Vous êtes sur la page 1sur 7

116 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, VOL. PAS-90, NO.

1, JANUARY/FEBRUARY 1971

more complete picture of maximum interchange, it would be slower, longer running times; a significant increase is not expected, although
and neither approach is totally appropriate because neither is a this remains to be proven. The ordering algorithm treats a remote
dynamic solution. control constraint, for instance, as separate and distinct from both
We share all the discussers' beliefs that optimization programs will buses and assigns it a number according to its own fill-in pattern.
prove to be an extremely powerful and useful tool. We hope to set Some constraints will certainly prove difficult to order, but others
up a program of this sort, but -when we do, we expect to find the may even be better than the normal bus ordering. Similarly, some
equality constraint forms presented here equally as useful in optimiza- constraints may prove easy to misapply, causing a high percentage of
tion for inequality constraints and objective functions. Only time will divergent cases. The results of these last two problems may be that
tell which algorithms are best for which problems. some models will prove more trouble than they are worth; only pro-
Solution efficiency can be degraded by more complex coding, less duction testing will determine this.
efficient ordering, or greater tendency toward nonconvergence. More The paper presents ideas, not laws of nature; it is hoped that others
sophisticated programs necessitate more complex codes which mean will be interested enough to experiment and expand these ideas.

Improved Methods for Load-Flow Calculations


YEHUDA WALLACH, MEMBER, IEEE, REED K. EVEN, MEMBER, IEEE, AND YAAKOV YAVIN

Abstract-Further modifications of Newton's method as applied The differences between the solution given in [1 ] (and refined
to the solution of load-flow problems are described. Considerable there) and the one described in' [3] are: 1) Cartesian, instead of
saving in computer time is achieved when using these methods. polar coordinates, are used; 2) in [3 ] the Jacobian matrix is fac-
The algorithms proposed can be adapted to the case when main tored into a product of two triangular matrices at each iteration
storage space is limited by using backing magnetic tape storage.
step, using the Gauss elimination method. In [1 ] and here, the in-
verse of the Jacobian matrix is computed once and then corrected
at each step or at each kth step (modified Newton-Raphson
INTRODUCTION method).
Four modifications of the Newton-Raphson method are
^TEWTON'S METHOD as applied to load-flow calculations described and the saving in computations achieved by them is
iN [1] is clearly superior in its convergence and ability to evaluated.
solve ill-conditioned problems to methods based on the Gauss-
Seidel algorithm. Unfortunately, when using this method one FORMULATION OF PROBLEM
has to invert the Jacobian matrix of a large system at each Kirchhoff's current law for a network of n + 1 busbars (nodes)
iteration. The reduction of the amount of computations is can be written in a way [1 ] that leads to the residual equations
therefore sought in inverting this Jacobian matrix, and some of n
the methods used to achieve this are presented.
The requirement of some of the new methods for even more
r = E (givi + bkivn+±) + Gko
storage space than that described in [1] is not prohibitive since:
1) the storage space of present-day computers is increasing fast- PkV0k - Pn±1Vn+k 0
Vk2 + Vn+k2
a point raised already in [1], [2]; 2) the load flow is computed
most often as a part of transient stability [3] or optimization [4] n

calculations and in both the network is substituted by a reduced rn+= i= 1 (bkiVi - qkivn+i) + Gk,n+l
one, with a much smaller number of busbars; 3) the proposed
methods can easily be adapted to the use of backing memory, Pn+±Vk + PkVn+± = 0 Ik (1(1))
the volume of which is practically unlimited. A program written
in ALGOL and using a limited main space storage is presented in
the Appendix. where
Pk +jPn+k power flowing into kth node
Paper 70 TP 162-PWR, recommended and approved by the Power Vk - JVn+k voltage of kth node
System Engineering Committee of the IEEE Power Group for gki + jbki admittance between nodes k and i, k i,
presentation at the IEEE Winter Power Meeting, New York, N. Y., i#dO
January 25-30, 1970. Manuscript submitted September 10, 1969;
made available for printing November 19, 1969. jGk+, + admittance between kth node and slack bus
G0o+
Y. Wallach was with the Faculty of Electrical Engineering, and the buses are numbered 0 to n with the slack bus being
Technion-Israel Institute of Technology, Haifa, Israel. He is now
with Purdue University, Lafayette, Ind. 47907. numbered 0.
R. K. Even is with the Faculty of Electrical Engineering, The 2n equations (1) are solved by Newton's method, where
Technion-Israel Institute of Technology, Haifa, Israel. in each iteration step 1 one corrects the voltage vector v accord-
Y. Yavin was with the Faculty of Electrical Engineering, Tech-
nion-Israel Institute of Technology, Haifa, Israel. He is now ing to
with the Departmeint of Electrical Engineering, Virginia Poly-
technic Institute, Blacksburg, Va. 24061. v(l) = u(1- l- [(-l - (-l 2
WALLACH et al.: IMPROVED LOAD-FLOW CALCULATIONS 117

It was shown in [1] that the Jacobian matrix J(P) can be Using definition (5), (13) becomes
written as
J-'(v + w) -

J-'(v) - J-'(v) -J(v + tw) J'(v).


J(') - JO + Jd4'z = G + D-1z D(l)j (3) dt ] t=o
(14)
where the G, B, D(l), and TV) are n X n admittance matrices In load-flow problems the Jacobian matrix is of a special
defined as kind, as can be seen in (3), (4); i.e.
G = [ge], B= [bfj], i,j = 1,2,... n
J(V + w) - J(v) = Jd(V + W) Jd(v) (15)
-
D(7) = diag [di(l)], TV1) = diag [d.+i(')]
and the difference on the right-hand side of this equation can be
Pi(Vi-( ) Vn+i(1) ) -
2p(+ivi(l)vn+((l) shown to be small in the vicinity of the solution by (4). There-
(Vj( ) 2+ Vn+i(Z))22 fore, the expansion of J(v + w) into a Taylor's series results in

- Pn_i(Vi() 2 - Vn+i(l)2) + 2pivi()V+f(l)


(4) J(v + w) -
J(v) =
[djJ(v + tw)] + a(w). (16)
(d()' + V+i(l) )
APPROXIMATED INVERSE OF JACOBIAN MATRIX Substitution of (15) and (16) into (14) yields
The Jacobian matrix as can be seen in (3) and (4) is a function Jk'(v 1+ w) J'1(v) - J1'(V) [Jd(v + w) - Jd(v) ]J '(v) (17)
of the vector of unknown voltages v. We are interested in the
value of its inverse at the point v + tw, 0 < t < 1: or
S(t) A J-(V + tw) (5) J '(v + w) ._ {I - J '(v) [Jd(v + W) - Jd(V) ]}JJ1(V) (18)
where where I is the unit matrix. This approximation to the inverse of
the Jacobian matrix will be used in the sequel.
w - -J-'(v) * r(v) (6)
and r(w) is the vector of residuals at the point w in (1). Gen- PROPOSED METHODS
erally p(x) means p at point x. Since the problem is that of applying Newton's method to
Denoting by sij the ijth element of S, by Vk, and Wk the kth load-flow calculations, the voltage vector in the lth iteration
elements of v and w, respectively, and noting that step is given by (2). Equation (18) can be used to compute the
ds~ 2nr 1 inverse of the Jacobian matrix in the lth iteration step if it is
dt
E dtz
kn [lLdsVkjWk
ldVk t
(7) assumed that
V(I) = V(1-1) + w (19)
and
rn -
in the following way:
8z(t = 1) - sjj(t =0) E [I- Wk
[I (J(1-1)) _l(Jd(l) Jdg(I-1)) ] IJ(1-1)1]-1- (20)
=
kc=l LOVk it=o [J(I) ]-' =

+ E E [ I WkWm + E(w) (8) Substitution of (20) into (2) yields an algorithm that will be
k=l m=lLVk d3Vm t=O called method 2, as compared to method 1 of [1].
It is interesting to note that algebraic manipulation of (20)
where
seen that
E(w)/||W112 tends to zero when the norm |w| 0, it is
leads to
[J()]1-l = [2I (J(--1))-'J(')][J(1-1)
- - (21)
sij(l) si(O)
- = d[-]
dt -t=0
+ a6(w)
and this resembles in form an equation for a successive improve-
ment of the inverse of a constant given matrix [5]. In contra-
where 6(w)/IIwI|
tends to zero as -* 0.
Equation (9) can be rewritten in matrix form as
IIwI distinction it should be noted that in (21) J(V) changes at each
iteration step 1.
The computation in method 2 starts with an initial guess of
[dt ]t=0 (10) v(°) and the calculation of the corresponding [J(o) ]-'. Equations
(2) and (20) serve for calculating successive values of v which
From the identity approach the solution.
d The problem of convergence is not dealt with here explicitly.
-_
dt S(t) S-1(t)] = 0 (11) Suffice it to say that convergence can be proved by using t6,
corollary 4.7 1.
it follows that The following methods are based on the assumption that the
Jacobian matrix changes slowly, which is justified by compu-
dS [dS'1 tations performed, e.g., in [1, table I]. In methods 3 and 4, (2) is
(12)
LdtJ L dt ]( rewritten as
Substitution of this into (10) yields V-= V(l-1) - (22)
S(1) -
(S) 0 -S(O)L dt } c S(o). (13) and the modifications of method 2 differ in the form of
only.
B(1)
118 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, JANUARY/FEBRUARY 1971

In method 3, B(') is derived from (20) as


B [I - (J(O))-1(Jd(l) - Jd( 1))] [J(O) -1. (23)
In method 4, J.(j) (which equals zero) is substituted for Jd -
in all the iteration steps. Equation (23) may then be rewritten as
B() [I - (J(o)) -lJd(1) I [J(O) ]-1 (24)
This somewhat unusual substitution is justified by results of
test problems.
The Newton-Raphson method yields good directions for the
successive corrections in the vector v. Since the methods proposed
thus far deviate from the Newton-Raphson one, an attempt is
made in method 5 to accelerate their rate of convergence by
choosing a better value for the correction vector without changing
its direction.
Following Broyden [7], a norm +(l) (X) is defined by
2n
4$1)(X>(>) = Ej rt2(v('1z-l)-A.B(t-l)l
~ X > 0 (25)

where rk(x) is the kth component of the residuals vector r,


computed at the point x. The aim is to choose X to minimize Fig. 1. Addition to network in [1]; admittance values in mhos.
+(l)(X). To avoid cumbersome computations, Broyden has
suggested to approximate (25) by
TABLE I
= [1 - X]20(l)(0) + XIO(')(1). (26)
Minimization of ;(') (X) (subject to X > 0) yields Methods
Network 1 2 3 4 5
xm() =V1 + 60 -1 1
_o (1) (27) 1 2 4 4 4
30 1
0 2 2 4 4 4 3
3 6 10 10 5 5
Method 5, as proposed here, is based on these results by
Broyden. At each iteration step two values of v are computed:
v (1) for X = 1, and vx(l) for X - A_(l) These are given by
vl( I) = (1 -1) B(1- 1) r( - 1) or Jd(1) - Jd(-1) and the fact that J-lr is a vector lead to
,/ = 16n2 multiplications. This is incomparably lower than
V(l) - - l) (28) any of the methods of [1] or [3] or method 2, and if the number
One of these is chosen according to of iterations is not prohibitively higher than in the other meth-
ods, method 3, 4, or 5 should definitely be preferred.
Vi if 0(') ((Xm) 5
.> 1 To find this out in practice, three runs were made. Network 1
uo =lVX(1), if 0(4)(Xm) < 0(/)(1). (29) is one of five bus bars as given in [8], network 2 is that shown in
[11 (14 buses), and network 3 (26 buses) is the same as network 2
COMPUTATIONAL CONSIDERATIONS AND RESULTS except that it has 12 new bus bars connected between old bus
bars 11-13 (Fig. 1). The iterative process is terminated when
The first consideration is that of the number of multiplications 2n
,u required in each of the methods proposed.
For method 2, the number ,u required to correct the Jacobian
a( - Ei rk(t)
k=1
< e 0.0001 (31)
matrix according to (18) equals 8(n3 + n2) 8n3 since Jd(') - -
The results are summarized in Table I, which lists the number
Jd(V-1) is a matrix whose columns contain two entries each, and of iterations required to obtain the above accuracy.
multiplication of two (2n X 2n) matrices requires , = 8n3. Insofar as Table I is indicative, methods 4 and 5 are the best.
This is about the same amount of computation as that required The times involved (with programs written in ALGOL and run on
by method 1, but the claim for greater accuracy is made for an Elliot 803 computer) for network 3 were: 12, 8, 6, 4, and 4
method 2. This is based on the fact that no division is required minutes, respectively. It seems that method 5 should be the most
and there is no real error accumulation (each one of the 8n2 efficient for large networks as indicated by values of X,?j in the
elements is calculated as a separate scalar product). solution of network 3. They were 0.9312, 0.9519, 0.9780, 0.9915,
In methods 3 and 4, matrix [J(o) ]-I is used. According to (3) and 0.9926 for the five iterations, a sequence clearly approaching

j(o) = [[B Bf
this matrix equals the value 1 in an ordered fashion. The calculation of the norms

G-G()
(30) 4 in method 5 does not increase the amount of work considerably,
since it is based on the calculation of 2r2 which is usually needed
anyway (and not a(l) as here).
and thus depends only on the network data, but not on the In all the proposed methods, the first step is to invert J(o)
loading. Its inverse may therefore be computed once for a given [(30)]. This matrix is symmetric, as can easily be verified by
network and given as part of the input to each one of the load- noting that the matrices G and B (being the conductance and
flow cases to be solved. The special kind of the matrices Jd(') the susceptance matrices, respectively, of the given network)
WALLACH et al.: IMPROVED LOAD-FLOW CALCULATIONS 119

for i:=1 step 1 uhtil N do-


be.gin
readnx(tempc);
d[i ] : =O;
hmhidettempc,PJriJ);
begin- integer i,j,n,N,M, , limit;
read n; N:-n+n; end;
begin varra,V.V,p,G,d,D,r,R,k,K,tempcrt:NI,
-read eps,limit;
for i:=1 step 1 until
tempaCi :nl; n do
integer array PJ[1 :NI,PG,P13[1 :ni; read v[i],vri+nl],p[i],Pri+n]G[i]EG[itn+].
for i:=i step - until N do V[i]:wv [i];
real xa xb,xc,xd,f,F,Fl,res,Res,tetaI 1: % ; fres tV rpfp,res ); dt (V, d); F: =f;
lambda,eps; for i:=1 step 1 until N do
real procedure dot(a,b); array a,b;
begin real m; integer ; m:=0; beginn
for j:=l step 1 until N do m:=m+a[j]*b[j1]; mxback(tempc,PJ[i]);
dot: m k[i]:=K[il:=dot(ten.pc,r);
end of dot product of vectors a and b; end;
p-rocedure fres(vr, f,res); array v,r; labl: 1: =1+1; i4f l>limit then goto out;
real f,res; for i:=1 step 1 until N do v[ih:=vCi]-K[i];
begin integer i, j; real z; f:=res:=-O; fres(v,r,f,res) ;teta:=f/F;
for i:=1 steP 1 until n do-
lambda :=(sqrt (1+6*teta) -1 )/t3*teta);
for i:=1 step 1 until N do
beLin M: =i-+n; wa : -f ri ]) t2+(v[M ) t 2; VEi 3: =Vfi ]-lambda*:KFi I;
mnxback(tempa,PG[i 1);
for J:=1 step 1 until n do fros(Vl,R,F1,.Ros);
temoe C -:=tempa(j]; if F1XF then
begin
mxback(tempa,.PB[i); F:=:,1; res:=Res;
for j:=1 step 1 until n do end else
tempc[j+nl :=temparj 1;
rti ] : =Gi l+(p[Ml*v(M1- for i:=1 stop
p[il*vti])/xa+dot(tempc, v); I until N do
for j*=1 step 1 until n do
VEi]:=v[ilJ; fl[il:=r[i];
end; F::=f;
z:=tempcEjl ;tempc[j]:=temp)c[J+nl ; end;
tempc [j+n]: =-z; if res<eps then goto out;
end; dtY, D);
rE1]:-=G[MI+(p[M]*vti 1+ for i:=l step 1 until
p.i]i*v[M])/xa+dotCtempc ,v); do
f:=f+(r[iJ)t2+(r[M])t2; begin nback(tempc pJril);
kri :=dot(tempc, R);
res:=res+abs(rEi3)+abs(r(MI);7 end;
end;
end of fres which calculated the vector of for i: = step 1 until n do
resiuals r, the sums of their squares (f) begin nixback(tempc, JiJ);
for j:=i step 1 until n do
and of their absolute values (res);
procedure dt(v,d); array v,d; begin xa: =tempcE.ii ;xb:--tempe[n+.l];
begin inteter i; xc:==-xa*d[j]+xb*d[j+nl;
for i:=1 step 1 until n do xd:==-xa*dfj÷nJ-xb*d[jl;
if j=i then xc:=1+xc;
begin M:=i+n 5 xa:=(vEiJ)t2;xb:=(v[M])t2; tempc j :Ac; teic Lj+n. : d-;
xc:=(xa+xb)t2; xb:xa-xb; end; KX[il:=dot(tempc,k) ; -
xa:2*vEM]*v[iJ; -rnxback(ternpc PJ[i+nJ);
dE[ih:=(pEi3*xb-,n[M]*xa)/xc; for j:=7 steI until n do
dEM]: =(pEM*xb+pi 1*xa)/xc; begin :xc :=tempcEjl ;xd: =tempc[j÷nJ;
end;
end of dt which computed vector d; tempo U]: =wcx*dEj]-xd*d[j+n];
for i:=1 step 1 until n do tempc [j+nl: =c*d[j+nj+xd*dji];
begin if j=i then
readmx ( tempa ); nrxhi de (tempaItPG[i ]);
tempcej+n]: 1+tempc[j+np;
end; K[i+n]:=dot(tempc,k);
end; end;
for i:=1 sten 1 until n do for i: =1 step 1
begin goto labl;
until N do dEi3:=DEil;
readmx(tempa); mxhi4-de(tempa,PB[i ]); out:end;
end ; (
a
) end; (b)
Fig. 2. ALGOL program corresponding to algorithm 5.
120 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, JANUARY/FEBRUARY 1971

are both symmetric. Thus the proposed mathematical formula- Discussion


tion [I] is superior in storage space and amount of computations
to the polar coordinate one proposed in [3], which yields a
"symmetric in pattern but unsymmetric in value" Jacobian Norris M. Peterson (Northern States Power Company, Minneapolis,
matrix. Minn.) and W. Scott Meyer (University of Minnesota, Minneapolis,
As already stated, the Jacobian matrix J(o) is inverted only Minn.): The authors are to be commended for their efforts to develop
a method for updating the Jacobian matrix of the Newton power
once. The amount of computations involved in this inversion can flow by application of advanced techniques of linear algebra. How-
be reduced by using the sparsity and the symmetry of J(O) and ever, a basic misconception concerning current Newton power flow
reordering it accordingly. Algorithms for this reordering based solution methods [3] appears throughout the paper, and this dras-
on graphs were developed recently [9], [10]. The blockwise tically affects the comparisons made by the authors, as shall be
structure of J(o) can be used as well to considerably reduce the shown.
Current Newton power flow programs [3], [12] have largely
storage space required. replaced all other solution procedures mainly because the Jacobian
matrix is not explicitly inverted. Rather, one uses sparsity-directed
CONCLUSIONS elimination or optimally ordered triangular factorization (OOTF),
to solve the linear Newton equations. This requires far fewer arith-
Modifications of the basic Newton-Raphson method for metic operations (for large n) than either the p = 8n3 or, - 16n2
solving the load-flow problem were developed. Numerical tests figures mentioned by the authors.
of several examples show that method 5, using an optimization As a first illustration, one might consider [13, table II]. This
table shows that the arithmetic operations required to triangularize
factor, promises to be the most efficient of those described the complex nodal admittance matrix of a 160-node power network
when large networks are solved. All the methods (3-5) given includes only 6110 multiply (and/or divide) operations. Thus the
lead to a considerable saving of computer time. similar operation count for triangularizing the Jacobian matrix of
the same 160-node network would typically be four times as large, or
24 440. The extra factor of 4 approximately reflects: 1) the use of
APPENDIX two real equations for each node of the system as opposed to the
PROGRAM FOR ALGORITHM 5 single complex equation of [13], and 2) the lack of symmetry in the
Jacobian matrix as opposed to the symmetric admittance matrix of
The ALGOL program given in Fig. 2 corresponds to algorithm 5. [131.
Programs corresponding to the other algorithms are similar to This figure of about 25 000 can be compared with the minimum
this, which uses magnetic tape backing storage. number of multiply operations mentioned for any of the authors'
Two commands for the transfer of blocks of data between the schemes; for the same 160-node system, the authors' multiply
count is seen to be u = 16n' = 409 600. Thus for this 160-node sys-
tape and the main storage are used. These are: tem; the complete OOTF of the power flow Jacobian matrix
1) mxhide (a,b); store vector a on the tape and assign to the would require only about 6-pereent of the u - 16n2 multiply
integer b the location on the tape of the first element of a. operations mentioned for the authors' approximate procedure. The
2) mxback (a,b); store in a the vector stored on the tape superiority of OOTF solutions further increases with system size,
since the computational effort for OOTF only varies linearly with
starting from tape location b. system size (,u o n) [141.
The data are read into the main storage from the input by the Tinney and Walker [11] have derived the following equation for
readmx (a) command. estimating the number of multiply operations required to triangu-
Vectors tempa and tempc are used for temporary storage larize an unsymmetric matrix of order n by OOTF:
during transfer of data. ?PJ[i], PG [i], and PB [i] show where the n-1
ith rows of the J(o), G, and B matrices, respectively, are stored. multiadditions = E rt2 (32)
A network of about 2000 busbars can be solved by this i=I
program on a computer with a 64K memory and a tape. where ri is the number of nonzero terms to the right of the diagonal
in row i of the table of factors. The triangularized Jacobian
REFERENCES matrix (table of factors) for a 1000-node power system (2000-order
matrix) will typically contain about 13 000 upper-triangular terms
[1] Y. Wallach and R. K. Even, "Application of Newton's method [3]. This indicates that, on the average, a row will contain about 6
to load-flow calculations," Proc. Inst. Elec. Eng. (London), or 7 terms to the right of the diagonal. If we further assuLme (for
vol. 114, pp. 372-374, March 1967. simplicity) that the ri are uniformly distributed about this mean,
[21 B. Stott, Y. Wallach, and R. K. Even, discussion to [1], then the minimum would be unity and the mnaximum would be 13.
Proc. Inst. Elec. Eng. (London), vol. 114, pp. 1163-1164,
August 1967. The operation-count estimate of (32) then yields
[31 W. F. Tinney and C. E. Hart, "Power-flow solution by New-
ton's method," IEEE Trans. Power App. Syst., vol. PAS-86 multiadditions 2000 [13S 1] = 170 000.
pp. 1449-1460, November 1967.
[4] H. W. Dommel and W. F. Tinney, "Optimal power-flow solu-
tions," IEEE Trans. Power App. Syst., vol. PAS-87, pp. 1866- This is to be compared with the = 16n2 -16 000 000 multiply
1876, October 1968. operations for the authors' method. Thus existing Newton power
[5] B. P. Demidovitch and J. A. Maron, Fundamentals of Compu- flow programs can "exactly" calculate the factored inverse (table
tational Mathematics (in Russian). Moscow: Nauka, 1966.
[6] B. Wendroff, Theoretical Nrumerical Analysis. New York: of factors of LU decomposition) of such a Jacobian matrix with
Academic Press, 1969. typically 1 percent of the effort required by the authors to approx-
[7] C. G. Broyden, "A class of methods for solving nonlinear imately update their explicitly stored inverse (upon which the , =
simultaneous equations," Math. Comput., vol. 19, p. 577, 1965. 16ft2 figure is based).
[8] Y. Wallach, "Gradient methods for load-flow problems," The preceding examples, as well as others [3, p. 1449], [12, p. 6],
IEEE Trans. Power App. Syst., vol. PAS-87, pp. 1314-1318, or [14, p. IV-13], clearly demonstrate the falsity of the authors'
May_ 1968. statement in the paragraph relating to (30) that their 16n2 multi-
[9] F. Harary, "A graphical-theoretical approach to matrix in- plication "is incomparably lower" than for existing Newton power
version by partitioning," Numer. Math. (in German), vol. 4,
p. 128. 1962. flow programs [3], [12].
[10] A. Nathan and R. K. Even, "The inversion of sparce matrices
by a strategy derived from their graphs," Comput. J., vol. 10,
pp. 190-194, August 1967. Manuscript received February 12, 1970.
WALLACH et al.: IMPROVED LOAD-FLOW CALCULATIONS 121

The authors' procedures appear to have other operational dis- TABLE II


advantages, too. It is stated in the paragraph relating to (30) that COMPARISON OF ITERATIVE METHODS
the Jacobian matrix need be inverted only once for a given system
and then may be supplied as input data for each case being solved.
It must be questioned how the inverted Jacobian matrix can be Relative
modified to reflect generator type switching (required to satisfy Running Time
generator var-inequality constraints), transformer tap changes, Method Iterations (seconds)
line outage conditions, and the like. Modern-day studies of a system 1) Modified Newton [16] 10 3.2
simply do not have either a constant admittance matrix or even a 2) Regular Newton [16] 4 4.2
Jacobian matrix with unchanging structure. 3) Modified extended
Newton (39) 9 16.7
REFERENCES 4) Modified multilinear
[11] W. F. Tinney and J. W. Walker, "Direct solutions of network update (36) 10 17.5
equations by optimally ordered triangular factorization," 5) Regular extended
Proc. IEEE, vol. 55, pp. 1801-1809, November 1967. Newton (38) 5 62.1
[12] H. W. Dommel, W. F. Tinney, and W. L. Powell, "Further 6) Regular multilinear
developments in Newton's method for power system applica- update (35) 6 74.7
tions," Paper 70 CP 161, presented at the IEEE Winter Power
Meeting, New York, N. Y., January 25-30, 1970.
[13] W. S. Meyer and V. D. Albertson, "Improved loss formula
computation by optimally ordered elimination techniques," r in a generalized Taylor series [15] through three terms gives
Paper 71 TP 136, presented at the IEEE Winter Power Meet-
ing, New York, N. Y., January 25-30, 1970. r(v + w) r(v) + Dr(v)w + - D2r(v)(w)2. (37)
[141 IBM Corporation, Power System Computer Feasibility Study 2
for the Department of Public Service of the State of New York,
vol. 1. San Jose, Calif.: IBM Research Division, December Assuming that v + w = v, (36) is equated to zero and solved for
1968. the correction vector w. Using the implicit function theorem, with
an additional series truncation, the regular extended Newton
algorithm is obtained [17] as follows.
V()= v(l) - [J()] -lrtl) - [J(l) ] -DlDr(l)I[J() ]-lrl) } 2. (38)
2
Notice that the regular Newton method simply neglects the last
term. In addition, a modified extended Newton algorithm is given by
J. Meisel and R. Barnard (Wayne State University, Detroit, Mich.
48202): In this discussion we give brief comparisons of regular, V(Z+I) = v -[J(O) ] -lr(l) - [J(O) ] l'D2r(0) I [J(O) ] -lr(l) }2 (39)
linearized, and extended Newton methods that are similar in concept 2
to those presented by the authors. At present, our conclusions where only r needs to be updated on each iteration.
favor the modified Newton method for the load-flow problem. These six algorithms have been applied to the Ward-Hale six-bus
The approximation introduced in (10) is meaningful only for small system [18] using (31) as a termination criterion. A conventional
w. If increments between iterations are large, then one must consider flat start was used. These tests were performed on a Burrough's
the more general first-order expansion [15, p. 186]: B5500 time-sharing system. The results are listed in Table II. No
J-'(v + w) = J-(v) + DJ-'(v)w special effort was made to optimize the FORTRAN program either on a
storage or a running-time basis; however, certain conclusions seem
+ (1 - X)D J-(v + Xw)(w, w) dX (33) clear.
The multilinear update (which corresponds to the author's
where v and w are regarded as elements of a Banach space B, D methods 2-4) should fall between the modified Newton and regular
denotes the Frechet derivative operator, and (w, w) is an element of Newton methods in number of iterations to obtain convergence.
the product space B X B. This was the case with the regular multilinear update algorithm with
If the remainder term is neglected, then (33) can be manipulated 6 iterations. However, the modified muiltilinear update took the same
into the form [15, p. 148, theorem 8.3.2J: 10 iterations as the modified Newton method. Since both these
multilinear update methods have much greater relative running
J-'(v + w) J-1(v) - J-'(v)[D2r(v)w]J-(v) (34) times than even the regular Newton method, they appear to offer
no practical advantage for the load-flow problem based on this
which is identical to (14). single experiment. Even if the running times were reduced by the
Corresponding to (20), we have established directly from (34) approximation used by the authors for the second Frechet derivative
the following algorithm for approximating J-: of r(v), the modified Newton method appears to have an adequate
convergence region and has the smallest running time.
[Jz]- I _ (ju -I) )_1D2r(1-1)(v(1)-ull)[(-)- (35) Similar conclusions apply to the extended Newton methods, both
regular and modified. For this problem, in fact, the regular extended
The algorithm in (35) is termed regular multilinear updating. Newton method took one more iteration than the regular Newton
Equation (35) differs from method 2, as given by (20), in that the method, with much greater running time. Again the running time
second Frechet derivative is used [16] rather than the approximation could be cut by an approximation for D r in (38). The only possible
given by (16). advantage to be gained with the regular extended Newton method
In addition, a modified multilinear updating algorithm is defined is a slightly larger region of convergence. We have not yet con-
by clusively shown this result.
[J(i) ] -l = [I - (J(O))-DJ2r()(v(1) - v(0)) [J() 3 -l (36) Even though the authors have formulated the load-flow constraints
as all load buses, this should not be looked upon as an essential
The modified multilinear updating algorithm of (36) is similar, but limitation of the methods presented. Our work includes the con-
clearly not identical, to methods 3 and 4 given by (23) and (24). ventional generator bus constraints. In fact the generalization of
In method 5, the authors are multiplying the correction vector w constraints as proposed by Britton [19] can be included with some
by a scalar factor X. We have also mnodified the correction vector of added programming complexity. The authors' statements concerning
the Newton method leading to an extended Newton method. Ac- symmetry in the Jacobian matrix must be modified if a generalization
cording to (1), the solution v such that r(v) = 0 is desired. Expanding of constraints are included.
The authors' statements concerning convergence of the algorithms
are true but of little value. The fact that there exists a neighborhood
Manuscript received February 26, 1970. S around v such that if v° E S, vl --+ v as co is valuable only if S
=,
122 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, JANUARY/FEBRUARY 1971

is a large defined region. Such information on S would allow for the discusser would like to know the authors' comments on the
guaranteed convergence for vu selected at some distance from v. mixed formulation employing rectangular coordinates for load buses
Knowledge of the region S would also give us an additional factor in and polar coordinates for generator buses with constant voltages.
comparing algorithms to go along with computing time and storage Finally the discusser wishes to thank the authors for making
requirements. available a copy of their paper.
In summary, our preliminary conclusions on a class of algorithms
quite sirmilar in concept to those presented in this paper are that
the modified Newton method seems best for the load-flow problem.
Where hard limits (such as generator var limits, etc.) are imposed,
and in effect v moves, an updating of J is required. The modified
method should be continued for as many iterations as possible. Yehuda Wallach, Reed K. Even, and Yaakov Yavin: The authors
wish to thank all those who have taken the trouble of writing down
REFERENCES
-their remarks on the paper.
The difference between our method and those mentioned by Mr.
[15] J. I)ieudonne, Foundations of Modern Analysis. New York: Peterson and Mr. Meyer is that in the first case the inverted matrix
Academic Press, 1960. is updated in every kth iteration only whereas in their methods a
[16] J. Meisel and R. D. Barnard, "Applications of fixed-point set of linear equations is solved by triangularization in each itera-
techniques to load-flow studies," IEEE Trans. Power App. tion. In the most general case such a solution requires I = 1n3
Syst., vol. PAS-89, pp. 136-140, January 1970. operations for I iterations, whereas in the method suggested in our
[17] R. Barnard, "Extended Newton algorithmns for calculatinig paper Al = (I/K)16n2 operations only are required. There is no
system singular points," Paper D-6:7, presented at the 3rd
Hawaii Conf. on System Sciences, Honolulu, Hawaii, January question that our method is clearly superior if these numbers are
1970. taken as the basis of discussion.
[18] J. B. Ward and H. W. Hale, "Digital computer solution of The methods proposed in our paper are not amenable to the
power-flow problems," AIEE Trans. (Power App. Syst.), application of sparsity considerations for lowering the number of
vol. 75, pp. 398-404, June 1956. operations. This is shown very convincingly by M'Ur. Peterson and
[19] J. P. Britton, "Improved load flow performance through a more Mr. Meyer. And yet when deciding what method to use one should
general equation form," this issue, pp. 109-116. pose the following considerations.
1) Is the structure of the network such that (32) holds true, i.e.,
that the "average" mentioned by the discussers is relevant to the
particular case?
2) How sparse is sparse? The measure of sparsity changes when
stability or optimization is considered.
D. K. Subramanian (Indian Institute of Science, Bangalore, India): 3) Taking advantage of sparsity means additional logic operations
The authors are to be congratuilated for developing four new methods for establishing the addresses of nonzero ternms. No comparable
for load-flow studies. Also, such mathematical rigors as norms and expenditure in computation time is required by our methods.
operators of fuLnctional analysis are introduced. The discusser would As the discussion by Mr. Meisel and Mr. Barnard shows, there are
like to know the opinions of authors on the following points. at least somiie indications that sparsity couild be used for uipdating
1) In method 4 Jd(°) is assumed to be a null matrix. Will the J-1. The whole comparison between the methlods will have to be done
convergence improve if instead Jd calculated during the first itera- fronm the beginining in this case. Our inethod can be imiproved, as
tion is used? Also Jd(-1) is a null matrix and not Jd(°). Mr. Meisel and Mr. Barnard show, by incluiding remainder terms,
2) Usually for normal well-conditioned systems the improvement thereby checking also on results. This can be done in triangulariza-
on voltage variables obtained during the first iteration of Newton's tion only at the penalty of additional time expendituire. MIoreover,
method is considerable compared to the values obtained in the there are indications in these discussers' notes that generator type
succeeding iterations. Hence if the methods 2-5 are used after one or switching and other changes in the structure of the iiimpedance matrix
two iterations using method 1, will the convergence improve? could be taken care of in our methods.
3) Equation (24) for BU') can also be obtained in the following It seems that there is no point in trying to imnprove our imethod 2
manner. Let Us use the AMonte Carlo method to invert [J(O) + Jd]: (as considered by Mr. Peterson and Mr. MIeyer) since we have dis-
carded it as an inefficient method.
[J(O) + Jd]- = [J(O)]-I[I + JdJ(°01] 1 (40) We have appended an ALGOL progranm to make the algorithm
[J(O) + -
Jd]l J(O)-1[I JdJ(O)-1]
-
(41) clearer. Of course the same algorithm can be programmed in ma-
chine language, thus saving the added penalty (which, contrary to
= J(O) - J(o) -JdJ(°) -1
-
what could be erroneously understood from Mr. Peterson and Mr.
which is the same as (21). Meyer's remark, is not inherent to the method) of using programmer-
Equation (41) is obtained from (40) using oriented language.
We did not have time to run the program as suggested in 1),
(I + AX)Y- = I-AX + AX2- (42) 2), and 5) of Mr. Subramanian's discussion, so we cannot give a
(I AX)-,
+ I- _
AX. (43) positive answer. We do feel that the proposed changes can speed up
the solution.
4) Another factor is sparsity, which is lost in these methods. We are thankful for the suggestions in the two last letters and
5) Since the authors very well explain the advantages of using think that it will still take some time before the controversy is
rectangular coordinates for voltages instead of polar coordinates, solved.

Manuscript received February 3, 1970. Manuscript received May 30, 1970.

Vous aimerez peut-être aussi