Sie sind auf Seite 1von 11

FAST DECOUPLED LOAD FLOW

B. STOTT 0. ALSAq

Power Systems Laboratory


University of Manchester Institute of Science and Technology, Manchester, U.K.

ABSTRACT

This paper describes a simple, very reliable and tion between MW and MVAR flows in a power system,
extremely fast load-flow solution method with a wide by mathematically decoupling the MW-F and MVAR-V
range of practical application. It is attractive for calculations 4-11
accurate or approximate off- and on-line routine and
contingency calculations for networks of any size,and The method described in this paper is a rational
can be implemented efficiently on computers with res- integration of some of these ideas. It combines many
trictive core-store capacities. The method is a dev- of the advantages of the existing "good" methods. The
elopment on other recent work employing the MW-(/ algorithm is simpler, faster and more reliable than
MVAR-V decoupling principle,and its precise algorith- Newton's method, and has lower storage requirements
mic form has been determined by extensive numerical for entirely in-core solutions. Using a small number
studies. The paper gives details of the method's per- of core-disk block transfers its core requirements are
formance on a series of practical problems of up to similar to those of the Gauss-Seidel method. The
1080 buses. A solution to within 0.01 MW/MVAR maxi- method is equally suitable for routine accurate load
mum bus mismatches is normally obtained in 4 to 7 flows as for outage-contingency evaluation studies
iterations, each iteration being equal in speed to 11 performed on- or off-line.
Gauss-Seidel iterations or 1/5th of a Newton itera-
tion. Correlations of general interest between the
power-mismatch convergence criterion and actual sol- NOTATION
ution accuracy are obtained.
APk + jAQk = complex power mismatch at bus k, where
INTRODUCT ION
AP k = P5s-V
k Vk
~Vm(Gkmcoskm+BkmsinOkm)
Gcs Bki (1)
Load-flowcalculations are performed in system mek
planning, operational planning and operation/control.
The choice of a solution method for practical appli- AQk =
9sVk- Vm(G sin km--B km cos km
k Vk X m km
mEk
(2)
cation is frequently difficult. It requires a care-
ful analysis of the comparative merits and demerits
of the many available methods1 in such respects as Pkp + i kp = scheduled complex power at bus k
storage, speed and convergence characteristics, to
name but the most obvious, and to relate these to the voltage angle, magnitude
requirements of the specific application and comput- Ok ' Vk = at bus k

ing facilities. The difficulties arise from the fact


that no one method possesses all the desirable fea- ( km (3k m
tures of the others. For routine solutions Newton's
method2 has now gained widespread popularity. However
Gkm + jBkm (k,m)th element of bus admittance
it is limited for small-core applications where the matrix [G] + j[BI
weakly-convergent Gauss-Seidel-type method is the
most economical, and it is not as fast as newer AO , AV = voltage angle, magnitude corrections
methods for approximate repetitive solutions such as
in AC security monitoring4'5. mek signifies that bus m is connected to bus k,
including the case m=k, and [ ] signifies vector or
Numerical methods are generally at their most matrix.
efficient when they take advantage of the physical
properties of the system being solved. Hence, for max AP ,maxIQ| = largest absolute elemnt of [APIJA Q]
example, the exploitation of network sparsity by
ordered elimination and skilful programming in the maxlev ,maxieSi
= largest absolute bus-voltage magni-
Newton and other methods has been a very important tude error, branch MVA-flow error, respectively.
advance3 . More recently, attention has been given
to the exploitation of the loose physical interac- DERIVATION OF BASIC ALGORITHM

The well-known polar power-mismatch Newton


method2 is taken as a convenient and meaningful star-
ting point for the derivation. The Newton method is
the formal application of a general algorithm for
solving nonlinear equations, and constitutes success-
ive solutions of the sparse real jacobian-matrix eqta-
tion
L AP iH N A

Paper T 73 463-7. recommended and approved by the IEEE Power System


Engineering Committee of the IEEE Power Engineering Society for presentation at
AQ L

The first step in applying the MW-O/MVAR-V decou-


AV/V
I (3)

the IEEE PES Summer Meeting & EHV/UHV Conference, Vancouver, B.C. Canada,
July 15-20, 1973. Manuscript submitted February 12, 1973; made available for pling principle is to neglect the coupling submatrices
printing May 15, 1973. [N] and [J] in (3), giving two separated equations

859

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
when [AP/VI and [AQ/V] are zero. A brief account of
[AP] = [H] [A)] (4) the alternative decoupled algorithms investigated is
given in Appendix 1.
[AQ] = [LI [AV/VI (5)
APPLICATION TO PRACTICAL LOAD FLOW SOLUTIONS
VkVm (G km sinkm
where H1m = LLkm = VkV km -Bkm cos km ) for mOk
Iteration scheme
2 2
Hkk = BkkVk-Qk and Lkk = -BkkVk+Qk Of the iteration strategies tried, undoubtedly
the best scheme for all applications is to solve (8)
Equations (4) and (5) may be solved alternately as a and (9) alternately, always using the most recent
decoupled Newton method11,re-evaluating and retriang- voltage values. Each iteration cycle comprises one
ulating [H] and [LI at each iteration,but further solution for [AO@ to update [01 and then one solutian
physically-justifiable simplifications may be made.In for [AV] to update [V],termed here the (l1,lV) scheme.
practical power systems the following assumptions are Separate convergence tests are used for (8) and (9)
almost always valid: with the criteria:

csOk 1; Gkmsinkm km; Qk BkkVk J


max AP < cp , maxiAQi < cq (10)

so that good approximations to (4) and (5) are: The flow diagram of the process is given in Fig.
1. The convergence-testing logic permits the calcul-
[AP] = [V.B'.V] [AO] (6) ation to terminate after a [(A] solution (called a 2
iteration). It is also possible, though unusual in
[AQ] = [V.B" .V] [AV/V] (7) practice, to terminate after more than one consecutive
[AO] or [AV] solution if [AQ] or [API respectively do
At this stage of the derivation the elements of not need converging further.
the matrices [B'] and [B"] are strictly elements of [-BI
The decoupling process and the final algorithmic forms
are now completed by:

(a) omitting from [B'I the representation of those net- KP=KQ=1l


work elements that predominantly affect MVARflows,
i.e. shunt reactances and off-nominal in-phase
transformer taps CALCULATE [AP/V]

(b) omitting from [B" I the angle-shifting effects of


phase shifters

(c) taking the left-hand V terms in (6) and (7) on


to the left-hand sides of the equations, and then
in (6) removing the influence of MVAR flows on the
calculation of [AO] by setting all the right-hand LKSOLVE (8) & UPDATE ol
V terms to 1 p.u. Note that the V terms on the
left-hand sides of (6) and (7) affect the behav-
iours of the defining functions and not the coup-
ling
? Yes ~ ~ -, ye
(d) neglecting series resistances in calculating the |CALCULATE [AQ/7V]| OUT
elements of [B' ], which then becomes the DC-approx-
imation load-flow matrix.This is of minor import-
ance, but is found experimentally to give slightly
improved results.
With the above modifications the final fast de-
coupled load-flow equations become |SOLVE (9) & UPDATE (VI Yes

[AP/V] = [B'] [AC] (8)


IKP=1 I to
[AQ/VI = [B" I [AV]
Both [BI] and [B"] are real,sparse and have the stru-
ctures of [HI and [L] respectively. Since they con-
tain only network admittances they are constant and Fig. 1 Flow diagram of the iteration scheme
need to be triangulated once only at the beginning of
the study. [B"] is symmetrical so that only its upper
triangular factor is stored,and if phase shifters are
absent or accounted for by alternative means [B'I] is
also symmetrical. Convergence characteristics

The immediate appeal of (8) and (9) is that very A typical convergence pattern for the method is
fast repeat solutions for [A] and [AV] can be cbtaiied shown in Fig. 2, in terms of the largest mismatches at
using the constant triangular factors of [B'] and [B" ]. every half iteration. Each solution of (8) and (9)
These solutions may be iterated with each other in produces rapid reduction of [AP] and [AQ] respectively
some defined manner towards the exact solution, i.e. towards the exact solution of the load-flow problem.

860

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
The effect of 0-changes on the MVAR flows is shown by were both monitored at every half iteration. In each
the increase produced in max[AQ[ after each solution case the process was iterated to the "exact" (very
of (8). In this and most other problems the effect accurate) solution. By storing the voltages, branch
of V-changes on MW flows is less pronounced, because MVA flows and losses at each stage of the calculation
the active losses due to MVAR flows are normally a post-solution history of the largest errors in thes
smaller than the reactive losses due to MW flows. quantities during the iterations was constructed.
Although this suggests that faster overall convergence
could be achieved by adopting a (2,1V) iteration The load-flow problems used were chosen as del-
scheme,this idea (among others) was not successful in iberately-severe tests. With the exception of the
any of the systems studied. Overconverging [AP] at any IEEE Standard Test Systems, the problems are prac-
stage produces a severe increase in [AQ], thereby tical systems from seven countries supplied by ind-
slowing down the overall convergence rate. ustry because they have presented convergence diff-
iculties. The systems contain a wide variety of
features, and cover the range of voltage levels from
EHV to distribution. They include long EHV line and
cable circuits, capacitive series branches, shunt
capacitors and reactors, and very small and very
large series impedances and X/R ratios. Some net-
X4 works are highly radial and others are highly meshed,
and the proportion of PV buses varies from 0 - 45%.
In some cases the voltage regulation is nearly 0.3
p.u.
:R
uz Table I gives results for unadjusted base-case
solutions of the test problems using the normal flat
U) voltage start. Details of the errors are shown for
the first three iterations. The largest bus voltage-
magnitude error maxievi at each iteration is given
m in percent. The largest branch-flow error maxies
E-4 at each iteration is given both in MVA and as a per-
U; centage of the branch flow. Except while the solu-
tion is still very inaccurate,the errors in the total
u
system MW and MVAR losses are insignificant, and are
not presented. Also, the maxiesi occurs on one of
the most heavily-loaded lines or transformers, and
is nearly all MW-flow error.
correspondence between max |A
The close
maxfes| gives a good guide to the specification of
PI
and
5 c in (10). Max|AQI and maxlev|
do not correlate so
NUMBER OF ITERATIONS closely, but by iteration 3 they are seen to be of
the same order as each other. Depending on the pur-
pose of the load-flow study, most of the solutions
shown in Table I are sufficiently accurate after 3
Fig. 2 Typical convergence pattern of the iterations, and even after 2 iterations in sewral
method (IEEE 118-bus system) cases. The number of iterations required is seen not
to be a function of problem size.
Apart from the results shown, the fast decoupled
program has been used on other problems. Noteworthy
Algorithms (8) and (9) each have geometric con- among these is a very difficult 43-bus system which
vergence, as explained more fully in Appendix 2. This fails with Newton's method and takes 22 (l0,lV) iter-
is not as fast as Newton's quadratic rate,but is more ations for an accurate solution. Otherwise, the 27-
than compensated for by the much-faster iteration bus African system of Table I takes the largest num-
speed. Unlike Newton's method, the fast decoupled ber of iterations of any problems tried. The method
method has not failed on any feasible problem. has also solved a 1080-bus system (see the section on
Solution Speed and Storage).
Accuracy requirements

The question as to what convergence tolerances ADJUJSTED SOLUTIONS


c and cq in (10) give an acceptably-accurate load-
flow solution is relevant not only to the present
method but more generally. Overconvergence, and The most common adjustments in routine load-flow
therefore unnecessary extra iterations,can be avoided solutions are single-criterion controls such as on-
by correlating the solution errors with the bus MW and load transformer taps, phase-shifters and area inter-
MVAR mismatches. Since power flows, voltage magni- changes,and generator Q limits and load-bus V limits.
tudes and losses are the primary objectives of a load- Conventionally, adjustments are made in between or
flow calculation,an investigation of their errors was during iteration cycles by some simple easily-pro-
included in the numerical tests performed. grammable parameter- or injection-changing error-
feedback logic or by bus-type switching12 . These
Numerical results schemes usually work satisfactorily with the slowly-
converging load-flow methods when the system is suf-
The studies carried out on the fast decoupled ficiently well-conditioned. With a powerful method
method included unadjusted, adjusted and sequential- such as Newton the number of extra iterations due to
outage solutions. In all cases maxiAP| and max|AQ| the adjustments can be large relative to the 3 or 4

861

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
TABLE I ERRORS DURING FAST DECOUPLED LOAD-FLOW SOLUTION

After Iteration 1 After Iteration 2 After Iteration 3 Accurate Solution


No. of
buses in max max max max|esl max max max max|esi max max max maxiesi max max no.
system |API IAQI levI API IAQI levi |API IAQI levl |API JAQJ of
MW MVAR MVA % MW MVAR % MVA % MW MVAR MVA % MW MVAR iters
13 87 2.5 0.72 105 13 11 0.4 0.08 9 1 1.4 0.04 0.00 1.4 0.2 0.002 0.001 52
14* 46 1.1 1.08 11 57 2 0.1 0.05 1 3 0.1 0.00 0.00 0.1 0.0 0.005 0.000 4
19 19 8.5 3.39 28 13 6 0.9 0.55 7 9 1.6 0.14 0.10 1.8 2.5 0.006 0.003 62
22 63 1.6 0.53 31 498 1 0.1 0.06 1 1 0.1 0.01 0.01 0.1 0.2 0.002 0.005 42
27 211 10.5 2.16 158 24 39 0.9 0.97 47 7 13.6 0.46 0.38 17.5 1.9 0.003 0.007 10;
30* 17 0.9 0.68 10 5 0.4 0.0 0.04 1 0.5 0.05 0.00 0.00 0.1 0.1 0.006 0.000 4
38 22 4.3 0.94 15 31 2 0.4 0.07 2 2 0.2 0.04 0.02 0.1 0.1 0.004 0.007 42
57* 24 1.0 1.24 9 9 1 0.1 0.29 1 0.7 0.1 0.01 0.03 0.1 0.01 0.004 0.006 42
107 204 10.1 1.10 165 8 19 1.3 0.22 9 10 4.8 0.24 0.06 2.4 2.6 0.007 0.006 7p
118* 58 0.9 0.37 34 726 3 0.1 0.02 2 3 0.3 0.00 0.00 0.3 0.5 0.009 0.003 42
125 257 12.5 1.56 77 20 6 0.8 0.28 3 3 0.9 0.11 0.06 0.5 0.5 0.007 0.001 6
180 225 7.1 1.56 153 21 9 0.7 0.29 4 2.6 1.8 0.13 0.06 0.9 0.6 0.002 0.004 62
205 20 1.4 3.52 25 14 5 0.2 0.84 4 4.5 1.1 0.05 0.21 0.8 0.9 0.003 0.004 6; 2
* IEEE Standard Test System

required for an unadjusted solution. For Q and V lim- works reliably, and for well-conditioned problems
its there seems to be no way of avoiding these extra a = 1 gives the most rapid convergence.
iterations.
When used with the new load-flow method, (11) is
The Newton formulation can be modified to give applied after calculating [V]. The choice of a and
automatic solutions for controlled parameters1 3-15 the total number of iterations for the solution are
but the benefits of this can be outweighed by the similar to those for Newton's method, since the tap
introduction of new problems. Not the least of these convergence rate is the main determining factor in
in some cases is the substantially-increased program- both methods. A (10,2V) iteration scheme is inferior
ming detail, for what is already a complicated method to the standard (lO,1V) scheme.
if efficiently sparsity-programmed. An organisation
may find that it is expensive or impractical to incor- Most important, the convergence rate is affected
porate these types of modifications into an existing little by reflecting the transformer Tr circuit para-
Newton program to cater for the individual and perhaps meter changes into [B"] at each tap correction.
changing requirements of their load-flow studies. Unless better initial tap estimates are available,
[B"] is therefore formed and factored at the start
With its very fast iteration time and reliable of the study assuming that all controlled taps are
convergence, the new method encourages the retention nominal, and it remains constant throughout the sol-
of the well-tried conventional adjustment schemes. ution.
Adjustments will be made before or after the solution
of (8) or (9) according to whether they primarily As an example, the fast decoupled method con-
affect MW flows or MVAR flows, respectively. Although verges to within 0.01 MW/MVAR maximum bus mismatches
the adjustments largely dictate the total number of in 11, 9 and 61 iterations for three different cases
iterations to convergence, the absolute increase in on a practical 22-bus system with 11 controlled tran-
computing time compared with the unadjusted solution sformers, 4 radially connected and 7 in network loops.
is not great. As a typical example, consider an This includes the enforcement of tap limits and cor-
unadjusted solution that takes 6 iterations. Adjust- rections to the nearest physical tap settings after
ments which add say, 6 extra iterations, represent initial convergence. Using (11) and the same a = 1,
little more time than one Newton iteration. This is the three cases require 10, 12 and 11 iterations, res-
insignificant for solutions requiring non-minimal pectively, with Newton's method.
input/output.
The development of single-criterion controls in
On-load transformer tap changing the fast decoupled program has as yet been limited to
in-phase transformer taps. It is conjectured that
A suitable adjustment algorithm for an in-phase with other controls it will also be found to be unnec-
off-nominal tapping ti p.u. controlling the voltage essary to reflect parameter changes into [B'] or [B"].
of bus k to V5P k p.u.
u iss
Generator Q limits

t.(new) -t.(old) -+ (V - Vk ) (11) Once a load-flow solution is moderately con-


11
verged, any PV-bus Q-limit violations can be correc-
ted. Provision must be made for subsequent inter-
where a is an empirical error-feedback factor. active effects, i.e. MVARs backing off limits and
Applied in between Newton iterations, this algorithm new Q violations.

862

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
Two approaches are available. In the first, each Transmission network outages
violating bus is explicitly converted to PQ type so
that the MVAR output is held at the limiting valUe. Line and transformer outages are simulated by an
The bus remains a PQ type during the rest of the sol- adaptation of the inverse-matrix modification tech-
ution unless at some stage it can be re-converted to nique applied to [B'] and [B"]. The experiments have
a PV type at the original voltage magnitude without shown that it is only necessary to simulate the re-
the violation. In the fast decoupled program, con- moval of series transmission elements from these
verting any number of bus types at one time involves matrices. Shunt capacitors and reactors, line charg-
retrianulating [B" ] . ing capacitance, and the shunt branches of off-
nominal-tap transformer equivalent circuits can remain
The second approach is to correct the voltage of in [B" ] without affecting convergence noticeably. All
each violating PV bus k by an amoUnt AVk at eac outages must of course be reflected correctly in the
following iteration to reduce the error AQk = (Qk calculation of [AP/VI and [txQ/V].
- Qk) to zero. The convergence of this process is
rapid when an approximate sensitivity factor Sk rela- For the outage of a series branch two non-sparse
ting AVk and AQk is used thus: vectors [X'] and [X" must be calculated,each requir-
ing one repeat solution using the factors of [B'] and
[B"], respectively. After each solution of (8), [AO]
AVk = Sk Ak/vk (12) is corrected by an amount

If Sk is defined according to (9) it is the diagonal -c' [X'] [M' [(AB] (13)
element corresponding to bus k in the inverse of
matrix [B" I augmented by the previously-absent row
and column for bus k. Appendix 3 shows that Sk can where c' is a scalar and [M'] is a highly-sparse row
be calculated easily and very rapidly without retri- vector containing one or two elements. This corre-
angulating [B"]. For an operational network each Sk tion requires about n multiplication-additions. Sim-
may be stored permanently and updated only for sig- ilarly, after each solution of (9), [AV] is corrected
nificant system configuration changes. The correc- by an amount
tion (12) ceases to be applied if at some stage in
the solution the value of Vk is restored to or goes -c" [X" [M"] [V] (14)
beyond its original value.
The details of this technique are given in Appendix 4,
Both approaches have been programmed and tested, including the case of multiple outages. The scheme
and are equally effective. The sensitivity method can equally be used for branch switch-in operations
usually takes more iterations, but against this it if desired.
requires no retriangulations for limit enforcement
and back-off, and is simpler programming-wise. Using Generator outages
a 0.01 MW/MVAR tolerance on bus-mismatches and Q-limit
enforcement, typical results for the IEEE 30-bus and A generator outage requires as input data the
118-bus systems are quoted. With two buses in the redistribution of scheduled MW generation to compen-
30-bus system violating their Q-limits by 17% and 6% sate for the loss of supply. It is usual to cater
after initial convergence,the solution takes 72 iter- for the partial loss of generation at a plant, in
ations by approach 1, and 9 iterations by approach 2. which case it is only necessary to reflect the changes
For the 118-bus system, with 12 buses initially vio- in scheduled MW in [AP/V]. The complete loss of a
lating their limits by an average of 12% and a maximum plant means that the net MVAR of the relevant bus
of 20%, the solutions take 72 and 121 iterations res- must be fixed by either of the methods used in enforc-
pectively. ing generator Q limits.
V-limits on PQ buses are handled in the inverse Since only one bus is involved, an alternative
manner, using either of the two approaches. is to augment the upper-triangular factors of [B']
and [B"] by an extra column each. This is exactly
APPLICATION TO OUTAGE STUDIES the same calculation as performed in obtaining the
sensitivity factor Sk used in the 9-limit adjustments,
Sequential branch- and generator-outage load- and is described in Appendix 3.
flow calculations are performed off-line for planning
and operational planning studies and on-line or semi- Experience with outages
on-line for' operation/control, to evaluate system
performance and particularly security following cred- Tests were conducted on several systems, includ-
ible outage contingencies. In all these applicatiohs ing a series of 35 line and transformer single-outage
a large and time-consuming number of load flows are cases on the heavily-loaded 180 bus problem which is
solved consecutively and automatically after a solu- part of the main British transmission system. Ten of
tion of the base-case problem, and branch flows, bus the outaged branches were chosen as the leAst-loaded
voltages and generator MVAR outputs are monitored to ones in the base-case, and 25 were chosen as the most
detect insecure and unsatisfactory conditions. For loaded ones, with base-case branch power flows of up
these purposes very accurate solutions are not norm- to 2370 MW and 240 MVAR.
ally required.
It was found that the (l,lV) iteration scheme
The approximate decoupled approach has already remains the best in this application, and it was con-
been developed successfully for the outage-study firmed that using the base-case solution to start the
application 4 . The present method can also be outage calculations is on the average distinctly
employed efficiently without retriangulating the base- better than using the normal flat voltage start (each
case [B'] or [B"] at any stage. time. The worst branch MVA-flow error at each itera-

863

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
TABLE II RESULTS FOR 35 SINGLE-BRANCH OUTAGE CASES ON 180-BUS SYSTEM

No. of (l0,lV) No. of ERRORS IN GROUPS


iterations for outage
cases in
m

max. branch-flow v s
error of 3% groups
largest average largest average largest average largest average

Group 1, 1 iteration 21 49 13 1.8 0.3 0.13 0.05 23 8


Group 2, 2 iterations 12 10 4 0.8 0.3 0.62 0.18 17 5
Group 3, 3 iterations 2 1 0.7 1.3 0.7 0.09 0.05 0.1 0.06

tion was obtained as before. Since this always seems or AQk and AQm. Also, it is noted that sinikm= -sinOmlk
to occur on one of the most heavily-loaded branches, and coskm = cosOmk. These trigonometrical functions
the performance of the method for security-monitoring are calculated using the approximations sinO % O-C3/6
can be meaningfully interpreted even though branch and cosO t 1_02/2+04/24, which have been found to give
MVA-rating data was not available. final load-flow results indistinguishable from those
obtained with the accurate function evaluations. If
In the tests, one (10,1V) iteration was adequate storage is not critical, the sine and cosine terms
in the majority of cases to give a maximum branch-flow calculated during the construction of [AQ/V] can be
error of an arbitrarily-chosen 3%, and in two cases stored for use in the next construction of [AP/V].From
three iterations were needed. Table II gives a break- a flat voltage start, each sine and cosine term can be
down of the 35 cases. initialised to 0 and 1 respectively for the first sol-
ution of (8).
SOLUTION SPEED AND STORAGE
During the experiments the constituent parts of
the solutions were timed on the CDC 7600 computer.
Central to the fast decoupled method is the use Results are given for a 1080 bus, 1862 branch system
of efficient sparsity-programmed routines for the tri- with 41 PV buses,and the IEEE 118 bus, 186 branch sys-
angulation of the sparse, real, symmetric matrices tem with 53 PV buses,using non-optimised FORTRAN com-
[B'] and [B"] by Crout elimination and the subsequent pilation.The inferior 'static' bus ordering scheme was
solutions of (8) and (9) by forward and backward sub- used and the sine and cosine terms were not stored.
stitutions 3. If, as is becoming more widespread, such
routines are available as standardpackages,the method Time, seconds
is not difficult to program efficiently. 1080-bus 118-bus
At the beginning of the load-flow study the sys- Bus ordering 0.062 0.007
tem buses are re-ordered to avoid excessive fill-up in formation & triangulation of [B'] 0.635 0.028
the table of factors during the triangulation of [B']. formation & triangulation of [B"] 0.515 0.007
It can be shown that this bus ordering remains equally calculation of [AP/VI & conver-
suitable for [B"] with the PV buses by-passed in the gence testing 0.059 0.006
ordering list. If the system is very large or if con- solution of (8) & updating [0] 0.063 0.005
secutive load-flow solutions are to be calculated then calculation of [AQ/V] & conver-
it is advantageous to minimise fill-up as far as poss- gence testing 0.053 0.005
ible by using a good ordering scheme 6.Since [B'] and solution of (9) & updating [V] 0.050 0.001
[B" ] remain unchanged, triangulation routines that per-
form the ordering during elimination are at no great For the 1080-bus problem, the solution to within 0.01MW
disadvantage. and MVAR bus-mismatch tolerances from a flat voltage
start required 7(10,lV) iterations plus two additional
The triangulation routine is programmed to take (10) iterations in the final stages. Excluding input/
account of matrix symmetry so that only the upper-tri- output, the total time was 3.2 seconds of which 1.98
angular factors of [B'] and [B"] are stored for use in seconds were for the iterative process. For the 118-
the solution routine. Using dynamic re-ordering accor- bus problem these times were 0.13 and 0.09 seconds
ding to row sparsity, the ordering, triangulation and respectively. The iteration times could be reduced
solution times each vary roughly linearly with network considerably by using a better ordering scheme and more
size3. For a typical well-developed network with n realistic mismatch tolerances.
buses and b branches the number of elements in the
upper-triangular factor of [B'] is about 3(n+b)/2. The storage requirements of the fast decoupled
Depending on the number and location of PV buses, the method are about 40% less than thosecf Newton's method.
triangulation and solution processes for [B"] are This saving is reduced somewhat if the sine and cosine
faster and require less storage than for [B'], in some terms are stored, and for outage cases where vector s
cases considerably so. [X'] and [X"] are needed.Each of the parts of the sol-
ution given in the above timing listis conveniertly per-
The calculations of the vectors [AP/VI and [AQ/V] formed in a separate program subroutine. Apart from
can each be performed rapidly in a single sweep of the simple subroutine overlaying,core storage can be econ-
branch admittance list.For each series-branch connec- omised by reading certain selected vectors from disk
ting buses k and m, it can be seen from (1) and (2) when they are required and writing them back after the
that where appropriate, the terms VkVmGkm and VkVmBkm relevant subroutine. These block transfers are per-
can be used directly in accumulating both APk and APm, formed a limited number of times during a solution,and

864

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
should not degrade the method's speed too severely. Having established this latter point, attention
Using this scheme,the core storage requirements of the was devoted to the decoupled Newton apprczcimations (6)
method are about the same as those of the Gauss-Seidel and (7) which can be re-written
method.
[AP/V] = [B'] [V.A] (16)
Network tearing for load-flow calculations is not
usually economical in computing speed when ordered [AQ/V] = [B"] [AV] (17)
elimination is used for solving the sparse network
equations. However, if a piecewise approach is desir- A question at this stage was whether the V terms on
able for other reasons, e.g. storage,any of the avail- the left-hand sides of (16) and (17) and on the right
able Y-matrix decomposition or diakoptical methods can of (16) should be omitted. The seven resulting combi-
be applied to give a series of smaller constant [B'] nations of 0 and V algorithms were compaied with each
and [B"] matrices. Note that since branches connected other on all the test systems,and the combination (8)
to PV buses are absent from [B"], a certain amount of with (9) emerged as the best, although one or two of
network tearing is already inherent in this matrix. the others were not much worse. (8) and (9) and other
versions can be rearranged to give direct solutions
CONCLUSIONS for [0] and [VI instead of [A] and [AV] ,but there is
no convergence difference unless further approximations
The fast decoupled method offers a uniquely attra- are made. The [A] and [AV] forms were preferred, in
ctive combination of advantages over the established spite of a few extra arithmetic operations, because
methods, including Newton's, in terms of speed,reliab- this form directly retains the familiar AP and AQ mis-
ility, simplicity and storage, for conventional load match quantities for convergence testing.
flow solutions. The basic algorithm remains unchanged
for a variety of different applications. Given a set All of the above tests were carried out using a
of good ordered elimination routines,the basic program (1,1V) iteration scheme,flat voltage starts, and no
is easy to code efficiently, and its speed and storage adjustments. Having finalised on the algorithms (8)
requirements are roughly proportional to system size. and (9), the problems of Table I were solved us ing
Auseful feature of the method is the ability to reduce each of the following schemes: (a) (2O,1V),(b) (2O,1V)
its core-storage requirements to approximately those of once and subsequently (l1,lV), (c) (2,2V) once and
The Gauss-Seidel method with a small rLunber of core- disk subsequently (10,lV)., (d) (lV,l1), and (e) (3,2V).
block transfers. The method performs well with conven- None of these schemes was as good as the sirple (l,lV )
tional adjustment algorithms,and solves network outage approach, which is also best for transformer-tap adjus-
security-check cases usually in one or two iteratiors. tment solutions, and branch outage solutions.
It is computationally suitable for optimal load-flow
calculations, and developments in this area are to be Several different attempts at acceleration, inc-
reported separately. luding block successive over-relaxation ard a heuristic
approach based on testing sign-changes in the bus mis-
ACKNOWLEDGEMENTS matches, proved to be unrewarding.

The authors are grateful to the University of Man- For all the branch-outage studies,the removal of
chester Regional Computing Centre for running the pro- line-charging and transformer nr-equivalent shunt ele-
grams. The Middle East Technical University, Ankara, ments from [B"] was found to affect the convergence
Turkey, is acknowledged by 0. Alsag for granting leave only minimally, provided that they are correctly re-
of absence. moved in the Calculations of [AP/VI and [AQ/V], and
similarly for shunt capacitor or inductor outages.
APPENDIX 1 However,failing to remove the outaged series elements
from [B"] and [B'] is totally unsuccessful.Appendix 4
Alternative decoupled versions investigated shows how these outages are carried out,and also 'that
there is no extra effort involved in removing trans-
This appendix gives a very brief outline of the former Tr-circuit shunts.
large number of experimental studies conducted on var-
iants of the decoupled approach. Each version was used APPENDIX 2
on all or a selected group of the load-flow test prob-
lems of Table I. In nearly all cases the problems were Notes on the convergence of the fast decoupled method
solved successfully by the different versions, demon-
strating the general effectiveness of decoupling the The iteration process of the fastdecoupled method
angle and magnitude calculations. The object of the has three distinct components, each with its own con-
studies was then to find the version that best combines vergence characteristics: (a) the solution of [AP/V] =
consistent and rapid success with other computational 0 for [0] using algorithm (8), (b) the solution of
advantages. [AQ/V] = 0 for [VI using algorithm (9), and (c) the
interactive effects of V-changes and 0-changes on the
Initially,the methods of refs.6,7 and 11 were com- defining functions [AP/VI and [AQ/V] respectively.
pared17 . The Despotovic and Decoupled Newton methods
converge well but require re-triangulations. Uemura's Algorithms (8) and (9) are both Newton-like,
method is not very satisfactory on some practical sys- except that instead of re-evaluating thetrue Jacobian-
tems because MVAR conditions are not excluded from the matrix tangent-slopes to the left-hand functions at
"P-0" matrix. The idea in ref. 18 of using the current each iteration, fixed approximated tangert-slopes [B']
mismatch [AI] in (9) and [B"] are used. The algorithms therefore corres-
pond to the generalised 'fixed-tangent' or 'constant-
[AI] = [B"] [AV] , (15) slope' method19 which has geometric convergence. For
reasonably-behaved functions, this method is very re-
and a variant in which each Vk is divided by cosOk, liable and if, as in the present application,the fixed
were less successful than anticipated. The other idea tangent-slopes correspond closely to the Jacobian
in ref. 18 of using the constant matrix [B'] in the matrix at the initial point (for a flat voltage start),
angle calculation of refs. 4 and 6 gives good conver- the initial convergence is very rapid.The process does
gence. not 'home in' as fast as the quadratic Newton method as

865

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
the exact solution is approached, but load-flow solu- the high sparsity of [C1] and [GI, the calculation be-
tions are rarely required to very high accuracy. comes extremely fast.

The fixed-tangent method is not thrown off-course The process can be expressed in more formal matrix
when it encounters 'humnps' in the defining functions, notation using the 'multi-product of inverse' approach
whereas Newton's method tends to be mis-directed even and is easily generalised for any non-symmetric matrix.
to the extent of divergence from the desired solution In the present case, the process can be used for the
in such cases. In the decoupled load-fbw problem,the original triangular factorisation of [B'] and [B" ],
changing V-values during the solution have the effect adding one column at a time.
of causing sometimes-oscillatory shifts in the shapes
of the multi-dimensional surfaces [AP/VI as functions
of [01] ,and likewise for 0-values on [AQ/VI as functions APPENDIX 4
of [VI. Since [B'] does not represent MVAR conditions
it corresponds to fixed tangent-directions that take r Branch outage calculation
cognisance of these shifts and are therefore not affec-
ted by V-oscillations. Likewise for [B"] in relation Let either (8) or (9) be represented in the base-
to MW conditions. case problem as the equation

[RI = [B] [E o]
.0
(22)
APPENDIX 3
0

Calculation of sensitivity factor for which a solution

The matrix factors of [B"] are expressed in the


symmetric form, so that (9) is [E0] = [B]
0
[R] (23)

[AQ/V] can be obtained using the factors of [B0 ]. In the most


= [U] [D] [U] [AV] (18)
general case, the outage of a line (neglecting charg-
ing capacitance) or transformer can be reflected in
where [D] 1 is a diagonal matrix and [U] is upper- [Bo] by modifying two elements in row k and two in row
triangular with unit diagonal elements. [Ult is not m. The new outage matrix is then
stored. The solution of (9) is then performed accord-
ing to the following three steps: (a) a forward-sub- [B1] = [B0] + b[M] t [M] (24)
stitution process on [AQ/V] and [U]t,giving an inter-
mediate vector [F]; (b) the trivial matrix multipli- where:
cation [G] = [D][F]; (c) a backward-substituti o n
process on [G] and [U] to give [AV]. b = line or nominal transformer series admittance
[M] = row vector which is null except for Mk = a,
If a new PQ-type bus is created in the network,
m - -1
an extra row and column must be added to [B"] .Letting
and M
the order of [B"] be (n-l), then the new PQ bus is a = off-nominal turns ratio referred to the bus
ordered n, i.e. last, and the enlarged matrix is corresponding to row m, for a transformer
= 1 for a line.

Depending on the types of the connected buses,only one


(19) row, k or m, might be present in [B'] or [B"],in which
case either Mk or Mm above is zero, as appropriate. If
1 2 both the connected buses are PV or slack, then [B"]
requires no modification.
To obtain the matrix factors of [B" ] it is only nec- It can be shown that
essary to enlarge the existing factors by adding an
nth column to [U] and a diagonal element to [D]. [B~0 ] -l - c[X][M][B0 I
-l -l
[B1] 1 = (25)
Replacing [AQ/V] in step (a) above by the sparse
[C1], [GI in step (b) gives the new sparse column of where:
[U] c = (l/b+[Ml [X]) and [X] = [B0o [M] (26)
[GI = [U lnU2 *...U ]t (20)
The solution vector [E1 ] to the outage problem is :

where U = 1. The new element D is given b


nn nn [E1] = [B 1] -1 [RI (27)

D = (C 2 - E Uu /D. ) (21) and from (23), (25) and (27) we have


nn j=l jn ii

= the sensitivity factor Sk used f


[E1] = [Eo] - c [X] [M] [E0] (28)

limits and generator outages. Hence, the solution to the base-case problem is easily
corrected, as in (13) and (14). [X] is calculated at
Using the normal solution routine, this calc the beginning of the outage case by a repeat solution
tion takes about the same time as a solution of using the factors of [B0] with [M]t as the independent
By reprogramming steps (a) and (b) to take accoun vector.

866

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
The above procedure can be applied recursively for problems", IEEE Trans. (Power Apparatus and Systems),
multiple simultaneous outages.The solution vector [El Vol. PAS-90, pp. 123-130, January/February 1971.
is corrected successively as the effect of each branch
outage is introduced one at a time. For a set of out- 7. K. Uemura, "Power flow solutions by a Z-m&3rixtype
aged branches 1 ..... m, the recursion is method and its application to contingency evalua-
tion", PICA Conference Record, p. 386, May 1971.
[E.] [E1i-1 ci [X.l[Mil [E.ilI for i=l...m (29) 8. K. Uemura, "Approximated Jacobians in Newton's pcwer
flow method", Proc. Power System Computation Con-
where [M.] is defined as before for the outagedbranzh ference, paper 1.3/2, September 1972.
introduced ith in the sequence.The scalars ci and the
vectors [Xi1 are calculated at the beginning of the 9. C.H. Jolissaint, N.V.Arvanitidis & D.G.Luenberger
case from the recursive algorithm: "Decomposition of real and reactive power flows: A
method suited for on-line applications",IEEE Trars
[Xi] = [B0I
M-1
[]
t (Power Apparatus and Systems) ,Vol.PAS-91, pp.661-
i
670, March/April 1972.
[Xi]
1
= [Xi]
1
- c.J XJ ] [M J ] [Xi] for j=l...i-l (30) 10. B. Stott, "Effective starting process for Newton-
Raphson load flows", Proc. IEE, Vol. 118, pp.983-
c. = (l/b 1 - x 1)-1
[M.1] [i) 987, August 1971.
1

-1 11. B. Stott, "Decoupled Newton load flows",IEEETrars.


all for i = 1... .m . If any ci= 0 a split network is (Power Apparatus and Systems) ,Vol.PAS-91,pp. 1955-
indicated. 1957, September/October 1972.

Although this scheme avoids the retriangulation of 12. G.W. Stagg and A.H. El-Abiad,"Computer methods in
the network matrices, it is faster for at most three power system analysis", McGraw-Hill, 1968.
simultaneous ly-outaged branches.
13. N.M.Peterson and W.S.Meyer, "Automatic adjustment
of transformer and phase shifter taps in the Newton
REFERENCES
power flow" ,IEEE Trans. (Power Apparatus and Syst-
1. D.A. Conner, "Representative bibliography on load-
ems), Vol. PAS-90, pp. 103-106, January 1971.
flow analysis and related topics",paper C73-104-7
14. J.P. Britton, "Improved load flow performance
presented at the 1973 IEEE Winter Power Meeting, through a more general equation form",IEEE Trans.
New York, N.Y., January 28-February 2, 1973.
(Power Apparatus and Systems) ,Vol.PAS-90, pp.109-
2. W.F. Tinney and C.E.Hart, "Power flow solution by
114, January 1971.
Newton's method", IEEE Trans. (Power Apparatus and
15. H.W. Dommel, W.F.Tinney and W.L. Powell, "Further
Systems), Vol.PAS-86, pp.1449-1460, November 1967. developments in Newton's method for power system
3. W.F. Tinney and J.W. Walker, "Direct solution of applications",paper 70CP-161-PWR presented at the
1970 IEEE Winter Power Meeting, New York, N.Y.,
sparse network equations by optimally ordered tri-
January 25-30, 1970.
angular factorization", Proc. IEEE, Vol.55,
pp. 1801-1809, November 1967.
16. B. Stott and E. Hobson, "Solution of large power-
system networks by ordered elimination: a compari-
4. N.M.Peterson, W.F.Tinney and D.W.Bree, "Iterative
son of ordering schemes", Proc. IEE, Vol. 118, pp.
linear AC power flow solution for fast approximate
125-135, January 1971.
outage studies", IEEE Trans. (Power Apparatus and
Systems), Vol.PAS-91,pp. 2048-2053, September/Oct-
17. J. Hassall, "Comparison of modern loa flowme1hods"
ober, 1972. B.Sc. Project Report, UMIST, April 1972.
5. W.F. Tinney, N.M.Petersen, "Steady state security
monitoring", Proc. Symposium on real time control 18. B. Stott, discussion of reference (4).
of elec. power systems, Brown, Boveri & Comp6Ltd.,
Baden, Switzerland, 1971.
19. C. -E. Froberg, "Introduction to Numerical Analysis",
6. S.T.Despotovic, B.S.Babic and V.P.Mastilovic, "A Addison-Wesley Publishing Comp., Inc., pp. 19-26
rapid and reliable method for solving load flow 1966.

867

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
Discussion to make it worthwhile?
The authors state that the current mismatch ideal8 did not work
E. Hobson (The Capricornia Institute of Advanced Education; Rock- out too well. Since we had more than just a passing interest in the cur-
hampton 4700; QLD, Australia): The authors have made significant rent mismatch formulation, what specifically was its weakness?
changes to previously described decoupled load flow methods in order We compliment and thank the authors for an excellent paper.
to improve solution speed and to-permit rapid reliability assessments to
be made. I am grateful to them for giving me the opportunity to ex-
periment with their methods before publication.
My comments refer to experience gained whilst incorporating
some of the authors' techniques into an operational interactive mini- A. K. Laha and K. E. Bollinger (University of Saskatchewan, Saskatoon,
computer load flow program written in Basic and employing the Gauss S7N OWO Canada): The authors are to be complimented for an inter-
Seidel algorithm l 1 ]. This program was designed primarily for small dis- esting paper. In solving the load flow problem there is always a trade-off
tribution authorities where versatility was considered more important of speed and storage. The formal Newton method has demonstrated
than speed. distinct advantages over other iterative schemes. This paper is a welcome
The claim that core requirements are similar to those of the Gauss addition to the different modifications of Newton's method, to ac-
Seidel method was found to be approximately true. Increased com- celerate the speed of convergence, which are already published in the
plexity of programming and the necessity to store the factorised forms literature. Others will no doubt further test the idea outlined in this
of both B' and B" matrices of equations (8) and (9) led to an increase paper thereby confirming the authors' claim about its relative superiority
in core requirements of approximately 2:1. This increase proved em- over other methods. In the meantime we would like to comment on
barrassing on an 8k PDP8 computer when attempting to replace the some of the salient points of the paper.
Gauss Seidel algorithm with the new decoupled method. In order to Although the idea of neglecting the off-diagonal matrices N and J
fit the new method onto the PDP8, certain changes were made to the was first suggested by Carpentier as early as 1963, this was later found
authors' approach. These modifications permitted the replacing of to suffer from weak convergence and poor reliability. As we under-
matrices B' and B" with a single matrix B, comprising the negated re- stand, the authors' decoupled method is a modification of equations (4)
actances of the nodal admittance matrix. For PV buses, the reactive and (5) in the sense that they introduced the (10, lv) successive itera-
mismatches of equation (9) were put to zero, and no adjustments made tion scheme, which in turn, increases the stability property of the
to V. Simple pre-ordering was considered adequate for the size of net- solution. The modified Newton method1 which we are developing at
works contemplated, hence no re-numbering scheme was employed. present, although completely mathematically different from the au-
With the above modifications, solutions were obtained for a range thors' method, follows the idea of a successive iteration scheme. The
of small distribution networks. In general the authors' claims were sub- modifications suggested in equations (6) - (9) of the authors' paper
stantiated. Each iteration took approximately 1.5 times a Gauss Seidel essentially increase the speed of computation time, but does not affect
iteration, and overall solution time for 'difficult' networks improved by the convergence properties of the algorithm nor will it affect the total
10:1 or more. The number of iterations was greater than for pure number of' iterations. Thus this algorithm apparently contradicts one of
Newton Raphson, but no speed comparisons were possible since dif- the authors' earlier comments regarding the algorithm2 that "this is
ferent computers were involved. Convergence was appreciably slowed also found not to be very satisfactory for most problems and the reason
and divergence problems were sometimes introduced in the simplified is that equation (4) (which is same as equation (5) of this paper) is a
method if in-phase transformer taps departed more than about 5% relative unstable algorithm at some distance from the exact solution."
from nominal settings. This was presumably because the effects of tap It may be interesting to know whether the authors' comment regarding
changing were being incorporated into the P - 0 coupling. A further the fast convergence property is based solely on the results as shown in
divergence problem arose with a very weak 9 bus radial system includ- Table I and figure 2 of this more recent paper or is there any mathe-
ing a line with R/X approaching 2. Convergence was achieved by making matical justification behind it.
R/X closer to unity. Removal of series resistive elements from B', as In Dr. Stott's earlier paper2, he introduced the idea of current
described by the authors, eliminated the convergence problem but did mismatch. Since that method and the one outlined in this paper are
not significantly reduce the number of iterations. Convergence could competitive to each other, we would appreciate it if the authors could
not be achieved by removing series resistive elements from the single give us a relative comparison of both of the methods from the point of
matrix B. view of speed, convergence characteristics, storage requirement etc. and
After experimenting with the authors' decoupled methods I con- which one, the authors think, is superior from an overall point of view.
clude that they have much to recommend them in terms of speed, One may note that the the 27th bus system, the two methods take 3 and
storage and simplicity in comparison with existing methods. 10/2 iterations respectively to arrive at the correct solution.
We fully agree with the authors' suggestion to approximate Cosine
REFERENCE: and Sine trigonmetric functions by the first few terms of the Series, as
it definitely increases the speed of computation. One should, however,
[1] E. Hobson "Using the PDP8/e as a Network Analyser". Proceed- use this approximation with caution as this may give errors if the angle
ings of the 1972 Decus Australia Symposium is quite large. These are cases where the phase angle of some buses are
600 or so out of phase with that of the slack bus.
Manuscript received August 1, 1973. Again we would like to commend the author on a well written im-
portant paper.
REFERENCES
T. E. Dy Liacco and K. A. Ramarao (Cleveland Electric Illuminating 1. R. Billinton, K. E. Bollinger and S. B. Dhar, "A New Modified
Company, Cleveland, Ohio 44101): This paper, which is a model of Newton's Method For Load-Flow Analysis", presented at the IEEE
clear exposition, reflects the thoroughness and extent of the efforts Winter Power Meeting, 1972.
made by Dr. Stott and his colleagues in developing and testing de- 2. Reference ( 11) of this paper.
coupled load flow methods. For the past two years we have been plan-
ning on using a decoupled load flow for both system operation and Manuscript received August 6, 1973.
system planning but have not had time to try out various techniques.
We did try Dr. Stott's first method1 1 on our system with very good re-
sults but could not pursue further experimentation with it. The authors
have done us and certainly many others in the industry a great service
by doing much of the testing for us. W. 0. Stadlin and B. F. Wollenberg (Leeds & Northrup Company,
The important result of all this work is the demonstration that a North Wales, Pennsylvania): The authors have performed an invaluable
decoupled load flow using a fixed susceptance matrix requiring only one service in their evaluation of alternate approximate forms of the
initial triangularization is more efficient and reliable than other existing Newton method. In order to further conserve computer memory we
methods. This idea had always been appealing to us since it was first would welcome the authors comments as to the feasibility of setting
proposed by Dr. Uemura7. We believe that the industry should try out [B"] equal to [B']. Shunt reactances and line charging would then be
this approach extensively in order to gain as wide an experience as considered as part of the Q load and the corresponding MVAR value of
possible. The author's refinement of the decoupling process by re- these shunt elements would be updated as part of the Q mismatch
moving all shunt reactances from the B' matrix is a remarkable idea. calculation based on the previously calculated voltage. Secondly, if such
Have the authors tried Uemura's method with the modified B'? Does a load flow is part of an optimal power flow package can approximate
the inclusion of the voltage magnitudes in the forcing functions make incremental losses be derived using similar decoupling techniques as
a significant difference in overall speed or convergence characteristics used in the solution - or must one revert to the normal Jacobian?
Manuscript received July 30, 1973. Manuscript received August 2, 1973.

868

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.
B. Stott and 0. Alsac: We are very grateful to the discussers for their approximated current-mismatch idea18, mentioned by Messrs. Dy
valuable comments and questions. Liacco and Ramarao, was because approximating each relevant Jacobian-
The main aim of the work reported in the paper was to identify a matrix element (Giksin0k+BikcosOk) by Bik is not valid for large bus
single and preferably simple version of the decoupled load-flow ap- angles. The approximations in the present method have no such limita-
proach that uses constant network matrices and performs well, in terms tions, since it is only the transmission angle Oik across each line that is
of speed and reliability, over a wide range of problems and applications. assumed not to be large. Likewise, in connection with the point made
From the many alternatives studied, the version embodied in eqs. (8) by Messrs. Laha and Bollinger, it is each sinOik and cosOik as per eqs.
and (9) emerged as the best according to these criteria. Other variants (1) and (2) that are computed using truncated series. The errors are ex-
differing slightly in form can perform very satisfactorily on specific tremely small for practical systems even up to 30-degree transmission
load-flow problems. Two of the chief points to come out of the work, angles.
the first of which replies to the relevant questions of Messrs. Dy Liacco In answer to the same discussers, we do not believe that the
and Ramarao, can be stated as follows. Newton (NM) or decoupled Newton (DNM) methods are at all com-
1. [AQ/V] is very much better than [AQ] as the defining func- petitive with the fast decoupled method (FDM) for the vast majority of
tion of the MVAR-V problem. It can be seen from eq. (2) that conventional load-flow applications. The computation per iteration of
AQi/Vi has only one term, QisP/Vi, that is non-linear in V. This term is DNM is only slightly less than that of NM, whereas for FDM it is five
significant only for impractically-small values of Vi and vanishes on times less, assuming very good programming in each case. Acceptable
no-load. Hence, for given values of 0 obtained from the solution of accuracy can frequently be obtained from a flat start with FDM in
eq. (8), each solution of eq. (9) for V is thoroughly reliable and normal- 2 or 3 iterations, as seen from Table I of the paper - the fact that NM
ly very accurate. In contrast, it is of secondary importance whether gives higher accuracy is seldom of practical importance. If the number
[AP] or [AP/V] is used for the MW-0 problem, since the non-linearities of iterations required by NM is 3 - 4, the break-even point in com-
here are due mainly to the sine and cosine terms in eq. (1). puting time is 15 - 20 iterations for FDM. The "constant slope" con-
2. Since shunt susceptances do not appear in [AP], nor in the vergence mechanism of FDM inhibits divergence for difficult problems
Newton matrix [H], large shunts represented in [B'] can cause poor including cases whose solution voltages are very low, with the effect
convergence of the MW-0 problem: we encountered this difficulty that the method is well-behaved even if a feasible solution is not pos-
when originally testing Uemura's version. Correspondingly, failure to sible. In certain difficult cases FDM converges rather slowly by its
represent such shunts in [B"] can give convergence difficulties of the normal standards, such as with the 27-bus, 43-bus and 9-bus problems
MVAR-V problem. mentioned in the paper or the discussion, but has a very high reliability
We thank Mr. Hobson for experimenting with the method at an in excess of other methods. Probably only NM with the starting process
early stage, and for the very useful ideas and information that he has of reference 10 is comparably reliable. Storage-wise, FDM requires a
presented. His results go some way towards answering Messrs. Stadlin little more than DNM but less than NM.
and Wollenberg about the possibility of using a single B matrix. This is The application of FDM for rapid sequential outage evaluations,
clearly an attractive prospect, but will be suitable only for some power such as in AC security monitoring, seems to be quite promising. For
systems, bearing in mind our point 2 above. The ideal application of a each outage case, a single calculation of the angles from eq. (8) can be
single B matrix is on lower-voltage networks with little or no shunt com- made initially. This is normally much more accurate than a pure DC
pensation and with one infeed point, where [B'] and [B"] are very load flow, since it starts from the base-case solution. At this stage the
similar or identical to each other. For instance, electrified railway net- MW flows are calculated and if none of them exceeds defined limits, the
works exhibit this property, and we have used a single B matrix with process proceeds to the next outage case. Otherwise, the FDM solution
complete success for many hundreds of load flows on these systems, in is continued to higher accuracy.
which the voltage magnitudes under extreme loading conditions go The subject of Messrs. Stadlin and Wollenberg's second question
down to less than 0.3 per-unit. Mr. Hobson's scheme for accommodat- has been of great interest to us. After developing our original Newton-
ing PV buses is analogous to that used in standard impedance-matrix based optimal load flow20, an obvious improvement was to replace the
load flow (setting AQi to zero is the same as using the calculated value NM load flow at each gradient step by FDM. To produce a program con-
of Qi in place of QisP). This will detract from the accuracy of the taining no matrix triangulations during the solution, we conducted
solution for V at each application of eq. (9). However, this may not many studies on eight different medium-size systems (up to 1 8 buses)
matter too much - as indicated in Fig. 2 of the paper, V convergence is to determine how the linear N-calculation can be accomplished using
usually well ahead of 0 convergence during the iterations, and some the factored [B'] and [B"] matrices, and how its accuracy affects the
slowing-down of the former may be tolerable, except for fast approxi- overall optimisation. We formulated a decoupled iteration scheme ex-
mate load flows taking one or two iterations. In practice, we think that actly equivalent to FDM with [AO] and [AV] in eqs. (8) and (9) re-
it will rest with individual users to determine whether a single B matrix placed by [AXp] and [ANq] respectively, and different left-hand sides,
is adequate for their systems. The difficult 9-bus problem mentioned by such that exact values of [N ] and [X ] could be obtained by iterating
Mr. Hobson, with solution voltages down to 50%, could be solved ac- to convergence. It was founS that if the last N-values are stored for use
curately by the normal (B' and B") method in 17 iterations. as starting values, a single iteration of this process at each gradient step
Messrs. Laha and Bollinger have referred to what is an interesting gives the accurate optimal solution obtained by the original Newton
apparent anomaly between the performances of eq. (9) and the de- program, usually in the same or slightly larger number of steps.21
coupled Newton eq. (5). The answer lies in our point 1 above. In We conclude by thanking the discussers again for their con-
reference 11 the difficulties experienced by Newton's method with the tributions and kind remarks.
non-linearities in the defining function [AQ] were overcome by using
the more-linear current-mismatch formulation. In hindsight, the simple REFERENCES
expedient of dividing each AQi by Vi would presumably have been just
as good. At this stage therefore, we have no reason to advocate the use 20. 0. Alsac and B. Stott, "Optimal load flow with steady-state
of decoupled current-mismatch formulations. Lack of success of the security", Paper T 73 484-3, presented at the IEEE PES Summer
Meeting, Vancouver, July 1973.
Manuscript received October 23, 1973. 21. 0. Alsac, Ph.D. Thesis, UMIST, Manchester, U.K., in preparation.

869

Authorized licensed use limited to: Eliud Cabrera Castillo. Downloaded on January 21, 2009 at 16:02 from IEEE Xplore. Restrictions apply.

Das könnte Ihnen auch gefallen