996 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO.

3, JULY 2000

An Efficient Brute-Force Solution to the

Network Reconfiguration Problem
Anthony B. Morton and Iven M. Y. Mareels, Senior Member, IEEE

AbstractWe suggest a method for determining a minimal-loss load currents by a simple equation involving the sensitivity
radial configuration for a power distribution network, using an ex- matrix of the network, which can be written down directly by
haustive search algorithm. While exhaustive, the method is highly inspection of the network topology. Knowing the feeder cur-
efficient, deriving its efficiency from the use of graph-theoretic
techniques involving semi-sparse transformations of a current rents and the (fixed) cable resistances, the network losses are
sensitivity matrix. The algorithm can be applied to networks easily calculated without the use of computationally intensive
of moderate size and has advantages over existing algorithms load-flow algorithms.
for network reconfiguration in that it guarantees a globally The use of a constant-current load model can be justified most
optimal solution (under appropriate modeling assumptions), and simply on the basis that it is no less arbitrary than a con-
is easily extended to take account of phase imbalance and network
operation constraints. A 33-bus example is used to demonstrate stant-power model. In reality, the load on a power system ex-
the operation of the algorithm. hibits a combination of ohmic, constant-current and constant-
Index TermsComputer applications, graph theory, power dis-
power characteristics, as well as switch-mode circuits that ap-
tribution, power distribution control. proximate one or other of these characteristics. The use of con-
stant-power models in load-flow studies is justified on the basis
of conservatism: given that these types of load have more severe
I. INTRODUCTION effects on stability and voltage drop than the others, it makes
sense to assume a worst-case scenario. If, on the other hand, one
T HE RECONFIGURATION problem for radial power
distribution networks may be stated as follows: given a
load profile for a distribution network with a number of tie lines
is concerned purely with comparing the network losses across
different configurations, there is no worst case: the question
and switching points, find a radial configuration for the network is purely whether the inevitable modeling error is sufficiently
which minimizes the network losses. The problem can be recast large to throw ones conclusions into doubt. Particularly when
into a graph-theoretic framework, where the problem is to find the anticipated voltage drop is small, a constant-power assump-
a spanning tree in a graph with weighted nodes and branches, tion provides no greater guarantee of a favorable response to this
such that an objective function of the weights, obtained by ref- question than a constant-current assumption.
erence to Ohms and Kirchoffs Laws, is minimized. Given an initial radial configuration or spanning tree, all other
A variety of approaches to this problem are surveyed by Sarfi trees can be generated, without repetition, by means of a re-
et al. in [1]. This survey commences by stating that [t]he gen- cursive branch-exchange algorithm. We show that a branch ex-
eralized reconfiguration problem presents a considerable com- change or elementary tree transformation involves a semi-
putational burden for a distribution system of even moderate sparse perturbation of the sensitivity matrix for a radial net-
proportions. This assumed computational burden follows from work. These two facts form the basis of a highly efficient brute
the observation that [t]he nonlinear nature of the distribution force algorithm for the solution of the network reconfiguration
system necessitates that at each iteration of an optimization al- problem.
gorithm a load flow operation be performed to determine a new
system operating point. If this is correct, it follows that a direct II. ELEMENTARY TREE TRANSFORMATIONS
or exhaustive solution is infeasible, so that a practical solution Let be the graph of a distribution network, whose nodes
must employ some heuristic search method, possibly guided by represent substations or buses and whose
a simplified optimization procedure. This indeed is the approach arcs represent cable segments. The rank of is
taken by all the methods surveyed, as well as more recent ap- , and the nullity is , equal to the number
proaches such as that of Haque and Khan [2]. of independent loops in the graph. If the network is radial, with
We suggest that this presumption of infeasibility cannot be no tie lines, then .
set aside so easily, if one is prepared to model the network Given a (spanning) tree in , the simplest way to obtain a
loads using a constant-current characteristic rather than the con- different tree (if ) is by exchanging one cable segment for
stant-power characteristic assumed in load-flow studies. With another in . The following definition formalizes this notion.
this assumption, the feeder currents can be obtained from the Definition 1: An elementary tree transformation in is a
transformation from a tree to a tree accom-
Manuscript received April 30, 1999. This work was financially assisted by the plished by removing one arc from and adding an arc
Australian Commonwealth Energy Research and Development Corporation.
The authors are with the Department of Electrical and Electronic Engineering,
. We abbreviate this operation by writing it in the form
University of Melbourne.
Publisher Item Identifier S 0885-8977(00)07213-7. (1)
08858977/00$10.00 2000 IEEE

Authorized licensed use limited to: Robert T Ah King. Downloaded on August 26, 2009 at 00:56 from IEEE Xplore. Restrictions apply.
MORTON AND MAREELS: AN EFFICIENT BRUTE-FORCE SOLUTION TO THE NETWORK RECONFIGURATION PROBLEM 997

Define the distance between two trees as do not. The more efficient algorithms are very difficult to de-
the number of arcs in but not in . We also denote by scribe, but it is possible in principle to substitute any of these
the fundamental cut set defined by with respect to ; that is, for the MayedaSeshu algorithm in our application, provided it
the set of all arcs in which connect the two components into operates by means of elementary transformations.
which is divided by the removal of .
We then have the following results, the proofs of which can III. APPLICATION TO LOSS MINIMIZATION
be found in [3]:
Let be a tree. Then is a tree if and only if Suppose we have an AC or DC distribution network which
and . we wish to operate in a radial manner (that is, as a tree net-
Trees and in are related by an elementary tree work) but which nevertheless contains a number of tie lines
transformation in if and only if . that allow the network topology to be reconfigured in various
Let be any trees in . If , then ways so as to minimize the cable losses. We may picture the
can be obtained from through a sequence of exactly network, including tie lines, as a graph having a nullity equal
elementary tree transformations. to the number of tie lines. The problem is then to find, from
Because the maximum possible distance between any two amongst all spanning trees of , a tree such that the tree net-
trees in is , it follows from the last of these statements that, work has minimal cable losses.
given any tree , we can generate any tree by In our analysis, we designate node in as a reference or
applying at most elementary tree transformations. This idea datum node. The node-arc incidence matrix of has rows and
forms the basis of an efficient algorithm for generating all the columns and is of rank ; removing the last row (corre-
trees in a graph from an initial tree . The basic procedure is sponding to the datum node) leaves a matrix which we call
as follows: the reduced incidence matrix. Selecting any columns of
corresponding to the arcs of a tree gives a nonsingular
1) Find an initial tree , having arcs. submatrix, and all such nonsingular submatrices are obtained in
2) For each arc , find the cut set and substi- this manner.
tute the arcs of one by one (if any) for to obtain In a radial distribution network with a single point of supply,
a new set of trees. When this is done for every arc of , it is convenient to identify this supply point with the datum node.
one obtains all trees of distance 1 from . We denote by the vector of currents (assumed fixed) leaving
3) Repeat the above process (subject to conditions detailed the network at each node through , excluding the datum
below) for each tree of distance 1 from , hence ob- node. The vector of cable currents is denoted . Given these def-
taining the trees of distance 2 from . initions, we have [5]
4) Repeat until one has all trees of distance up to or
(whichever is the lesser) from . One then has all the
trees in the graph.
To this basic procedure, of course, must be added some checks for some matrix , called the sensitivity
to ensure no backtracking occurs. The following conditions are matrix of the network . Furthermore, if is a tree then
sufficient [3] to ensure that each tree in the graph is generated
once and once only by the above procedure:
1) The arcs forming the tree are ordered such that every
initial subsequence forms a connected subgraph. where is defined as follows: let if is traversed in
2) Arcs are substituted in strict ascending order as they ap- its own direction on the (unique) directed path from to
pear in the arc sequence comprising . (In other words, if is traversed in the opposite direction, and 0 if is not on
if and is substituted for to form a the path.
new tree , then neither nor may be substituted to To determine the network losses, we first calculate and
generate a new tree from . then the losses as the sum . Our brute-force approach in-
3) When generating new trees from a tree by substituting volves generating all the spanning trees of and comparing the
an arc , a new are is substituted if and only if it belongs above sum for each. Though simple, this method can be practi-
both to the set and to the set . [Note that cable as long as the process of recalulating , and the losses
by Condition 2) above.) at each step is not too complex. That this is so is established in
Let denote the total number of spanning trees in often the following section.
called the complexity or tree-number. The execution time for
the above algorithm, due to Mayeda and Seshu, is
IV. TRANSFORMATION OF SENSITIVITY MATRIX
, and the storage space required is . Smith [4]
outlines a number of alternative algorithms for generation of Let be a tree with sensitivity matrix , and let
spanning trees, the most recent of which have execution time be an elementary tree transformation involving arcs
. (This is referred to as constant amortized time, . We denote by the sensitivity matrix of , and
the amortized time being the execution time divided by the seek to understand the relationship between and . This in-
number of trees generated.) Some of these algorithms generate formation allows us to relate the cable currents and hence the
trees using successive elementary tree transformations, others losses in the two subtrees and .

Authorized licensed use limited to: Robert T Ah King. Downloaded on August 26, 2009 at 00:56 from IEEE Xplore. Restrictions apply.
998 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 3, JULY 2000

for all , and (if constant-current models are assumed) that

Now suppose and are indices corresponding (with some

abuse of notation) to an arc and a node . Let
be the path from to in , and let be the corresponding
path in . Then and (viewed as arc-sets) form
a partition of . It follows that

if (5)
Fig. 1. An elementary tree transformation.
If then there are two cases to consider, according as
or . In the former case must include
Consider , a subgraph of of nullity 1. Then
, giving . In the latter case on the other hand
is a cut set of partitioning it into two components
passes through , but in the opposite sense to ; thus
, such that . We take to be the
. It turns out in fact that we may write
component containing the datum node . (See Fig. 1.)
Let and be the nodes incident with , where we take if (6)
and . Similarly, let be the nodes incident
with . To take account of the orientation of we define to without regard for whether or . For we already
be 1 if is directed from to , and 1 otherwise. know that lies entirely within , hence that if
Now, has a row corresponding to every arc of and a . It is also readily seen that for all
column corresponding to each node of other than the datum , hence that if , and , then for all
node . To simplify matters here we augment by a zero , and in particular .
column corresponding to the datum node, so that for Equations (2)(6) completely describe the effect of an ele-
all by our convention. One row of will correspond to the arc mentary tree transformation on the sensitivity matrix for a tree
being substituted; let be the index of this row. network. For convenience we collect these into a theorem.
Our first observation regarding the transition from to is Theorem 1: Let be the graph of a distribution network, let
that the th row of is unchanged if the corresponding arc is be a tree and let be an elementary tree
not in the unique loop of . For, if is the path from transformation involving arcs . Let and be the
to some node , then replacing by does not affect that sensitivity matrices of the tree networks and respectively,
part of that does not lie in the loop . It follows that, if the in which the th row corresponds to the arc , respectively ,
external circuits have constant-current characteristics, the cable and other rows correspond to like arcs. Let , be
current is also unchanged. An arc lies in if and only if the nodes incident with , and let if is directed from
it belongs to exactly one of the paths and in ; if it to and 1 otherwise. Then
belongs to both then it has the same orientation with respect to 1) if or .
both. Accordingly, 2) .
3) If and , with , then
if (2) if , otherwise .

We observe also that the th column of is unchanged if the V. AN ALGORITHM FOR TOPOLOGICAL LOSS OPTIMIZATION
node lies in . For in this case, the path from to in
must lie entirely within , so that no part of this path is affected Theorem 1 forms the basis for an exhaustive yet efficient al-
by the transformation. Equivalently, the path in does not gorithm to solve the network reconfiguration problem. The al-
include , giving gorithm takes as input the reduced incidence matrix of the
distribution network with tie lines included, the resistances
if (3) of each cable segment, and the external circuit current for
the first nodes. (Since the sensitivity matrix and hence
We may therefore restrict our attention to rows of corre- the cable currents in a tree network do not depend on the cable
sponding to arcs in , and to elements of these rows corre- impedances, it is not necessary to specify the reactances of the
sponding to nodes in . Consider first the row corresponding cables as input to the algorithm.) It begins by determining an
to itself. In this will be replaced by a row corresponding initial spanning tree for the given network, and the sensi-
to ; it is apparent that in feeds precisely those nodes tivity matrix for the corresponding tree network (obtained
(in ) that are fed by in . Thus the elements of row are by inverting the appropriate submatrix of ). The initial cable
unchanged except in sign. A sign change occurs if the cut set currents and losses are also calculated at this point.
is not uniform, alternatively if (defined above) and Proceeding from these initial conditions, the algorithm gen-
differ in sign. It follows that erates all the spanning trees for the network, using the algorithm
of Section II, and at each step modifies the sensitivity matrix ac-
(4) cording to Theorem 1. The sensitivity matrix is modified row by

Authorized licensed use limited to: Robert T Ah King. Downloaded on August 26, 2009 at 00:56 from IEEE Xplore. Restrictions apply.
MORTON AND MAREELS: AN EFFICIENT BRUTE-FORCE SOLUTION TO THE NETWORK RECONFIGURATION PROBLEM 999

row; whenever a row of the matrix changes, the corresponding TABLE I

cable current is recalculated and the losses perturbed by the dif- THE 33-BUS TEST SYSTEM
ference between the squares of the old and new currents, multi-
plied by the appropriate cable resistance:

In the case of the row corresponding to the arc being substituted,

the magnitude of the current is unchanged, and the loss need
only be altered to take account of the physical cable substitution:

After each step the loss is compared with the minimum loss
obtained so far, and if not higher, the current configuration is
recorded. In those cases where the minimal loss configuration
is not unique, the algorithm is capable of reporting all configu-
rations having the same minimal loss.

VI. AN EXAMPLE
As an example of the algorithm outlined above, we choose
the 33-bus system of Baran and Wu [6]. In our terminology,
this system comprises a cable network having nodes
(or buses), cable segments, and nullity . This
network is usually taken to model an AC distribution network
having one energy input at bus 0 and specified current demand
at buses 1 through 32.
To fit in our framework, we relabel the supply bus 0 as bus 33,
and take it as the datum node. The cable network topology and
resistances are given in Table I below. The 32 load circuits
are assumed to have constant currents whose real and imagi-
nary parts are provided in Table I, next to the data for the first 32
cable segments. (These are derived respectively from the quoted
values of real and reactive power in [6, Table 1]. The power
quantities were used directly, as the unit of current is of no im-
portance to the algorithm.) Cable segments 33 through 37 are
the tie lines. We ran the algorithm both on the usual AC net-
work and on a DC network having the same topology and the
same RMS current magnitude at each node.
A MATLAB implementation of the minimal-topology algo-
rithm took around two hours on a multi-user UNIX work-station and Wu [6] and Haque and Khan [2], though an exact compar-
to process the 50 751 spanning trees for the network of Table I. ison is not possible owing to the different assumptions on the ex-
As the subnetwork comprising branches 1 through 32 is con- ternal circuit models. The aforementioned studies assume con-
nected and therefore a tree network, it was automatically se- stant-power characteristics, in keeping with the load-flow ap-
lected as the starting tree . For the AC network with external proach, whereas we assume constant current; one might there-
currents specified as in Table I, it was found that the following fore expect slightly different results to be obtained due to the
elementary tree transformations gave a network with minimal effect of voltage drop.
losses of 20.413 MW, as compared with 28.266 MW for the ini-
tial network . VII. CONCLUSION

Replace branch With branch We have presented the outlines of an efficient brute-force al-
gorithm for the solution of the network reconfiguration problem,
and illustrated its use with a typical distribution network of mod-
erate complexity.
At present the algorithm will solve the network reconfigu-
ration problem for a DC, single-phase AC, or balanced three-
This corresponds to closing four out of the five tie lines, and phase AC network, though it takes no account of operational
opening four branches of the original network to maintain the ra- limits, such as maximum current or power-transfer limits on in-
dial structure. The result may be compared with those of Baran dividual cable segments, or node voltage constraints. It is not in

Authorized licensed use limited to: Robert T Ah King. Downloaded on August 26, 2009 at 00:56 from IEEE Xplore. Restrictions apply.
1000 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 3, JULY 2000

principle difficult to incorporate checks for such limits into the REFERENCES
algorithm, and to reject any particular configuration that fails [1] R. J. Sarfi, M. M. A. Salama, and A. Y. Chikhani, A survey of the
to satisfy these checks. The algorithm could also be modified state of the art in distribution system reconfiguration for system loss
to handle unbalanced three-phase networks simply by treating reduction, Electric Power Systems Research, vol. 31, pp. 6170, 1994.
[2] M. H. Haque and M. E. Khan, Loss minimization in distribution
each current quantity as a vector containing either the line cur- systems, Australasian Universities Power Engineering Conference
rents or the magnitudes of their symmetrical components. A (AUPEC96), vol. 1, pp. 8388, 1996.
symmetrical-component representation would be necessary if [3] W. Mayeda, Graph Theory: Wiley-Interscience, 1972.
[4] M. J. Smith, Generating spanning trees, M.Eng.Sc. thesis, Department
the cable resistances depended significantly on phase sequence, of Computer Science, University of Victoria, 1997.
but this dependence is generally much less for resistance than [5] A. B. Morton and I. M. Y. Mareels, Overload prevention and loss min-
for reactance quantities. imization in managed distribution networks, Australasian Universities
Power Engineering Conference (AUPEC98), vol. 1, pp. 107112, Oc-
In assuming constant-current characteristics for the external tober 1998.
circuits, our approach follows that of Liu et al. [7], who recog- [6] M. E. Baran and F. F. Wu, Network reconfiguration in distribution
nized that in this case the problem could be formulated purely in systems for loss reduction and load balancing, IEEE Transactions on
Power Delivery, vol. 4, no. 2, pp. 14011407, 1989.
terms of currents. The graphical approach we use is more akin [7] C. C. Liu, S. J. Lee, and K. Vu, Loss minimization of distribution
to the original study of Merlin and Back [8], and particularly to feeders: Optimality and algorithms, IEEE Transactions on Power De-
the work of Civanlar et al. [9] whose branch exchanges corre- livery, vol. 4, pp. 12811289, 1989.
[8] A. Merlin and H. Back, Search for a minimal-loss operating spanning
spond to our elementary tree transformations. The recognition tree configuration in an urban power distribution system, in Proceed-
that the problem is essentially combinatorial in nature, and that ings of the 5th Power System Computation Conference (PSCC), Cam-
a generic switching operation may be represented as a series of bridge, UK, 1975, pp. 118.
[9] S. Civanlar, J. J. Grainger, H. Yin, and S. S. H. Lee, Distribution recon-
branch exchanges, is due to Baran and Wu [6]. figuration for loss reduction, IEEE Transactions on Power Delivery,
The computational burden of the algorithm is significant, but vol. 3, pp. 12171223, 1988.
not prohibitively so for practical networks where the number of [10] N. Biggs, Algebraic Graph Theory, second ed: Cambridge, 1993.
tie lines (the nullity) is small by comparison with the number of
buses. To obtain asymptotic bounds for the computational com-
plexity of the algorithm it would be necessary to have a formula
Anthony B. Morton is a Ph.D. student with the Department of Electrical and
for the number of spanning trees in a given graph. Unfortunately, Electronic Engineering at the University of Melbourne. His interests include
the only known general expression for this quantity is as the de- DC power reticulation, nonlinear control, dynamical systems and renewable en-
terminant of the matrix [10], and there is no known ex- ergy. Anthonys research into the modeling and control of the Managed DC Bus
earned him a Master of Engineering Science in 1996.
pression in terms of topological properties such as the number
of arcs in each fundamental loop, though expressions can be de-
rived in the simplest cases.
Iven M. Y. Mareels is Professor of electrical and electronic engineering and
ACKNOWLEDGMENT Head of the Department of Electrical and Electronic Engineering at the Univer-
sity of Melbourne. He has held senior academic positions at a number of univer-
The authors wish to acknowledge the support of the Depart- sities including the Australian National University, the University of Newcastle
ment of Electrical and Electronic Engineering at the Univer- (Australia) and the University of Twente in The Netherlands. Ivens current re-
sity of Melbourne, and of the Managed DC Bus research team search interests are in the area of modeling, simulation and control of nonlinear
dynamical systems, with applications to power systems, robotics and aircraft
at Otis Elevators Australian Engineering Centre, Dingley, Vic- dynamics. He has ongoing interests in the mathematical modeling of biological
toria, Australia. processes and learning systems.

Authorized licensed use limited to: Robert T Ah King. Downloaded on August 26, 2009 at 00:56 from IEEE Xplore. Restrictions apply.