IEEE Transactions on Power Delivery, Vol. 4, No.

2, April 1989
ss 2
U - - - ' - -
1
1401
NETWORK RECONFIGURATION IN DISTRIBUTION SYSTEMS
FOR LOSS REDUCTION AND LOAD BALANCING
M e w E. Baran Felix F. Wu
Department of Electrical Engineering and Computer Sciences
University of California, Berkeley
Berkeley, CA 94720
Abstract - Network reconfiguration in distribution systems is realized by
changing the status of sectionalizing switches, and is usually donefor loss
reduction or for load balancing in thesystem. In t hi s paper, general for-
mulation and solution methods areproposed for theseproblems. In net-
work reconfiguration for loss reduction, thesolution involves a search over
relevant radial configurations. To aid thesearch, two approximatepower
flow methods with varying degreeof accuracy havebeen developed. The
methods are computationally attractive and in general giveconservative
estimates of loss reduction. For load balancing, a load balanceindex is
defined and it is shown that the proposed solution method for loss reduc-
tion can also be used for load balancing. Test results are included to show
theperformanceof the proposed method.
Keywords: distribution automation, distribution systemoperation, distri-
bution systemplanning, power flow analysis, combinatorial optimization.
j I. INTRODUCTION
In primary distribution systems, sectionalizing switches are used for
both protection, to isolate a fault, and for configuration management, to
reconfigurethenetwork. Fig.1 shows a schematic diagramof a simplified
primary circuit of a distribution system together with sectionalizing
switches.
'7 CB7
U"
I
Figure1: Schematic diagramof a primary circuit of a distribution system
In the figure. load points, wherethe distribution transformers aretapped
off fromtheprimary circuit, is marked by dots, " 0 ". As alsoshown in the
figure, there are two types of switches in the system: normally closed
switches connecting the line sections (CB1 - CB6 ), and normally open
switches on thetie-lines connecting either two primary feeders (CB7), or
two substations (CB8), or loop-typelaterals (CB9).
Distribution systems are normally operated as radial networks; how-
ever, configuration is changed during operation by changing thestate of
somesectionalizing switches. For example, in Fig.1, switches CB7 and
CB8 can beclosed and CB3 and CB6 can be opened to transfer load from
onefeeder to another.
Especially with theintroduction of remotecontrol capability to the
switches, on-line configuration management becomean important part of
88 SM 556-3 A paper recommended and approved
by the I EEE Transmission and Distribution Committee
of the I EEE Power Engineering Society for presentat-
ion at the I EEE/PES 1988 Summer Meeting, Portland,
Oregon, July 24 - 29, 1988.
February 1 , 1988; available for printing
Xay 4, 1988.
Manuscript submitted
distribution automation. An important operation problemin configuration
management is network reconfiguration. As the operating conditions
change, thenetwork is reconfigured for two purposes: (i) to reducethesys-
tempower loss, (ii) to relievetheoverloads in thenetwork. Wewill refer
to the first problemas network reconfiguration for loss reduction and the
second as load balancing. Another configuration management operation
involves the restoration of serviceto as many costumers as possibleduring
a restorativestate following a fault. This problemis called servicerestora-
tion apd can betreated as a special load balancing problem. The network
reconfiguration for loss reduction can also beused in planning studies with
a different intelpretation; namely, to decidethrough which feeders thenew
customers areto besupplied.
The early studies on thenetwork reconfiguration weredirected to the
planning stage [3-51. In planning, the main objectiveis to minimizethe
cost of construction. An early work on network reconfiguration for loss
reduction is presented by A. Merlyn [6]. His solution schemestarts with a
meshed distribution systemobtained by considering all switches closed
and then theswitches are opened successively to eliminatetheloops. An
equivalent linear resistive network model is used to determine the
branches to be opened. In a study doneby Ross et al. [7], two different
search algorithms are giventor feeder reconfiguration. They developed
someindices to measurethe degreeof constraint violations andused them
to obtain a feasiblepoint when the operation point is not feasible. These
indices arealso used to check theoptimality of thesolution for power loss
reduction. Recently, Civanlar et al. [lo] presented a computatiody
attractive solution procedure for power loss reduction through network
reconfiguration. A simpleformula was derived based on somesimplifying
assumptions to calculatethe loss reduction as a result of a load transfer
between two feeders.
In [ 7] , a detailed recipefor servicerestoration is given by using the
indices developed for feasibility. In [8]. Castro et al. proposed simple
search techniques for sewice restoration and load balancing considering
data baseand implementation requirements for on-linedistribution auto-
mation applications. Castro and Franca [9] recently proposed modified
search algorithms for service restoration and for load balancing. A
modified Fast Decoupled Load Flow is used to check theoperating con-
straints. Aoki et al. proposed moredetailed search methods for service
restoration and feeder load balancing in [ll] and [12] respectively. They
consider the capacity and voltage constraints and use an approximate
power flow solution algorithmto determinethe loads to be transferred
between the two feeders/transformers.
In t hi s paper, weconsider thenetwork reconfiguration problemfor
both loss reduction and load balancing. Wefollow thesolution approach
proposed by Civanlar et al. However, here we introducetwo different
methods, with varying degreeof accuracy, to approximatepower flow in
the system after a losd transfer between two substations, feeders, or
laterals. The methods make use of a new set of power flow equations
which have been developed specially for radial distribution feeders and
used in the capacitor placement problem[13]. Weusetheseapproximate
power Bow methods to estimateboth loss reduction and load balance in
the system. Becausereactivepower flows are explicitly included in the
equations, themethods can also be used for systems that arenot well com-
pensated.
The organization of therest of thepaper is as follows. A general for-
mulation of theproblemis given in the next section and a general search
algorithmis presented in section 3. In section 4 and 5. the estimation
methods for loss reduction and load balanceare given respectively. The
proposed methods have been programmed and tested and the results are
given in section 6. Conclusions aregiven in section 7.
0885-8977/89/04OO- 140 1 $0 1 .oO 0 1989 IEEE
1402
II. FORMULATION OF THE PROBLEM
In this section, the network reconfiguration problems for both loss
reduction and load balancing are formulated and their similarities are
pointed out.
2.1 ProblemStatement
To simplify the presentation, wewill represent thesystemon a per
phase basis and theloads along a feeder section as constant P,Q loads
placed at theend of thel i es. Weal soassumethat every switch is associ-
ated with a line in thesystem. For example, weassumethat thesystemof
Fig.1 can be translated to an equivalent network shown in Fig.2.
ss1
m
Figure2: Oneline diagramof a small distribution system
In the figure, solid branches represent thelines that are in service and con-
stitute thebase radial configuration. Dotted branches (branches 20,2 1,22)
represent the lines with open switches.
The base network can be reconfigured by first closing an open
branch, say branch 21 in thefigux. Sincethis switching will createa loop
in thesystem, (composed of branches 1,2,3,21, 11, 10,9,8,7, and 15), a
branch in theloop containing a switch has to be opened, say branch 7, to
restoretheradial structureof thesystem. As a result of this switching, the
loads between thebranches 7-1 1 will be transferred fromone feeder to the
Other. Wewill usethe sameterminology used in [7] and call t h i s basic
switching operation a brunch exchange between branches 21 and 7. In
general, as illustrated in the introduction, more complex switching
schemes are possible; we will simulate such cases by applying several
branch exchanges successively.
The load transfer between different substations can besimulated by
branch-exchangetypeswitchings too. In this case, substation nodes (node
SSI and SS2 in the figure) will beconsidered as a common nodealthough
they are not thesamenode. The methods to bepresented in this paper can
handleboth cases. This is an important property of theproposed methods.
The network rewnfiguration problems for loss reduction and load
balancing involve the same type of operation, namely the load transfer
between thefeeders or substations by changing thepositions of switches.
They only differ in their objective. Other factors, such as the voltage
profileof thesystem, capacities of theIinWtransformers, reliability con-
.
straints can beconsidered as constraints.
To state these problems as optimization pmblems, note that the
radial configuration corresponds to a "spanning tree" of a graph represent-
ing thenetwork topology. Thus, wehavea so-called minimal spanning
tree problem which can bestated as follows. Given a graph, 6nd a span-
ning tree such that the objectivefunction is minimized whilethe following
constraints are satisfied: (i) voltageconstraints, (ii) capacity constrains of
liies/transformers. (iii) reliability constraints.
This is a combinatorial optimization problem since the solution
involves theconsideration of all possiblespanning tms.
2.2 Power Flow Equations
To calculatethe terms in optimization problemdefined in theprevious
section, wewill usea set of power flow equations that arestructurally rich
and conducive to computationally efficient solution schemes [13]. To
illustratethem, consider theradial network in Fig.3.
...... T'1k..'....- i-1 i i+1 ( - *
P,.Q, - P; f Qr - i mi i+i*Qi+i pn 'Qn
PL ,Qti
Figure3 : Oneline diagramof a radial network
Werepresent the lines with impedances zl =r, +jx, , and loads as constant
power sinks, SL =PL +jQL .
Power flow in a radial distribution network canhe described by a set of
recursiveequations, called Di s t Fh branch equatwns , that use the real
power, reactive power, and voltage magnitude at the sendhg end of a
branch - Pi,Qi,Vi respectively to express the same quantities at the
receiving end of thebranch as follows.
p .2+Q 2
Vi'
Pi,] =pi - ri - - "LI+I
(1.i)
P?+Qi'
Viz
VLl =Vi2 - 2(r i Pi +xi Qi ) +(r? +x?)- (1 .iii)
Hence, if Po .Qo , Vo at the Erst node of the network is known or
estimated, then thesamequantities at theother nodes can becalculated by
applying the abovebranch equations successively. Weshall =fer to this
procedureas a forward update.
DistFlow branch equations can bewritten backward too. i.e., by using
the real power, reactivepower, and thevoltagemagnitudeat the receiving
end of a branch, Pi , Q, , Vi to express thesamequantities at thesending
end of thebranch. The result is thefollowing recursiveequations, called
thebackward branch equations,
Pl - l =Pl + f i T + P P , pl'2+Q,2 (2.i)
v,
(2.ii)
where, PI' =Pi +Pti , Q; =Q, +QL; .
Similar to forward update, a backward update can bedefined: start
updating from the last node of the network assuming the variables
Pn , Qn , Vn at that point are given and proceed backwards calculating the
samequantities at the other nodes by applying Eq.(2) successively. Updat-
ing process ends at the first node(node0) and will providethe new esti-
mateof thepower injections into thenetwork, PO .Qp
Notethat by applying backward and forward updateschemes succes-
sively onecan get a power flow solution as explained in [131.
23 Calculation of the Objective Terms
Having a network model, now we can express thepower loss and
measurethe load balancein thesystemin terms of systemvariables.
For loss reduction, theobjectiveis to minimizethetotal i2r losses in
the system, which can becalculated as follows.
(3)
This will be theobjectivefunction, cp of network reconfiguration for loss
reduction.
For load balancing, we will use theratio of complex power at the
sending end of a branch, SI over its kVA capacity, Si"" as a measureof
how much that branch is loaded. The branch can bea transformer. a tie-
linewith a sectionaliiing switch or simply a line section. Then wedefine
theload 'balanceindex for thewholesystemas the sumof thesemeasures,
i.e.,
This will be theobjectivefunction, cb of load balancing.
As noted before, the two problems aresimilar. They both requirethe
same data (systemparameters and load) and load flow calculation to
evaluatetheobjectives for a given network topology.
1403
III. A SEARCH METHOD USING BRANCH EXCHANGES
The radial distribution network reconfiguration problems are formu-
lated as combinatorial, nonlinear optimization problems in the previous
section. The solution involves theselection, among all possible ms , of
the best feasible one (i.e., the one that has an operating point satisfying the
constraints and minimizing the objective). Of course, a search examining
all possible spanning trees will give the solution, but it would becomputa-
tionally formidable; since, for one thing, thenumber of possiblespanning
trees that can be generated by branch exchanges will be exhaustive for
practical size problems, and another, examining a spanning tree requires a
power flow solution of the corresponding system to determine
the associated objective. Therefore, an efficient search scheme needs to be
developed. In this section, we introduce a simple, heuristic solution
method to search over relevant spanning trees systematically by using
branch exchanges.
As exemplified in the previous section, branch exchanges can be
used to create relevant spanning trees starting froma base spanning tree.
In general, given a spanning tree To. weassociate a loop with every open
branch in the network by considering as if thebranch wereclosed. Fig.4
shows such a loop associated with open branch b . Branch exchange
creates a new treeby closing an open branch, @ranch b in the figure) and
by opening a closed branch in the loop (say branch min the figure).
Figure 4: The loop associated with open branch b
The basic idea of the search scheme using branch exchanges is to
start with a (feasible) tree and then create new ones successively by imple-
menting one branch-exchange at a time. At each level, the branch-
exchange to be implemented is chosen to bethe best one (the one that
improves the objective function the most without any constraint viola-
tions) among all the possible trees (children) that can be generated f
the current incumbent spanning tree (parent) by branch exchanges. The
method can bedescribed as an algorithmwith the following steps.
Step 1: Given a feasible tree To (parent),
run a Power Flow to determine the operating point.
Step 2: Examine all the children of the parent as follows.
For each open branch b
- find a new candidate tree, T by
- identifying the loop
- deciding on thebranch, mto be removed
- for thecandidate W. T
- calculate thereduction in objective, Aq,,,,
Step 3: Sort thechildren (trees examined) by using Ach s
Step 4: Find the tree T* which has thegreatest Ach z 0
Step 5: If there is such a T ,
and satisfies the feasibility constraints.
then choose T as To and go to step 1; else stop.
Wenote the following comments about thesearch.
0 This search does not examine all the possible mes and hence thesolu-
tion will belocally optimal.
0 Computational efficiency of this algorithmhinges on two things; the
selection of branch n to be opened, since it eventually effects the
number of searches to be performed, and the calculation of objective
terms, ACS , for each calculation requires a new power flow solution.
Although the power flow solution can be obtained by DiszFfow
efficiently, nevertheless, it is desirable to be able to estimate thepower
flows faster without actually running a DistFlow for each branch
exchange considered. This will reduce the DistFlow solutions to one
for each search level (iteration).
0 The estimated power flows a~ used in ranking open branches. There-
fore. emrs in estimated figures may lead to a different search than that
of using an exact power flow.
In the next section, two different power flow approximation methods,
with varying degree of accuracy, ye given for loss reduction. In section 5,
it is shown that these methods can also beused for load balancing.
IV. APPROXIMATE POWER FLOW METHODS
FOR POWER LOSS ESTIMATION
In this section, wepresent two methods to determine the power flow in
a radial distribution systemapproximately. The methods will be used to
estimate the power loss reduction due to a branch exchange.
4.1 Method 1: Simplified DistFlow Method
Estimation of Power Flow
Wecan simplify the DistFlow branch equations, Eq.(l) by noting that
the quadratic terms in the equations represent the losses on the branches
and hence they much smaller than the branch power terms Pi and Qi .
Therefore, by dropping these second order terms wecan get a new set of
branch equations of the following form.
Pi+l =Pi - PG+l (5.i)
(5.iii)
V& =Vi2 - 2(ri Pi +xi Qi )
Since the network is radial, the solution for the simplified Di mow qua-
tions can be obtained easily; for a radial network of the type shown in
Fig.3, the solution is of the following form.
Qi +l = Q, - QG+I (5.ii)
(6.i)
n
(6.ii)
(6.iii)
We will call Eq.(6) simplifred DistFlaw equations and use themin this
section for power flow solution of a given network configuration.
The power loss on a branch can now be approximated as
Qi +l = QLk
k=+2
Vi:l =Vi2 - 2(ri Pi +xi Qi )
(7)
where, we have used the fact that Vi2 =1 p.u. Then the total systemloss is
simply the sum of all branch losses, i.e.,
0 = ri(P?+Q?) p.u. (8)
n-1
1 4
Estimation of Power Loss Reduction due to a Branch Exchange
Now consider the branch exchange between branches b (onginally
open) and m(originally closed) in Fig.4. As a result of the simplifying
assumptions made above, power flow will change only in the branches
constituting the loop shown in the figure. Let the branches in the loop that
extends between nodes 0 , . . . , k-1 and k be denoted by the set L and the
ones on the other side ( 0 , . . . ,n-1.n and k )by the the set R. Then. as
shown in Appendix A, power loss reduction due to thi s branch exchange
can be calculated as
U, - =2p, ( Cr l PI - C . r ~ P ~ ) + 2 Q m( ~ r ~ Q l - Cr1Ql ) (9)
l eL I s R I PL I cR
-(P,+Q:)[ C. r~ 1
I e R U
Eq.(9) is a quadratic function of the power transfer P,,, , Q,,, , i.e.,
U-h(P,,,,Q,,J =2drp.P,,, +2drq.Qm - tr. (P, +Q,) (10)
where, the coefficients drp , drq , tr are independent of the branch mcon-
sidered and can be calculated by using only the original branch flows,
PI , Ql . The relationship between the loss reduction ALP and the power
transfer @,Q) is illustrated in Fig.5 assuming P andQ are continuous vari-
ables. In the figure, the circle defined by ALP-b =0 divides th? P-Q plane
into two regions such that for any point inside circle m b (P , Q) >0
(positive loss reduction, i.e. losses are reduced), and for any points outside
the circle U- b ( P , Q) c 0 (negative loss reduction, i.e. losses are
increased).
1404
Figure5 : Loss reduction as a function of power transfer
This property of Eq.(lO) can beused to avoid checking every branch
around the loop for branch exchange. Let us first consider thebranch
exchange between branches b and k in Fig.4. and call it the nominal
branch exchange. Thecorresponding power transfer will be 9. Qk and
let this point (Pk ,Qk) beinsidethecircleon theP,Q planein Fig.5. Then
the points corresponding to the other branch exchanges pi e (Pk-1,Qk-I)
in the figure) will be funher away fromthe origin on theP-Q planethan
conclusions.
(Pk ,e,) Since9- 1 >P k and Qk. Therefore, Wehavethefollowing
O E F i k c o then
ALPM < O I E L andhencethertisnobranchinL thatcanbeacan-
didatefor branch exchange.
there is a branch in L that can bea candidatefor branch exchangeand
the branch to be opened should bethe onethat optimizes ALP-. This
can bechecked by star$g frombranch k and searching thebranches
backward in L until ALP*, is maximum.
Wehavethe following comments about themethod.
0 This method is efficient computationally. Both the calculation of
power loss terms. ALP and identification of branches to beexchanged
I.equires only simplecalculations.
0 Accuracy analysis of the method in Appendix B shows that a weak
bound on the error in estimating loss reduction around the loop,
ep =ALP -ALP is
.I f Wi k > O then
where. fii and 6; denotethe power losses on thebranches in R and
L respectively after the branch exchange.
The emr bounds inEq(l1) indicate that the estimate is conservativein
thesensethat when the loss reduction is large (ALP- 3 0). emr will
tend to be positive, (i.e., ALP- >0 --). ALP -ALP- 2 0 ). Similarly
when the loss reduction is negative, emr will tend to benegative, (i.e.,
U- 0 +ALP - U- I O ). However, when the loss reduction
figures are small, theemr will betwo sided. This emr analysis shows
that there may besome"misses," (i.e., a branch exchge with positive
loss reduction may beidentified as theone with negativeloss reduc-
tion) and there may be some "mislabeling," (i.e., a branch exchange
with negativeloss reduction may be identified as the one with positive
loss reduction).
4.2 Method 2: Backward and Forward Update of DistFlow
Power Flow Update
The second method makes useof the backward and forward updates of
DistFlow, introduced in section 2. to updatepower flow around theloop of
a branch exchange. For the nominal branch exchangeb-k of Fig.4. the
method comprises thefollowing steps.
Step 1: Backward Update
Updatethe power flow mund the loop by backward updatestarting
fromthenodes C and n of theloop and by canying out the power and
voltageupdates separately (i.e., use Eq(2.i) and Eq(2ii) with original
voltages, Vi to updatethe powers, and use Eq42.iii) to update thevol-
tages). Let theupdated powers be
and thevoltageupdates at the common nodebe $ : and $;
Step 2 Forward Update
Comparethe voltagedifferences at nodeo (differencebetween V, and
rid,, $; ) . If the voltage difference i s too large (larger than a
predefined value. Emu), go through a forward updateto ducf,the>mr
(this timestarting fromthe common nodeo and using V, , Pd, P, as
initial, given values and applying theforward update). Let the updated
powers be
(13)
use the difference betw~? the P k , P, and P k , FR"as power
mismatches and comt P, and Pok by adding themthe mismatches.
i.e.,
pi".di", i=ok+l,. . . ,k ; pii",Q,", i=on+l,. . . ,n
Step 3: Correct thePower Estimateat the Coqmo?,Node
(14)
L" A. .." 1"
P& = +(p; -PJ ; P, 4; +(P, - P, )
Details of development of this algorithmis given in Appendix C.
Note that backward and forward update constitutes an itemtion of
power flow solution using DistFlow branch equations. Here., weexploit
the method by localizing it to the loop of branch exchangeand performing
a special iteration. Therefore:
0 themethod is computationally more. efficient than a full power flow,
0 accuracy of themethod will mainly depend on load transfer P k , Qk.
Calculation of Power Loss Reduction
For power loss estimation, notethat
L n
Pok -Fo; =w k + w L ; P, -Pon =-@k + u R (15)
where, ALPp and ALPL represent thepower loss reductions on theR and
L sides of theloop respectively. Therefore, thetotal power loss reduction
can beapproximated as
(16)
A "
A d =U L +U R =(Pok - Pok) +(P, - F:)
V. LOAD BALANCING WITH BRANCH EXCHANGES
When the general search algorithmintroduced in section 3 is used for
load balancing, the calculations will besimilar to that of the loss reduction
case. The only differencewill bein thecalculation of the objective; for
load balancing. weneed to estimatethevalueof thenew objective, load
balanceindex, cb for every branch exchangeconsidered during thesearch.
The objective, given by Eq.(4), can however becalculated by using
the two approximate pwer flow methods introduced in Sec.IV - the
simplified DistFlow method and the forward and backwardupdate method,
becauseboth of themethods givethe approximatepower flows in thesys-
tem following a branch exchange. Once the new power flow in the
branches, P,' , Qi are estimated then thenew load balance index can be
computed by employing Eq.(4), i.e.,
P? +Q?
Cb =z-
s,-Z
When thetwo methods are compared for load balancing, simplified
DistFlow methad seems moreattractivebecause of thefollowing reasons.
0 Since the index of load balance is relative, the accuracy of simplified
DistFlow method should beade@mte.
0 Simplified DistFlow provides a quick and c ~d e estimate of thepower
flows without requiring data on network parameters.
W. TESTRESULTS
The proposed solution method Bas been implemented in Fo~tran-77.
The approximate power flow methods described in k.W, (M1) -
simplified DistFlow. and (M2) - backward and forward updates of Dist-
Flaw. are used to guidethesearch In addition, exact power flow method,
DistFlow is also usedas another method (M3), to check the accuracy of
M1 andM2. Thetest d t s for loss reduction will be presented here to
illustrate theperformanceof the proposed method.
The test systemis a hypothetical 12.66 kV systemwith a 2 feeder subs-
tation, 32 busses. and 5 looping branches (tielines). Me systemdata is
given in Table1 together with the voltageprofileof the base configuration.
The total substation loads for the base configuration 5084.26 kW and
1405
satisfy the conservative property and get closer to the exact values,
accuracy and conservative property of M1 weakens (estimation
becomes two sided). This is particularly true for estimates of branch
exchange37-28.
ii. The thnx methods lead to thesamesearches at the upper levels (up to
level 4). At level 4, whilethesearches with M2 and M3 convergeto a
local optimumas expected, thesearch with M1, which uses less accu-
rateloss reduction figures, mislabels brancli exchange37-28 (i.e., esti-
mates a negativeloss reduction as positive), and performs two more
searches leading to the global optimumpoint. However, thetwo solu-
tion points have close objective values (total loss reductions at the
solution points of M1 and M2. M3 are 125 kW and 118 kW respec-
tivel y).
iii. Branch exchanges occur on thelower voltageside of theloop.
iv. The voltage profileof the systemincreases as the loss is minimized.
(the minimumbus voltageof the syst em raises f 0.88 p.u. to 0.92
The last two observations wereused as heuristic rules in thesearch in [7]
and [lo]. However, theobservations may not always be true; a counter-
exampleis given in Fig.6, wherebranch exchange7-3 is a switching on
the higher side and results in a positiveloss reduction.
P.U.).
level
1
2547.32 kvar. The systemis not well-compensated and lossy (thetotal
loss is about 8% of the total load). A lossy systemis selectedbecause the
loss reduction is expected to be appreciable.
In the test runs, theconstraints mentioned in Sec.II are not imposed.
In fact, the voltageprofileof the basesystemconfiguration is lower t han
theusual lower limit of 0.9pa.; which shows that the system is not well
configured. Also, it is assumed that every branch in the systemis avail-
ablefor branch-exchange.
A summary of the test run is given in Table 2. In the table, each row
cormponds to a branch exchange. Branch exchanges are d e w by a pair
of numbemin second column. The other columns labeled, M1. M2. M3
correspond to searches guided by methods M1. M2. M3 respectively. The
row after a search level, row six for example. shows thebranch-exchange
chosen based on theordering of the loss reduction figures obtained in that
search level.
bout M1 M2 M3
34-14 05.34 06.64 06.79
35-7 66.30 77.25 83.43
33-6 67.70 77.89 84.23
Table 2 Test Results
lscarch I branch I lossredudon in kW
37-28
br-tx
6-7
34-14
61.31 70.15 73.17
33-6 33-6 33-6
-01.71 -02.05 -02.67
05.08 05.92 06.00
1 I 36-32 I 01.6 I 02.32 I 02.36
36-31
br-ex
6-7
37-28
11.76 13.63 15.77
08.35 05.12 0522
36-31 36-31 36-31
-01.11 -01.33 -01.58
1 1 1 i3 1 f.6. 1 17.82 1 18.30
03.00 03.06
37-28 08.35 05.12 05.22
brex 35-11 35-11 35-11
-05.19 -06.42 -07.07
34-14 -00.65 -00.80 m. 82
11-10 -00.16 -00.19 -00.22
34-14 -0044 -01 12 -01.14
I 4 I 11-10 1 m: 76 I -W:90 I -01.04
1 1 ;I. 1 37r 1
-05.02 1 -05.16
37-28 01.39 -01.35 -01.42
o7.m
1 1 ;I. 1 37r 1 -05.02 1 -05.16
37-28 01.39 -01.35 -01.42
o7.m
1 6 / ! ! I I
40.92
28-27 -03.93
i. In the Erst search level, theloss reductions are big. they satisfy thecon-
servative property discussed in Sec.IIl andthe emm in estimations are
small enough so that thesamebranch exchangeis chosen by all three
methods. However, as wego down the search levels, the loss reduc-
tions get smaller and thedifference between the two estimated values
becomes more. visible; while estimation figures of M2 consistently
v:=1 V i =.897
z , =( r +j x ) ohms
sL =pL +j QL WA
figure 6: Exampleof a branch exchangeon thehigher voltageside
From these observations. we have the following suggestions to
i mpme computational and convergencecharacteristics of the method.
0 In general backward and forward updatemethod is morereliablet han the
simplified DistFlow method in estimating the power loss reduction due to
a branch exchange, especially as the loss reduction figures get smaller.
Therefore. the decision as to which method to choose should be madeby
considering the magnitudeof loss reduction figures. A schemecomprom-
izing between accuracy and computation would be to st art with the
simplified Di mow method and then switch to backward and forward
updatemethod as the loss reduction figures get smaller.
0 The search schemegives acceptablesolution for practical purposes since
even if the solution converges to a local optimal point, the difference
between thelocal solution and the global solution will besmall. Further-
more, the convergencecharacteristics of thesearch can beimproved by
checking the locality of the solution. A possibleschemewould beto do
another quick search implementing morethan one branch exchangewith
big loss reduction at each search leveL Then the two solutions can be
compared to see if they convergeto the samepoint.
M. CONCLUSIONS
In thi s paper a general formulation of the feeder reconfiguration
problemfor loss reduction andload balancing is given and a new solution
method is presented. The solution employs a search over different radial
configurations created by considering branch exchangetype switchings.
Table 1: Network data of the test system
TIE LINES ------
------
Br. Rc. Sn. B r . Pm.- __ Sn. Node - Br. Rc. Sn. B r . Pm.- - Sn. Node -
NO Nd. Nd. r(ohm] x(ohm) PL(kW) QL(kvar)IVI**2 NO Nd. Nd. r(ohm) X(0hm) PL(kW) QL(kvar)IVI**2
1 0 1
2 1 2
3 2 3
4 3 4
5 4 5
6 5 6
7 6 7
8 7 8
9 8 9
10 9 10
11 10 11
12 11 12
13 12 13
14 13 14
15 14 15
16 15 16
17 16 17
0.0922
0.4930
0.3660
0.3811
0.8190
0.1872
0.7114
1.0300
1.0440
0.1966
0.3744
1.4680
0.5416
0.5910
0.7463
1.2890
0.7320
0.0470
0.2511
0.1864
0.1941
0.7070
0.6188
0.2351
0.7400
0.7400
0.0650
0.1238
1.1550
0.7129
0.5260
0.5450
1.7210
0.5740
100.00
90.00
120.00
60.00
60.00
200.00
200.00
60.00
60.00
45.00
60.00
60.00
120.00
60.00
60.00
60.00
90.00
60.00
40.00
80.00
30.00
20.00
100.00
100.00
20.00
20.00
30.00
35.00
35.00
80.00
10.00
20.00
20.00
40.00
0.9927
0.9574
0.9374
0.9176
0.8707
0.8641
0.8550
0.8432
0.8324
0.8308
0.8280
0.8161
0.8125
0.8099
0.8074
0.8037
0.8026
18 1 18
19 18 19
20 19 20
21 20 21
22 2 22
23 22 23
24 23 24
25 5 25
26 25 26
27 26 27
28 21 28
29 28 29
30 29 30
31 30 31
32 31 32
0.1640
1.5042
0.4095
0.7089
0.4512
0.8980
0.8960
0.2030
0.2842
1.0590
0.8042
0.5075
0.9744
0.3105
0.3410
0.1565
1.3554
0.4784
0.9373
0.3083
0.7091
0.7011
0.1034
0.1447
0.9337
0.7006
0.2585
0.9630
0.3619
0.5302
90.00
90.00
90.00
90.00
90.00
420.00
420.00
60.00
60. 00
60.00
120.00
200.00
150.00
210.00
60.00
40.00
40.00
40.00
40.00
50.00
200.00
200 .oo
25.00
25.00
20.00
10.00
600.00
70.00
100.00
40.00
0.9916
0.9845
0.9831
0.9818
0.9504
0.9373
0.9309
0.8643
0 .E557
0.8201
0.7945
0.7816
0.7739
0.7723
0.7717
Br. Rc. Sn.
No Nd. Nd.
33 7 20
34 8 14
35 11 21
36 17 32
31 24 28
Br.
r (ohm)
2.0000
2.0000
2.0000
0.5000
0.5000
-
Prm.-
x (ohm)
2.0000
2.0000
2.0000
0.5000
0.5000
1406
To guide the search, two different power flow approximation
methods with varying degree of accuracy have been developed and tested.
The methods are used to calculate the new power flow in the systemafter a
branch-exchange and they make use of the power flow equations
developed for radial distribution systems.
0 Estimation methods axe computationally very efficient and in general
give consenfative results. They also consider both real and reactive power
flows. Therefore. they can be used in searches to reconfigurea given sys-
temeven if the systemis not well compensated and reconfiguring involves
load transfer between different subtations.
.The search method introduced in this paper, as well as in [7], [lo], [ll].
has the following appealing properties: it is not exhaustive, it is of order
m2 (mis the number of open switches), and it involves about mpower
flow solutions. Its convergence characteristics is acceptable although it
does not guarantee convergepce to the global optimum. However,
modifications to this basic search is proposed to improve its computational
and convergence characteristics whenever it is needed.
For load balancing, a load balance index is defined and it is shown
that the search and power flow estimation methods developed for power
loss reduction can also be used for load balancing since thetwo problems
are similar. Between the two estimation methods, thesecond - simplified
DistFlow method, seems to be more appropriate for load balancing
because of therelative nature of load balancing concept.
Acknowledgements
Science Foundation under grant ECS-8715132.
Both accuracy analysis and the test results show that:
This research is supported by TUBlTAK-TURKEY and by National
REFERENCES
Distribution Systems, Westinghouse Electric Corp., East Pittsburg,
PA, 1965.
"Bibliography on htribution Automation" IEEE Transactions on
Power Apparatus and Systems, vol. PAS 103. June 1984, pp. 1176-
1182.
T. Gonen, A. A. Mahmoud and H. W. Golbum, "Bibliography of
Power Distribution SystemPlanning", IEEE Transactions on Power
Apparatus andSystems, vol. PAS 102, June 1983, pp. 1178-1787.
D. I. Sun, D. R. Fanis et al., "Optimal Distribution Substation and
Primary Feeder Planning via the Fixed Charge Network Formula-
tion", IEEE Transactions on Power Apparatus and System, vol.
E. Lakervi and J. Partenen, "Linear Voltage Regulation and Loss
Calculation Methods for Low Voltage Network Design", IEEE Tran-
sactions on Power Apparatus ~ n d Systems, vol. PAS 103, Aug.
A. Merlin and H. Back, "Search for a Minimal-Loss Operating Span-
ning Tree Configuration in Urban Power Distribution Systems",
Proc. of5 th Power Systems Comp. Con., Cambridge, U.K., Sept.
D. W. Ross, M. Carson, A. Cohen, et al., "Development of Advanced
Methods for Planning Electric Energy Distribution Systems", DE0
final report no SCI-5263, Feb 1980.
C. H. Castro, J. B. Bunch, and T. M. Topka, "Generalized Algo-
rithms for Distribution Feeder Dep1oyn;ent and Sectionaliiing",
IEEE Transactions on Power Apparatus and System, vol. PAS 99,
C. A. Castro Jr. and A. L. M. Franca, "Automatic Power Distribution
Reconfiguration AlgorithmIncluding Operating Constraints", Proc.
of IFAC on Electric Energy Systems, Rio de J aneiro, Brazil, 1985.
PAS 101, March 1982. pp. 602-609.
1984, pp. 2249-2253,
1-5,1975.
MarCh/April1980, pp. 549-557.
[lo] S. Civanlar, J. J. Grain&r,-and S. H. Lee. "Distribution Feeder
Reconfiguration for Loss Reduction", IEEE PES Winter Meeting,
Feb. 1987, paper no: 87WM 140-7.
111J K. Aoki, H. Kuwabara, T. Satoh, and M. Kanezashi. " An Efficient
Algorithm for Load Balancing of Transformers and Feeders by
Switch Operation in Large Scale Distribution Systems", IEEE PES
Summer Meeting, 1987, paper no: 87SM 543-2
[12] K. Aoki, T. Satoh, M. Itoh, et al., "Voltage Drop Constrained Res-
toration of Supply by Switch Operation in Distribution Systems",
IEEE PES Summer Meeting, 1987, paper no: 87SM 544-0
[13] M. E. Baran, and F. F. Wu, "Optimal Sizing of Capacitors Placed on
a Radial Distribution System", presented at IEEE PES Winter Meet-
i ng, Feb. 1-6,1988, paper no 88WM 065-5
APPENDIX
by Simplified DistFlow Method
TOcalculate the power loss reduction due to a branch exchange, we
need to estimate new branch flows after the switching around the loop
which is defined by the branch exchange. Note that the branch flows
before. the switching are known.
First consider the nominal branch exchange between branch b (on@-
nally open) and k (originally closed) of Fig.4. By simplified DistFlow
equation of (6). branch flows around the loop, Pi , Qi will change by
Pk , Qk amount, i.e..
P:=Pi - P& Q ; = Q ~ - Q ~ i c L (a1.i)
P; =Pi +Pk Q, ' =Qi +Qk i e R (a.13)
For the general case - branch exchange between branches b and m, it
can easily be verified that the branch flows around the loop will change by
P, ' =P~- P, , , Q; = Q~ - Q, i c L (a.2.i)
P; =Pi +P, , , Q; =Qi +Q, , , i e R (a2.ii)
Now wecan calculate the real power loss reduction due to branch
exchange b-mas follows. By Eq.(8), the original total power losson L
and R sides of the loop will be
A. Estimation of Power Loss Reduction due to a Branch Exchange
P , , Q, , i.e.,
OL =BdP?+Qt 2) OR =Cr dP?+Q?l (a.3)
ISL I d
These terms can be updated after the branch exchange by using the
updated power flows. P( , Q;. Let the updated loss terms be L F ~ ,G;.
Then the real power loss reduction. ALP-b, due to branch exchange b-m
will be
ALP-^ =ALP~+AL& =(OL-O~)+(LJsI-O~) (a.4)
When the loss terms QL .OR are substituted in above equation and the
terms are rearranged, power loss reduction can be written only in terms of
original branch flows, Pi .Qi as in Eq.(9).
B. Accuracy Analysis of Simplified DistFlow Method
Wewill derive the relation between the actual loss reduction, ALP and
the estimated value, ALP- due to a branch exchange by simplified DistFlow
method in two steps; first power loss reduction along a radial network will
be studied, then these results will beused to evaluate the power loss reduc-
tion around the loop of a branch exchange.
Power LossReduction on a Radial Network
Consider the radial network shownin Fig.3. As noted before, such a
network represents one side, say side R , of the loop of a branch exchange.
Let the power change at the end of thenetwork due to branch exchange be
w,, , AQ,, >0.
Lets first comider the change in voltage profile as a result of change of
power at the end node n, Af, , AQ,,. A good esdmate can beobtained by
using the simplified DistFlow equations of (6). FromEq.(6), the change in
branch power will be
U. 1 - p : - p . - t =Af n AQi =Q/ - Qi =AQ. i =O, ..JI 0.1)
Assuming V i =V, , change in voltages can beobtained by Eq.(6.iii) as
0.2)
k=o kco
Therefore, the voltage profile along the network will drop with an increas-
ing magnitude towards the end of the network, i.e.,
0.3)
' AV. SAV,,-lS ... <AVo = O
Now consider the new power flow in the network and let the power at
the sending end of branch i be as follows.
P; =pi +U,, +wi r ; Q,: =ei +AQ, +A L Q ~ , ~ 0.4)
Then by using the backward equations of (2), it can beshown that
Pi-1 =Pi - l +APa+ALPi - l , ; Qi-1 =Qj - l +AQa+UQi - l r 0. 5)
1407
where, ALPi-,,, =ALPi,, 0. 6)
1
The last termin 0.6) is due to theloss reduction on branch i , ALPi . Con-
sequently, the terms - ALP;,, , - ALQi, represent the actual loss reduction
between nodes i and n and havethe following property.
(pi +P~)' +(e! +QLi12 (pi +pli12 +(ei +e,)*
Vi2 Vi2
+ri [
0 d ALP,,-l, S . . . S ALP.,. ; 0 5 ALQ,-l,, 5 . . . 5 ALQ,,
0.7)
Theseterms are of second order, i.e.,
l APml > lALf';,,I ; I AQe. l ~I ALQi , , I i = O J ,..., n-1 0. 8)
Now consider the loss reduction on the network. Power loss reduction
on line I , ALPl can becalculated as
ALP1 =LP,-LP; 0.9)
f'? +Q? (PI +U n +-I,)' +@I +AQn +MQ1.m1'
=r, --
v: r' V;2
The simplified DistFlow method estimates this quantity as
UI =PI -P;=rl (f'? +Q?) - rt [(PI +APn)z+ (Q, +AQ.)2] 0.10)
Combining Eq.0.9) and Eq.0.10) will give
1 1
ALP1 = - -LP1'+ &PI
Vl Vl2
0.12)
1
where, &!PI =-{ r1 [(Pl+AP,$+ (Ql+AQm)']
Vi
1
-- 71 [(Pt+dp,+ALPt,,)2+ (QI+AQ.+ALQ~,)~I) 50
V?
Eq(b.12) indicates that 6Lpl represents the extra loss reduction due to
correction terms ALPI, and ALQl,, and it is of second order, i.e.,
16Lpl I << IALPI I . Dropping these second order terms, Eq.0.11) can be
bounded as
Since the total loss on the section is the sumof branch losses, the
bound for thewholesection will be
where, f? and fz represent aggregated voltages, i.e..
0.14)
0.15)
Therefore,
V,2Sf.<V,2 ; V2Sf , ZI V?. 0. 16)
By letting
f&?=v,z ; f&v2 0.17)
a weak bound on the error of estimation can beobtained as
Power Loss Reduction Around the Loop of a Branch Exchange
Consider the nominal branch exchangeof Rg.4. Power loss reduction
on theR side of loop is given in theprevious section. Now consider the
other side. side L , of theb w h exchangeloop. The samederivation steps
of theprevious section can beapplied for this side also and thebounds on
loss reduction error can be obtained by mimicking the derivation of R
side. The result will be
Finally, adding Eq. 0. 18) and Eq.(b.l9), bounds on the total loss reduc-
tion error can bewritten as
C. Development of Method 2
three different schemes that can beused to approximatethe
new power flow in a systemafter a branch exchangeby making useof the
backward and forward updates of DistFlow introduced in Sec.11.
Scheme 1: Consider the nominal branch exchangebetween branches b
and k in Fig.4. As shown in Appendix B, thevoltageprofilearound the
loop will changein such a way that it will behigher on theL sideand
lower on the R side than their original values and it satisfies thefollowhg
(c.li)
(c. lii)
If weassumethat V; =Vk , Vi =V, and also that thechangein power
leaving theloop, such as P, , Q, shown in thefigure, are negligible, then
thepower flow in the loop can beupdated by using thebackward update
starting f" thenodes k and n of theloop. Since thevoltageand power
at thesepoints are known, the corresponding quantities can becalculated
at thepoints, k-1. n-1, by using thebackward equations. Repeating this
process up to node0, the common node, weget theupdated values
There
relationship.
AV? =v:- V; L AV; 2 . . . L AV: =o
AV; =V2 - V: 5 AV,,-, 5 . . . I AV,' =0
f,' , Q,' , $,', i ELM (c.2)
Scheme 2: The approximations for thenew power flows obtained by
scheme1 usetheexact branch equations except the voltages at thetermi-
nal nodes (V, instead of V;, V, instead of Vi). However, theerror in vol-
tages, AV:, AV:, Wi g the biggest as indicated by Eq.(c.l), is carried back
to theother nodes by backward voltageupdateof Eq.(2.iii). Therefore,
this is not good. A better approach will bedropping thevoltageupdate
and instead using the original voltages; thus only the power updates (
Eq.(2.i) and Eq.(2.ii) ) arecarried out.
Scheme 3: Since the accuracy of thebackward updatedepends criti-
cally on thevoltage estimate at theloop terminal nodes n and k , it is
desirableto check on thevalueof voltagedifference. AVk , AV. and mmct
the approximation, if necessary, to improvethe estimation.
To estimate AVk , AV., wewill use thefollowing property. Consider
the radial network of Ag.3 again and let us calculatethechange in total
voltage drop across the network. V:-,, =V2 -V> as a result of a power
changeat theend of the network, AP,, , AQ. >0, by using simplified Dist-
Flow equations in two different ways. Fist, as it is donein the Appendix
B. by assuming that Vi =V, and by using the simplified forward voltage
equation of Eq.(6.iii). Then the voltage drop can be calculated as
(AV:-.)f =-AV:. Second, by assuming Vi =V,, and by using thesimplified
backward equations V,ll =V: +2(r,P,'+ x, Q,'), and calculating the
comsponding voltage change as (AV,'-)b =AV,'. It can be shown that
thesetwo estimates are equal to each other, Le,
(AV,'-,,,,' = 6.3)
The results indicatethat a good estimate of voltagechangeat the ter-
minal nodes n and k can beobtained by canying out the voltageupdate
separately from thepower update, whileperforming the backward update
and comparing the voltagedifferenceat nodeo (differencebetween V, and
the updated Vb). If the differenceis too large ( larger than a predefined
value, PX), onemay go through a forward updateto reducetheemr (thi s
time starring fromthecommon nodeo and using V.,@L,&, as initial,
given values and applying theforward update).
Weuse scheme3 as the second method of updating the power flows
around theloop of a branch exchange.
Mesut E. Baran received his B.S. and M.S. formMiddleEast Technical
University, Turkey. Heis currently a Ph.D. student at the University of
Califomia, Berkeley. His research interests includedistribution and power
systems, optimization, systemtheory, and control theory.
Felix F. Wu received his B.S. fromNational Taiwan University. M.S.
fromtheUniversity of Pittsburgh, and Ph.D. fromtheUniversity of Cali-
fornia, Bekeley. Heis a Professor of Electrical Engineering and Com-
puter Sciences at the University of California, Berkeley.