BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

Coordinated Control of FACTS Devices in Power Systems for Security Enhancement

Abstract FACTS devices provide the possibility to control voltages and power ows in a power system and therefore to improve the security of the system. In order to make use of this possibility, the set values of the FACTS controllers have to be chosen appropriately. A valuable option is the application of Optimal Power Flow control, where the set values are determined such that an objective function is minimized given the model of the system. But due to the large size of power systems, it is often difcult for different reasons to include the entire system into the optimization process. In this paper, sensitivity analysis is used to determine the area on which the FACTS device has considerable inuence and then only this limited area is included in the Optimal Power Flow control. If there are several devices placed in the same system, the areas assigned to these devices might overlap indicating mutual inuences. Therefore, a coordination of the control entities is needed in order to avoid conicting behavior of the devices rising the issue of Multi-Area Control. Here, the method based on Approximate Newton Directions is extended for the case of overlapping areas which are determined by sensitivity analysis.

I. I NTRODUCTION Todays power grids are driven closer to their transfer capacities due to the increased consumption and power transfers, endangering the security of the system. On the other hand, FACTS devices are a powerful technology that can solve many outstanding problems in power systems. They provide the opportunity to inuence power ows and voltages and therefore are able to enhance the system security, e.g. by improving the voltage prole or increasing the transfer capacity of a system without the need of new lines [1]. In order to prot from the benets of FACTS devices, appropriate set values for the controllers of these devices have to be determined. A possible method is the application of Optimal Power Flow control. As the main objective in this paper is to improve system security, the objective function includes the improvement of the voltage prole by minimizing voltage deviations from given reference values and the resolution of congestions by relieving overloaded lines. As economic aspect, also the minimization of active power losses is taken into account. But the need of modelling and including the entire system into the optimization process results, for realistic grids, in an optimization problem of considerable size. However, even though the inuence of a FACTS device is not constrained to the bus or line where it is placed, the area in which the inuence is signicant is limited. Therefore, the

problem size is reduced by taking only this area of inuence into account in the optimization [2]. If the number of FACTS devices in a system increases, the areas of inuence of several devices may overlap. Such an overlapping indicates that mutual inuences among the devices exist. Hence, the control of the devices has to be coordinated, otherwise, devices with conicting objectives might start ghting against each other. Due to the large size of realistic grids and also because of the fact that interconnected power systems often comprise several countries, or control areas, each with its own regulation entity, many studies for various applications using Multi-Area control in power systems have been carried out. But to the authors knowledge, none of them covers the case where the areas are determined independently from each other possibly yielding overlapping areas and buses and lines which are not included in any of the areas. In this paper, the Multi-Area control method based on Approximate Newton Directions [3] is extended to these cases. First, the models for the FACTS devices applied in this paper are given. Sect. III describes the determination of the area of inuence by carrying out sensitivity analysis. In Sect. IV, the Optimal Power Flow control problem is introduced and it is explained how the Unlimited Point algorithm is used to solve it. The sensitivities are then used in Sect. V to reduce the size of the OPF problem taking into account only the area of inuence of a device. In the following section, the concept of the method based on Approximate Newton Directions for non-overlapping areas is given and in Sect. VII, the extensions to this method are elaborated. Finally, results and a conclusion are given. II. M ODELLING OF FACTS DEVICES As the intention is to improve the steady-state security, the power system as well as the FACTS devices are modelled using static equations. The FACTS devices considered in this paper are Static Var Compensators (SVC) and Thyristor Controlled Series Compensators (TCSC), since these FACTS devices, besides Phase Angle Regulators, are used most frequent in power systems [4]. An SVC is shunt-connected to a bus and inuences the voltage VSV C at the bus to which it is connected by injecting or absorbing reactive power QSV C [1]. This characteristic is modelled by a shunt-connected variable susceptance BSV C

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

R + jX VSVC jBSVC (a) Fig. 1. jB 2 (b)

jXTCSC

jB 2

(a) Model of an SVC and (b) of a TCSC.

(Fig. 1(a)) for which the injected reactive power QSV C results in 2 QSV C = VSV C BSV C , (1) As the size of the SVC is limited, a lower and an upper bound, BSVC,min and BSVC,max , exist for the effective susceptance BSV C , thus, BSVC,min BSVC BSVC,max . (2)

is a line between bus i and j, Pij and Qij are the active and reactive power ows on this line at bus i. The input variable u depends on the considered device. For the SVC the sensitivity values are determined either with respect to the effective susceptance BSV C or the voltage VSV C and for the TCSC either with respect to the effective reactance XT CSC or to the active power ow PT CSC . The equality constraints g(z, u) correspond to the power ow equations. For large power systems, the usage of (5) is computationally too expensive. In such cases, the sensitivities can be determined numerically by using power ow simulations at different operating points of the FACTS device, thus, Ky = y y1 y0 dy = , du u u1 u0 (8)

A TCSC is connected in series with a transmission line and is able to inuence the active power ow PT CSC through the line by adapting the reactance of the line [1]. Hence, the device is modelled as a variable reactance XT CSC connected in series with the line, as shown in Fig. 1(b). The total reactance Xline of the line including the TCSC is therefore Xline = X + XT CSC , (3)

where y0 , u0 and y1 , u1 are the system and control variables for the initial operating point and for an operating point with a slightly changed control setting, respectively. As power systems are nonlinear systems, the sensitivity values are also dependent on the specic operating point for which the sensitivities are determined. But for all considered variables, it has been found that the sensitivities can be approximated sufcient accurately by Ky = ay u2 + by u + cy , (9)

where X is the reactance of the line itself. The effective reactance XT CSC of the TCSC is limited to a range between XTCSC,min and XTCSC,max determined by the size of the device but also by the allowed line compensation, set to 80% capacitive and 40% inductive in this paper, resulting in XTCSC,min XTCSC XTCSC,max . III. S ENSITIVITY A NALYSIS The inuence of FACTS devices is constrained to a limited area around the devices. Therefore, only this limited area has to be taken into account in the optimization process. In order to determine the area of inuence of a device, sensitivity analysis is applied. A. Determination of Sensitivity Values In general, the sensitivities Ky of variables y(z, u) with respect to changes in variables u and subject to the equation system g(z, u) are given by [5] Ky = y y dy = du u z g z
1

where the parameters ay , by and cy have to be determined appropriately. As the course of the sensitivities over the range of operating points is close to linear, the parameter ay for the quadratic term is rather small. B. Determination of Area of Inuence Having identied the sensitivity values, the area of inuence is determined such that only buses and lines are included in the area for which the sensitivities Ky are larger than a certain limit Ky,lim . But as already mentioned, the sensitivity values are also dependent on the specic operating point. Therefore, the question arises which operating point is to be taken into account and the sensitivities of which variables, thus, if for lines active or reactive power ows and for buses voltage magnitudes or angles are used. In addition, the sensitivities, e.g. for the SVC, can be determined either with respect to the effective susceptance BSV C of the device or with respect to the voltage VSV C at the bus where the device is placed. Investigations have shown negligible variations over the range of the device settings for

(4)

g . u

(5)

Determining the sensitivities in power systems for the considered application, the state variables z and the variable y are dened as z y = [Vi , i ] = [Pij , Qij , Vi , i ] (6) (7)

SVC: sensitivities for reactive power ows and voltage magnitudes with respect to the voltage VSV C , and for TCSC: sensitivities for active power ows and voltage angles with respect to the active power ow PT CSC through the device.

where for a network with n buses for each bus i {1, . . . , n}, Vi and i denote voltage magnitude and angle and where for all (i, j), i {1, . . . , n}, j {1, . . . , n} for which there

In addition, an SVC mainly inuences reactive power ows and voltage magnitudes whereas a TCSC mainly inuences active power ows and voltage angles. Therefore, these sensitivities for the operating point BSV C = 0 and XT CSC = 0, respectively, are used to determine the area of inuence.

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

IV. O PTIMAL P OWER F LOW C ONTROL In a general OPF problem, the goal is to determine control settings such that a particular objective is minimized, taking into account equality and inequality constraints according to the model of the system. The optimal settings are then determined by solving this optimization problem by a given optimization algorithm. Here, the Unlimited Point algorithm [6] is used for this purpose. A. Denition of Optimization Problem For the general formulation of an optimization problem, the objective function f (x), the equality and the inequality constraints, g(x) and h(x), have to be dened which form the optimization problem min
x

B. Solving the Optimization Problem In order to nd the solution of this optimization problem, the Unlimited Point algorithm introduced in [6] is applied. The resulting Lagrange function is given by L(x) = f (x) + T g(x) + (2 )T h(x) (15)

and the corresponding rst-order optimality conditions by L() x g() x h() + 2 x = 0 (16) (17) (18) (19)

= 0 = 0 0

diag{} =

f (x) g(x) = 0 h(x) 0

(10) (11) (12)

s.t.

where x includes state variables z as dened in (6) as well as control variables u from now on dened as BSV C for SVCs and XT CSC for TCSCs and also additional slack variables used to dene soft constraints. For the Optimal Power Flow control with the goal to determine the optimal settings for FACTS devices, the equality constraints g(x) correspond to the power ow equations. The inequality constraints h(x) in this paper include on one hand the constraints on the FACTS device settings (2) and (4) and on the other hand constraints to prevent the lines from overloading dened as |Iij | Iij,lim + s2 , (13) ij where |Iij | is the current on the line between buses i and j and Iij,lim is the capacity of the line. The slack variables sij are heavily penalized in the objective function such that the controller has a strong incentive to set them to zero. Like this the constraints are turned into soft constraints and it is avoided to end up with an unsolvable system. The objective function is composed of three different components: 1) minimization of bus voltage deviations from given references to improve the voltage prole, 2) prevention of overloaded lines, 3) and minimization of active power losses, thus,
n

where are slack variables used to transform the inequality constraints into equality constraints and the hat indicates the values of the variables in the optimal point. By using the squares of and , the condition that the Lagrange multipliers have to be positive and that h(x) 0 has to hold are fullled. The only modication to the conditions given in the original Unlimited Point algorithm is that in [6] equation (19) is formulated with squared variables and . The modication has been applied in order to avoid that the matrix of the linearized system becomes singular. As this equation is used to force either or to zero, the result is the same. By applying the Newton-Raphson method, this system of nonlinear equations is rst transformed into a system of linear equations, i.e. 2 x L g T hT diag(2) 0 g 0 0 0 h 0 0 diag(2) 0 0 diag() diag() L g . (20) = h + 2 diag() Starting with initial values x(0) , (0) , (0) and (0) and iteration step k = 1, the variables are updated by x(k) (k) (k) (k) = x(k1) + x(k) = (k1) + (k) = (k1) + (k) = (k1) + (k) (21) (22) (23) (24)

f (x) =
i=1

v (Vi Vi,ref ) +
(i,j)I

s s2 ij (14)

until the changes are smaller than a pre-dened threshold yielding the optimal values x, , and . V. L IMITED A REA C ONTROL In the optimization problem dened in Sect. IV-A, the entire grid of the system is taken into account to determine the optimal settings for the FACTS devices. For large power systems, the computational effort to solve this problem is tremendous and also the data of the entire system is often not available. Therefore, the limited area control is derived where only the area of inuence of the device is taken into account in the optimization process.

+
(i,j)I

l Pij,loss (Vi , i , Vj , j ),

where Vi,ref is the reference voltage for bus i, where I denotes all (i, j), i {1, . . . , n}, j {1, . . . , n} for which there is a line between bus i and j, and where for each (i, j) I, Pij,loss () is the active power loss. The weighting parameters v , s and l are chosen according to the desired importance of each term.

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

6 1 3 9 5 2 4
SV

14 18 13 17 12 7 11 16 15 19
C

20

Fig. 2.

Illustration of the limited area control

In Fig. 2, an example of a grid with an SVC at bus 9 and a possible area of inuence is shown. For buses which are at the border of the area and for which there exist adjacent buses not included in the area, the power ow equations cannot be set up because the voltage magnitude and angle of the buses outside of the area are not included in the state variables of the optimization problem any more. For such buses, e.g. bus 8, the voltage magnitude and angle are determined using the sensitivity values to approximate the inuence of the FACTS device yielding Vj j = Vj,0 + KVj u, = j,0 + Kj u, (25) (26)

where Vj,0 is the voltage magnitude if the set point of the FACTS device is zero and KVj is the sensitivity of the voltage magnitude at bus j with respect to u. For the voltage angle the notation is accordant. It was already discussed in Sect. III that the sensitivities with respect to the effective susceptance and reactance of SVC and TCSC vary signicantly over the range of the device settings but can be approximated as a function in BSV C and XT CSC by a second-order polynomial given in (9). In order to have the best possible approximation of the voltage magnitude and angle, the average between the sensitivity value at the initial point BSV C = 0 or XT CSC = 0 and the set point for u is taken to approximate the inuence of the FACTS device. Hence, the sensitivities KVj and Kj are given by 1 1 (27) aVj u2 + bVj u + cVj , 2 2 1 1 K j = (28) aj u2 + bj u + cj . 2 2 Having several FACTS devices placed in the same system, the areas of these devices may overlap indicating mutual inuences among and necessitating a coordination between them. Such a coordination is achieved by multi-area control. KVj = VI. M ULTI A REA C ONTROL In multi-area control, the overall problem dened in (10)(12) is decomposed into several subproblems, each associated with an area, and solved in an iterative and coordinated procedure. There exist various decomposition techniques which have been proposed and applied in the literature. Each of these techniques has its advantages and disadvantages. Comparisons of a range of decomposition techniques can be found in [7], [8], [9].

In this paper, the method based on Approximate Newton Directions proposed in [3] and applied for OPF in [10] is used. In this method, each bus and the associated variables are assigned to one distinct area p. These variables are socalled local variables and are the decision variables of area p. All other variables in the system are external variables for area p and are assumed to be xed to the values obtained from the other areas. Concerning constraints, two types are distinguished: constraints which include only local variables and constraints which include local as well as external variables, e.g. the power ow constraints associated to a bus at the border of area p. The latter ones are called interconnecting constraints and are included as regular hard constraints in area p and as soft constraints in the neighboring area m for which the used external variables are decision variables; i.e. these constraints are included in the objective function of area m penalized by the corresponding Lagrange multiplier given from area p. Regarding the objective function, the overall objective function is assigned completely to each area with external variables xed. The values for external variables and Lagrange multipliers are communicated among the areas after each iteration step. The concept of the method based on Approximate Newton Directions emerges from the rst-order optimality conditions for an optimization problem as given in (16)-(19). Applying the described decomposition, the rst-order optimality conditions for all areas combined are equivalent to the rst-order optimality conditions for the overall problem [3]. As with the Unlimited Point algorithm the inequality constraints are transformed into equality constraints, the following description will focus only on equality constraints. The procedure to set up the subproblems is as follows: 1) Determine which buses, and thus which variables, are included in which area. 2) Assign the overall objective function to each area p and dene the external variables as xed. 3) For each bus i in each area p, set up the power ow equality constraints and include them into the constraint set of area p. 4) Determine for each constraint whether it is an interconnecting constraint gp,int (xp , xpe ) which involves local variables xp as well as external variables xpe or whether it is a constraint gp (xp ) only using local variables. 5) Include gp,int (xp , xpe ) in each area m as a soft constraint in the objective function for which a variable xm appears as external variable in this equation. For M areas, the subproblem to be solved by area p {1, . . . , M } with decision variables xp at a particular iteration step is therefore given by min f (xp , xpe ) + (pe ,int )T gpe ,int (xp , xpe )
xp

(29)

subject to gp,int (xp , xpe ) = gp (xp ) = 0 0 (30) (31)

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

where the subscript p denotes local variables and equations associated with buses in area p. The subscript pe is used accordingly for external variables and equations associated with external buses. The bar and subscript pe notation for a variable, e.g., vpe , indicates that the value of v is set to the value determined for v in the previous iteration. The subscript int indicates interconnecting constraints that include local variables as well as external variables. The variables pe ,int are the Lagrange multipliers for the interconnecting constraints gpe ,int included as soft constraints in area p and as hard constraints in the area where the bus to which this constraint is associated to is located. Using this problem setup for each area, the iterative optimization process is started. Instead of solving the subproblems until optimality only the rst Newton-Raphson step is applied. Then the updated variables are exchanged and used in the next iteration. The outline of the scheme is therefore as follows: 1) Each area p initializes its variables xp and p,int by setting the variables for voltage magnitudes and angles and the manipulated variable to the current steady-state values and the Lagrangian multipliers to a common value, e.g. to 0.1. Iteration counter l is set to 1. 2) Given xpe and pe ,int from the initialization or iteration step l 1, each area p {1, . . . , M } determines in parallel with the other areas the rst iteration step of the Newton method for its subproblem given by equations (l) (l) (29)-(31) to obtain xp and p,int . (l) (l1) (l) 3) Each area updates its variables by xp = xp + xp (l) (l1) (l) and p,int = p,int + p,int . 4) The areas exchange the requested values resulting from their optimization problem with their neighbors. 5) Unless a stopping condition is satised, e.g., the absolute changes in all variables from step l 1 to l are smaller than a pre-dened threshold, the next iteration is started by increasing l and going back to step 2. The second step of this scheme actually corresponds to solving C (l) (l) = d(l) (32) where C (l) contains on its diagonal the system matrices according to (20) for the subproblems of all areas 1 . . . M , the vector (l) is the update vector for the system variables and Lagrange multipliers and d(l) is the composition of all righthand vectors as given in (20), both ordered by areas 1 . . . M. The condition for convergence to the optimal point is derived in [3] and is given by = I C 1 C < 1 (33)

diagonals of C and C are equal but the off-diagonal elements of C are non-zero whereas these element are zero in C. For further explanations, refer to [10]. If this condition is not fullled, additionally a Preconditioned Conjugate Gradient method has to be applied in order to reach local convergence. As the optimization problem is possibly non-convex and non-linear, the Generalized Minimal Residual method (GMRES) [11] is chosen in [3] with C (l) as pre-conditioner. The GMRES algorithm is used to solve a linear system of equations, e.g. determining the rst Newton-Raphson step of the overall problem by solving (34), without having to take the inverse of the system matrix. The exact solution is reached after the number of steps equal to the number of variables in the system. But if an appropriate pre-conditioner is used, already after a few steps a quite accurate solution is obtained. In the considered application, the goal is to bring the solution (l) (l) closer to . In order to simplify notation, C , C and d are used for (l) C , C (l) and d(l) in the following algorithm. The update (l) determined by the areas is chosen as initial point. The GMRES algorithm is then given by:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 (0) := (l) r := C 1 d C (0) w1 := r/||r|| z1 := (w1 , r) q := 0 while ||C (q) d|| > tq ||d|| q := q + 1 r := C 1 C wq for i := 1 to q do hi,q := (wi , r) r := r hi,q wi end hq+1,q := ||r|| wq+1 := r/hq+1,q for i := 1 to q 1 hi,q := i+1 hi,q + i+1 hi+1,q hi+1,q := i+1 hi,q + i+1 hi+1,q end := h2 + h2 q,q q+1,q q+1 := hq+1,q /; q+1 := hq,q /; hq,q := zq+1 := q+1 zq ; zq := q+1 zq vq := zq /hq,q for i := q 1 down to 1 do vi := zi q j=i+1 hi,j vj /hi,i end (q) := (0) + q vi wi i=1

19 20 21 22 23 24 25 26 27 end

where (A) is the spectral radius of matrix A. The matrix C corresponds to C at the optimal point. The matrix C is the system matrix for the overall problem in the optimal point, rows and columns ordered such that C = d(l)
(l) (l)

(34)

gives the Newton-Raphson step at iteration l for the overall system, thus, C is built according to C. Therefore, the block

The only parameter for which a coordination among the areas is needed is the residual vector r determined in row 8. As C (l) is block diagonal, the inverse of this matrix consists of the inverses of each block, each determined by the corresponding area. Each area is responsible for a distinct part of the vector and is able to set the vector together if it gets the rest of the vector from the other areas. Therefore, at each iteration step within the GMRES algorithm communication is needed only in row 8. An overview over the entire iteration scheme is shown in

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

Initialize xp , p,int Set l := 1

(0)

(0)

and how to include constraints concerning the common part of the areas (see also [12]). A. Decision Variables From the point of view of a particular area p four types of buses can be determined: 1) local buses: buses which are only included in area p; 2) common buses: buses included in area p and some other area m, hence, located in the common area; 3) external buses: buses not included in area p but included in some other area m; 4) omitted buses: buses of the entire system not included in any of the M control areas. The terms local, common, external and omitted are also used for the variables associated with the respective buses. The local variables for area p are denoted by xp , the common variables by xpc , and the external variables by xpe . The decision variables of area p include the local as well as the common variables whereas the external variables are assumed xed and given by the corresponding area. After each outer iteration step l, the required values of common and external variables are exchanged among the areas. As will be explained in the following, the omitted variables do not appear in any constraint.

Apply rst NR-step in all areas (l) Set q := 0 a)

Apply one GMRES-step in all areas (q) b)

Update xp , p,int

(l)

(l)

Exchange data among areas

B. Constraints In a power system, two real valued power ow equations Check stopping are associated to each bus. These constraints correspond either condition to active power balance, reactive power balance, xed voltage magnitude or xed voltage angle setting. Taking the areas in Fig. 4 as an example, it is obvious that for the gray colored a) GMRES entering criterion buses, the areas need omitted variables to set up the active or fullled c) Set l := l + 1 b) while GMRES criterion reactive power balances. But other than for external variables, fullled these omitted variables are not decision variables of any area c) stopping condition fullled and therefore cannot be assumed to be decided by an other area. Hence, the power ow equations for these buses have to be adjusted. Fig. 3. Flow chart for the optimization scheme. In order to nd the voltage magnitudes and angles at these buses the formula (25) and (26) derived in Sect. V are used. Fig. 3. Applying the Approximate Newton Direction method in For buses included only in area p, only the inuence of the combination with the GMRES method, the solution for many FACTS device of this area has to be taken into account, optimization problems in power systems can be found in a resulting in 1 1 distributed way. Vj = Vj,0 + aVj ,p u3 + bVj ,p u2 + cVj ,p up ,(35) p p 2 2 1 1 j = j,0 + aj ,p u3 + bj ,p u2 + cj ,p up . (36) p p VII. E XTENSIONS 2 2 In the case where a bus which is connected to omitted buses In the described method, it is assumed that each bus is is included in the common area, this indicates that the devices included in exactly one area, i.e. the areas are non-overlapping. of both areas have signicant inuence on voltage magnitude For the application described in the previous sections, this and angle at this bus. Thus, the corresponding equations are method has to be extended for the case where the areas are 1 1 overlapping, thus, some buses are included in more than one Vj = Vj,0 + aVj ,p u3 + bVj ,p u2 + cVj ,p up + p p 2 2 area, and where there exist buses not included in any of the 1 1 + aVj ,m u3 + bVj ,m u2 + cVj ,m um , (37) areas. m m 2 2 In case of overlapping areas, there are buses which cannot 1 1 j = j,0 + aj ,p u3 + bj ,p u2 + cj ,p up + uniquely be assigned to one particular area but are common p p 2 2 to at least two areas. This raises questions on how to treat the 1 1 + aj ,m u3 + bj ,m u2 + cj ,m um , (38) m m common variables, how to decompose the objective function 2 2

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

t1 Area p Area m t4 t4 t2 t1 t2 t1 t1 Fig. 4. t1 t2 t1

t3 t4 t4

TABLE II D ECOMPOSITION OF THE OBJECTIVE FUNCTION FOR AREA p term in overall f () f1 (xp ) f2 (xp , xpc ) f3 (xpe , xpc ) f4 (xpc ) f5 (xp , xpe ) how to include in fp () fp () = + f1 (xp ) fp () = + f2 (xp , xpc ) fp () = + 0 fp () = + 1/ f4 (xpc ) fp () = + f5 (xp , xpe )

t4

t4 t4 t4 t4

C. Objective Function The overall objective function consists of terms which from the view point of a given area involve local, common, external and/or omitted variables. Terms including omitted variables are discarded as the fact that the corresponding buses are not included in any area implies that the FACTS devices do not have signicant inuence on these terms. This yields a reduced objective function fred . Terms including only local and/or external variables, can be treated as within the basic method but terms involving common variables need an additional consideration. Coming forth from the rst-order optimality conditions, the reduced overall objective function fred has to be decomposed such that it holds that the gradient of this reduced overall objective function is equal to the sum of the gradients of the objective functions of the areas, i.e. fred fp = . x x p=1
M

Illustration of different bus types. TABLE I

T REATMENT OF CONSTRAINTS ASSOCIATED WITH DIFFERENT BUS TYPES FOR AREA p type t1 t2 t3 t4 bus location local, common common external external, omitted how to deal with constraints gt1 (xp , xpe , xpc ) gt2 (xp , xpe , xpc ) fp = + pc gt2 (xp , xpe , xpc ) fp = + pe gt3 (xp , xpe , xpc ) -

where um is an external variable for area p and vice-versa. Now it is possible to set up power ow equations for each bus using only local, external and common variables. As there are constraints involving common variables, which is not the case in the basic method, it has to be discussed on how such constraints have to be treated. The rst step is to distinguish between four different types of buses depending on where they are located and to which buses they are connected. The bus types t1 -t4 can be described as: 1) bus type t1 : all local buses and in addition, common buses that are not connected to any local bus; 2) bus type t2 : common buses that are connected to at least one local bus; 3) bus type t3 : external buses that are connected to at least one local bus; 4) bus type t4 : omitted buses as well as external buses that are connected to only external, common and omitted buses. In Fig. 4, an illustration of these types from the view point of area p is given. Depending on the bus type, the constraints associated to the bus are treated differently in the problem formulation. In general, each area includes all constraints associated to buses included in the area as hard constraints with external variables xed. If an area m uses local variables of area p to formulate a hard constraint, this constraint has to be included in the objective function of area p as soft constraint with external and common variables xed and weighted with the Lagrange multiplier of area m for this constraint. Table I gives an overview on how the constraints are taken into account. Thus, for a particular area, its constraints of type t1 and t2 are included as hard constraints with external variables xed, while its constraints of t2 and t3 are included as soft constraints in the objective function with external and common variables xed. Constraints associated with buses of type t4 are not taken into account in the considered area.

(39)

Table II shows how terms depending on local, common and external variables, and combinations of these are taken into account in the decomposed system from the view point of area p fullling criteria (39). The number of areas that include the common variables appearing in the considered term is denoted by . D. Extended Procedure Having dened how constraints and objective function are formulated in the overlapping case, the adapted procedure to set up the subproblems is as follows: 1) Determine which buses, and thus which variables, are included in which area and distinguish between local, common, external and omitted variables. 2) Dene the objective function of each area p including the terms of the reduced overall objective function as dened in Table II. 3) For each bus i in each area p, set up the power ow equality constraints and include them into the constraint set of area p. 4) Determine the type t1 , . . . , t4 for each bus. 5) For buses of type t2 and t3 include the constraints associated with these buses according to Table I as soft constraints into the objective function. The scheme for the optimization process stays the same as for the non-overlapping case including the application of the GMRES algorithm. It has to be noted, that the case with the limited areas does not converge exactly to the solution of the overall optimization

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

BSV C [p.u.]

because with the approximations of the inuences of the FACTS devices on the voltage magnitudes and angles at the border buses some errors are introduced and terms on which the devices have little inuence, i.e. terms concerning omitted values, are neglected. But as has been shown in [2], these errors are small if the areas have an appropriate size. VIII. R ESULTS The described method is applied to the control of FACTS devices in the IEEE 57 bus grid. In the following, three different test cases are considered: two cases with SVCs and one with TCSCs. The parameters in the objective function (14) are assigned to s = 10, l = 0.2 and v = 1 such that the terms concerning voltage deviations and active power losses contribute roughly the same to the objective function value and that overloadings of lines are penalized heavily. The starting point x(0) for the iteration process is set to the base case where no FACTS device is in operation. A. Test Case 1 In the rst test case, an SVC is placed at bus 14 and another SVC at bus 29. The areas of inuence for the two devices are shown in Fig. 5. The limits of the lines are chosen such that no line is overloaded when the setting of the FACTS devices are set to zero. Therefore, the focus lies on improving the voltage prole and decreasing the active power losses. The courses of the effective susceptances BSV C over the outer iterations l are shown in Fig. 6. Already after a few iteration steps, the SVC settings converge to their nal values. As the spectral radius is 0.84, the Preconditioned Conjugate Gradient method is not needed and therefore no additional inner iterations q have to be carried out. The performance of the limited multi-area control is determined by comparing the resulting value for the overall objective function f () and the FACTS device settings BSV C with the base case when no FACTS device is in use and with the results for the optimal control where the entire grid is taken into account in the optimization. In Table III, these values are

SVC settings 1 0.8 0.6 0.4 0.2 2 4 6 Iterations 8 10 Area 1 Area 2

Fig. 6. Course of the effective susceptances BSV C of areas 1 and 2 for test case 1. TABLE III O BJECTIVE FUNCTION VALUES AND EFFECTIVE SUSCEPTANCES OF SVC S FOR TEST CASE 1. Reference 0.09536 0.0 0.0 Optimal 0.09206 0.33664 0.17801 Limited 0.09208 0.32316 0.15512

f () BSV C1 BSV C2

given for this rst test case. The small deviations in device settings and consequently also in the objective function value between the limited multi-area and the optimal case result from the fact that by using approximations for the voltage magnitudes and angles at the border of the areas, some errors are introduced. B. Test Case 2 The second test case differs from the rst case just by the values for the line limits. The line limit for line 33 is chosen such that it is slightly overloaded in the base case. This problem now yields a spectral radius of 1.49. Hence, some inner iterations q, i.e. GMRES iterations have to be carried out. In Fig. 7, the course of the SVC settings over the outer iterations l are shown. Additionally, 62 additional inner iterations of the GMRES algorithm are necessary in total to make the settings converge to their nal values. The resulting values for the objective function and the effective susceptances of the SVCs are listed in Table IV. In the second row, the values of the objective function term concerning the line loadings are given. In the optimal as well
SVC settings

5
4

1
2

16

16

17
18 21
5

20

19 29 30
19 20

15
17

45

58
15

26 27

14
13

18 6

31

28
14

25

12

SV
21 44

13

72
46

59

60
47

32 57
26 24

66 61 79
48
12

34
23

33
22

38

62 78 50
49 39 57

38

36

35

63

50

BSV C [p.u.]

37

53

Area 1
76
42

0 -0.1 -0.2 -0.3 2 4 6 Iterations 8 10 Area 1 Area 2

27 25 33 37

51
49 46 47
34 35

43

24 39 42 45 43
28 30 31

77
75 52 73
40 56
74

55
56

71

44
32

48
36

54

11
11

41
10

23 64

40 41
7 29 52 53 54 55

22
8

SV

67

68

69

70

80
9

C
65

Area 2 Fig. 5. Areas of inuence for SVCs at buses 14 and 29.

10

51

Fig. 7. Course of the effective susceptances BSV C of areas 1 and 2 for test case 2.

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

TABLE IV O BJECTIVE FUNCTION VALUES AND EFFECTIVE SUSCEPTANCES OF SVC S

TCSC settings 0.02 Area 1 Area 2

2.

XT CSC [p.u.]

f ()
ijI

s s2 ij

BSV C1 BSV C2

Reference 0.11293 0.01757 0.0 0.0

Optimal 0.09821 0.0 -0.26981 0.04565

Limited 0.09855 0.0 -0.26388 0.01998

0 -0.02 -0.04

Line Loadings w/o FACTS with FACTS

6 Iterations

10

Fig. 10. Course of the effective reactance XT CSC of areas 1 and 2 for test case 3. TABLE V O BJECTIVE FUNCTION VALUES AND EFFECTIVE REACTANCES OF TCSC S
FOR TEST CASE

Iij /Iij,lim

0.5

3. Limited 0.09713 0.00171 -0.01844 0.00997

f () 20 40 Line Number 60 80
2 ijI s sij XT CSC1 XT CSC2

Fig. 8.

Line loadings without and with SVCs for test case 2.

Reference 0.32843 0.23306 0.0 0.0

Optimal 0.09542 0.0 -0.01756 0.01003

Line Loadings

Iij /Iij,lim

as in the limited mulit-area control all line loadings are brought below 100 %. In Fig. 8, these line loadings are shown for the reference and the limited multi-area case, thus, the overloaded line 33 is relieved and brought below its limit. C. Test Case 3 In the third test case, two TCSCs are placed in the IEEE 57 bus grid, i.e. in lines 22 and 72. The areas of inuence determined by sensitivity analysis are shown in Fig. 9. The line limits are chosen such that lines 6 and 60 are overloaded in the base case, when the FACTS devices are out of operation. Figure 10 shows the evolution of the FACTS device settings. As the spectral radius for this problem is lower than 1, the Preconditioned Conjugate Gradient method is not needed and therefore no inner iterations are carried out. After only a few steps the FACTS device settings converge to their nal values.

w/o FACTS with FACTS

0.5

20

40 Line Number

60

80

Fig. 11.

Line loadings without and with TCSCs for test case 3.

5
4

1
2

16

16

17
18 21
5

20

19 29 30
19 20

15
17

45

58
15

26 27

14
13

18 6

31

28
14

25

12

13

21 44

72
46

59

Area 1
66

60
47

32 57
26 24

61 34
23

33
22

38

79
48

12

37
7
6

53 78 50

62 63
49 39 57 50

38

36

35

27 25 33 37

51
49 46 47
34 35

76

42

43

24 39 42 45 43
28 30 31

77
75 52 73
40 56
74

55
56

71

44
32

48
36

54

11
11

41
10

23 64

40 41
7 29 52 53 54 55

In Table V, the objective function values and the FACTS device settings for the base, the optimal and the limited multi-area case are listed. The TCSC settings in the limited multi-area control are close to the values determined by the optimal control and the same also hold for the objective function value. The reason for the deviations again lies in the approximations of the voltage magnitudes and angles at the border of the areas using sensitivity values. Consequently, also small errors in the determination of the line loadings are introduced. The controller in the limited multi-area case sets the effective reactances such that using these approximations the line loadings are all below 100% but in reality line 6 is still slightly overloaded. Thus, the term corresponding to the line loadings, given in the second row of Table V, is not exactly zero. But looking at the line loadings for the base case and for the limited multi-area control as shown in Fig. 11, the line loadings are reduced signicantly. IX. C ONCLUSION In this paper, a control for FACTS devices is developed which is also feasible for large power systems. The Optimal

TCSC

67 22

68

69

70

80
9

Fig. 9.

TCSC
8

65

10

51

Area 2 Areas of inuence for TCSCs in lines 22 and 72.

BULK POWER SYSTEM DYNAMICS AND CONTROL - VII, AUGUST 19-24, 2007, CHARLESTON, SOUTH CAROLINA, USA

10

Power Flow control is applied to a limited area including only buses and lines on which the device has considerable inuence. The method based on Approximate Newton Directions with additional Preconditioned Conjugate Gradient method is extended for the case where the areas are independently determined by sensitivity analysis, which may lead to overlapping areas as well as situations where part of the grid is entirely left out from the optimization. The simulation results show fast convergence for the device settings to the nal values. The control performance of the limited multi-area case is close to the performance of the optimal control taking the entire grid into account. R EFERENCES
[1] N. G. Hingorani and L. Gyugyi. Understanding FACTS concepts and technology of exible AC transmission systems. IEEE Press, New York, 2000. [2] G. Glanzmann and G. Andersson. FACTS control for large power systems incorporating security aspects. In Proc. of X SEPOPE, IP012, Florianopolis, Brazil, 2006. [3] A. J. Conejo, F. J. Nogales, and F. J. Prieto. A decomposition procedure based on approximate newton directions. Mathematical Programming, (Ser. A 93):495515, 2002. [4] A. Edris, R. Adapa, M.H. Baker, L. Bohmann, K. Clark, K. Habashi, L. Gyugyi, J. Lemay, A.S. Mehraban, A.K. Meyers, J. Reeve, F. Sener, D.R. Torgerson, and R.R. Wood. Proposed terms and denitions for exible ac transmission system (FACTS). Power Delivery, IEEE Transactions on, 12(4):1848, 1997. [5] T.S. Karakatsanis and N.D. Hatziargyriou. Probabilistic constrained load ow based on sensitivity analysis. Power Systems, IEEE Transactions on, 9(4):18531860, 1994. [6] Giorgio Tognola. Unlimited point-Ansatz zur L sung des optimalen o Lastussproblems. PhD thesis, Diss. ETH Nr. 12437, ETH Zurich, 1998. [7] B.H. Kim and R. Baldick. A comparison of distributed optimal power ow algorithms. IEEE Transactions on Power Systems, 15(2):599604, 2000. [8] R. R. Negenborn, B. De Schutter, and J. Hellendoorn. Multi-agent model predictive control for transportation networks: Serial versus parallel schemes. In Proc. of Proceedings of the 12th IFAC Symposium on Information Control Problems in Manufacturing (INCOM 2006), pages 339344, Saint-Etienne, France, 2006. [9] M. Arnold, S. Kn pi, and G. Andersson. Improvement of OPF o decomposition methods applied to multi-area power systems. In Proc. of Power Tech 2007, Lausanne, Switzerland, 2007. [10] F. J. Nogales, F. J. Prieto, and A. J. Conejo. A decomposition methodology applied to the multi-area optimal power ow problem. Annals of Operations Research, (120):99116, 2003. [11] Yousef Saad. Iterative methods for sparse linear systems. Siam, Philadelphia, 2nd edition, 2003. [12] G. Hug-Glanzmann, R. R. Negenborn, G. Andersson, B. De Schutter, and H. Hellendoorn. Multi-area control of overlapping areas in power systems for facts control. In Proc. of Power Tech 2007, Lausanne, Switzerland, 2007.