Anda di halaman 1dari 7

616 IEEE Transactions on Power Systems, Vol. 10, No.

2,M a y 1995

An approach to corrective control of voltage instability


using simulation and sensitivity
T. Van Cutsem

Senior Research Associate, FNRS - Dept. of Elec. Eng., University of Liege


lnstitut Montefiore, Sart Timan B 28 - B-4000 Liege, Belgium
vct@rnontefiore.ulg.ac.be

Abstract. This paper deals with the diagnosis of voltage collapse trigger automatic protecting actions (e.g lower limit on grid voltage
situations, following large disturbances and/or load increases. A to trigger OLTC blocking). On-line, because mid-term voltage in-
method is proposed to identify the set of buses where load restora- stability leaves some time to react from a control center and, while
tion is responsible for the collapse and to determine the correspond- too short for a human operator, this time allows some computer-
ing corrective actions. It has been implemented in a fast voltage ized corrective control to be thought o f in a near future. Another
stability simulator, using sensitivity techniques. Tap changer block- on-line application is the “study mode” where corrective actions
ing and load shedding are illustrated on a real-life 410-bus system. can be suggested in answer to anticipated scenarios proposed by
operators.
In recent years, significant progress has been made in the identi-
1. INTRODUCTION fication o f critical system buses. Sensitivities, although poor mea-
sures of the global system robustness, can be used to the purpose
Voltage stability has become a major concern for the secure of this identification. Among them, the sensitivities o f reactive gen-
operation o f many power systems [I]. eration to load powers, originally proposed in [2], can be computed
In this field much attention has been paid to defining appro- very efficiently. They have been used e.g. in [3] t o identify reactive
priate analytical frameworks and improving computing tools o f the power support locations for improving stability under progressive
preventive type (calculation o f security margins and voltage stability system loading. Another approach consists in performing Jacobian
indicators, planning o f reactive power support, etc.) eigenvector analysis around the point where a saddlenode bifurca-
Comparatively less efforts have been devoted to corrective con- tion occurs (i.e. where this matrix is singular). In Ref. [4], the
trol, in particular t o determining the best action to stop a developing left eigenvector relative to the zero eigenvalue was derived as a La-
voltage collapse. Although admittedly more complex, it is however grange multiplier vector, a by-product o f a fast optimization-based
a needed approach, complementary to the preventive one. method for reactive power margin calculation. It is also provided
In this paper, “mid-term’’ voltage instability is of concern. Sim- by the direct method o f Ref. [5]. An enlightening geometric inter-
ply stated, this is the scenario where, following some disturbance, pretation of this eigenvector is available in Ref. [6]. More recently,
transformer On-Load Tap Changers (OLTCs) try to restore their Ref. [7] proposed to extend the eigenvalue analysis to other than
secondary voltages and hence the corresponding load powers, while critical operating points, in order to find directions of system stress,
it has become impossible for the generation and transmission system corresponding to several most critical eigenvectors.
to meet this demand. Part of this impossibility originates from lim- A first limitation of the present approaches is that they show
itations on generator reactive power, imposed mainly by Maximum where the problem is, but do generally not tell how much correc-
Excitation Limiters (MELs), acting in the same time frame. Beside tive control is needed. A second limitation is that they apply to a
OLTCs, partial load restoration may result from a “spontaneous” progressive loading scenario and do not really deal with large dis-
trend of the load itself (e.g. “thermostatic” loads). turbances. It would be most useful t o couple these approaches with
When a power system approaches or enters such an instability, the time-domain simulation of voltage instability. This paper paper
efficient corrective controls are (non exhaustively) : switching in of offers some contributions into these directions.
shunt capacitors, start up o f f a s t units, blocking/reversed control of
OLTCs (or changing their voltage set-points), and in the last resort
INSTABILITY
2. ANALYSIS OF VOLTAGE
load shedding. A combination of them is often advisable : e.g. tap
changer blocking to slow down the system degradation and leave
some time for another action, maybe load shedding. 2.1. System modelling
Corrective control requires improved diagnosis methods, espe-
cially in view o f the seriousness of some of the above listed actions. A general power system modelling relevant t o voltage stability
Improved methods and tools are needed both off-line and on-line. studies takes on the form :
Off-line, t o better set up the criteria and alarm thresholds which will

Equation (1) represents the transient dynamics of machines,


SVCs, etc. with the corresponding state variables x.
Equation (2) represents the network relationships and involves
the vector y of bus voltage magnitudes and phase angles. Loads
are represented as voltage-dependent active ( P , ) and reactive (Q,)
This paper was presented at the 1993 Athens Power Tech powers. Although the derivations accomodate t o any voltage de-
Conference held in Athens, Greece, September 5-8, 1993. pendence, we adopt the exponential modelling :

0885-8950/95/$04.00 0 1993 IEEE


617

whose behaviour around the equilibrium Vz = hois qualitatively


K
P, = P,"(,>"* and Q, = & p (K~ ) ~ * (4)
the same, except that the dead-band is ignored. Assuming that
v, v, there are n OLTCs, all equations (10) are written in matrix form :
t o depict the transient behaviour o f the i-th load. We call Pp and d
QP the d e m a n d coefficients. Load sensitivity t o frequency or the -r = K ( V i - V,) (11)
dt
summation over several m,'s and j3,'s have not been considered for
the sake o f simplicity only. Thus the equations (2) relative t o the with r = [ T ~ ... T , ] ~ ,K = diag[K,], V: = [ V i . .. V
n:'] and Vz =
i-th load bus take on the form : [V,, . . . VZnIT. Moreover, Vz is a subvector o f y and hence :

vz = u y (12)
where U is a matrix o f 0's and t l ' s .
The equilibrium points o f the system (2, 9, 11) are solutions o f :

where gPI and gQ, are the traditional power injection relationships
v;-vz=v;-uy = 0 (13)
obtained from the network. f(X,Y) = 0 (14)
The discrete-time equation (3) captures controls and protecting g(x,y,r) = 0 (15)
actions taking place on a slower time scale : OLTCs, MELs, au-
Around such an equilibrium, the linearized behaviour is given by :
tomatically switched capacitors/reactors, secondary voltage control
[ 8 ] , etc.; z includes transformer ratios, field current limits, shunt d
susceptances, AVR and SVC voltage set-points, etc. The transition -Ar = A Ar (16)
dt
times k are dictated by built-in delays. For instance, the equation
relative t o the variable ratio T o f an OLTC can be written : in which the Jacobian A can be obtained directly from (11) :

T(k + 1)
= ~ ( k-
= T(k)
) AT if V, > V: + E
+
if IV, - 51
' <E
= ~ ( k ) AT if Vz < 5'- E
) (7)
or alternatively, by linearizing Eqs. (2, 9, 11 and 12) :

- d
where (see Fig. 1) V, the secondary, controlled voltage, 5' the
corresponding set-point, E half the dead-band and AT the ratio
step size. It is common t o have the first change in T delayed by 0 - Ax
more than the subsequent changes.
__......... 0
1 I JXY 1
K R X 6 K
"primary? ~ - - - - ? - - - L1- "secondary"
J
and eliminating A x and Ay, which yields :
5k
Figure 1: transformer OLTC modelling

Remark. We assume that load restoration originates from


,OLTCs and do not include differential equations describing sponta- Stable OLTC operation requires the eigenvalues o f A t o have neg-
neous load restoration. The latter could be also dealt with by the ative real parts.
method described in this paper. Remark. The above derivation requires J,, t o be regular. The
singularity o f J,, corresponds t o a saddle-node bifurcation (or loss
2.2. Linearized analysis o f equilibrium) o f Eqs. (1,2) for the current z. In such a situation,
some faster system degradation takes place and the transient dy-
In the sequel we focus on the mid-term voltage instability caused namics x can no longer be neglected. In many publications, voltage
by OLTCs. Within this context, it is acceptable to neglect the collapse has been associated with the singularity o f Jxy.; in the mid-
transient dynamics x and replace Eq. (1) by : term time frame however, the singularity o f concern IS that o f J .
Now, it may occur that an OLTC instability yields a so depressed
voltage profile that a singularity o f J,, results [9, lo]. The prac-
tical problem o f concern is t o take a corrective action fast enough
A deeper analysis o f voltage instability requires t o linearize the t o prevent the system from reaching this dangerous point.
system dynamics around a given operating point. For the purpose
o f this linearized analysis (not o f simulation !) : (i) we focus on 2.3. System portrait in load power space
OLTCs and consider the other components o f z constant around
the operating point o f interest (z will generally change from one More insight on voltage collapse is given by the system portrait
operating point t o another); (ii) it is convenient to replace the in the load power space. This is the space o f the active and reac-
discrete time dynamics (7) by the continuous one : tive powers (4) o f loads controlled by OLTCJ. These are nonlinear
functions o f the y vector and thus they are implicit nonlinear func-
tions o f the state vector r. We are going t o consider the system
trajectories in this space.
618

We group all demand coefficients into :

So = [Pf Qy ... P," 911' -t---------- ---- --------


and all load powers into :

For the sake o f simplicity, we neglect OLTC dead-bands and we


assume that the reference voltage yo of the load model (4) co-
incides with the OLTC set-point Go (see Eq. (7)). Under these
assumptions, So is the mid-term equilibrium point to which a (sta-
ble) trajectory S ( t ) converges after a disturbance.
Now not all values of So yield a solution t o the equilibrium equa-
i--
s,o

I
S(n+'
" I
11 s(o-) = so",
... A .........f ...........
s ( B ) ............ .
/& I \
h B
-.>
..'
-" '.

tions (13-15). This impossibility to satisfy any demand is precisely Figure 2: Portrait of collapse in load power space
a cause o f voltage collapse. In other words, in the load power
space, there exists a feasible set S outside which the system has
no mid-term equilibrium. 1 line 12 P(v2) = P(' Vi)
We show next that following some disturbance, the trajectory I * Q=O
S ( t ) cannot "escape" from this set. Indeed, assuming that it can, v,
at any point outside S we could trivially write for all loads controlled
by OLTCs : Figure 3: Elementary 2-bus system
vz",
PI = PI(a)Q1 = QI
Q vz",p
(a) (22)
v,, v,, curves refer to bus 1. The solid lines are the famous network PV
curves before and after the initiating disturbance. The mid-term
These equations also express that, for a (larger) demand P,'" = load characteristics is the dashed vertical line (constant power due
8 , Qp' = Ql, the outside point would a the mid-term equilibrium to the OLTC restoring &). In the post-contingency configuration
point (since V,, = Vz",).This is impossible by definition of S. there is no mid-term equilibrium since the dashed line does not
The feasible set is limited by a hypersurface C. Each point cross the new network PV curve. The dotted lines are the transient
o f C corresponds t o a saddle-node bifurcation of the continuous- load characteristics for successive values of the OLTC ratio. In i t s
time modelling (11). In practice, different parts of C correspond to attempt to restore V, = Go the OLTC makes the system point
different generators under reactive power limitation. move from B to C to D.
At each point o f C, A has one zero eigenvalue, i.e. it is singular.
Now, since :
det J = det A . det J,, (23) Vl
J is also singular.
The classical scenario of voltage collapse is when smooth in-
creases in demand cause So to go out o f S . Once So reaches the
hypersurface C, S(t) cannot follow it any longer. Relationships be-
tween the resulting system behaviour and the geometry of C have
been detailed in [6].
In the large disturbance scenario, a loss of generation or trans-
mission equipments causes C t o shrink so significantly that So ends
up outside the new S.This is sketched in Fig. 2, where S1 and S,
are the active or reactive powers o f the same or different loads. The
dashed (resp. solid) line refers to the pre- (resp. post-)disturbance
configuration. Assuming that equilibrium had taken place before P,,,
the disturbance occurrence at t = 0, the system experiences a Figure 4: P V curves showing voltage instability
jump from S(O-) = S o to S(O+) due to the load sensitivity to volt-
age. S(O+) is necessarily within S. From there on, the OLTCs act
In this example, A is a scalar : A = -K a&/dr. It is negative
to move S(t) towards S o . However, S ( t ) cannot escape from S ;
therefore, after "hitting' C, the trajectory goes backwards. Com- (stable) above point C, positive (unstable) below. Intuitively the
paring the Jacobian A (or J ) right before and,right after the hitting stability can be checked by slightly disturbing the state variable T.
of C. one real eigenvalue has changed from negative t o positive. A negative value a t point B, for instance, means that if B were
a mid-term equilibrium point, it would be stable. In the shown
Remark. Note that this eigenvalue analysis is performed around
the current operating point which is not an equilibrium point (there scenario, there is no equilibrium any longer but one can figure out
is no mid-term equilibrium any longer). Here, the objective o f eigen-
B as an equilibrium by merely lowering the set-point V;. This shifts
the vertical dashed line to the left up t o crossing the PV curve at
value analysis is not to check local stability but merely to detect
the hitting of C. B, Note that V: does not appear in J and that all system variables
are left unchanged.
2.4. A simple illustrative example A is zero a t point C. This is a saddle-node bifurcation point in
the sense that for a smaller demand Po or set-point V;, the mid-
We consider the elementary system o f Fig. 3. The load fed term load characteristic (a vertical line here) is tangent to the PV
through the OLTC is purely resistive (cr = 2, Q0 = 0) for simplicity. curve a t point C. In the sequel, we rather call C the critical point.
Figure 4 illustrates a typical instability scenario where a large In this one-dimensional example, the load power space is simply
disturbance causes the system to loose mid-term equilibrium. All the P-axis of Fig. 4. In the post-disturbance configuration, S is
619

the interval ] - 00, P,,,,,] while C amounts to the point P,,,. In S show how sensitivities can be used to detect the critical point and
the post-disturbance trajectory goes t o P,, and then comes back obtain the normal vector.
to the left.
4.1. System dynamics simulation
3.
As far as mid-term voltage instability is of concern, a fast sim-
T O CORRECTIVE CONTROL
ulation method consists in approximating the system trajectory
through Successive Transient Equilibrium Calculations (STEC). The
From the system portrait in the load power space, it follows
principle is as follows :
that corrective control of voltage instability will aim a t bringing the
- the system evolution following some perturbation is viewed as the
demand point So back into the (post-disturbance) feasible set S .
response of the transient dynamics (1,2) to the discrete changes in
This can be done through :
controls described by Eq. (3);
0 a decrease in load : e.g. by modified tap changer control or load
shedding;
- as the transient dynamics x plays a marginal role in the very volt-
age collapse mechanism, the dynamic equations (1,2) are replaced
0 an increase in generation : e.g. by switching in capacitors,
by the equilibrium equations (2,9);
switching out inductors or starting fast units.
- the mid-term dynamics is rendered by considering that each equi-
There are many ways to move the demand point into the feasible
librium "lasts" until a new transition takes place; the transition
set. In practice, however we would like to minimize the amount of
times are dictated by OLTC delays, inverse time characteristics of
unrestored load, the amount of committed generation, etc. This is
MELs, sampling period of secondary voltage control, etc.
equivalent to moving So to a point S* of the feasible set such that :
With respect to full numerical integration, STEC is consider-
ably faster and, even during large disturbance voltage instability,
2
,=I
S; - is minimum (24) it automatically provides the equilibrium points around which the
linearized analysis of 82.2 has to be performed. The computa-
with - Sp 50 i=l,. . . ,n (25) tional efficiency comes from the absence o f numerical integration :
rather, Eqs. (2,9) are solved for the successive values of z. A New-
which corresponds to an L1-norm criterion. ton method is used whose Jacobian is the sparse J,, matrix of Eq.
This yields a diagnosis of considerable help in setting up ade- (18). Moreover, the factors of J,, are updated as rarely as pos-
quate corrective actions, since it gives the minimal action required sible : for instance, changes in Pp's and Qp's, transformer ratios,
to have the system kept a t the stability limit instead of collapsing. shunt capacitor/reactor susceptances or AVR voltage set-points do
Clearly, in practice, a somewhat stronger corrective action will be NOT trigger a J,, update. For further details and examples, please
taken, to cover uncertainties in modelling, etc. refer to [lo] and Section 5.
The next key idea is t o approximate the boundary hypersurface C
linearly, i.e. by its tangent hyperplane. In Section 4, we will show 4.2. Eigenvalue monitoring
how the critical point (where C is hit) can be located along the
trajectory and the normal vector n (see Fig. 2) can be computed. Consider now the sensitivities of the total reactive power gener-
Therefrom it is trivial t o build the tangent hyperplane. ation Qc to the active eo (resp. reactive Qp) demand coefficients
To further illustrate the idea, we consider the case where active of the various load buses, a t a system mid-term equilibrium (i.e.
and reactive power injections could be changed individually and with V,, = b:, i = 1, ... ,n).This is a concept introduced in the
with no limit. In this case, S* is the point of C where all loads early Ref. [2], that we adapt to our equilibrium equations (13-15).
are restored (i.e. S: = Sp) except t h e o n e corresponding t o the Rewriting these equations for convenience in compact form :
largest c o m p o n e n t of the n o r m a l vector n. This can be shown
from the general Kuhn-Tucker optimality conditions. Consider for F(X, S o ) = 0
instance the 2D case of Fig. 2. At point A, load S, is restored where X is the generalized state vector : X = [rT xT yTIT, a
while load S1 is not; a t point B, the converse holds. The points general sensitivity relationship yields :
of C satisfying (25) lie in between A and B. This portion of C
is almost a straight line, with a normal vector n more directed
along the S, axis than along the SI one. Hence the point which
+
minimizes (Sf - SI) (S: - S2) is B and the smallest unrestored
load is Si - Sz(B). The point which minimizes the L2 norm Note that QG is a function of network voltages only while So appear
[(S! - + (S: - S2)2]1'2is the normal projection of So on C . in network equations (2) only. Therefore :
It corresponds to restoring less load.
In practice, active and reactive power injections cannot be con-
trolled apart from each other (except for shunt compensation) and
there are limits on the change of a single bus injection (one can-
not shed more than the whole load, switch more than the whole
available compensation, etc.). All these practical considerations A modal decomposition of J-' yields
are easily taken into account in the very simple procedure of bring-
ing the point So on the other side o f the locally tangent hyperplane
while minimizing the L1 norm (24).

4. I M P L EM ENTAT10 N CONSIDERATIONS where A, is the i-th eigenvalue of J and v, (resp. w,)the corre-
sponding right (resp. left) column eigenvector. Similarly, BQG/aX
The method described in the previous section is general and can can be expressed in terms of the w,'s as :
be implemented in various ways. We have chosen to incorporate it
into the fast voltage stability simulation program described in [lo],
whose principle and advantages will be first recalled. Then we will
620

Introducing ( 2 7 ) and (28) into (26) gives after simplification :

In a close neighbourhood o f hypersurface C, ( 2 9 ) is dominated


by the critical, almost zero eigenvalue Ac and the sensitivities can
be approximated by :

where w, is the left eigenvector relative t o A., In practice, in a rea-


sonably large neighbourhood of C, w, does not change significantly
(care must be only taken o f abrupt changes caused by machines go-
ing into reactive limitation). In particular, there is no reason for the
components o f w, nor for a, t o change sign. Therefore, when
the trajectory hits C, the above sensitivities all change f r o m large
positive to large negative numbers.
Remark. It is easy to show that the left eigenvector relative to
the zero eigenvalue of Jacobian A is a subvector of w,. The same
holds for the reduced Jacobian of reactive power vs. voltage mag-
nitudes, considered in [7] as well as its sparse approximation used
in [4]. Therefore, in the neighbourhood of C, all these eigenvectors
and the above sensitivities bring essentially the same information. e
,.... .....................................
3
i..........................MV
L
(20 kV,...)
.................
4.3. Normal vector computation
Figure 5: Western part of EDF 380/225 kV system
The normal vector n is obtained by adapting the derivation in
[6]. For a small displacement dSo along hypersurface C, one has :
close to the study region have a detailed modelling with 3 internal
aF BF aF z variables and 3 equations o f the type (9). Other generators are
- dX+ -dSo = J dX+ -
ax dSO
dSo = 0
dSO simply represented as PV buses with governor effects. The overall
system frequency is also an z variable.
where J is singular. Premultiplying by wT yields : The following dynamic components are considered : N 90
OLTCs. 30 HV capacitor switching devices, MELs on a l l detailed
generators, and the heart o f the so-called Coordinated Secondary
Voltage Control (CSVC), acting in parallel in two regions. CSVC
As this holds for any dSo along E, (dF/&So)T~, is the sought is a new controller [8], presently under implementation, aimed at
normal vector n. Equation (30) can then be rewritten as : keeping bus voltages close to specified values while maximizing re-
active power reserves, within a region. To this purpose, each 10 sec-
onds, measurements of "pilot node" voltages and generator reactive
power productions are collected, from which set-point corrections
are computed and sent a t the next time step to generator AVRs.
Hence, close to where the sensitivities change sign, their relative These corrections minimize the sum of squared pilot node voltage
magnitudes bring the s a m e information as the n o r m a l vector. deviations and (normalized) machine reactive productions, taking
It must be emphasized that the computational effort t o obtain into account inequality constraints on controls, pilot node voltages,
these sensitivities along the system trajectory is quite moderate.
generator reactive outputs and sensitive bus voltages. This o p
According t o Eq. (26), it amounts to computing and factorizing
timization uses quadratic programming apd relies on a load flow
the sparse J matrix, then solving the linear system with vector '
sensitivity model of the EHV system.
aQc/aX as (sparse) independent term. The structure of matrix
aF/aSo makes the remaining calculations trivial. 5.2. The voltage collapse scenario

5. ILLUSTRATIVE SIMULATION RESULTS Figure 6 shows the system response t o a steep load increase fol-
lowed by a line tripping, as obtained from STEC. Starting from a
5.1. System description and modelling high load level, the regional demand So is increased by 100 MW/24
Mvar a t 4 successive times. Figure 6.a shows the small voltage d e
The results hereafter involve the Western region of the Electricit6 clines due to OLTCs maintaining HV and M V voltages. They are
De France (EDF) system, which has experienced voltage problems counteracted by the CSVC which increases AVR voltage set-points
in the past. as shown in Fig. 6.b. At t = 310 s, line 4-5 is tripped, resulting in
The model includes 410 buses, 610 branches and 55 generators. a voltage collapse. This causes a stronger CSVC response, up the
It covers the 380 kV grid of the overall EDF system, the 225 kV point where operating conditions have so deteriorated that the o p
grid of the Western region (see Fig. 5), a HV and M V level repre- timization problem can n6 longer be solved (flat part of the curves).
sentation according to the lower left scheme in Fig. 5, in which all Figure 6.c shows the field currents o f four generators. Machine A,
transformers are equipped with automatic OLTCs. heavily stressed during the load pick-up, is switched under rotor
In the study region, loads are modelled a t the M V level, accord- current limit right after the line tripping, while machines B and C
ing t o Eq. (4) with cy = 1.5 and /3 = 2.4. Thirty generators in or are switched during the subsequent (tentative) load restoration.
621

1.I I I I
identical vectors, equivalent to n, according t o Eq. (30).
I I I
At the so obtained critical point, the total unrestored active load
1.05 is 328 M W (unrestored reactive loads are not shown due to space
limitations). Its distribution over the 25 M V buses is given in the
1 4th column o f Table 1. Number between parentheses indicate buses
whose voltage is within the OLTC dead-band.
0.95

Table 1
0.9 -
unrestc d active loads (MW)
0.85 bus W,tl Qp at at after
nb. t = 350- t = 350+ t = 350 2nd II 3rd
0.8 'C rele
- -
6 118.7 -306.2 39.9 -
35.9
0.75 7 115.7 -298.7 13.8 -
12.2
0.7 1 I I I
I I 1 8
9
109.9
108.1
-282.7
-277.5
9.2
8.1
0 100 200 300 400 500 600

c
1.25 I I I I 1
10 105.5 -270.9 15.7
volt set-pt AVR gener A - -- -- -
, 11
12
102.5
101.4
-262.8 20.4
volt set-pt AVR gener B -260.0 14.4
volt set-pt AVR gener C *- 13 101.1 -259.5 5.4
1.2 volt s e t 4 AVR dener D 14 100.8 -258.4 21.0
15 100.4 -257.1 24.8
16 96.3 -246.8 8.2
1.15
17 90.9 -231.6 16.3
18 72.0 -181.3 20.3
19 70.8 -176.5 1.9
1.1
20 68.4 -170.5 3.8
21 67.4 -169.5 5.3
22 67.3 -169.0 (5.8)
1.05
23 62.9 -157.2 10.5
24 61.5 -150.3 (10.3)
25 58.0 -144.4 (-4.4)
1
26 53.3 -132.7 (0.7)
27 52.6 -125.3 7.2
3.4 1 28
29
49.6
47.6
-121.7
-113.2
(1-5)
8.4
-
30 47.5 -115.9 6.4

5.4. Determination of minimal OLTC blocking


,-
I
I
We seek to determine how to best block OLTCs in order to keep
_---I
2.6 the system a t the stability limit instead o f collapsing. We could

1
decide to have tap positions blocked a t the values corresponding to
2.4 ,---- fieldcur. generA U -
the above obtained critical point. However this is not optimal in
---J
---
.e*
*
f i l d c u r generB U ---.
fieldcur:generCgu/ ---- the sense that we should maximize the restored load, i.e. we should
,*+ field cur. gener D @U) find the smallest set of OLTCs to be blocked. This corresponds
---I-

in Fig. 2 t o moving from the point where C is hit t o the point B

1.8 I I I t(s) I I
where the restored load is maximal. The idea o f the tests reported
hereafter is to have this task performed by the OLTCs themselves
0 100 200 300 400 500 600
in a modified system simulation.
Figure 6: collapse following load increase and line tripping To this purpose, we restart from the identified critical point
and block all OLTCs at their current positions. Then, OLTCs cor-
responding to the smallest sensitivities (i.e. smallest components
5.3. Critical point identification of n) are released in order the restore the corresponding loads. The
subsequent system evolution is determined up to equilibrium. Note
The sensitivities (26) are computed along the system trajectory. that this trajectory is fictitious : it has nothing to do with the real
A change in sign is detected a t t = 350 s. Table 1 shows the corrective control o f the system, which we will show later on. Se-
last positive and the first negative values, for the 25 most sensitive quences of OLTC release and time simulation are repeated until the
M V buses controlled by OLTCs (bus numbers refer to EHV level total unrestored load does no longer decrease significantly. Clearly,
for an easy reference t o Fig. 5). Only reactive sensitivities are the final system operating point does not lie on C exactly but is
shown; these are dimensionless (pu/pu or Mvar/Mvar). A typical somewhat inside the feasible set, which is desirable anyway.
value before disturbance is 2, a t most 3. One easily checks that Coming back to the above example, we restart from the critical
that dividing each vector by - say - its first component yields two point a t t = 350 s and block all OLTCs. We then release OLTCs
622

corresponding t o sensitivities lower than half the maximum value, 1.08


i.e. below 60 in this case. We simulate the resulting system evo-
I I I I 1
! I--*_
I

...- >...---.a
---_._
lution up t o equilibrium. At this point, the unrestored load has
significantly reduced t o 165 MW. Its distributiorr over the 25 M V 1.06 -
--"LL.
-
c---_
,--I
buses is shown in the 5th column o f Table 1. Underlined results I
refer to OLTCs still blocked (it happens that some o f them have
their voltage within the dead-band). For released OLTCs, the un-
,.04 -
restored load is not exactly zero owing t o the dead-bands. Besides,
loads behind blocked OLTCs have somewhat recovered due t o the 1.02 -
CSVC increasing the EHV (primary) voltages. The possibility o f
taking such controls into account is precisely an advantage o f time
domain simulation.
Two more groups of OLTCs are then successively released, yield-
ing the last two equilibria detailed in Table 1 and a total unrestored
load o f 156 and 152 M W , respectively. The l a s t gain in total load
o.98 -
being marginal, the procedure is stopped. An attempt t o further
restore some loads would artificially impute instability to the 0th- 0.96
ers, while they are equally responsible for the collapse, as indicated 0 100 200 300 400 500 600 700 a00
by the close values o f the sensitivities (incidentally, loads behind Figure 7: system saved by OLTC blocking and load shedding
still blocked OLTCs start slightly decreasing). The smallest set o f
OLTCs to be blocked has been identified, along with the corre- increases.Itcombines timedomain with sensitivity
anal-
sponding positions, primary grid voltages, etc. ysis. The method has been shown very efficient on a real-life sys-
For validation purposes* a corrective based On the above tern. The resulting fast and versatile computing tool contributes to
diagnosis is The system evolution is unchanged UP to the improving voltage stability analysis and corrective control determi-
line tripping a t t = 310 s as well as during the next 60 seconds. This nation, especiallyfor real-time applications.
delay is supposed t o be the time needed to detect the emergency
conditions and block OLTCS. Only the above identified OLTCS Acknowledgement. This work was the preliminary phase of a col-
are blocked, on their current positions. The resulting EHV voltage laboration with the "Direction des Etudes et Recherches" of E.D.F.
evolution can be seen in Fig. 7 from 2 = 370 to 1 = 600 s (the
remaining o f the figure refers to the next section) : there is no
collapse. Rather, the CSVC makes voltages slightly increase. 7. REFERENCES
5.5. Determination o f minimal load shedding
[I] IEEE System Dynamic Performance Subcommittee, "Voltage
The results of Table 1 show that, for the scenario of concern, stability o f power systems : concepts, analytical tools, and in-
there is a 152 MW excess load, located a t the first 12 buses o f dustry experience", IEEE Document 90TH0358-2-PWR, 1990.
Table 1. This corresponds to the minimal load shedding required to [2] J. Carpentier, R. Girard and E, Scano, "Voltage collapse prox-
maintain the system a t the stability limit. For already mentioned imity indicators computed from an optimal load flow", Proc. 8th
reasons, among those buses, there is some freedom in choosing the p s c c , Aug. 1984, pp. 671-678
loads to be shed. w e choose for instance to shed 155 M W (and 30 [3] M. and A. phadke, *scontrolof voltage stability using
Mvar) of the load at bus 15. sensitivity analysis", IEEE Trans. on Power Systems, Vol. 7, No
This load shedding takes place at t = 600 s in the simulation 1, Feb. 1992, pp. 114-123.
of Fig. 7. Moreover, a t the same time, we release all OLTCs. As [41 T. Van Cutsem, method to compute reactive power mar-
can be seen, voltages corne back to an acceptable value instead gins with respect to voltage collapser8, IEEE on Po,,,er
of collapsing, while all but the shed loads do finally recover their Systems, vol. 6, No. 2, Feb. 1991, pp. 145-156
pre-disturbance values.
~ ~ This last
~ release
~ of O ~L T C ~is kperformed
, for validation 151 C. Cafiizares, F. A h a d o , C. DeMarco, 1. Dobson and W.Long,
"Point Of collapse Inethods applied to power systems",
purposes and would not be recommended in practice. On the con-
IEEE TranJ. O n Power Systemsp 7 t No *' May 1992t pp.
trary, tap changer blocking before load shedding is motivated by the
673-683
necessity t o stop the system degradation, before shedding loads. In
the shown case, tap changer blocking would be sufficient but its [6] I. Dobson, "Observation on the geometry of saddle node bifurca-
effects are known to be limited intime, due to e.g. thermostatic tion and voltage collapse in electric power systems", IEEE Trans.
load restoration. -
on CAS Part I, Vol. 39, No 3, March 1992, pp. 240-243
(71 G.K. Morison, B. Gao and P. Kundur, "Voltage stability analysis
5.6. Computational efficiency using static and dynamic approaches", IEEE Trans. on Power
Systems, Vol. PWRS-8, No 3, Aug. 1993, pp. 1159-1171
To fix ideas, the simulation of Fig. 6 takes about 12 seconds [8] J.-p. Paul, J.Y. Least and J.M. Tesseron, "Survey of the sec-
On a 28-MIPS S U N SPARCstation 2- On the Other ondary voltage control in France : present realization and inves-
a full sensitivity calculation takes about 0.25 second CPU. Hence,
together, detecting the change in sign o f sensitivities and perform-
tigationst*, IEEE T ~ on power
~ ~systems,
~ vol.
. p w ~ s - 2 ,N~
2, M~~ 1987, pp. 505- 511
ing the modified, non-collapsing simulation takes no more than 15
[9] T. Van Cutsem, "Voltage Collapse Mechanisms: A Case Study",
seconds CPU, on the average. The method is thus very efficient.
prOc. Bzl,k Power System Voltage Phenomena II seminar,
Deep Creek Lake, Md, Aug. 1991, pp. 85-101
6. CONCLUSION
1101 T. Van Cutsem, "Analysis of emergency voltage situations",
This paper has proposed a general method for diagnosing mid- Proc. 1993 11th PSCC, Avignofi, France,, Vol. 1, pp. 323-330
term voltage instability following large disturbances and/or load

Anda mungkin juga menyukai