Advanced Contingency Selection Methodology

George K. Stefopoulos, Fang Yang, George J. Cokkinides and A. P. Sakis Meliopoulos

subnetwork solutions which solve for a subset of state vector methods. A typical application of subnetwork solutions to contingency selection is the bounding method [3], [4]. The subnetwork solution method had been also generalized with the zero mismatch approach [9], which is an iterative AC power flow solution method and is effective for both contingency screening and analysis. Typical contingency selection methods consist of either ranking methods using a performance index (PI) or screening methods based on approximate power flow solutions. It is widely recognized that PI based methods are efficient but vulnerable to misrankings, while screening methods are more accurate but inefficient. It has been also identified that the inaccuracies of the P1 based methods are mainly due to (a) nonlinearities of the system model [10], and (b) discontinuities of the system model arising from generator reactive power limits and regulator tap limits [11], [12]. A hybrid contingency selection method had been proposed in the past which takes advantage of the best features of the two approaches [11]. This method utilizes a procedure based on the concept of contingency stiffness index to identify 'nonlinear' contingencies and a performance index method to identify contingencies causing discontinuities. While the hybrid method performs very well, the computational burden is many times greater than pure PI based methods. This paper describes the development and implementation of contingency ranking and selection algorithms as part of a power system analysis program. The implementation is based on PI-based methods and on the use of the quadratized power flow system model, which is described in the Appendix. Each contingency is modeled with the introduction of a contingency or outage control variable. The ranking is based on the value of the sensitivity of the performance index with respect to the contingency control variable for each contingency. The computation of the sensitivities is performed using the very efficient co-state method. Moreover, the paper introduces some new concepts as a way for improving the accuracy of the P1-based contingency ranking methods. The basic concept is the use of state-linearization rather than performance-indexlinearization with respect to the outage variable and it appears to provide more accurate results in contingency ranking and selection, at the expense of slightly increased computational time. However, with the appropriate use of sparsity techniques this increase can become minimal. The methodology implementation is demonstrated with two simple systems. Numerical experiments are proposed to further investigate and quantify its performance in large scale systems. The Plcontingency selection is implemented as part of a power
Abstract- This paper describes the development and implementation of contingency ranking and selection algorithms, as part of a power system security assessment program. The work concentrates on performance-index-based algorithms and uses a contingency control variable for precise contingency representation. The ranking is based on the value of the sensitivity of the performance index with respect to the contingency control variable for each outage. The computation of the sensitivities is performed using the very efficient co-state method. Furthermore an approach for improving the accuracy of performance-index-based contingency ranking methods is introduced. This approach is based on state rather than performance index linearization with respect to the contingency variable and it provides more accurate results in contingency ranking and selection. The effectiveness of the proposed method in identifying critical contingencies is illustrated using some small test systems. The ultimate goal is to achieve fast and accurate contingency selection, without having to solve the full load-flow problem for each contingency (as is the current utility practice). Index Terms- Contingency Analysis, Contingency Ranking, Performance Index.

variables, using sparse

I. INTRODUCTION

ONE of the main computational issues in power system steady state security assessment as well as in reliability studies is contingency ranking and selection. Specifically, the critical contingencies from all possible contingencies need to be identified and analyzed. For large scale systems, the process imposes a substantial computational burden. For this reason, there have been consistent efforts to invent fast contingency selection algorithms and subsequent contingency analysis. Significant developments in the area of contingency selection include (a) contingency ranking with performance indices (PI) [1], (b) local solutions based on concentric relaxation [2], (c) bounding methods [3,4], etc. In addition, significant developments for contingency analysis include the introduction of the fast decoupled power flow [5], sparsityoriented compensation method [6], sparse vector methods [7], partial refactorization [8], etc. Much attention on this issue has been focused on
This work was supported by the Power System Engineering Research Center (PSERC). George K. Stefopoulos is with the Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: gstefopg ece.gatech.edu). Fang Yang is with the Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: gtg65 lj mail.gatech.edu). George J. Cokkinides is with the Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: george.cokkinides@ ece.gatech.edu). Sakis Meliopoulos is with the Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: sakis.meliopoulos@ ece.gatech.edu).

0-7803-9255-8/05/$20.00 2005 IEEE

67

system analysis program and it is used in combination with compensation-based contingency analysis method for steady state security assessment, and as a pre-filter for contingency ranking in power system reliability analysis.

Jc =YEw X(-

~~'CIN

(2)

where

II. PI METHODS FOR CONTINGENCY RANKING PI-based contingency ranking methods reported in the literature are based on the evaluation of the PI gradient with respect to an outage. In [12] a rigorous definition of an outage has been proposed with the use of the outage (or contingency) control variable Uc that has the following property [12]:
1.0, if the component is in operation 10.0, if the component is outaged (1)

wj

weighting factor, 0 < wj <I,

I.
n

IN,j

: current-based thermal limit of the line, : magnitude of actual current through circuit j, positive integer parameter defining the exponent.

* Circuit power-based index:

(3)
where

The outage control variable is used in the component modeling of a power system, as illustrated in Fig I and 2.
BUS k
u. (gkm
+

wj

weighting factor, O < Wj < I,

jbkm)

BUS m

SN,J
n

Si

: power-based thermal limit of the line, : apparent power through circuit ], positive integer parameter defining the exponent.

jbkms U

Fig. 1. Circuit outage control variable

Uc.
k
YlkmUc

Fig. 3. Representation of common mode outages with control variable

F. G ut
oPower
System

Uc.

* Fig. 2. Generating unit outage control variable

Voltage index:
=

Uc~

Jv

Note that in the case of a generator outage the outage control variable not only affects the power produced by the outaged unit, but also indicates the re-dispatch of its produced power to the remaining units, via a re-dispatch algorithm (Fig 2 illustrates a linear re-dispatch, based on participation factors). The contingency control variable can not only model independent contingencies, like the ones indicated above, but can be also effectively used to model common mode contingencies, i.e., contingencies that are dependent upon each other. Such a situation is illustrated in Fig 3.

)
0

(4)
<

where
Wk
:

weighting factor,

<

I,

vk mezm nominal bus voltage value (typically I p.u.),

(it is in general the i.e. I (Vma + 7i.
mean

value in the desired

range,

Vk*,,p :voltage deviation tolerance (i.e. (Vkmax - Vknin )),

I

Vk

: : :

Depending on the purpose of contingency selection, variety of performance indices can be applied to it: * Circuit current-based index:

actual voltage magnitude at bus k, positive integer parameter defining the exponent,

total number of PQ buses.

68

* Generation reactive power index:

JQ

T -T

QQmean Jmea j

~~~~~~~~5)J(x, U() ag(x, U,) =Ax x

(10)

where
Wj

weighting factor, 0 < w1 < 1,

:

Q.i,mean

expected generated reactive power value, This is the mean value is the allowable range for each generator, i.e., I (Qmax + Qmin )
reactive power deviation tolerance,

Qj?,ep
Qj
n

flow equations, with the outage control variable incorporated in them, uC is the vector of the control variables of interest, x is the state vector, xT is the co-state vector. Note that the performance index may depend explicitly on the control variables and it also depends implicitly on them through the power flow equations. The explicit dependence is captured by the partial derivative ai and the implicit by the
term

J(x,u,) is the performance index, g(x,u,) = 0 are the power

This is half of the allowable range,

I(Qrmax _ Qrmini)

i.e.,

rT. ag(x, '0 . Furthermore,

auc

auc

the co-state vector is

:actual reactive power generated by unit j, : positive integer parameter defining the exponent, : total number of generating units.

Several other performance indices can be defined. In this paper the circuit current-based index and the voltage index are mainly considered. By including the outage control variable in the system modeling the defined performance index J can be expressed as a function of the state vector and of the control variables,J(x,uU). The change of the performance index due to the contingency is:
AJ =J(x

',u. =0.0)-J(x0,UL

= 1.0),

(6)

where x is the initial state, prior to the contingency, xnew is the state after the contingency and the control variable u. changes from 1.0 to 0.0, modeling the component outage. The first order approximation of the performance index variation is provided by the derivative of the PI with respect to the control variable:
dl dil7 AJ = - Au =-.(U-l) duc duc

(7)

and for a change in

uc from 1.0 to 0.0:

tJ =(8) A^r=_ duc

It is therefore expected that the derivative of the performance index with respect to the outage control variable at the present operating point will provide a measure of the severity of a disturbance. Therefore, contingencies are ranked based on the values of dJ , which expresses the first order change of the

invariant for all contingencies, therefore it is pre-computed at the present operating condition, resulting in extremely fast computations, even for large scale power systems. After the co-state vector is computed the sensitivity of the performance index for each contingency is simply a vector-vector multiplication, with one of the vectors being very sparse. The contingency ranking algorithm based on the co-state method is efficient and precise as a first order method. As a matter of fact the computational requirements are equivalent to one iteration of the power flow algorithm for the entire set of contingencies (cost of computing the co-state vector). However, for contingencies that trigger severe nonlinearities, the method may lead to misrankings. This is because the behavior of the performance index around the present operating point may be significantly different from the behavior as we move away from the current operating point. This issue has been addressed with the hybrid method which separates contingencies into those that trigger sever nonlinearities and those that do not. The former are processed with more accurate and computationally demanding contingency selection methods and the latter (which represent the majority of contingencies) are ranked with the above described PI based method. The processing of a small number of contingencies via selection methods adds to the computational burden. Therefore it is important to be able to use PI based methods on all contingencies. The following proposed method in this paper provides a promising approach towards this goal.
CONTINGENCY RANKING The proposed state linearization approach is a variation of the PI-based contingency ranking algorithm. In this method, instead of linearizing the performance indices directly, the system states are linearized with respect to the contingency control variable; the performance index J is then calculated as follows:
III. HIGHER ORDER STATE-LINEARIZATION PI-BASED

performance index. The values of these derivatives can be calculated using the co-state method: (9) aJ AT * ag(X,uC) 9 ddl _ =u-X Ou c duc auc

duc

J = J(x +-(u

duc

1)U )c

(1 1)

69

where x : present operating condition, X : system state vector, u : contingency control variable. The utilization of the linearized system states in calculating the system performance index provides higher order terms in Taylor's series. The unique potential of this method has been proven in some preliminary work by the authors described in [13]. The state-linearization sensitivity method provides the traces of indices with curvature, which can follow the highly nonlinear variations of the original indices to some extent, while the PI-linearization method provides only the straight line. Therefore, the higher order sensitivity method is superior to the simple PI-linearization-based method. The contingency selection is based on the computation of the performance index change due to a contingency and subsequent ranking of the contingencies on the basis of the change. Mathematically one can view the outage of a circuit as a reduction of the admittance of the circuit to zero. We use again the outage control variable, uc, as illustrated in Figures 1 through 3. Consider the performance index, J. The change of the performance index due to the contingency is:
AJ = Jrx
+

control variable. The higher order analysis curve is the state linearization curve with respect to the contingency control variable.
1.5
1.4
~~I

l-l

Plot of Jc vS u
~~I
I

,-r

--

1.2-R-t----|------I-------r-------------, ctua ciurw te,

f

1.2 -----

i ugI el !dlrl-l ljl14

- - - --------

- - -4 - - -

9-

- - -

0.
a)

'ana

1sc

-F

AI

,_
__

__

__

I_
0.7

_
0.8

anlrtbre ralysis
0

cru_____________
0.4 0.5 0.6 Control Variable u

0.1

0.2

0.3

0.9-

Fig. 4. Plots of circuit-loading index vs. the contingency control variable uc.
Plot ofJvVs u

du,

duic

1),Uc

- J(X , uc =

1.0)

(12)

where x is the present operating condition. The sensitivity of the state with respect to the control variable can be easily computed as:

2I

dx
duc

aG(x, u) -aaG(x, u)
Ax
.

(13)

a)
C:

4)1

Note that aG(x, u) is the Jacobian of the system and therefore

a.

aCI I
!

it is pre-computed at the present operating condition and remains invariant for all contingencies. Thus for each contingency we have to only compute the partial derivatives of the power flow equation G(x,u) with respect to the contingency control variable. This vector has only few nonzero entries and therefore the computations are extremely fast. Taking into account the sparsity of this vector can greatly improve the efficiency of the method. It should also be noted that dx is a vector of the same size as the state vector each

ax

0.4 0.5 0.6 Control Varable u

Fig. 5. Plots of voltage index vs. the contingency control variable uc.

element of which is the derivative of the corresponding state with respect to the control variable. Once the new state is computed via this linear approximation, the calculation of the new value of the performance index is a straightforward operation. The concept of the approach is illustrated graphically in Figures 4 and 5 based on results obtained from the application of the method to a test system. The first order analysis curve represents the PI-sensitivity linear curve after performing the linearization of the index with respect to the contingency 70

duc

IV. METHODOLOGY IMPLEMENTATION

The proposed methodology is implemented in a Visual C++ environment using object oriented techniques. It is created as an expansion of existing power system analysis software, developed by the power systems laboratory of Georgia Tech. Both the PI-linearization and the state-linearization approaches are integrated in the same environment, and the user can pick the one to chose. An additional option of calculating the sensitivity of the PI wrt to the outage control variable numerically, using a secant approximation is also available and mainly used for debugging purposes. The methods and data needed to perform the PI-based contingency

ranking are implemented in a separate class, thus facilitating the module implementation and its re-usability from other parts of the program (like contingency analysis, security assessment, reliability analysis, etc). The linearization of the performance indices under consideration are performed by objects of a different class, the Linearization class, making the program more generic and easily expandable. A graphical user interface has been developed to facilitate the input of data. The user defines the performance index to be used, the contingencies to be considered and the desired solution approach (i.e. first order analysis with Pllinearization, higher-order analysis with state-linearization and if numerical check with secant approximation will be performed) along with the required solution parameters. The results provided are the estimation of the performance index change, AJ, based on which the ranking is performed. The results are presented to the user in a numerical form. The user interface of the program is shown in Fig 6.
n
Close
__

using the PI-linearization approach, the state-linearization approach and the full load flow solution, for the current-based circuit loading index. Table II presents similar results for the voltage index.

Fig. 7. Test system used for contingency ranking evaluation.

RANKING RESULTS FOR CURRENT-BASED CIRCUIT LOADING INDEX FOR FOURBUS SYSTEM.

TABLE I

TW CX BCsed Cxcu ktdex Loading oogs" index2 Acietiv Power Ba.d Ciouk Loading Index Cwwation Reactive Power hidex

Exponent n

iSme Device
Contmml 3 (spao) Comnat 4 (Spamn

AM Deyvt

Outaged Line 10-30 20-40 30-40

Ranking

Actual 2
3

P1-linearization Ranking
2 4

State-linearization
Ranking

2
3

RANKING RESULTS FOR VOLTAGE INDEX FOR FOUR-BUS SYSTEM.

TABLE II

CktIl0 ak.e CVdw naf go

I*, ghw O.drMUelyxe

Delta u: 'O Q,
Fig. 6. Main user interface for probabilistic power flow analysis.

>

Outaged Line 10-30 20-40 10-20 30-40

Actual Ranking 1 2 3 4

PI-linearization Ranking 4 1 2 3

State-linearization Ranking 1 2 3 4

V. NUMERICAL EXAMPLES The method has been applied to two small power systems, a four-bus system and the IEEE 24-bus reliability test system (RTS). The developed power system analysis software is used to perform all the computations, including the base-case power flow analysis. The quadratic power flow (QPF) model is used in both cases. The proposed state-linearization method is compared to the PI-linearization contingency selection algorithm. The cases and the system loading is chosen so that the traditional index-linearization approach results in some misrankings. A. Four-bus system The four-bus test system is depicted in Fig. 7. It consists of two constant power loads, two generators and four transmission lines represented with pi-equivalent circuits. Table I presents comparative results from contingency ranking

The state-linearization approach shows considerably improved performances and provides the correct ranking in both cases. The traditional PI-sensitivity approach results in one misranking in each case, especially in the voltage index, where it ranks the most severe contingency as least severe (the relative ranking of the rest of the contingencies is correct). This is related to the nature of the index and the system. Doing a homotopy continuation plot of the value of the performance index for values of the outage variable uC from one to zero we can see that the value of the index is reducing when we are not far away from the base-case operating point (where the sensitivity if computed) and then starts increasing rapidly as uC approaches zero. This effect cannot be captured by the first order sensitivity, it is, though, partially captured by the higher order state-linearization. It should also be noted that there is no load flow solution for an outage of line 10-30, and this is the reason why this contingency is ranked as the most severe in the actual ranking.

71

B. IEEE 24-bus reliability test system (RTS) The IEEE 24-bus test system is shown in Fig 8. System data can be found in [14].

TABLE III RANKING RESULTS FOR VOLTAGE INDEX FOR THE TOP 10 CONTINGENCIES OF THE IEEE 24-BUS SYSTEM

Outaged Branch 60-100 C 20-60 10-110 T 150-240 50-100 100-120 T 80-100 10-50 240-30 110-140 30-90

Actual Ranking 1 2 3 4 5 6 7 8 9 10 11

PI-linearization Ranking 39 1 2
19 5 3 4 6 31 7 8

State-linearization Ranking 1 2 3 8 6 4 5 7 32 9 10

T: Transformer branch, C: Underground cable

Fig. 8. The IEEE 24-bus reliability test system.

The eleven most severe contingencies in terms of voltage deviation (as identified by full power flow solution of the system) are presented in Table III, along with the ranking obtained by the PI-linearization and the state-linearization approach. The traditional PI-linearization approach results in several misrakings. Though most of them are not significant, there are three very important misrankings, with severe contingencies belonging to the top-ten list being ranked as much less severe. The most severe contingency is among them and it is in fact identified as the least severe. The reason for that is again similar, as in the four-bus system and is related to the nature of the performance index, as explained in the previous paragraph. Most of the other severe misrankings are, nevertheless, identified as being in the top-ten list. The statelinearization approach performs much better and all the top contingencies are identified. There is only one severe misranking, with one contingency being ranked significantly lower than it actually is.

VI. CONCLUSION This paper presents the basic concepts and an implementation of a computationally efficient performanceindex-based contingency ranking methodology along with example results. The presented methodology utilizes multiple performance indices and it can capture various effects on the system. Furthermore, the paper introduces some basic concepts for improving the accuracy of the P1 contingency selection methods. The basic idea used is linearization of state rather then direct linearization of the PI with respect to the contingency variable. Preliminary results of this approach are presented and they indicate satisfactory and promising behavior. It should be noted, however, that since both ranking approaches are based on linearization around a base-case operating point, and because of the highly non-linear nature of the power systems, some misrankings are always unavoidable in such approaches. A contingency may result in significant, non-linear deviation from the base case which may not be appropriately captured by an approach based on linearization. The contingency ranking methodology is implemented as a module of a general power system analysis software. The module is tested with some small size test-systems with satisfactory results in terms of accuracy and efficiency. Further improvements in computational issues will allow the adequate and efficient application of the methodology for large scale systems. The module is to be used as contingency selection pre-filter in the implementation static security assessment methodologies, where combined with compensation-method-based algorithms, will be part of security assessment software. In addition it is also used in reliability studies for contingency selection [15].
APPENDIX: QUADRATIC POWER FLOW[13,161

The quadratic power flow (QPF) model is based on modeling any power system component as a set of equations of order no greater than 2 (quadratic). This can be always achieved with the introduction of additional state variables without any simplifications. Application of connectivity constraints (Kirchoff s current law) yields the quadratized power flow equations:

72

G(x, u() = A[X, u,

where

Ix, u'. k?i [x, u ]I+ b

O,

(A. 1)

[5]
[6]

: the state vector, in Cartesian (rectangular) form, x Uc : the vector of control variables, A ,Q : constant matrices, b : constant vector. The solution to the quadratic equations is obtained using Newton's method:

[7]

[8]
[9]

xk1
k

=k

J(xk ) 'G(xk)

(A.2)
[10]

where

j(xk ) the Jacobian matrix at iteration k The Iterative procedure terminates when the norm of the QPF equations is less than a certain tolerance. Therefore the QPF equations G(x, uc) = 0 comprise a different mathematical system of nonlinear algebraic equations from the traditional power flow (TPF) equations. The state vector consists of the real and imaginary part of the voltage at each bus and of additional internal state variables for each device. The system G(x, u,) = 0 consists of the current balance equations at each bus, plus additional internal equations for each one of the nonlinear devices that exist in the system. All the equations are of order at most quadratic. The application of the co-state method in the QPF fornulation is practically identical to the TPF formulation. Assuming the performance index:
J = F(X,u')
I

: the step of iterations

[11]

[12] [13]

[14] [15]

[16]

(A.3)

the derivative of J with respect to Uc is given by:

dJ dUt,
dF dui

[17]

aF(X, u,) auc

'TOaG(X,uc) auc
)

(A.4)

[18]

B Stott and 0. Alsac, "Fast Decoupled Power Flow," IEEE Trans. on Power Apparatus and Systems, vol. PAS-93, No. 3, pp. 859-869, May/June 1974. 0. Alsac, B. Stott and M. F. Tinney, "Sparsity-oriented Compensation Mathods for Modified Network Solutions," IEEE Trans. on Power Apparatus and Systems, vol. PAS-102, No. 5, pp. 1050-1060, May 1983. W. F. Tinney, V. Brandwajn and S. M. Chan, "Sparse vector method," IEEE Trans. on Power Apparatus and Svstems, vol. PAS-104, No. 2, pp. 295-301, Feb. 1985. S. M. Chan and V. Brandwajn, "Partial matrix refactorization," IEEE Trans. on Power Systems, vol. PWRS-l, No. 1, pp. 193-200, Feb. 1986. R. Bacher and W. F. Tinney, "Faster local power solutions: The zero mismatch approach," IEEE Trans. on Power Systems, vol. PWRS-4, No. 4,pp. 1345-1354,Feb. 1989. G. Irisarri, A. M. Sasson and D. Levner, "Automatic Contingency Selection for On-Line Security Analysis - Real Time Tests," IEEE Trans. on Power Apparatus and Systems, vol. PAS-98, No. 5, pp. 15521559, Sept./Oct. 1979. A. P. Sakis Meliopoulos and C. Cheng, "A hybrid contingency selection method," in Proceedings of the 10th Power System Computation Conference, Graz, Austria, Aug. 1990, pp. 605-612. A. P. Sakis Meliopoulos, Carol S. Cheng, Feng Xia, "Performance Evaluation of Static Security Analysis Methods," IEEE Trans. on Power Systems, vol. 9, No. 3, pp. 1441-1449, Aug. 1994. Sun Wook Kang, A. P. Meliopoulos, "Contingency Selection via Quadratized Power Flow Sensitivity Analysis," in Proceedings of the IEEE 2002 Power Engineering Society Summer Meeting, vol.3, pp.1494-1499. IEEE Committee Report, "IEEE Reliability Test System," IEEE Trans. Power Apparatus and Systems, vol. PAS-98, No. 6, pp. 2047-2054, Nov./Dec. 1979. Fang Yang, A. P. Sakis Meliopoulos, George J. Cokkinides and George Stefopoulos, "A Bulk Power System Reliability Assessment Methodology," in Proceedings of the 8th Int. Conf on Probabilistic Methods Applied to Power Svstems, Ames, IA, Sept. 12-16, 2004, pp. 44-49. S. Kang and A. P. Meliopoulos, "Analytical approach for the evaluation of Actual Transfer Capability in a deregulated environment," in Proceedings of the 32nd Annual North American Power Symposium, 1999. A. P. Meliopoulos, George Contaxis, R. R. Kovacs, N. D. Reppen, N. Balu, "Power system remedial Action Methodology," IEEE Trans. Power System, vol. 3, No. 2, pp. 500-509, 1988. George Stefopoulos, Fang Yang and A. P. Sakis Meliopoulos, "An Improved Contingency Ranking Method," in Proceedings of the 35"h Annual North American Power Symposium, Oct. 20-21, 2003, Rolla,

where

MO.

XT aF(X,u,) a(G(x,c ) -' is the co-state vector of the set of

ax

ax

QPF equations and aG(X, uC) is the Jacobian matrix of the set ax of QPF equations.
REFERENCES
[1] G.C. Ejebe and B. F. Wollenberg, "Automatic Contingency Selection," IEEE Trans. on Power Apparatus and Systems, vol. PAS-98, No. 1, pp. 92-104, Jan./Feb. 1979. [2] J. Zaborszky, F.W. Whang and Prasad, "Fast contingency evaluation using concentric relaxation," IEEE Trans. on Power Appa-atus and Systems, vol. PAS-99, No. 1, pp. 28-36, Jan./Feb. 1980. [3] V. Brandwajn, "Efficient bounding method for linear contingency analysis," IEEE Trans. on Power Svstems, vol. PWRS-3, No. 1, pp. 3843,Feb. 1988. [4] V. Brandwajn and M. G. Lauby, "Complete bounding for ac contingency analysis," IEEE Trans. on Power Systems, vol. PWRS-4, No. 2, pp. 724-729, May 1990.

73