Вы находитесь на странице: 1из 10

IEEE Transactions on Power Apparatus and Systems, Vol. PAS-97, No.

6, Nov/Dec 197b 2031


PENALTYFACTORS FROM NEWTON'S METHOD
F. L. Alvarado, Member, IEEE
Department of Electrical and Computer Engineering
The University of Wisconsin-Madison
ABSTRACT The procedure was probably first utilized by
Dy Liacco, who briefly mentions it in papers
The penalty factors calculated by the during 1971 [9], 1974 [10] and 1975 [11]. A
classical B-coefficients method are shown to slightly more detailed presentation of the
be proportional (but not equal) to the penalty method has been recently presented also by
factors calculated directly from the trans- Dy Liacco as a discussion to a paper by Happ
posed Jacobian. It is also shown that the [12, 13]. There is evidence that other re-
economic operating point is independent of the searchers are also aware of the method [14,
load and generation distribution factors Z. 15]. Nevertheless, a paper entirely dedicated
Expected computational times in examples with to this important method, and in particular
up to 5873 buses and 943 generators are given. dedicated to contrasting this method with the
The evidence presented supports conclusions classicalB-coefficientsmethod, is not avail-
also reached by others indicating that the able.
transposed Jacobian approach is superior to
classical B-coefficients based on its compu- One feature of this paper is its use of
tational speed. the distribution factors L to provide the link
with classical B-coefficients methods. These
INTRODUCTION distribution factors are interpreted in terms
of the frequency participation factors K of
On line economic dispatch is based on the loads and generators. It is shown that the
knowledge of the generator incremental costs economic operating point is (as it should be)
and the transmission penalty factors. The in- independent of these factors. It is also shown
tent of this paper is to establish a clear that the transposed Jacobian method does not
relationship between classical B-coefficient produce penalty factors that are equal to the
methods for the calculation of these penalty classical ones, only proportional to them.
factors and the newer "transposed Jacobian" However, knowledge of the distribution factors
approach for the direct calculation of these does in fact allow the calculation of factors
same factors. A second goal of this paper is that are equal to the classical ones and there-
to serve as a tutorial paper on the details fore correct incremental losses can be eval-
and subtleties of transposed Jacobian penalty uated if so desired.
factor calculations.
It is not the intent of this paper to do
The B-coefficients method is based on an a survey of related work in economic dispatch
approximation of the system losses as a quad- nor to argue the virtues of the penalty fac-
ratic function of the generation powers [1]. tors approach. For these and related matters
Classic calculation of B-coefficients requires the reader is referred to a recent and thor-
resorting to Kron's power-invariant transfor- ough survey paper by Happ [12].
nations. Kirchmayer's calculation method [2]
was followed by later digital implementations FUNDAMENTAL CONCEPTS
by Kirchmayer,et al. [3] and Happ, et al. [4].
More recently, the calculation of B-coeffi- Consider a system with n buses. Let buses
cients has been improved upon by Meyer [5], 1,2, ...,m be generation buses. Let bus 1 be
who not only streamlined and clarified the the bus that would be selected as slack bus in
calculations, but also made them compatible a powerflow study. Let Ai be the incremental
with sparsity techniques. cost of operating the generator at bus i . In
the absence of network losses, the most eco-
An alternate approach to the B-coeffi- nominal operating conditions for the network
cients methods is the direct exact calculation are when:
of the penalty factors without the need for an
approximate loss formula. The first such pro- _ X (1)
cedure was introduced in 1956 by Shipley and 1 -2 - m
Hochdorf [6]. Exact calculation of penalty
factors for systems of practical size remained, Let PL be the network losses and Pqi the
however, impractical until after the introduc- power injection at generation node i. If the
tion of the sparsity-oriented Newton power flow incremental losses are known, the optimal op-
by Tinney and Walker in 1967 (7]. The exact erating point occurs when:
calculation of penalty factors from the trans-
posed Newton power flow Jacobian was first sug- L X = L X L X (2)
gested by Dommel and Tinney's work in 1968 (8]. 1 1 2 2 m m

where Li- is the penalty factor for generator


i. These penalty factors can be determined
from:

F 78 266-9. A paper recarvrnded and approved by L. - 1 3


the IzZEEEPc r System Engieering Camiittee of the i dPJ
IEEE Pc*wer Engineering Society for presentation at the 1 - dP.
IEEE *PES Winter Maeting, New York, NY, January 29 -g
February 3, 1978. Manuscrpt submitted January 10,
1977; made availale for printing Novnuser 30, 1977.

0018-9510 /78/l100 -2031$00. 75 ©) 1978 IEEE


2032
As mentioned in the introduction, classi- Therefore
cal methods call for the calculation of the
incremental losses by first calculating an ap- _P_ _
proximate quadratic expression for the losses: (f - f ) = n (8)

P -Pt[B
g GG IP g
+ BtGO P g + B
00 (4) i=l
L Define the normalized load coefficients 9Q as:

From this approximate expression the vec- p


tor of incremental losses can be found from: p AA K.
1 (9a)

_L
dP
-
2Pt[B
g GG
] + BtO
GO (5) i-l
g
K9
This paper does not use (4) and (5). In- kgI A 1 (9b)
stead, it calculates dPL/dfg directly. Before n P
proceeding with this calculation, it is useful L i
to gain some physical insight into the nature i
of these incremental losses.
It should be apparent that
Since the early days of load flow analysis
it has been recognized that it is not realistic
to specify the power at all generators and at
all loads. One generator is normally chosen
1Pi
9 = 1 (10)
as "slack" generator and its power Pgl is left i
deliberately unspecified. The reason for this
is that the losses of the system are not pre- Substituting (8) and (9) into (6) and simplifying
cisely known prior to starting the calculations.
Once the problem is solved, the power at this p
slack generator can be readily determined. di =
zpi D (lla)
In an actual system there is always agree-
ment between input power (total of all genera- Qodi + p (llb)
tion) and output power (total of all loads plus Qdi =
k
i D
losses). If the penalty factors are calculated
using a mathematical model of the system it Equation (11) merely states that the load
must also be required that generation match at any bus can be expressed as a constant load
total load plus losses,
in addition to a fraction of the system's total
. .
The derivative of the losses with respect ~~~~~~frequency-sensitive load. The explicit depen-
to a single generator power requires an infin- dency on the frequency has been eliminated,
itesimal incrementation of oone~gi while hold- since frequency itself is of no importance.
ingsmall theremaininggenerator
ing all the powers
remaining generator powersconstant. eonsant Every bus (including generation and
allowed to have a load. The
slack) is
constants are 9Q
Even though infinitesimal, this power must be intimately related to the k-factors of classi-
matched by an equal increment in losses and cal B-coefficients. They can be determined by
loads. In a practical system this is accom- running two case studies for the same system.
plished by an infinitesimal increment in fre-
whc.eut runnin
They can two be studied by a
sam e of
quency which results innicrae
quency, increased loads
od.BtBoth the type of be determined
alsoload present atby each bus. As of
a knowledge in
active and reactive loads might increase as a all previous papers on the subject, this paper
result. In addition, the power supplied by any assumes that these values are either available
generator operating on a frequency-bias mode . .
may also be affected. Each bus of the system cor
willcan
be be reasonably
shown that the estimated. bpoint
Ultimately
economic operating it
can therefore be assumed to have a frequency- is independent of these factors anyway.
sensitive component to its load (which includes
any frequency-sensitive generation). For the CALCULATION OF THE PENALTY FACTORS
infinitesimal increments of interest, this
function, whatever it might be, can be linear-
ized. Thus, the load at a bus can be expressed The losses of the system can be determined
as follows: from:
P di = Po.
di + KPi (f - f ) (6a) m
.
n
.(12)
iL l gi _ di

= 0di + KQ(f - f ) (6b) Substitution from (12) and (11) yields:

Let PD be the total frequency-sensitive m


load of the system. That is, let: P = I P. - LI di ~D (13)

-
KP(f
D~~0=f° ()(7
p During the operation of an actual system
certain quantities are either known or can be
2033
controlled independently. Other quantities are The following equations provide the neces-
only indirectly controllable. This distinction sary equality constraints. Notice that no
is also possible in the mathematical model of special treatment is given to the slack bus,
the system. The reduced gradient approach to since Kirchhoff's law must be applicable to it
the optimum power flow problem makes such a too. Also notice that reactive power balance
distinction [6]. Among the quantities that are equations at generation buses are not neces-
either constant or independently controllable sary.
are the following: + -
a) Parameters of the linear passive portion AP1 Pg +d Z1PD + P1(v) = 0
of the system (such as transmission lines,
transformers or shunt capacitors). These
parameters include equivalent fr imped- AP - -p PO + P ) 0
ances, tap settings, phase shifts, etc. 2 g2 d2 +2D 2 0
Knowledge of these parameters permits * g
finding the nodal admittance matrix. ,)
b) Active power at generators (P i, for i =1,
2, ..,m). g P
AP =-P + p0 + k + p (v) =0
c) The frequency-independent component of the n gm dm m D m
active and reactive loads (P0 and Q0. for
i l, 2,..,n). di di o P
d) The frequency-sensitivity coefficients for
i =m+l = +Pdm+l +Rm+l D P m+(v) 0 (15)
all the loads nd for i= 1, 2,.., + Q P + Q () =

m~
e) n).
The voltage magnitude at
buses (IViI for i =1, 2,...,
all generation
m)
AQ + dm+l +km+l PD +Qm+l
tv a
f) The voltage phase angle at one bus chosen
as reference, generally the so-called slack
bus (61).
A +Po
dn
+ P
nD
+ P(v)=
n V

Once the parameters and independent vari- + Q


ables mentioned above are specified, the fol-
lowing variables can be determined or measured:
AQ Qdnn + nPD + Qn(v) = 0

a) The voltage magnitude at all load buses Let Pg be the vector of all generation
(IV i for i =m+l, m+2, ...., n) . powers:

b) The voltage phase angle at all buses ex- Pt A [p p P ] (16 )


cept reference (6.i for i= 2, 3, .., n). g gl g2l 'gm
c) The total frequency-sensitive load (PD). The calculation of the penalty factors
requires knowledge of the derivative of PL
d) The reactive power supplied by each one of with respect to the independent variables P9
the generators (Q i for i =1, 2, ...,m). subject to the equality constraints (15). This
gi is simply the reduced gradient of the loss PL.
Making an inventory of dependentvariables, The best way to determine this reduced gradi-
we see that there is a total of 2n variables. ent is to define the Lagrangian:
However, since the reactive powers supplied by
the generators (item d) are of no interest, the A p + PAP + AP + P AP
variables of interest are 2n-m. It must, there- L 1 1 222 m+l m+l
fore, be possible to establish 2n-m equations.
These equations come from Kirchhoff's current Q A
law. The net power injected into a bus (gener- +m+l Qm+l +--
ation minus load) must be equal to the total
power Pi flowing into-the linear passive por- where theres are the Lagrange multipliers and
tion of the system at each bus. This power P are interpreted further later on. Replacing
can be readily expressed in terms of the un- L (
known bus voltages and phase angles. Since the m n n
manner in which this is done is well known [9] P - P - PD + iX AP.
it will not be repeated here. - g i d
Let v be the vector of all unknown volt-
age angles and magnitudes: n7 Q
+ RAQ. (17)
i=m+l

v [62 '6m' dm lml' 'n' ln] The reduced gradient can then be found by
(14) evaluating:
2034

to aa
a- . 0l Let A be the vector of all normalized
subject av- = 0 and DP = ° load coefficients excluding reference:
g
-t Qkp
The partials of £ with respect to gener- ==I .2* ..
m+l m+l ... N Q
N] (24)
24

ation powers Pgi are readily found from (17):


as 1 -
P= i1, 2, ...,m (18) Based on the notation established in (22)-
aPgi 1 , (24) it is possible to write equations (19)-
The partials of £ with respect to voltages, (21) in a much more compact form (notice that
a.C a c = 0)
angles and PD are also found from (17) and set
equal to zero: alvIi = 0 implies IV. i
|alv'I
a
=
P
3AP 1 p 3APn Q DAQm+1 P"
t 1
a6. = l a6.l + + Pn 36. + 3m+l JM
6. + - -= 0 (25a)

+ 3Qnad 3AQ~
+***+~Q t- ~p
P
n 2, (25b)
+
6
=
0 i ==2
3, ..., n (19) t ff ++ l
3 = 1

Solving for 6l from equation (25b):

a___ al a+ P + + Sn ; P+ m+l a 1 = 1 (1- It ) (26)

a AQ
+ +Q n - 0 Substituting into equation (25a) and re-
n - ......arranging:

i =m+l, ,n (20) + ' [ =t X lpl(27)i

asn n
___

D
-
1 + i t +i
1,+
ili=m+l
= 0
(21) which can also be expressed as:

Let J denote the Jacobian used for an aJ H p P [ (28)


ordinary Newton power flow. This Jacobian is L l1 1vli av
given by:
If this paper were to conclude here, it
would have already achieved an important re-
3AP2 1
MpP MP2 sult. Equation (28) can be used to solve for
as
2 ... '| ad m+i Iv-j
|. I m+l the Lagrange multipliers
used in (26) and (18) to find all the desiredbe , which can then
_ - _-- - - - - - - - - - - r - - incremental loss factors; these can in turn be
. , , . substituted directly into (3) to find the pen-
. , . alty factors L. A remarkable simplification is
. . possible, however. This simplification is based
_ - - - -
-
r _ _ on the well known Matrix Inversion Lemma [16],
J = 3AP | CAP aAP 1 also known in the power literature as network
adM+l IV I alw
IVm+1 compensation 17]and a variety of other names.
2I ad The form of the Matrix Inversion Lemma used in
, , paper is as follows.
this
3AQ IaAQM~'AQm~
M+1m+ ll
M+11 Let be the solution to:
26
2

_,_ _ _,_ _ _ _
I IVI
_ _ _ _ _ _ _ _ _ _,_ _ _
x

Ax = b (29)
., , . Let (x+AR) be the solution to:
(22) (A+a r c t)(x
+Ax) = b (30)

It may be observed that this Jacobian where a and c are vectors and r is a scalar.
does not contain a row for the reference bus If x is known, the value of the correction
active power mismatch. Let 6 be the vector of term Ax can be found from:
all Lagrange multipliers excluding reference:
Ax = w(ctx) (31)
L
[2 63 *-. m Sm+l Sml *- 'n~rn 2) t
2035
where the vector w is determined from: The calculation of the penalty factor at
the slack generator must follow a slightly
Aw = a (32) different procedure. From (26):

Using equation (27) it is possible to p 1 -t -


identify x+ AR with B, A with Jt,bwith (3P,/ 1 (1 - Q ) (26)
av)t, r with (1/i) and c with 9. The solution
to (27) or (28) can thus be found by first Substitution into (18) and (3) yields:
solving:

it i, = _pa~lj (33) L1 =P + it (42)

An important remark is in order at this


Compensating this result by AS: point. It should be apparent that using the
from (33) directly without compensating it
-wr(- ) + it[t (34) does not result in the traditional penalty
factors unless ky
= 1 and all the other kP's
are zero. This is equivalent to saying that
where w is determined from: all the frequency dependent load of the system
is located at the slack bus. Experimentation
t UP lj t (35)
by the author has shown that substantial dis-
W (35)
crepancies can be incurred if such an assump-
av tion is made. However, if the penalty factors
are to be used only for economic dispatch under
equal modified incremental generation cost
Comparing (35) and (33), it can be ob- LfXi, then it does not make any difference if
served that: ail the penalty factors are multiplied by the
same constant factor. If all the L's are di-
w -a' (36) vided by L the following new penalty factors
L' are obtained:
Therefore no separate solution for w is LI 1 (43a)
required and (34) can be reduced to: 1

AS = I
-t i )37) L! 1 i=2,
3,...,m (43b)
-1Pt 0'
1

The value of 6' can therefore be compen- There are several implications of this
sated to yield 6", which is the solution to observation. The most striking one is that the
(27) but scaled by a factor 9. distribution factors ., are, in fact, not re-
quired at all. Neither, of course, are the s-
.. =t .. (38)factors
')(Q _ 6 ' of classical B-coefficients.
P + (38) Another remark of interest pertains to
the accuracy of the calculations. The crucial
which reduces to step (the solution of (33)) involves the Newton
power flow Jacobian (without a slack bus). This
__P matrix is generally diagonal dominant and is
_
_ 1 known to be well behaved. The result is that,
6= 6' -Pt (39) provided the Jacobian has been correctly cal-
culated, little numerical difficulties can be
expected. This is in contrast to other methods
where nearly singular matrices result.
Therefore the solution to (27) is:
Regarding computational requirements of
_~ 1-the method, known results can be used to esti-
1
R = P -t ~~ --IS (40) mate these requirements [18,19]. The method
(. + i9 't has, nevertheless, been tested successfully in
a variety of systems as indicated in a later
section. Some additional computational consi-
Once the Lagrange multipliers 6 are known, derations are also given there.
the penalty factors can be determined from
(18) and (3). Equation (18) states that:
SUMMARY OF THE ALGORITHM
L= 1 - Si (18 ) A very concise review of the algorithm is
a~~~gi 1 ~~~now given.
A. Preliminary steps
Substitution into (3) yields
These steps may be done as part of a
L. = 1 (41) power flow prior to the calculation of the
1 SPpenalty factors.
2036
Table I. Approximate Time Required for Various Operations (sec. CPU Univac 1110)

System Size Calculation of


Jacobian Penalty Factors
Buses Generators Ordering Calculation Factorization (Transposed Solution)
30 5 .035 .011 .048 .022
118 54 .172 .048 .203 .089
662 69 1.70 .328 2.36 .602
5837 943 15.1 2.90 21.0 5.59

1) Knowing the topology of the system, order The penalty factors were then calculated
the rows and columns of the Jacobian to using a transposed repeat solution and modify-
reduce fill-in. ing the results as indicated earlier in this
paper. The factored Jacobian from the last
2) Solve the power flow problem by Newton s Newton power flow iteration was used. If this
method. At each iteration: factored Jacobian had not been available, it
would have been necessary to first construct
a) Construct the Jacobian J ,
it and factor it. Furthermore, if it had not
b) Factor J been ordered, it would have been necessary to
do so.
c) Solve JAv = Mismatch
The CPU requirements for each of the test
systems are given in Table I. The times are
B. Penalty factor steps: those for a Univac 1110 computer.

to optimize execution time was


1) Solve Jt = _| ja 1 made. No Inattempt
particular, considerable savings in
computation time could be realized by two sim-
ple modifications:
2) Let: L1= 91
P + it
1) All buses connected to the slack generator
L could be renumbered toward the bottom of
3) Let: L. = for all generators the Jacobian. This could reduce the
a of interest . amount of computation required for the
forward elimination step of the transposed
repeat solution to a nearly trivial num-
As outlined in the previous sections, it is ber of operations.
possible to simply let L1= 1 and dispatch based
on Li = l/6{lP. The resultant penalty factors 2) All generation buses of interest would
will not be the traditional ones, but the re- also be located more or less toward the
sultant operating conditions will be the same. bottom of the Jacobian. This could reduce
Furthermore, no c distribution factors are the amount of computation required for
needed the back substitution step of the trans-
posed repeat solution to a relatively
small number of'operations.
CONCLUS IONS
Classical B-coefficients require an amount
of computation proportionalin to of
square The
EXAMPLES the number of generators the the
system.
The penalty factors for a variety of sys- transposed Jacobian method, on the other hand,
teams have been calculated using the method requires an amount of computation that is less
described in this paper. The systems tested than quadratic on the total number of buses
range from 30 buses to 5837 buses. In each [18]. If the ratio of generators to total
case an ordinary Newton Power Flow was first buses is approximately constant, it can be
required: safely concluded that the transposed Jacobian
solved.This operation .
will be faster in all systems except small sys-
ordering the Jacobian tems or those systems with few generators. In
addition, of course, the transposed Jacobian
2) 2)
Evaluating the Jacobian
Evaluatin
each the Jacobiaemethod
iteration . does
. not require the knowledge of the
~~~~~~~~~~distribution
factors and yields exact rather
3) Factoring the Jacobian each iteration than approximate values.
theIt is thus hoped that this paper has pre-
4) Performing. arepeat
acoban ech
factred
solutionwith
ieraton.the
sented the necessary evidence to help establish
transposed Jacobian method as the method
of choice for all future economic dispatch.
2037
REFERENCES
[1] W. D. Stevenson, Jr., "Elements of Power [10] T. E. DyLiacco, "Real Time Computer Control
System Analysis", Chapter 10, Third Ed., of Power Systems", Proceedings of the IEEE,
McGraw-Hill, 1975. July 1974, pp. 884-891.
[2] L. K. Kirchmayer, "Economic Operation of [11] T. E. DyLiacco, "System Control Center
Power Systems", Wiley, 1958. Design", Proceedings 1975 Henniker Confer-
ence, pp. 196-230.
[3] L. K. Kirchmayer, H. H. Happ, G. W. Stagg,
J. F. Hohenstein, "Direct Calculation of [12] H. H. Happ, "Optimal Power Dispatch-A Com-
Transmission Loss Formula-I", AIEE Trans., prehensive Survey", IEEE Transactions on
Part III, Vol. 79, December 1960, pp. 962- Power Apparatus and Systems, Vol. PAS-96,
969. No. 3, May/June 1977, pp. 841-854.
[4] H. H. Happ, J. F. Hohenstein, L. K. Kirch- [13] T. E. DyLiacco, discussion to [12].
mayer, G. W. Staff, "DirectCalculationof
Transmission Loss Formula-II", IEEE Trans., [14] B. F. Wollenberg and W. O. Stadlin, "A Real
Vol. PAS-83, 1964, pp. 702-707. Time Optimizer for Security Dispatch",
IEEE Trans., Vol. PAS-93, Sept.-Oct. 1974,
[5] W. S. Meyer, "Efficient Computer Solution pp. 1640-1649.
for Kron and Kron-Early Loss Formulas",
PICA Conference Proceedings, Minneapolis, [15] C. L. Wadhwa, J. Nanda, "New Approach to
Minn., June 3-6, 1973, pp. 428-432. Modified Coordination Equations for Eco-
nomic Load Dispatch", Proc. IEE, Vol. 123,
[6] R. B. Shipley and M. Hochdorf, "Exact Eco- No. 9, Sept. 1976, pp. 923-925.
nomic Dispatch-Digital Computer Solution",
AIEETransactions, Part III (PAS), Vol. 75, [16] A. P. Sage and J. L. Melsa, "Estimation
December 1956, pp. 1147-1153. TheorywithApplications to Communications
and Control", Appendix A, McGraw-Hill, 1971.
[71 W. F. Tinney and C. E. Hart, "Power Flow
Solution by Newton's Method", IEEE Trans- [17] W. F. Tinney, "Compensation Methods for
actions on Power Apparatus and Systems, Network Solutions by Optimally Ordered
Vol. PAS-86, pp. 1449-1460, November 1967. Triangular Factorization", IEEE Trans.,
Vol. PAS-95, July-August 1976, pp. 123-127.
[8] H. W. Dommel and W. F. Tinney, "Optimal
Power Flow Solutions", IEEE Transactions [18] F. L. Alvarado, "Computational Complexity
on Power Apparatus and Systems, Vol. PAS- in Power Systems", IEEE Trans., Vol. PAS-
87, pp. 1866-1876, October, 1968. 95, July-August 1976, pp. 1028-1037.
[9] T. E. DyLiacco, B. F. Wirtz, D. A. Wheeler, [19] F. L. Alvarado, "A Comparative Study of
"Automationof the CEI System for Security," Power Network Algorithms", 14th Annual
1971 PICA Conference Proceedings, pp. 93- Allerton Conference, Sept., 29-Oct. 1, 1976,
102. Monticello, Illinois.

Fernando L. Alvarado (M'67)


ws born in Lima, Peru on
September 15, 1945. He re-
ceived the B.E.E. and Pro-
fessional Engineering degrees
from the National University
of Engineering in Lima, Peru,
in 1966 and 1967 respectively.
He received the M.S. degree
| lSi from Clarkson College of Tech-
nology in 1968 and the Ph.D.
degree from the University of
Michigan in 1972.
From 1972 to 1974 he was with the Univer-
sity of Toledo, Toledo, Ohio, where he was in
charge of courses and laboratories in power
systems and electromechanical energy conver-
sion. In 1975 he joined the University of Wis-
consin at Madison, where he is currently an
Associate Professor. His research activities
have been centered around computational prob-
lems in power systems.
Dr. Alvarado is a member of Eta Kappa Nu
and Sigma Xi.
2038
Discussion
(D) K. Takahashi, J. Fagan, M. Chen, "Formation of a sparse Z-bus
R. P. Sood (Regional Engineering College, Kurukshetra, India): The impedance matrix and its application in short circuit studies,"
author is to be complimented for presenting a novel and useful IEEE PICA Conference Proceedings, Minneapolis, pp. 63-71,
algorithm for economic dispatch of large scale power systems. The 1973.
discusser would offer the following comments and seek some clarifica-
tions. Manuscript received February 2, 1978.
The economic dispatch of a power system is based upon scheduling
generation so that the cost of supplying the system load is minimised
while accounting for the effect of generator operating limitations and
transmission losses. Earlier loss formulas involved the application of Walter O. Stadlin (Leeds & Northrup Company, North Wales, PA):
co-ordination equations using Kron's classic B coefficient method to The author has shown how the classical definition of incremental losses
represent system losses. However, the calculation of loss formula itself can be related to the loss sensitivity coefficients derived from the load
involves considerable computational effort and the simplified form of flow Jacobian. Generation schedules obtained from either an accurate
loss formulas introduce appreciable errors due to assumptions made in loss formula or from the Jacobian will be very nearly the same. The
the network behaviour. Practical procedures avoid this difficulty by classical approach computes a system Lambda (incremental cost of
using a set of B matrices for different network structures and loadings. delivering power to the loads) while the contemporary approach, utiliz-
Meyer [5] has presented efficient computer solution method using ing the Jacobian, computes an incremental cost of delivering power to a
Kron's loss formulas. Kirchmayer [2, 3, 4] and Happ [12] also have specified reference bus. When the Jacobian is used to derive transmis-
made very valuable contributions in this area. sion loss penalty factors it is not necessary to compute a system Lambda.
Podmore [A] has developed a simplified and improved method The manner in which system Lambda and the reference bus in-
from computational point of view for economic power dispatch using cremental cost are related depends on ones viewpoint. Dr. Alvarado has
D.C. load flow loss formula. In the D.C. load flow loss formula, the chosen to relate his derivations to system frequency. Another popular
generation on the slack bus is the dependent variable whereas in Kron's point of view is to relate individual bus load changes to total system
B coefficient loss formula, the total system load is dependent variable at load change. Most loss formulas are derived on this basis, ignoring fre-
a hypothetical load centre. Consequently calculation of dP/6P8,i from the quency effects. The following equations summarize the resulting rela-
D.C. load flow loss coefficients give the incremental loss when an incre- tionship between System Lambda and the Reference Bus Lambda.
ment of power is delivered from generator i to the slack generator. On System Lambda Method-
the other hand, calculation of dPL/tAPi from Kron's B coefficients ac a P cPL
give the incremental loss when an increment of power is delivered from ?\s = - -&= (& l- )
generator i to the total load. s PD Pg D Pg \
The development of the modern nonlinear programming methods for any generator
for optimal economic dispatch are based on the Carpentier's formula- fo a g t
tion of maximising or minimising a function of several variables subject p CR bCR aPR ?
to constraints. The present advances in this area are based on the im- s =PD - = -PD -PD
PR
=

provement of Sasson's [B], Dommel and Tinney [8] approaches in the


area of power system static optimization methods, to include security for reference bus generator
and other constraints.
Recent interest is centered around refinement and extension of Reference Bus Lambda Method: ?\R =
aCA
p
(ap
Dommel and Tinney's approach. This is evident, as in the last 3-4 Cg g
years, significant advances have been made for calculation of incremen- Incremental Generating Cost:
tal losses and penalty factors using transposed Jacobian of Newton's g
load flow and sparsity techniques for maximum overall efficiency as
well as economy of programming. Its usefulness both as an on-line and Loss Sensitivity Vector:
off-line computational tool is quite promising. bPR t3-1 jR =
The author has demonstrated its application for medium to very
large systems consisting of as many as 5837 buses and 943 generators.
P
g $
-= [J] J-R -2d,_q
Table 1 of the paper presents time required for the various operations of Reference Bus Versus
the proposed algorithm. The discusser would be interested to know Total Load Sensitivity:
whether the author has applied the above algorithm to decoupled N
Newton load flow method to these Systems, since fast decoupled load _PR _PR bPdi +PR_Qdi
flow algorithm [C] in Dommel and Tinney's approach achieves signifi-
cant savings in computer storage. If so, how do the computational times
-D =
___

di TD
aP
aP a d

of the proposed algorithm compare with the decoupled version? It


would also be interesting to compare the overall storage requirements
Bus Load Factors:
6Pdi aQdi
and arithmetic operations in both methods as well as with D.C. load -P and a
flow loss formula.
Would the author please comment about the extension of the effi- In summary the numerical value of system Lambda will depend on
cient sparse Z-Bus [DI for computing B coefficients as they can be con- -how the incremental losses are defined. The Jacobian or reference bus
sidered as an equivalent network model containing generator nodes, ho the definedneedto
other aredoesn't To acobian incrental
boundary nodes and an equivalent load point for application to
economic dispatch.
approach, onghthebtherhand, assume any incremental
relationship among the bus loads. Only the magnitudes of the bus loads
I once again commend the author for presenting a valuable are required to solve for the Jacobian.
theoretical and a practical contribution in the area of economic dispatch Manuscript received February 15, 1978.
of large scale power systems.

REFERENCES B. Stott and 0. Alsac (CEPEL, Rio de Janeiro, Brazil): The paper re-
publicizes a technique for computing penalty factors directly from the
(A) Podmore, R., "A simplified and improved method for calculating bus voltages. Thus the method is applicable in modern control centers
transmission loss formulas," in Proc. 8th PICA Conf., 1973, pp. where up-to-date values of these voltages are provided by on-line state
428-432. estimation/load flow. The technique, which is very simple and elegant,
(B) Sasson, A. M., "Nonlinear programming solutions for load flow, has received frequent brief treatments (many not referenced) in the
minimum loss, and economic dispatching problems" IEEE literature. It is not essential to employ Lagrange functions or multipliers
PAS-88, pp. 399-409, 1969. in the derivation. The direct approach used in say, Refs. A and 1, seems
(C) Shoults, R. R., "A simplified economic dispatch algorithm using to be clearer and more obvious. The present derivation is quite unusual.
decoupled network models" IEEE paper No. A 77 738-8. It is constructed around the assumption that the load-frequency
2039
characteristics are relevant to the basic economy-dispatch calculation, A crucial point about which clarification would be appreciated is
and ends up showing this assumption to be incorrect. In the process, the this:
technique is considerably more obscured than elucidated. Since PD essentially acts as a slack generator thereby allowing all Pg,
The industry is interested in quantifying the benefits of deriving (including Pre) to be independently variable, then the best strategy
penalty factors on-line. In some cases, large fuel savings have already (although unrealistic) to minimize F appears to be to decrease all Pi to
been estimated (see Ref. B, for example). On-line penalty factors have zero. We can see this by minimizing
the important advantage over off-line B-coefficients that they corres- m m
pond to the actual configuration and loading of the network. The ac- L = I Fj (Pa) - A[ I Pl - PD - PLU
curacy with which the penalty factors need to be computed on-line may j=1 =1
not itself be too critical: simpler and/or faster variants on the technique
under discussion may well be quite adequate.
The direct use of the Jacobian matrix is most appropriate when its
triangular factors (this means storing the lower one) are automatically
available from an ordinary or optimal Newton power flow. Ref. C with respect to all Pg. It can be argued that PD is a dependent variable
shows how to compute the penalty factors from a fast decoupled load (depending on all Pg,), hence differentiating partially with respect to Pg.
flow, avoiding the explicit Jacobian matrix. The method is iterative in holding all other Pe constant yields,
principle, but starting with previous values, a single iteration should be
adequate in practice. dFi/dP8i - A [1 - a PD/ a Pgi - a PL/ a Pgi] = 0
The most computationally attractive version of the technique is ob-
tained where it is justified to assume that following generator MW However from (A) it follows that
rescheduling, the change in active losses is nearly all due to the change
in MW flows in the network. Then three quarters of the Jacobian 1 - a PD/ a Pgi - a PL/ a P8i = 0
matrix drops out of the analysis, leaving the formula:
a
[PL/ a P] = [HI]-' a PL/ a ej so that the conditions for a minimum become,
where [H] is the Jacobian submatrix [ a P/ a 0e. This version, presented
in Ref. 1 (page 263), simplifies further for EHV systems, replacing [H] dF,/dPgi = 0 (B)
by the dc load flow matrix as used in the fast decoupled load flow.
for i = 1, 2, ..., m. One can see from (B) that the usual coordination
REFERENCES equations do not hold under the assumption of a free PD, since no cost
was placed on this variable as is the case when Pgref is a dependent
A. H. H. Happ, "Optimal Power Dispatch", IEEE Trans. Power variable. The strategy is therefore to decrease P.i and let PD, for which
App. Syst., vol. PAS-93, pp. 820-830, May/June 1974. no cost is incurred, take up the slack.
B. J. L. Scheidt, "A Survey of Power System Control Center Such an ambiguity does not appear if the load is constant and the
Justifications", paper F 78 016-8, presented at IEEE PES Winter reference generator is allowed to be a dependent variable, interpreting
Meeting, New York, 1978. a PL/ a P8g as the partial derivative of PL with respect to Pgi holding all
C. 0. Alsac and B. Stott, "Decoupled Algorithms in Optimal Load other Pj, constant except for Pgrefw
Flow Calculations", paper A 75 545-4, IEEE PES Summer To summarize then, this paper describes an elegant method of
Meeting, San Francisco, July 1975. determining aPL/aPPi when all other P,,, including that of the swing
bus, are held constant (i.e. all Pgi are independent). However the
Manuscript received February 21, 1978. assumption needed to derive these quantities (frequency dependent
load, PD) makes the usual coordination equations not applicable. In
fact, it appears, that unless an additional cost or constraint are imposed
E. F. Richards (University of Missouri-Rolla, Rolla, MO) and A. B. R. on PD, no minimum of F will exist, except for the trivial case of F = 0
Kumar (Michigan Technological University, Houghton, MI): This when Pg, = 0.
paper is an excellent application of using the transposed Jacobian ap-
proach for obtaining a fast method of calculating penalty factors. The Manuscript received February 23, 1978.
author has made effective use of the information available from the
load flow program which, of course, is always used as a part of an
economic dispatch program.
We are a bit confused from the author's explanation of the Kavuru A. Ramarao (Cleveland Electric Illuminating Co., Cleveland,
algorithm. In the paper, the author states on Page 6, "The factored OH): This is an interesting tutorial paper.
Jacobian from the last Newton power flow iteration was used". Does
this imply that the penalty factors and generator powers are recomputed
as
This tecu eas first implemented in an off-line computer pro-
gram as a result of research in Ref. 1 and the results used in the ADS
in every iteration of the load flow program? (analog) system at the Cleveland Electric Illuminating Co. Subsequently
Would the author explain why he chose to use both refactorization the same technique was implemented in real-time at the System Opera-
and compensation? Isn't it true that normally one or the other would be tion Center and has been in operation since 1973.
used? With the subsequent replacement of the Newton load flow by the
Fast Load flow we found the calculation of A' by the eq 33 to be ineffi-
Manuscript received February 23, 1978. cient in the use of time and core.
A decoupled version of solving the eq 33 to obtain (' is possible by
the use of the same triangular factors of (' and (3" used in Fast Load
F. D. Galiana (McGill University, Montreal, Quebec, Canada): I found flow instead of triangularizing the Jacobian transpose matrix.
this paper by Dr. Alvarado very stimulating and well written. He has Without being aware of this decoupled algorithm first reported in
discussed an established area in a very clear manner, while at the same Ref. 2, we obtained the same algorithm in our independent research.
time developing some very interesting results. This has been implemented in real-time and has been in operation since
In this paper Dr. Alvarado very neatly calculates the partial November 1977.
derivatives aPL/.aaP under the assumption that all other generations
including that of the swing bus are constant, in addition to the base REFERENCES
loads and the generation bus voltages. This derivative is well defined in
this case by the introduction of a frequency dependent load component, (1) T. B. Dy Liacco, "Control of Power Systems Via the Multi-Level
PD, which acts as a new dependent variable or state, adjusting its value Concept" SRC-68-19, Systems Research Center, Case Western
to compensate for the change in P., and PL through Reserve University, Cleveland, 1968.
8 ~~~~~~~~(2)
0. Alsac and B. Stott, "Decoupled Algorithms in Optimal Load
m (A) Flow Calculations" A75 545-4, IEEE Summer Meeting 1975.
I dP,. = dPL + dPD
i-1 Manuscript received February 27, 1978.
2040
F. L. Alvarado: The author thanks the discussers for their interest in Both Dr. Sood and Dr. Ramarao mention the fast decoupled ver-
this paper. sion of the algorithm, which was originally reported by Alsac and Stott
Mr. Stadlin's comments provide a valuable observation, evident (Reference [C] of Stott and [21 of Ramarao). It is gratifying to see that
from the paper itself: that the calculation of the traditional incremental real-time operation of the algorithm has been achieved.
losses (as well as system A) is dependent on the incremental "bus load Dr. Galiana's comment is most interesting and worthy of detailed
factors", whether these factors are assumed to be a result of frequency- consideration. Recall that the equations that impose the physical system
dependence of loads and generators (as done in the paper) or any other constraints are (15). They can be compactly written as:
dependence mechanism. The formulas provided by Mr. Stadlin provide h(P, v PD) = 0 (A 1)
an alternate direct means of comparison between the "system A" and TicdisThe i t fAl
the "reference A". These formulas are a valuable addition to the stated objective of economic dispatch is to find the total incremental
objective of the paper, which was to illustrate these relationships. It dodtA df(P,)/dP, (A.2)
should be perhaps noted that some of Mr. Stadlin's partial derivatives
would be better denoted as total derivatives. Subject to the additional constraint that the load should not drift
In reference to Dr. Richards and Dr. Kumar's discussion, it should from its present value:
be mentioned that the calculation of penalty factors is not done every PD = O (A.3)
iteration but only after a "converged solution" is obtained, therefore T a c
explicit, but it is, of
the Jacobian in question is the one for the last iteration. Network com Thc s additional constraint tsseldomaremade
pensation (also known as the Matrix Inversion Lemma) is used merely course, very real. At this point there total (m-d)
only independent
for the purposes of the derivation and not for any computational pur- variables Therefore evaluation of the derivative (Ao2)aerniael
poses; its use illustrates the power and generality of the concept. As in- resortingeither to the "chain rule" of differentiation ort alteri notI
that P ie e t
dicated in the paper, the computational algorithm uses a correctly fac- (and preferably) to Lagrange multipliers, Noticea also one
tored Jacobian and no explicit use of network compensation is needed. treated as a known quantity. Rather (Ad3) is treated as is tsubl
Dr. Stott is of the opinion that the "direct" approach to proving tconc whilePa remains that
among the dependent variables It is this subtle
leads to the apparent paradox outlined by
the method is preferable. There is evidence, however, that what may conceptual difference
seem obvious now hasn't always been so. The use of "distribution fac- A new variable (the losses) along with a new equation (equation 13)
tors" (whether those factors are interpreted as "frequency originated", can be appended to the original set without altering the problem in any
as in the paper, or not) is still widely regarded as a necessary step for cay
way:
economic dispatch calculations. A direct path leading from one for- (13)
mulation to the other appears to have had positive tutorial value, as m n PO
evidenced by other comments received by the author. Although no at- PL = E - 1: di - PD
tempt at an exhaustive list of references was made, we agree that all 1 i 1
three references provided by Dr. Stott are very pertinent.
The approximations mentioned by Dr. Stott involve using only ac- Thus, the original problem of evaluating (A.2) subject to (A. 1) and
tive powers and replacing [HI with the dc load flow matrix. These are (A.3) has been embedded into a larger equivalent problem. This larger
clearly desirable if further computational and/or storage gains are problem can then be solved without (at first) taking into consideration
desired at the expense of an (admittedly small) sacrifice in accuracy. the constraint imposed by (A.3); only at the end is the constraint (A.3)
These all follow rather directly from the general procedure presented in imposed. It is easy to verify that the evaluation of (A.2) is not affected
the paper, while the converse is not true. Thus, only the most general by (A. 3) as long as PD is, in fact, zero. This is precisely what is implicitly
case was deemed worthy of detailed study. done during "traditional" economic dispatch; this is also done in the
The paper by Podmore mentioned by Dr. Sood is indeed a most in- paper for illustrative purposes. A more straightforward approach to the
teresting one, particularly when one realizes that it was presented entire problemlisaof course, not to introduce(3)andtomerely replace
alongside [5]. It can now, in retrospect, be interpreted as an approxima- the known value for PD in (A. 1). Thus (A-2) can be evaluated subject to
tion to the more general transposed Jacobian procedure, as mentioned (A. 1) alone. This approach leads to the more direct but less enlightening
by Dr. Stott. With regard to the computation of B-coefficients using the procedure referred to by Mr. Stadlin and Dr. Stott and previously used
sparse inversion algorithm of Takahashi, this author remains somewhat by Happ and others to introduce the transposed Jacobian method.
skeptical as to its practicality and efficiency for the intended purpose.
Manuscript received May 11, 1978.

Вам также может понравиться