Vous êtes sur la page 1sur 2

PROCEEDINGS OF THE IEEE, APRIL

712

1975

Proceedings Letters
TABLE I
This section is intended primarily
for rapid dissemination of brief
CONVERGENCE COMPARISON BETWEEN D L F A N D
reports on new research results in technical areas of interest to IEEE
NLF; 5 3 0 0 - B u s CASE
members. Contributions are reviewed immediately, and acceptance is
determined by timeliness and importance of the subject, and brevity
and clarity of the presentation. Research letters must contain a clear
Decoupled Load Flow
Newton
Load
Flow
concise statement of the problem studied, identifv new
results, and Unsolved
Unsolved
make evident their utility, importance, or relevance to electrical engiIteration XIAPI x l a Q l Equations XIAPI X I A Q I
neering. Key references to related literature must be given.
Conmbutions should besubmitted
in triplicate to theEditor,
1
402000 476 000
9172
414
000
409 000
PROCEEDINGS
OF THE IEEE, 345 &st 47 Street, New York, N. Y. 1001 7.
2
69 700 10 300
6339
26400
50000
4303
m e length should be limited to five double-spaced typewritten pages,
880
1690
3560
3
3970
6920
counting each illustmtion as a half page. An abstract of 50 wonis or
4
390
94
44
45
66
less and the original figures should be included. Instructions covering
71
40
4
10
5
31
abbrevkrtions, the form for references, general style, and prepamtion
6
24
30
2
o f figures are found in Information for IEEE Authors, available on
I
12
25
0
request from the IEEE Editorial Department. Authors are invited t o
suggest the categories in the table of contents under which their letters
best fit.
After a letter has been accepted, the authors company or institution
T A B L E I1
d l 1 be requested to pay a voluntary charge o f $70 per printed page,
TIMECOMPARISON (SECONDS)
BETWEEN D L F AND
calculated to the nearest whole page and with a$70 minimum, to cover
NLF; 5 3 0 0 - B u s CASE
part of the costo f publication.
~

Comments on Review of Load-Flow Calculation Methods


A. M. SASSON, W. SNYDER,

AND

M. FLAM

The load-flow program is certainly the most frequently


used computerprogramforpowersystemsapplications.
It consumes the
largest computer time per year. This time has been growing as power
system planners see the need forsolving ever largercases in a succession
of
of base case andmanycontingentsolutions.Thedevelopment
has greatly simplified data handling
adequate data management schemes
for the user. Typically, a large number of cases are saved (data plus
solution) for future use. Through the use of auxiliary programs, such
as a Common Format program that interfaces a utilitys data files with
those of other utilities, a Load-Flow Equivalent program that sets up
portions of systemsnotwantedexplicitlyfora
anequivalentfor
certain study, and Change Case programs that incorporate many data
changes to an existing saved &e., an experienced user can quickly set
up a variety of cases. In our installation, most load-flow cases are in
excess of lo00 buses, some being in the 5000-7000 range.
Thegreatdemandthat
load-flow applications place oncomputing
resources accounts for the continued interest in the development of
techniques that are faster, exhibit quicker convergence for a variety of
use of core. We thus seein
systemconditions,andmakeefilcient
Stotts paper a history of load-flow methods that are now 20 years
old. When Newtons method was introduced seven years ago, it seemed
to be the defdtive step which would replace previous programs once
andfor all. Inour discussion, we wouldlike to offersomerecent
experiences obtained with Stotts decoupled load flow which appear
has comeforreplacement
of the classical
to suggest thatthetime
Newtons method for those utilities solving very large w s .
After writing a Fortran version of the decoupled load flow (DLF),
confvming the results of the original paper, andcomparing it to a
Fortran version of a Newton load flow (NLF), it was decided to write
an assemblybguage routine of the DLF for proper comparison with
our production NLF program. This work has recently been completed.
Table I presents a comparison of the convergence characteristicsof the
Manuscript received October 1, 1974.
The authors are with the American Electric Power Service Corporatiqn, New York, N.Y. 10004.
B. Stott, Proc. IEEE, vol. 62, pp. 916-929, July 1974.

Modules

DLF

3 assembly
Input and data
Voltage solution
2
Line flow calculation
Output routines 202

3
19
2
202

~~

Buses

5201
488
10
0

NLF

61

two methods for a typically large system having 5300 buses and 8722
branches. The case contained no tapchangingunderload(TCUL)
transformers, interchange control, or reactive limit constraints which
are normallyaccountedforbycorrectionsbetweeniterations.
A
tolerance of 1 MW and 1 Mvar was used in both cases. The difference
in slack bus injection between the two solutions
was of 4.9 MW and
0.3 Mvar, whichshowed thattheywerequite
close toeachother.
The Unsolved Equations column refers to the sumof both P and Q
equationsineachhalf-iteration
of DLF, while the UnsolvedBuses
column refers to buses where either P,Q,.or both had not converged.
The results of Table I can be considered typical in that DLF requires
one or two iterations more than NLF, although there are cases where
they both converge in thesame number of iterations.
Great care was exercised in making a proper time comparison. The
NLF production program is a modular program for which the voltage
solution is a module by itself. This module is apart from all input and
output modules and those that form the admittance matrix and determine the optimal ordering. Some parts of the voltage solution module,
which shall be referred to as pre- and postsolution, were identical in
530O-bus case were
thetwo programs. Corerequirementsforthe
650K forthe NLF. Forameaningfulcomparison,
thereductionin
corerequirementsfortheDLF
was compensatedbykeepingthe
admittance matrix in core between the P and Q half-iterations and by
storing in the same locations sine and cosine terms between the following Qand P half-iterations.CorerequirementsforDLFwerethen
648K. Table I1 compares timings in seconds for the various modules
of both programs for an IBM 370/168 machine. This is task time or
actual central processing unit (CPU) time.
Table I1 shows that the voltage solution has been made faster by a
factor greater than 3. It has been found that, for large cases, a factor
in the 3-4 range is typical. For smaller cases, the advantage of DLF
685-bus 1393over NLF is reduced considerably. For example, for a
line case, the voltage solution timings were 3.1 and 5.4 s, respectively.
It is of some interest to further break up the voltage solution time
19 s
for DLF. Table 111 presents the components that went into the
quoted in Table 11. Given that the programs have identical pre- and
postsolution routines which took
4 s, DLF reduced a 56-s figure of
NLF to 15 s. Table 111 shows that an iteration of DLF is of the order
of 1 s for the 5300-bus case, with 80 percent of this time going into
mismatch computations anddata handling.

PROCEEDINGS LETTERS

71 3

TABLE Ill
VOLTAGESOLUTION
COMPONENTTIMINGS
FOR

DLF; 5300-BusCASE

Time (s)

Function

2
Resolution
5
Triangularization of B
Triangularization
3
of
B
7
Timetakenfor
seven iterations
80-percent mismatch computation and data handling
10-percent forward-backward substitutions
10-percent voltage corrections
Postsolution
2

19

Total
time

Thethree tablespresentedshowadefinite
advantage in replacing
V F by DLF in those
installationswhere large cases arecommonly
solved. In the 5000-bus case range, DLF does not leave much room for
furtherimprovements in reducing significantly computer timings. A
savings to be achieved by
conservative projection of yearly estimated
the conversion to DLFis in our installation in excess of $100 000.
We have followed Stottssuggestion of using a Taylor-series expansion
for cosine and sine computations within mismatch calculations. It was
found that in some large cases the 3-term cosine, 2-term sine expansion
affects convergence by requiring one extra iteration as compared with
using the4-term cosine,3-termsine
expansion. These cases tend to
lines with largeangle differences. Alincludeagreaternumberof
though retaining an extra term in the expansion has a negligible effect
on timings, one alternative lies in the use of a Chebyshev economization of the Taylor series. The Chebyshev polynomialhasauniform
errorcharacteristicwithin
its defined range,while the Taylorpolynomial is better at the center point of the range and gets progressively
worse toward the end points. For example, the *45 range Chebyshev
and the Taylor sine expansions retaining two terms are

c = 0.9990211 e - 0.1603402 e3
T = e - 0.1666666 e 3 .
Themaximumerror of C in the range -45,45 is 1.54 X
while
the error of Tat +45is 23.71 X
Some convergence problems have been detected in certain cases with
the NLF programs. In these cases, NLF voltages
both the DLF and
dropped more severely than DLF voltages, but it is too early to form
any definite conclusions on this point.
to timings of outputroutines presented in
A final pointrefers
Table 11. It reveals quite clearly that further payoffs lie in this area,
possibly by reducing actual output by moreanalysis with less printing.

Comments on Review of Load-Flow Calculation Methods


R. D. MASIELLO

AND

B. F. WOLLENBERG

Until very recently, automatic generation control and economic


dispatch were the major applications programs utilized in real-time control
computers for electric power systems. The growingawareness of the
generalusefulness of real-time contingencystudieshas led to the inclusion of load-flow-based functions in manyreal-time control systems.
These programs now often dominate machine sizing and timing design
considerations, overshadowing the actual control functions. The
pressure to reducecore usage, central processing unit (CPU) time,and
generalmachine loading of any applications program can be intense,
given all theconsiderations involved inthe design of suchsystems.
Theshadowcosts
involvedin thesettingofconstraintsonthese
Manuscript received October 2 5 , 1974.
The authors are with Lee& & Northrup Company, North Wales, Pa.

TABLE IV
12WBus 1667-LINE LOAD-FLOW
TIME
~

Function

Time

Newton Load Flow


1-0
Nevton Load Flow
Solution
Fast Lkcoupled SolutionTotal
F a c t o r i z a t i o n of B&B
Solution
Repeat Contingency
Solutions
8-11

25 s
45 (I
18 s
7 s

11

usage parameters are very high in that increased usage in one area then
forcesapainfuleconomictradeoffbetweensoftwareeffort
insome
other area versus overall hardware improvement costs.
Stotts fast decoupled load flow is therefore an attractive alternative
to Newton load-flow algorithms for use in such systems. We would like
to share our experiences with the fast decoupled load flow anddescribe
someadvantages of using it whichhave not beenmentioned asof
this writing.
I. TIMING
First, our comparative timings are presented in Table IV. These times
are foraFortranfastdecoupled
load flowandapartially
Fortran,
partially optimized assembly-language Newton load flow, running on an
XDS Sigma 5 computer.
The relative improvement realized has been obtained with load flows
as small as 246 buses.
11. PROGRAMMING
A carefully coded Newton load-flow solution algorithm (partially in
assemblylanguage) requires roughly 2000 32-bit words of storage. A
fast decoupled program performing the identical tasks requires
about
1000 words.These
savings mayseem trivial whenextremely large
systems (upwards of several thousand buses) are involved, but for typical small-to-medium-sized systems it becomes significant. For instance,
a straightforward implementation of a 1200-bus Newton load flow requiresonly about 30Kwordsof core.Most important,thefastdecoupled method i s an inherently simpler algorithm.
Also, the coding of the fast decoupled method is readily modularized
(relative to aNewton) so thatstructuredprogramming can be used
without serious loss of efficiency. A structured implementation of the
simplerlogic results indecreasedoverall costs for design, coding, debugging, and code optimization. These last points should be of interest
to anyone who has avoided implementing a Newton load flow for fear
of the complexity involved.

111. APPROXIMATIONS
TO THE TRIGONOMETRIC
FUNCTIONS
It has been suggested that truncated polynomial expressions be used
in computing the mismatch vectors where sine and cosine functions are
required. This improves machine speed periteration;howwer,too
severe an approximation hinders iterativeconvergence.
Asuperiormethod
is tostorethe
rectangularcoordinatesofthe
voltage vector and thevoltage magnitudes, updating these for magnitude
corrections by the use of 1) and updating these for angular corrections
by the use of 2). A considerablesavingin machine time is achieved.
Sinceangular corrections quickly become small, efficient polynomial
expansions can be used without loss of accuracy. Themismatch,of
course, is exact to machine resolution. The cost is the use of a vector
of bus voltage magnitude.

I ) Magnitude Corrections:
Let AVi be the solved voltage magnitude correction at bus i. Then the
rectangular components are corrected by
~ + AVi/Vi)
a) V R =~ V R (1
b) Vri = Vzi ( 1 + A Vi/ Vi)
and the magnitude, trivially, by
C) Vi = Vi + A Vi.
2) Angle Corrections:
Mi is the solved correction to busi angle:
a) C = cos M i
S = sin Mi
B. Stott, P~oc.IEEE, VOI.
6 2 , pp.

916-929,

July

1974.

Vous aimerez peut-être aussi