1 Power system small signal stability analysis using genetic optimization techniques Zhao Yang Dong Department of Electrical Engineering Yuri V. Makarov David J. Hill The University of Sydney NSW 2006, Australia Abstract Power system small signal stability analysis aims to explore dierent small signal stability conditions and controls, namely, 1) exploring the power system security domains and boundaries in the space of power system parameters of interest, including load ow feasibility, saddle node and Hopf bifurcation ones, 2) nding the maximum and minimum damping conditions, and 3) determining control actions to provide and increase small signal stability. These problems are presented in the paper as dierent modications of a general optimization problem, and each of them has multiple minima and maxima. The usual optimization procedures converge to a minimum/maximum depending on the initial guesses of variables and numerical methods used. In the considered problems, all the extreme points are of interest. Additionally, there are diculties with nding the derivatives of the objective functions with respect to parameters. Numerical computations of derivatives in traditional optimization procedures are time consuming. In the paper, we propose a new black box genetic optimization technique for comprehensive small signal stability analysis, which can eectively cope with highly nonlinear objective functions with multiple minima and maxima and derivatives which can not be expressed analytically. The optimization result then can be used to provide such important informations as system optimal control decision making, assessment of the maximum network's transmission capacity, etc. Keywords: Genetic Algorithms; Power System Security; Bifurcations; Stability 1. Introduction In the open access environment, the power utilities sometimes are forced to work far away from their predesigned conditions. In this situation, it is necessary to re-approach the problems related to power system security, stability and transfer capability [1]. On the other hand, several recent major system blackouts in dierent countries and voltage collapses require additional attempts in power system stability area. Several power system oscillatory instability problems occurred recently again put forward attention in the research area of small signal stability. In the space of power system parameters, the small signal stability domain is restricted by complicated surfaces of dierent kinds. These surfaces can be load ow feasibility, aperiodic and oscillatory boundaries. The last two are often referred to as Saddle node and Hopf bifurcation boundaries. One of the most important tasks is to obtain the system security measure or, in other words, the adequate stability margin. There are many denitions for the stability margin [2] -[11]. But the denition based on the distances from the system's state point to the stability domain boundary in the space of power system controlled parameters seems to be more appropriate [12], [13]. Due to complexity of the boundary, it is quite dicult to nd out the critical shortest distance which is used as stability margin. Additionally, their are normally several subcritical distances which are close to the critical one. So the security margin must be assessed in several critical and subcritical directions. One of the approaches developed recently is the analytical approach [12], [14] -[17]. Corresponding to the general method proposed in [18] the optimization problem can be depicted by Fig. 1. Then in each direction dened by input y, Input Φ’ ∆y 2. Small Signal Stability Analytical Approaches In certain practical cases, it is necessary to analysis the maximum transfer capability in a certain loading direction. For example, the problem can consist in assessment of the maximum power transfer from a particular generator to a particular load. To locate the saddle node and Hopf bifurcations as well as the load ow feasibility boundary points in a given loading direction within one procedure, the following constrained optimization problem is proposed, see [18], ) max=min (1) 2 subject to f (x; y0 + y) t ~ J (x; y0 + y)l0 ? l0 + !l00 J~t (x; y0 + y)l00 ? l00 ? !l0 li0 ? 1 = = = = li00 = 0 0 0 0 0 (2) (3) (4) (5) (6) where is the real part of an eigenvalue of interest, y0 is the current operation point, and y denes the distance and direction to security boundaries from y0 , f (x; y0 + y) = 0 represents the load ow condition, and J~ is the system Jacobian with l = l0 + jl00 as its eigenvector corresponding to the eigenvalue = + j!. To consider the load ow constraint, this constrained optimization problem can be represented with Lagrange function form as = 2 + f t (x; y0 + y) ) minx;; (7) = 2 + 0 ) minx;; (8) The problem has many solutions including the load ow feasibility, saddle node and Hopf bifurcations and minimum and maximum damping points [18]. The result depends on the initial guesses of variables and selected eigenvalue. The goal is to determine the small signal stability boundary point which is closest to the point y0 . To get this point, it is necessary to solve the problem for dierent eigenvalues and initial values of the loading parameter , which is a very time consuming task. Φ State Matrix x Output f Eigenvalues τ η State Variables Optimization procedure Rotation Fig. 1. System Model Diagram for Small Signal Stability General Method Optimization with considering of the state variables shown in Fig. 1, the load ow constraint is computed rst, while the state variables, which are to be optimized, provide input value to set up the state matrix. The eigenvalue computation is then based on this matrix. In the algorithm, the real part of the critical eigenvalue(s) will be used to form the objective function . The value of the objective function f , as the output, will then provide information used to proceed the optimization procedure and the state variables are to be adjusted accordingly. When the optimization process converged, all characteristic points along the direction dened by y can be located. To located all these points in the whole plane, a loop is needed to rotate the direction and repeat the optimization procedure locating all characteristic points in the whole plane/space of interest. Generally, this plane/space is a cut set of the space spanned by all parameters of interest [19]. In case of exploring the shortest distances, the problem becomes even more complicated. In principle, it is possible to use the problem represented by equations (1) - (6) to get the critical distance vector. Ideas similar to those put forward in [12] can be tried. For example, after obtaining the closest instability point along the given direction y, it is possible to analyze the angle between the loading and left eigenvector of the corresponding matrix [load ow Jacobian or state matrix] in this point. As this vector shows the normal direction with respect to the stability boundary, the angle computed can be used to rotate y in the direction where the distance decrease. A more ecient way consists in the direct computation of the closest instability point. To get this point, the following modication of the general optimization problem (1) -(6) can be used, jjy ? y0 jj2 ) min (9) subject to f (x; y) = 0 (10) t 0 0 00 J~ (x; y)l ? l + !l = 0 (11) t 00 00 0 ~ J (x; y)l ? l ? !l = 0 (12) 0l ? 1 = 0 (13) i li00 = 0 (14) Once again, the result depends upon the initial guesses of variables and selected eigenvalues. The choice of these parameters is a complicated task. One can use some practical ideas regarding the most dangerous loading direction and critical eigenvalues [for example, corresponding to the inter-area oscillatory modes.]. Nevertheless, it looks quite dicult to get all of the critical and subcritical distances by this approach. One of the most confusing diculties in all above mentioned procedures is computing the small signal characteristic points in view of breaks in objective function and constraints. For example, to take account of the reactive power limits, for generators reactive power limits, we must use dierent models in the constraint set. A sudden change in the model causes sever problems for the optimization procedures. In fact, it can lead to an instant instability [12], [20]. To nalize, we should conclude that the analytical approaches to the small signal stability analysis have many problems which can be hardly solved by traditional optimization techniques. That is why we are attempting to apply the genetic optimization procedures. 3. Genetic Algorithms with Sharing Function Optimization Method Genetic Algorithms(GAs) [21] are heuristic probabilistic optimization techniques inspired by natural evolution process. In genetic algorithms, the tness function is used instead of objective function as in the traditional optimization procedures. Each concrete value of variables to be optimized is called as an individual. Then a certain number of individuals composes the generation. In the process of GAs, individuals with better tness survive and those with lower tness die o, so to nally locate individual with the best tness as the nal solution. They are capable of locating the global optimum of a tness function in a bounded search domain, provided a sucient population size is given. The GA sharing function method is able to locate the multiple local maxima as well. Genetic algorithms have been already eectively applied in complicated multidimensional optimization problems, which can be hardly solved by traditional optimization methods. Genetic algorithms, which mimic the natural evolution, usually contain the following steps. Firstly, produce the initial population; secondly, evaluate tnesses for all individuals in the current population; thirdly, perform such operations as crossover, reproduction and mutation depending on the existing generation tnesses, and form a new generation. Then the procedure is repeated till some termination criterion is met and the optimum is thus obtained [21]. 3.1. Sharing Function Method To compute multiple maxima of the tness function, the genetic algorithm sharing function method, [21], [22], can be applied. The method decreases the tnesses for similar individuals by the \niche count", m0 (i). For each individual i, the \niche count" is computed as a sum of sharing function values between the individual and all individuals j in generation, see Eqn. (16). The similarity of individuals is evaluated by the distance, d(i; j ) from each other, see Eqn. (15). The resulting shared tness 0 is changed through dividing the original tness by the corresponding niche count, see Eqn. (17), [21] -[23]. d(i; j ) = d(xi ; xj ) n X sh[d(i; j )] m0 (i) = j =1 i) 0 (i) = Pn ( sh j =1 [d(i; j )] (15) (16) (17) The sharing function is dened so that it fullls, 8 < sh(d) = : 0 sh(d) 1 sh(0) = 1 limd!1 sh(d) = 0 (18) For example the sharing function can have the form, sh(d) = 1 ? ( d ) ; if d < 0; otherwise (19) where is a constant, and is the given sharing factor. By doing so, an individual receives its full tness value if it is the only one in its own niche, otherwise its shared tness decreases due to the number and closeness of the neighboring individuals. In this paper we apply the genetic algorithm with sharing to small signal stability analysis. This optimization problem is highly non-linear and some times, even non-dierential-able, which makes it very dicult to be solved by traditional optimization methods. 3.2. Black Box System Model For the analytical approaches, the optimization problems is highly non-linear and some times, even non-dierential-able. It is known that the traditional optimization methods meet serious diculties with convergence while solving such problems. Besides the rest, the constraint sets in Eqn. (2) -(6) take account of only one eigenvalue during the optimization. To get the stability margin for all eigenvalues of interest, as well as the critical load ow feasibility conditions, it is necessary to vary the initial guesses and repeat the optimization. Additionally, the functions in Eqn. (2) -(6) can have breaks due to dierent limitations applied to power system parameters. For example, the generator current limiters may cause sudden changes in the model, and consequently breaks in the constraint functions. This makes the analytical optimization problem even more complicated. In the genetic optimization procedures, those diculties can be overcome by using the black box power system model given in Fig. 2, described in the sequel. Unlike the model used in the analytical form of small Black Box System Model γ Load flow calculation converged not converged State variable State matrix Eigenvalue calculation formation calculation Fitness Function Φ α=0 Fig. 2. Black Box System Model for Optimization signal stability problems, this black box has control parameters as inputs, and the tness function as outputs. Inside the black box, the load ow is computed rst. If it converges, then the state variables and matrix are computed, and then the eigenvalues of the state matrix are obtained. Thereafter, the critical eigenvalue is chosen for analysis. The critical eigenvalue's real part is used to compute a particular value of the tness function. If the load ow does not not converge, which means that a load ow solution does not exist, we put the critical eigenvalue real part to zero. By such a way, the load ow feasibility points are treated in the same way as the saddle node and Hopf bifurcation points. The tness function can be changed quite exibly depending on the concrete task to be solved, see the next section for explanation. To demonstrate the advantages of the black box model, let's consider the tasks in Abstract. To reveal all characteristic small signal stability points, such as maximum loadability, saddle and Hopf bifurcation etc. along a given ray y0 + y in the space of power system control parameters, y, the general small stability problem Eqn. (1) -(6) can be used. If the above problem is solved by traditional optimization methods, the solution obtained depends on initial selection of the eigenvalue traced, and variables x. Moreover, even for one eigenvalue selected, it is not possible to get all the characteristic points in one optimization procedure. By applying the black box model and GA techniques all the problem characteristic points can be found within one optimization procedure. In this case, the input is the loading parameter, = , and the tness function is 2 for maximization and 12 for maximization. To compute the function, the load ow is computed for a given value of . If the load ow converges, then the state matrix and its eigenvalues are computed, an eigenvalue of interest is selected (for example, the critical eigenvalue with the minimum real part.), and used to get the tness function. The black box model has only one input and one output, and is used in the standard GA optimization. To nd out all the critical distances to the load ow feasibility and bifurcation boundaries in the problem, the same black box system model can be used. In this case, the inputs are y and , and the tness function is increased when the distance decrease and the critical eigenvalue real part tends to zero. It is understood that the shape of power system small signal stability boundaries can be very complicated, and there exist many niches or in other words, local maxima/minima. In order to ensure GA to locate the multiple maxima of the tness function, and to avoid the noise induced by genetic draft, sucient population size should be considered. However, too large population size will result in slow convergence. Techniques for choosing the population size can be found in [23]. In our test systems, the population size in the range from 30 to 160 was selected. It has been discovered that this population size is sucient to locate the maxima in the space of power system variables. 3.3. Fitness Function Formulation 1 0.9 (; d) = 1 (d)2 () (20) where is the real part of the system critical eigenvalue, d = jj(y ?y0)jj is the distance from the current operating point y0. The diagonal matrix scales the power system parameters of dierent physical nature and range of variation. The rst multiplier in Eqn. (20) reects the inuence of distance, and the second one keeps the point y close to the small signal stability boundary. For example, the following expressions for 1 and 2 can be exploited, Eqn. (22) and (22), 1 (d) = 1=d 2 () = ek?2 (21) (22) where k is the factor dening the range of critical values of . The second multiplier acts as a lter. If k is very large, say about 1000 or more, then only those s which are very close to zero can pass the lter and survive during the genetic optimization process. The lter eliminates a large number of negative s, to force the GA select individuals close to the small signal stability boundaries. The lter function takes the shape as shown in Fig. 3 0.8 0.7 Filter Function Values The tness function plays an important role in genetic algorithms. The optimization result depends extensively on the tness function's ability to reveal the inuence of the factors of interests. There are many ways to convert a practical objective function into a tness function, [21]. Here, since the small signal stability properties of the system are of prime importance, the tness function should be selected to reect the inuence of the critical eigenvalue and the critical distance to the small signal stability boundaries. In this paper, we suggest the following general form of the tness function, see Eqn. (20) 0.6 0.5 0.4 0.3 0.2 0.1 0 −0.01 −0.008 −0.006 −0.004 −0.002 0 0.002 Input variable 0.006 0.008 0.01 Fig. 3. The Filter Function for the Fitness Function 4. The Optimum Operation Direction As a result of the former parts of the paper, critical and sub-critical distances are computed by GAs. In a wider meaning, these distances can also be any most inuential directions of operation in the space of any power system parameters of interest. For example, they can associated with critical/subcritical distances, minimum damping conditions, Saddle node or Hopf bifurcations, load ow feasibility boundaries etc. Upon obtaining these directions, the optimum operation direction can be dened thereafter. The approach can be visualize by Fig. 4, where any two or γ2 V1 3.4. Population Size The shape of power system small signal stability boundaries can be very complicated, and there are many niches existing. In order to ensure GA to locate the multiple maxima of the tness function, and to avoid the noise induced by genetic draft, the sucient population size should be considered. Techniques for choosing the population size can be found in [23]. In our test systems, the population size in the range from 30 to 200 was selected. It has been discovered that this population size is sucient to locate the maxima in the space of power system variables. 0.004 V3 γ1 O V2 γ3 Fig. 4. The Cutset of Power System Security Space more vectors, for example, the critical and subcritical distance vectors, in the parameter space are located by GA sharing function method as V~1 and V~i , where i = 2,...,m are subcritical distance vectors. They form a cutset in the space and dene a new vector m X V~k = ?( kvi V~i ); i = 1, 2, ..., m i=1 (23) where kvi and kvj are weighting factors depending on the inuence of parameter sensitivity as well as the dierence between dierent jjVi jjs. For example, to reveal the inuence of the critical and subcritical distance toward instability, they can be taken as kvi / 1=jjVi jj: Pj Closest Small signal stability boundariies subcritical solution 4 5. GAs with Classical Optimization Methods for Optimization In cases where more detailed information about the small signal stability boundaries are required, both genetic algorithm and classical optimization methods should be used provided that the problem can be handled by classic optimization methods. Since the stability boundaries are very complicated, we are only interested in those cut-sets where most inuential stability conditions exist, i.e. areas close to critical and subcritical distances, minimum damping etc. In such cases, GAs can be used rstly to locate the approximate optima, then classic method will be applied in the vicinity of these approximate optima to explore only these critical and subcritical areas in detail. Note that classic method will be used only if the problem formulation can be handled by it, i.e. dierentiable, convex etc. As shown in Fig. 5, these solutions obtained by GA is the initial guessed solution for classic optimization, and the area close to them is the area where exact optima calculation is required. Since GA can locate only approximate optimum, classic method will have to be used to get the exact optimum depending on the problem accuracy requirement and problem dimension. Note that classic method will be used only if the problem formulation can be handled by it, i.e. dierentiable, convex etc. As shown in Fig. 5, these solutions obtained by GA Critical Solution 0 γ2 (25) Then V~k gives the direction of optimum operation which enables, at least in the meaning of system parameters involved, safest operation direction. Stability problems of other kinds can be investigated accordingly. γ1 γ4 (24) Or in the sense of parameter sensitivity, they can be, kvi = @Vi =@pi : is the initial guessed solution for classic optimization, and the area close to them is the area where exact optima is required. When the most evident γ3 Pi subcritical solution 3 subcritical solution 2 Fig. 5. The most evident small signal stability condition points stability conditions or critical & subcritical distances have been located, the initial most inuential scope of these characteristic conditions will be explored by classic optimization method in the vicinities if necessary and possible. By combination of these two methods the computation costs will be reduced because un-important conditions will not need to be computed, and subsequent control activities can be based on the information obtained from the optimization procedures. Similar step size/direction control algorithms - the multi-rate algorithms can be found in [24]. 6. Power System Model for Small Signal Stability Analysis 6.1. Single Machine Innite Bus Model Analysis A well known single machine innite bus system is studied here with GAs and sharing function method to locate its critical and subcritical distances to instability. The model consists of four dierential equations, which cover both generator and load dynamics [25], see Fig. 6. The mathematical model of the system is the following: m_ = ! (26) M !_ = ?dm ! + Pm + +Em ymV sin( ? m ? m ) + Eo ~ 0 yo (−θ o- π/2) V δ Em ym (−θ m- π/2) δm 20 ~ 15 10 Pd+jQd Fig. 6. The single-machine innite-bus power system model +Em2 ym sinm (27) Kqw _ = ?Kqv2 2 V 2 ? Kq vV + +E00 y00 V cos( + 00 ) + +Em ym V cos( ? m + m ) ? ?(y00 cos0 + ym cosm )V 2 ? ?Q0 ? Q1 (28) 2 2 _ k4 V = Kpw Kqv V + (Kpw Kqv ? Kqw Kpv )V + q 2 + K 2 [?E 0 y 0 V cos( + ? h) ? + Kqw 0 pw 0 0 ?Em ym V cos( ? m + m ? h) + +(y00 cos(0 ? h) + ym cos(m ? h))V 2 ] ? ?Kqw (P0 + P1 ) + Kpw (Q0 + Q1 ) (29) qw ). Paramwhere k4 = TKqw Kpv and h = tan?1 ( KKpw eters of the system are the following [25]: Kpw = 0:4, Kpv = 0:3, Kqw = ?0:03, Kqv = ?2:8, Kqv2 = 2:1, T = 8:5, P0 = 0:6, Q0 = 1:3; P1 and Q1 are taken zero at the initial operating point. Network and generator values are: y0 = 20:0, q0 = ?5:0, E0 = 1:0, C = 12:0, y00 = 8:0, 00 = ?12:0, E00 = 2:5, ym = 5:0, m = ?5:0, Em = 1:0, Pm = 1:0, M = 0:3, m = 0:05. All parameters are given in per unit except for angles, which are in degrees. The active and reactive loads are featured by the following equations: Pd = P0 + P1 + Kpw + Kpv (V + T V_ ) (30) Qd = Q0 + Q1 + Kqw + Kqv V + Kqv2 V 2(31) The system (26) -(29) depends on four state variables , m , !, V . Their values at the initial load ow point are the following: = 2:75, m = 11:37, ! = 0, and V = 1:79. Note that the initial point is not a physical solution as the voltage V is too high as Q1 is zero. To show the results in the parameter plane, which have been obtained formerly by the authors in [18], [26], the boundaries are plotted in solid line in Fig. 7. In the same gure, critical and subcritical stability points located by Genetic Algorithm with sharing function method are plotted in . It is evident that this method with the black box model can locate the Reactive Load Power, p.u. C 5 0 −5 −10 −15 −20 −25 −30 −30 −20 −10 0 10 Active Load Power, p.u. 20 30 Fig. 7. The Closest Stability Boundary Points desired results. They can be used to nd out optimum operation directions by Eqn. (23) given in former section. 6.2. Model Analysis for Optimal Control Direction This classical power system model composed of 3 machines and 9 buses is taken from [27] as shown in Fig. 8. Stability studies for a similar system can be C 2 3 8 2 3 7 9 5 6 A B 4 1 1 Fig. 8. The 3-machine 9-bus System found in [28], [29]. The machines of the system are modeled by using the classical model for machine 1 and two-axis model for machines 2 and 3 - see equations (32) {(37). j1 !_1 = Tm1 ? E1 Iq1 ? D1 !1 (32) !1 ?Edi0 ? (xqi ? x0i )Iqi EFDi ? Eqi0 + (xdi ? x0i )Idi Tmi ? Di !i ? Idi0 Edi0 ?Iqi0 Eqi0 ? Edi0 0 Idi ? Eqi0 0 Iqi _i = !i i = 2; 3 = = = = (33) (34) (35) (36) (37) Unlike [28] and [29], in our tests we neglected the excitation system dynamics, so our state variables were the following: , !, Eq0 , and Ed0 . Algebraic variables were Id , Iq , , and V ; bifurcation parameters were Pld and Qld . . The notations, parameter values and description of the system (32) {(37) can be found in [27]. In the general case of small signal stability analysis, the load dynamics should be denitely taken into consideration, but nevertheless some stability aspects can be studied with the constant load model [30]. This 3 machine 9 bus model considers constant load models only. In the preliminary examination, eigenvalue minimization approach is applied. The loads Qld at buses 5, 6 and 8 were chosen as varying parameter to dene the hyper-space within which the security boundaries lay. The loading was repeatedly tried by GAs with sharing function till the point where load ow did not converge where the tness function was dened as zero so the individuals associated will die o in the next generation. At each step, eigenvalues of the state matrix were computed to nd the bifurcation points and minimum and maximum damping conditions. In the problem of locating the critical distance to stability boundaries, the entry representing the real part of system critical eigenvalue, was set to zero instead of the tness function itself in case the load ow does not converge. Initially two most evident solution points were found in the space dened by load powers at buses 5, 6 and 8. They form two vector, V~1 and V~2 starting from the initial operation point, and thus the plane was dened as a cut-set in the whole parameter space. The result was shown in Fig. 9, from these two vector, the optimum operation direction dened as V~3 which ensures maximum loading increase while maintaining system security can be calculated as V~3 = ?(k1 V~1 + k2 V~2 ). The load ow and stability properties of the system alone that direction were checked to verify the transmission capacity. The result of GA optimization was given in table 1. Note in table 1. the entries 1; 2; 3 till 6 are GA optimization results indicating the solutions on a comparative base. Optimal Loading Direction 40 Active Load Power at Bus 8, p.u. _1 0 _ q0i Edi0 d0 0i E_ qi0 ji !_i Vec3 30 20 10 0 −10 Vec2 −20 0.34 Vec1 0.52 0.32 0.51 0.3 0.5 0.49 0.28 0.48 0.26 Active Load Power at Bus 6, p.u. 0.47 Active Load Power at Bus 5, p.u. Fig. 9. The Critical Directions ( 1 2 ) to Small Signal Stability Boundaries and its Corresponding Optimum ~2 ~1 V ec Direction ( 3 ), where 3 = ? ( VV ec ~ 1 + V ec ~ 2 ). ec V~ ec ; V ~ ec V~ ec No: 1 2 3 4 5 6 V~ ec k jj jj jj Table 1 Load direction where fitness function reveals the most evident small signal stability points rad: ( ) -1.5708 -1.5708 1.5704 1.5708 1.5708 -2.5417 rad: ( ) -1.5264 -0.7645 -1.0531 0.1141 2.2257 2.3514 p:u: ( ) -0.3503 -0.5057 -0.4029 3.0745 0.4413 0.6378 jj Critical Eignevalue 0.0368 + j0 0.0300 + j0 0.0186 + j0 0.0158 + j0 7.2969 10 04 + j0 -0.0021+ j0 ? , and dene the loading variation as mentioned before, i.e. Pi = cos cos , Pj = cos sin , Pk = sin . These result is from GA with sharing function method, and the tness function is on closest distance to critical and sub-critical eigenvalues. However, to make the solution more practical, more powerful GA algorithms are needed to provide better solution reliability and faster computation speed. 7. Conclusion The GAs based approach to small signal stability analysis method using a black box power system model was addressed in this paper aimed at providing means to increase transmission capacity while ensuring the system stability. With appropriate system model the method provided here can locate global as well as local optimum solutions in the form of critical and subcritical distances to to the small signal stability boundaries. They can be used to make decision on control actions. Ideas of dening the optimum operation direction based on critical and subcritical directions or other conditions were put forward here. To make the techniques more practical and ecient, further research on GAs with sharing function method and simplied eigenvalue computation algorithms are required. Acknowledgment This work was sponsored by the ERDC/ESAA Research Program Contract PN2420/94120. Z. Y. Dong's work was supported by Sydney University Electrical Engineering Postgraduate Scholarship. The authors would like to thank Mr. Haining Liu of Tianjin University for his helpful discussion on Genetic Algorithms. References [1] M. Ilic, F. Galiana, L. Fink, A. Bose, P. Mallet and H. Othman, \Transmission capacity in power networks", Proc. of the /em 12th Power System Computation Conference, Dresden, August 19-23, 1996, pp. 5{21. [2] \Proposed terms & denitions for power system stability", Task Force on Term & Denitions, System Dynamic Performance Subcommittee, Power System Engineering Committee, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-101, No. 7, July 1982, pp. 1894 {1898. [3] S. Greene, I. Dobson and F. L. Alvarado, \Sensitivity of the loading margin to voltage collapse with respect to arbitrary parameters", IEEE PES Winter Power Meeting, 96 WM 278-2-PWRS. [4] J. Barquin, T. Gomez and F. Luis Pagola, \Estimating the loading limit margin taking into account voltage collapse areas", IEEE PES Winter Power Meeting, Paper 95 WM 183-4 PWRS, 1995. [5] A. Berizzi, P. Bresesti, P. Marannino, G. P. Granelli and M. Montagna, \System-area operating margin assessment and security enhancement against voltage collapse", IEEE PES Summer Power Meeting, Paper 95 SM 584-3 PWRS. [6] A. C. Zambroni and V. H. Quintana, \New techniques of network partitioning for voltage collapse margin calculations", IEE Proceedings of Part C., Generations, Transmission, and Distribution, Vol. 141, pp. 630-6, Nov. 1994. [7] I. Dobson and L. Lu, \New methods for computing a closest saddle node bifurcation and worst case load power margin for voltage collapse", IEEE Trans. on Power Systems, Vol. 8, No. 3, pp. 905 {913, August 1993. [8] N. Flatabo, O. B. Fosso, R. Ognedal, T. Carlsen, and K. R. Heggland, \Method for calculation of margins to voltage instability applied on the Norwegian system for maintaining required security levels," IEEE Trans. on Power Systems, Vol. 8, pp. 920{928, Aug. 1993. [9] O. B. Fosso, N. Flatabo, T. Carlsen. O. Gjerde, and M. Jostad, \Margins to voltage instability calculated for normal and outage conditions", CIGRE, paper 38-209, 1992. [10] Y. Sekine, K. Takahashi, Y. Ichida, Y. Ohura, and N. Tsuchimori, "Method of analysis and assessment on power system voltage phenomena, and improvements including control strategies for greater voltage stability margins", CIGRE, paper 38-206, 1992. [11] T. V. Cutsem, \A method to compute reactive power margins with respect to voltage collapse," IEEE Trans. on Power Systems, Vol. 6, pp. 145{ 156, Feb. 1991. [12] I. Dobson, \Computing a closest bifurcation in stability multidimensional parameter space", Journal of Nonlinear Science, Vol. 3., No. 3., pp. 307{327, 1993. [13] Y. V. Makarov, D. J. Hill and J. V. Milanovic, \Eect of load uncertainty on small disturbance stability margins in open-access power systems", Proc. Hawaii International Conference on System Sciences HICSS-30, Kihei, Maui, Hawaii, January 7-10, 1997. [14] Y. V. Makarov, V. A. Maslennikov, and D. J. Hill, \Calculation of oscillatory stability margins in the space of power system controlled parameters", Proc. of the International Symposium on Electric Power Engineering Stockholm Power Tech: Power Systems, Stockholm, Sweden, 18-22 June, 1995, pp. 416-422 [15] I. Dobson and L. Lu, \Computing an optimum direction in control space to avoid saddle node bifurcation and voltage collapse in electric power systems", IEE Trans. on Automatic Control, Vol. 37, No. 10, Oct. 1992, pp.1616 {1620. [16] I. Dobson, \An iterative method to compute a closest saddle node or Hopf bifurcation instability in multidimensional parameter dpace", Proc. of the IEEE International Symposium on Circuits & Systems, San Diego, CA, May 1992, pp. 2513 { 2516. [17] I. Dobson, \Observations on geometry of saddle node bifurcation and voltage collapse in electrical power systems", IEEE Trans. on Circuits and Systems - I: Fundamental Theory and Applications, Vol. 39, No. 3, March 1992, pp. 240 {243. [18] Y. V. Makarov, Z. Y. Dong and D. J. Hill, \A general method for power system small signal sta- bility analysis", Proc. the 1997 Power Industry Computer Application conference, PICA'97, May 11-16, Columbus, Ohio, pp. 280 {286. (to appear in IEEE Trans. on Power Systems) [19] Y. V. Makarov, D.J . Hill, and Z. Y. Dong, \Computation of bifurcation boundaries for power systems: a new -plane method", to be submitted to IEEE Trans. on Circuits and Systems. [20] H. G. Kwatny, R. F. Fischl and C. Nwankpa, \Local bifurcation in power systems: theory, computation and application", D. J. Hill (ed), Special Issue on Nonlinear Phenomena in Power Systems: Theory and Practical Implications, IEEE Proceedings, Vol. 83, No. 11, November, 1995, pp. 1456{ 1483. [21] D. E. Goldberg, Genetic algorithms in search, optimization, and machine learning, AddisonWesley Publishing Co. Inc., 1989. [22] D. E. Goldberg and J. Richardson, \Genetic algorithms with sharing for multimodal function optimization", Genetic Algorithms and Their Applications: Proc. of the 2nd international Conference on GAs, July 28-31, 1987, MIT, pp. 41{49. [23] S. W. Mahfoud, \Population size and genetic drift in tness sharing", L. D. Whitile and M. D. Vose edt. Foundations of Genetic Algorithms 3, Morgan Kaufmann Publishing, Inc. 1995, pp. 185{ 223. [24] M. L. Crow and J. G. Chen, \The multirate simulation of FACTS devices in power system dynamics", Proc. of the 19th PICA Conference, May 7 12, 1995, pp. 290{ 296. [25] H. -D. Chiang, I. Dobson, et al. \On voltage collapse in electric power systems". IEEE Trans. Power Systems, Vol. 5, No. 2, May 1990. [26] Z. Y. Dong, Y. V. Makarov and D. J. Hill, \Computing the aperiodic and oscillatory small signal stability boundaries in the modern power grids", Proc. Hawaii International Conference on System Sciences HICSS-30, Kihei, Maui, Hawaii, January 7-10, 1997. [27] P. M. Anderson, A. A. Fouad, Power system control and stability, Iowa State University Press, Ames, IA. 1977. [28] P. W. Sauer, B. C. Lesieutre, and M. A. Pai, \Maximum loadability and voltage stability in power systems," International Journal of Electrical Power and Energy Systems, Vol. 15, pp. 145{ 154, June 1993. [29] M. K. Pal, \Voltage stability analysis needs, modeling requirement, and modeling adequacy," IEE Proceedings of Part C., Vol. 140, pp. 279{286, July 1993. [30] J. H. Chow and A. Gebreselassie, \Dynamic voltage stability analysis of a single machine constant power load system", Proc. 29th Conference on Decision and Control, Honolulu Hawaii, December 1990, pp. 3057{3062. Zhao Yang Dong (Student Membe IEEE'96) was born in China, 1971. He received his BSEE degree as rst class honor in July, 1993. Since 1994, he had been studying in Tianjin University (Peiyang University), Tianjin, China for his MSEE degree. Now, he is continuing his study as a postgraduate research student in the Department of Electrical Engineering, the University of Sydney, Australia. His research interests include power system analysis and control, genetic algorithms, electric machine and drive systems areas. Yuri V. Makarov received the M.Sc. in Computer Engineering (1979), and the Ph.D. in Electrical Networks and Systems (1984) from the St. Petersburg State Technical University (former Leningrad Polytechnic Institute), Russia. He is an Associate Professor at Department of Power Systems and Networks at the same University. Now he is conducting his research work at the Department of Electrical Engineering in the University of Sydney, Australia. His research interests are mainly in the eld of power system analysis, stability and control with emphasis on mathematical aspects and numerical methods. David J. Hill (M IEEE'76, SM IEEE'91, F IEEE'93) received his B.E. and B.Sc. degrees from the University of Queensland, Australia in 1972 and 1974 respectively. In 1976 he received his Ph.D. degree in Electrical Engineering from the University of Newcastle, Australia. He currently occupies the Chair in Electrical Engineering at the University of Sydney. Previous appointments include research positions at the University of California, Berkeley and the Department of Automatic Control, Lund Institute of Technology, Sweden. From 1982 to 1993 he held various academic positions at the University of Newcastle. His research interests are mainly in nonlinear systems and control, stability theory and power system dynamics and security. His resent applied work consists of various projects in power system stabilization and power plant control carried out in collaboration with utilities in Australia and Sweden.