 5. Logarithm scale has been adopted on
the vertical axis so that the abrupt transition around the bifurcation point is more clearly visualized.
For the simpler two-bus system analyzed in the previous section, Fig. 5 shows the sequence of solution points converging to
the bifurcation point, as computed by the FLF-based algorithm.
In this particular example, with
, the bisection phase
starts between points 4 (feasible) and 5 (infeasible). Notice that,
in both test cases (two and 30 buses), by far a majority of points
are computed during the bisection phase.
Fig. 6 shows a close-up view of Fig. 5 near the bifurcation
point, where the results of the bisection phase are better visualized.
V. TEST RESULTS
This section is aimed at experimentally checking the performance of the FLF-based maximum loading point computation.
Results for several benchmark cases will be separately provided
when reactive power limits are ignored and enforced, respectively.
The following common settings have been chosen:
During the rst phase,
. As the accuracy of the
FLF process is not very relevant at this preliminary stage,
the convergence threshold is
. Each load ow solution is initialized with the (real) solution of the previously
computed (feasible) point, except for a small imaginary
component
which is added to every load voltage

Fig. 6. Zoomed view of Fig. 5 around the bifurcation point.

magnitude. As explained earlier, this is intended to help

the FLF reach the next point, in case it is infeasible.
During the bisection phase,
while
and
afterwards. The stopping criterion is
. During this stage each load ow solution is
initialized with a state vector whose real component is
the latest feasible point while the imaginary component is
taken from the latest infeasible point.
A. Reactive Power Limits Ignored
Table II collects the test results obtained with a Matlab prototype of the proposed method when applied to several networks
ranging from 14 to over 2000 buses, along with those obtained
by the following software in the public domain:
The predictor-corrector continuation power ow implemented in Matpower [43].
The predictor-corrector continuation power ow implemented in PSAT [23].
The PoC method, as implemented in UWPFLOW package
[16], which is used through the interface provided in PSAT.
Perpendicular intersection and adaptive step size (initial
value set to 0.5) options were adopted for the continuation
power ow, as they were found experimentally to provide faster
solutions.

GMEZ-QUILES et al.: COMPUTATION OF MAXIMUM LOADING POINTS VIA THE FACTORED LOAD FLOW

MAXIMUM LOADING POINTS

4133

TABLE II
, NUMBER OF STEPS AND SOLUTION TIMES (REACTIVE POWER LIMITS IGNORED)

From left to right, the table shows, for each network and
tested method, the maximum loading point
, the number of
steps or intermediate points computed (in the proposed method,
real and complex points, corresponding to feasible and infeasible cases respectively, are separately recorded, as well as the
total number of FLF iterations involved throughout the process)
and the solution times on an Intel i7 with 8Gb of RAM running
at 2.8 GHz.
The following conclusions can be reached from the gures
shown in the table:
Unlike competing methods based on continuation power
ow or PoC, the number of intermediate points computed
by the proposed method is independent of the network size.
In fact, once the bisection phase begins, the number of
additional points computed is only a function of the initial
interval width,
, and the convergence tolerance
specied for the maximum
. The following expression
can be easily derived:

where is the number of points that will be computed

during the bisection phase. For instance, with
and
,
. This explains why 15 points
(
and
during the preliminary phase, and
then 13 extra points during the bisection phase) are sufcient when
, as happens with three out of the ve
systems tested.
According to the fourth column, the average number of iterations required by the FLF-based procedure to converge
to a point (feasible or infeasible) is 3.1, which is an outstanding gure considering that a majority of computed
load ow solutions are relatively close to the bifurcation
point. This further conrms the robustness of the FLF procedure, already visualized in Fig. 3.
Even though solution times depend to a large extent on how
professionally the code is written, the proposed method,
being conceptually very simple (the same load ow solver
is repeatedly called), is also much faster than competing
methods, particularly for large systems.
Except for the UWPFLOW method, which tends to provide
larger values of (sometimes leading clearly to infeasible
solutions), basically the same critical point is obtained by
the remaining tested techniques.
B. Reactive Power Limits Considered
Table III collects the test results when reactive power limits
are enforced. In this case, Matpower is not shown as the current
version does not support this function when running continuation power ows. Some difculties were found in isolated cases

TABLE III
AND SOLUTION TIMES
MAXIMUM LOADING POINTS
(REACTIVE POWER LIMITS ENFORCED)

when applying the remaining competing methods, as noted by

footnotes in the table.
As expected, a signicant reduction in the values of
can
be noticed, along with an increase in solution times, particularly
for the PSAT tool.
VI. CONCLUSION
This paper deals with the classical problem of determining
the maximum loading point of a power system in steady state,
which is of paramount importance in a context characterized by
increasingly stressed scenarios, facing transmission of massive
amounts of renewables to longer distances than ever. The proposed method resorts to the recently introduced factored load
ow solution procedure, which has shown a unique capability
to reliably converge to complex solutions when the maximum
loading point is exceeded. This feature is exploited in this paper
to devise a simple and fast procedure for determining the maximum loading points by means of bisection searches between
feasible and infeasible load ow cases. The proposed method
has been successfully tested on a diversity of benchmark cases.
When compared with existing and more elaborated continuation
power ow or PoC methods, the new method has shown to be
computationally superior.
Future research efforts will be directed to improve the
search procedure so that the number of intermediate load ow
points obtained, and hence the computational effort, are further
reduced.
ACKNOWLEDGMENT
The authors thank Prof. F. Milano for his valuable help when
using UWPFLOW from PSAT interface.

4134

REFERENCES
[1] A. J. Korsak, On the question of uniqueness of stable load-ow
solutions, IEEE Trans. Power App. Syst., vol. PAS-91, no. 3, pp.
10931100, May 1972.
[2] C. J. Tavora, Smith, and J. M. Otto, Equilibrium analysis of power
systems, IEEE Trans. Power App. Syst., vol. PAS-91, no. 3, pp.
11311137, May 1972.
[3] V. A. Venikov, V. A. Stroev, V. I. Idelchick, and V. I. Tarasov, Estimation of electrical power system steady-state stability in load ow
calculations, IEEE Trans. Power App. Syst., vol. PAS-94, no. 3, pp.
10341041, May 1975.
[4] B. K. Johnson, Extraneous and false load ow solutions, IEEE Trans.
Power App. Syst., vol. PAS-96, no. 2, pp. 524534, Mar. 1977.
[5] R. D. Evans and R. C. Bergvall, Experimental analysis of stability
and power limitations, Trans. Amer. Inst. Elect. Eng., vol. XLIII, pp.
3958, Jan. 1924.
[6] E. B. Shand, The limitations of output of a power system involving
long transmission lines, Trans. Amer. Inst. Elect. Eng., vol. XLIII, pp.
59103, Jan. 1924.
[7] R. E. Doherty and H. H. Dewey, Fundamental considerations of power
limits of transmission systems, Trans. Amer. Inst. Elect. Eng., vol.
XLIV, pp. 972984, Jan. 1925.
[8] A. Arapostathis, S. Sastry, and P. Varaiya, Bifurcation analysis of the
load ow equations, in Proc. 19th IEEE Conf. Decision and Control
Including Symp. Adaptive Processes, Dec. 1012, 1980, pp. 641644.
[9] A. Arapostathis, S. Sastry, and P. Varaiya, On the solutions of the load
ow equations in power systems, in Proc. 19th IEEE Conf. Decision
and Control Including Symp. Adaptive Processes, Dec. 1012, 1980,
pp. 152153.
[10] J. Baillieul and C. I. Byrnes, Geometric critical point analysis of lossless power system models, IEEE Trans. Circuits Syst., vol. 29, no. 11,
pp. 724737, Nov. 1982.
[11] J. Jarjis and F. D. Galiana, Quantitative analysis of steady state stability in power networks, IEEE Trans. Power App. Syst., vol. PAS100, no. 1, pp. 318326, Jan. 1981.
[12] P. Kessel and H. Glavitsch, Estimating the voltage stability of a
power system, IEEE Trans. Power Del., vol. 1, no. 3, pp. 346354,
Jul. 1986.
[13] T. Van Cutsem, A method to compute reactive power margins with
respect to voltage collapse, IEEE Trans. Power Syst., vol. 6, no. 1, pp.
145156, Feb. 1991.
[14] K. Iba, H. Suzuki, M. Egawa, and T. Watanabe, Calculation of critical loading condition with nose curve using homotopy continuation
method, IEEE Trans. Power Syst., vol. 6, no. 2, pp. 584593, May
1991.
[15] V. Ajjarapu and C. Christy, The continuation power ow: A tool for
steady state voltage stability analysis, IEEE Trans. Power Syst., vol.
7, no. 1, pp. 416423, Feb. 1992.
[16] C. A. Canizares and F. L. Alvarado, Point of collapse and continuation
methods for large AC/DC systems, IEEE Trans. Power Syst., vol. 8,
no. 1, pp. 17, Feb. 1993.
[17] Z.-C. Zeng, F. D. Galiana, B. T. Ooi, and N. Yorino, A simplied
approach to estimate maximum loading conditions in the load ow
problem, IEEE Trans. Power Syst., vol. 8, no. 2, pp. 646654, May
1993.
[18] N. Yorino, S. Harada, and H. Cheng, A method to approximate a
closest loadability limit using multiple load ow solutions, IEEE
Trans. Power Syst., vol. 12, no. 1, pp. 424429, Feb. 1997.
[19] G. D. Irisarri, X. Wang, J. Tong, and S. Mokhtari, Maximum loadability of power systems using interior point nonlinear optimization
method, IEEE Trans. Power Syst., vol. 12, no. 1, pp. 162172, Feb.
1997.
[20] G. C. Ejebe, J. Tong, J. G. Waight, J. G. Frame, X. Wang, and W.
F. Tinney, Available transfer capability calculations, IEEE Trans.
Power Syst., vol. 13, no. 4, pp. 15211527, Nov. 1998.
[21] C. Canizares, Voltage Stability Assessment: Concepts, Practices and
Tools, Power System Stability Subcommittee of the Power System Dynamic Performance Committee, PES-TR9, Aug. 2002.
[22] C. Canizares, L. Rouco, and G. Andersson, Electric Energy Systems:
Analysis and Operation. Boca Raton, FL, USA: CRC Press, 2009,
ch. 10.
[23] F. Milano, Power System Modelling and Script. New York, NY,
USA: Springer, 2010, ch. 5.
[24] C. W. Taylor, Power System Voltage Stability. New York, NY, USA:
McGraw-Hill, 1994.
[25] T. V. Cutsem and C. Vournas, Voltage Stability of Electric Power Systems. Norwell, MA, USA: Kluwer, 1998.

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

[26] Y. Dai, J. D. McCalley, and V. Vittal, Simplication, expansion

and enhancement of direct interior point algorithm for power system
maximum loadability, IEEE Trans. Power Syst., vol. 15, no. 3, pp.
10141021, Aug. 2000.
[27] I. A. Hiskens and R. J. Davy, Exploring the power ow solution space
boundary, IEEE Trans. Power Syst., vol. 16, no. 3, pp. 389395, Aug.
2001.
[28] C.-W. Liu, C.-S. Chang, J.-A. Jiang, and G.-H. Yeh, Toward a
CPFLOW-based algorithm to compute all the type-1 load-ow solutions in electric power systems, IEEE Trans. Circuits Syst. I: Reg.
Papers, vol. 52, no. 3, pp. 625630, Mar. 2005.
[29] P. Xu, X. Wang, and V. Ajjarapu, Continuation power ow with adaptive stepsize control via convergence monitor, IET Gener., Transm.,
Distrib., vol. 6, no. 7, pp. 673679, Jul. 2012.
[30] R. J. Avalos, C. A. Canizares, F. Milano, and A. J. Conejo, Equivalency of continuation and optimization methods to determine saddlenode and limit-induced bifurcations in power systems, IEEE Trans.
Circuits Syst. I: Reg. Papers, vol. 56, no. 1, pp. 210223, Jan. 2009.
[31] T. J. Overbye, A power ow measure for unsolvable cases, IEEE
Trans. Power Syst., vol. 9, no. 3, pp. 13591365, Aug. 1994.
[32] T. J. Overbye, Computation of a practical method to restore power
ow solvability, IEEE Trans. Power Syst., vol. 10, no. 1, pp. 280287,
Feb. 1995.
[33] L. V. Barboza, A. A. P. Lerm, and R. Salgado, Unsolvable power
ow-restoring solutions of the electric power network equations, in
Proc. IEEE Int. Symp. Circuits and Systems (ISCAS), May 2326, 2005,
vol. 5, pp. 52945297.
[34] Z. Hu and X. Wang, Efcient computation of maximum loading point
by load ow method with optimal multiplier, IEEE Trans. Power
Syst., vol. 23, no. 2, pp. 804806, May 2008.
[35] S. Iwamoto and Y. Tamura, A load ow method for ill-conditioned
power systems, IEEE Trans. Power App. Syst., vol. PAS-93, no. 4,
pp. 17361743, Apr. 1981.
[36] A. Gmez-Expsito and C. Gmez-Quiles, Factorized load ow,
IEEE Trans. Power Syst., vol. 28, no. 4, pp. 46074614, Nov. 2013.
[37] A. Gmez-Expsito, C. Gomez-Quiles, and A. de la Villa Jan, Bilinear power system state estimation, IEEE Trans. Power Syst., vol.
27, no. 1, pp. 493501, Feb. 2012.
[38] C. Gmez-Quiles, A. de la Villa Jan, H. Gil, and A. Gmez-Expsito,
Equality-constrained bilinear state estimation, IEEE Trans. Power
Syst., vol. 28, no. 2, pp. 902910, May 2013.
[39] A. Gmez-Expsito, Factored solution of nonlinear equation systems, Proc. Roy. Soc.A, 2014, 470, 20140236.
[40] A. Gmez-Expsito, C. Gmez-Quiles, and W. Vargas, Factored solution of infeasible load ow cases, in Proc. Power Systems Computation Conf. (PSCC), Wroclaw, Poland, Aug. 2014.
[41] Y. Sekine, Y. , and A. Yokoyama, Multisolutions for load ow
problem of power system and their physical stability, in Proc. 7th
Power Systems Computation Conf. (PSCC), 1981, pp. 819826.
[42] A. Gmez-Expsito and E. R. Ramos, Reliable load ow technique
for radial distribution networks, IEEE Trans. Power Syst., vol. 14, no.
3, pp. 10631069, Aug. 1999.
[43] R. D. Zimmerman, C. E. Murillo-Snchez, and D. Gan, MatPower [Online]. Available: http://www.pserc.cornell.edu/matpower/

Catalina Gmez-Quiles (M'12) received the engineering degree from the University of Seville, Spain, and the M.Sc. degree from McGill University, Montreal, Canada, both in Electrical Engineering. In 2012 she got the PhD degree
from the University of Seville. Her research interests relate to mathematical and
computer models for power system analysis.

Antonio Gmez-Expsito (F'05) received the electrical engineering and doctor

degrees from the University of Seville, Spain, where he is currently the Endesa
Red Industrial Chair Professor. His primary areas of interest are optimal power
system operation, state estimation, digital signal processing and control of exible ac transmission system devices.

Walter Vargas received the electrical engineering degree, with Power specialty
in 2007 from the Escuela Superior Politcnica del Litoral, Ecuador, and the
M.Eng degree in Electric Energy Systems in 2013 from the University of
Seville, Spain. He is currently part of the Department of Electrical Studies at
the Operation Center of CELEC EP TRANSELECTRIC.