Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO.

2, APRIL 2005 1031

A Decomposition Approach to Optimal Remote


Controlled Switch Allocation in Distribution Systems
P. M. S. Carvalho, L. A. F. M. Ferreira, and A. J. Cerejo da Silva

Abstract—The growing demand for improved quality of service decomposition and convex analysis. We propose a simple and
increases the importance of network automation, namely the in- effective algorithm to divide the solution space into smaller in-
vestment in remote-controlled switch (RCS) devices. These allow dependent subspaces and this way take advantage from the weak
improving the fault isolation and reconfiguration time and there-
fore increasing the system quality of service. The investment in interconnectivity of most distribution network feeders. We then
switch devices comes at a cost and thus must be optimized. The project the subspace objective function into a convex direction
problem of determining the optimal number of devices and their (the number of RCS direction) and use convexity to reduce the
optimal location is a difficult problem: the solution space is combi- search effort necessary to find the optimum in each subspace.
natorial and the objective function is nonanalytical. We propose to The paper is organized as follows. In Section II the optimiza-
address this problem in a two-stage decomposition approach. Re-
sults are presented to i) divide the solution space into independent tion problem is formulated, the complexity of the problem is
subspaces, and then ii) solve the optimization problems in each sub- evaluated, and two problem properties are derived. In Section III
space. The solution approach is illustrated for a real distribution the solution approach is presented. We use the derived properties
network problem. to propose effective algorithms to decompose the problem into
Index Terms—Circuit topology, optimization methods, power subproblems and then solve each subproblem independently. In
distribution economics, power distribution planning, power dis- Section IV the proposed approach is illustrated with an example
tribution reliability. from a real urban distribution network. In Section V we con-
clude the paper.
I. INTRODUCTION
II. OPTIMAL RCS ALLOCATION PROBLEM

P OWER interruptions impose costs on society by inter-


rupting the normal flow of commerce and disrupting
the provision of important services. One way of reducing the
Distribution networks are constituted by several feeders from
different substations. Each feeder defines a radial configura-
interruption costs is to reduce the outage time by investing in tion from the substation transformers to the loads. Frequently,
facilities such as remote controlled switches (RCS) devices. feeder-ends are connected to each other within the same feeder
RCS installation allows improving i) the fault isolation time and to other feeders through normally open switches. Hence,
when installed in normally closed branches and ii) the fault the distribution network topology consists in a graph over which
(island) reconfiguration time when installed in normally open some open switches define a radial configuration.
branches. By improving operation times, RCS devices improve
outage time and thus increase the overall system quality of A. Notation
service. As RCS installation comes at a considerable cost [1], The following notation is used in the formulation of the op-
the selection of the RCS installation sites must be selected in timal RCS allocation problem:
order to optimize its benefit/cost tradeoff. This optimization reliability benefit function;
problem is the subject matter of the paper. G branches of the distribution network graph;
Several authors addressed the problem of sectionalizing S branches for possible RCS location;
switch allocation in distribution networks [2]. They proposed unitary cost of the RCS devices;
different heuristics based on different techniques to address the selection of branches from S.
problem. Techniques proposed include genetic algorithms [3], Other notation is used throughout the paper and defined when
[4], simulated annealing [1], and other heuristics based on the necessary.
Bellman’s principle of optimality [5].
In this paper we propose to address the problem of RCS al- B. Formulation
location with classical optimization. The techniques used are The unitary cost of RCS devices may include installation and
maintenance costs. In the following we will refer to such costs
Manuscript received July 31, 2003; revised December 18, 2003. Paper no. as the RCS investment cost.
TPWRD-00409-2003. The optimal RCS allocation problem consists in finding the
P. M. S. Carvalho and L. A. F. M. Ferreira are with the Instituto Superior
Técnico, Technical University of Lisbon, Lisbon, 1049–001 Portugal (e-mail: best possible tradeoff between RCS investment costs and reli-
pcarvalho@ist.utl.pt; lmf@ist.utl.pt). ability benefits. Let the reliability benefit function be given by
A. J. Cerejo da Silva is with Escola Superior de Tecnologia, Instituto , where stands for a reference reliability
Politécnico de Castelo Branco, Castelo Branco, 6000-767, Portugal (e-mail:
cerejo@est.ipcb.pt). cost and refers to the reliability cost function. The relia-
Digital Object Identifier 10.1109/TPWRD.2004.838470 bility cost function must represent the system outage cost. One
0885-8977/$20.00 © 2005 IEEE
1032 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005

Techniques have been proposed to reduce the problem com-


plexity by using the Bellman’s principle of optimality [5]. How-
ever, the problem, as we formulate it, is not suitable for the appli-
cation of the Bellman’s principle of optimality. The Bellman’s
principle of optimality states that the optimal solution with
Fig. 1. Illustration of the problem solution representation. The solution is
represented by a binary array x indexed to an array of branches from S. The set RCS, say , must be achieved by adding one RCS device to a
of branches selected by x in S is x(S) = fa; b; d; f; g; i; k g. solution, say , in a way that solves the optimal problem
of getting from to . The principle does not state that the
solution is the optimal solution with RCS—in
usual approach to evaluate the system outage cost is to compute
general it is not [7]. Because solutions of (1) are sets of branches
the cost of the energy not supplied (ENS). Other costs imposed
where order does not matter, and the optimal solution cannot
by the regulator as penalties for quality of service below defined
be directly related to optimal solutions with different number of
standards can also be added [6].
RCS, the Bellman’s principle of optimality is not used to reduce
The problem solution may be represented by binary branch
problem complexity.
decisions: either a RCS is to be placed in a branch or not. We
We reduce problem complexity by (i) using a polynomial-
take a binary array to select branches from S (see Fig. 1) and
time partitioning algorithm to divide the set S into several in-
refer to as the set of braches selected by in S.
dependent subsets, say and (ii) decomposing the
Considering the binary representation of the solution , the
problem into a set of convex independent subproblems to be
problem of finding the best possible tradeoff between RCS in-
solved independently. The approach consists in the decompo-
vestment costs and reliability costs may be formulated as in the
sition of (1) into the set of subproblems of (2), and further pro-
following:
jection of subproblems into the convex maximization problems
of (3)
overall (1)
overall (2.1)
To compute the reliability cost the expected value of the ENS
overall (2.2)
is computed. The expected value of the ENS is computed numer-
ically, not analytically. Branch faults are enumerated exhaus- overall (2.3)
tively (for each and every branch of the network graph G) and
the correspondent post-fault configurations analyzed for
overloads and undervoltages. For each fault, the processes of where
isolation (with time ), reconfiguration (with time ) or repair
(2.4)
(with time ) are analyzed in G and the corresponding ENS
computed for given branch failure rates. The reliability cost is (2.5)
then computed from the value of the ENS for a given ENS price. The projection of each maximization problem in into a max-
The existence of a RCS device: (i) changes the time to isolate imization problem in the number of RCS, say , results in the
faults if the RCS is placed in a normally closed branch; following problem formulation:
and (ii) changes the time to reconfigure faults in other branches
if the RCS is placed in a normally opened branch. When (3.1)
more than one RCS can be closed to restore power in an iso-
lated network section (island), the ENS is evaluated for the best where
switching operation. The best switching operation is chosen
(3.2)
with the following rules:
R1) The best operation is the one that leads to fewer circuit The theoretical framework for problem decomposition into a
overloads if all operations do lead to overloads; set of independent convex subproblems relies on two problem
R2) The best operation is the one that leads to the better properties given in the following section.
voltage profile if two (or more) operations can be done
without circuit overloads; D. Properties
R3) The best operation is the faster operation if two (or
Here we state two problem properties. The first property re-
more) operations can be done without circuit overloads
lates the benefits of different sets of RCS. The second property
and without node undervoltages. The faster operation
relates two different concepts of subproblem dependence. The
is usually a RCS operation.
two properties constitute the theoretical framework for problem
decomposition and subproblem solution.
C. Complexity Property 1: Given any two sets of branches and :
The problem formulated in (1) is a difficult problem both be- .
cause i) the reliability function is difficult to evaluate, and ii) the Proof: For each possible faulted section, one of two pos-
space dimensionality is very high. If we designate by the size sible situations occurs: (i) either the island may be restored by
of S, the solution space is where stands connecting a RCS from one set only, and thus
for the number of -combinations of elements. ; or (ii) the island may be restored by connecting
CARVALHO et al.: A DECOMPOSITION APPROACH TO OPTIMAL REMOTE CONTROLLED SWITCH ALLOCATION IN DISTRIBUTION SYSTEMS 1033

occurs when the faults of cannot be restored by (at


least) one of the sets (say ) because all switching operations
lead to circuit overloads or undervoltages. Hence, subset inde-
pendence should be investigated analytically, not topologically.
Fig. 2 illustrates analytical independence among topology de-
pendent sets.
In the following section, we use the stated problem properties
Fig. 2. Illustration of analytical independence among topology dependent sets. to propose (i) a space partitioning polynomial-time algorithm to
The figure shows a network with one substation (circle), 13 loads (triangles), divide the set S into analytically independent subsets, and (ii) an
and two RCS (in branches x and y ). x and y are topologically dependent as
fa; bg =L L
\ . If the faults in fa; bg can be restored without overloads optimization algorithm to solve the subproblems related to each
independent subset.
or undervoltages only by connecting x (and never y ) then a RCS in y cannot
change ENS due to faults in fa; bg (see rule R3). So, x and y are analytically
independent despite being topologically dependent.

RCS from both sets and , and in such case the benefit III. SOLUTION APPROACH
of being able to restore the fault is reported both in and
and thus .
The proposed solution approach consists of three steps:
Note that Property 1 generalization to more than two sets
is straightforward Step 1) Evaluate the benefit of placing each single RCS
. alone. If the benefit is negative, remove the RCS
The concepts of topological and analytical dependence are from S.
defined in the following. Step 2) Run the space partitioning algorithm to divide S into
Definition 1: The set of loops is defined to be the subset several independent subsets.
of G defined by the branches in the paths between every two Step 3) Run the optimization algorithm over each indepen-
nodes adjacent to branches of . dent subset.
Definition 2: Two sets of branches and are said to be In the following, we give details of Step 2 and Step 3.
topologically independent iff .
Definition 3: Two sets of branches are said to be analytically
independent iff . A. Space Partitioning Algorithm
Take Property 2 to relate the two concepts of dependence.
Property 2: (First form). If two sets are analytically depen- The space-partitioning algorithm consists in testing and clus-
dent then the sets are also topologically dependent tering RCS branch possibilities. We propose to test analytical
dependency with Definition 3 and cluster branches when de-
pendent. By doing so, we may divide the set of all RCS branch
(4.1) possibilities into several clusters. Such clusters are the indepen-
dent subsets, as (i) within each cluster RCS branches are de-
(Second form). If two sets are topologically independent, then pendent; and (ii) between different clusters RCS branches are
the sets are also analytically independent independent. The test and cluster algorithm can be designed to
be polynomial-time. We propose the following design:
Algorithm 1
(4.2)
Take sets of single RCS from S.
Their corresponding solutions may be represented by
Proof: Analytical dependence results from the existence
of more than one RCS able to restore the same post-fault isolated
section (the island). A set of loops define a network subgraph
of the faults possible to be restored with branches of . If
two sets are analytically dependent, then there must be at least
one branch-fault possible to be restored either with a branch
of or with a branch of . To be so, such branch must
belong to both and . Hence, and cannot be mutually The procedure starts with sets of single RCS and follows
exclusive. The equivalence between the first and second forms a given order to visit them. Then, the benefit of each visited set
of Property 2 derives directly from De Morgan’s basic law: is added to the benefit of a downstream set and compared with
. the benefit of the sets’ union. If the benefit of the sets’ union is
The second form of Property 2 allows using topology analysis lower than the sum of the individual benefits, then the second
to divide the set of possible RCS locations S into analytically set is clustered to the visited one . When the
independent subsets. However, our experience shows that an- visited set is not the top one, the procedure must also try to
alytical independence is found frequently among topology de- cluster the upstream sets. The proposed procedure is presented
pendent sets—(4.1) is an implication, not an equivalence. That in the following flowchart:
1034 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005

Fig. 3. Geographic view of the real 15 kV urban distribution network used for
illustration. Possible RCS locations are shown with squares an identified with
characters a, b, c, . . . k. The four HV/MV substations are shown as circles.

The final values of , represent each of the in-


dependent subsets. For instance if and
then the two independent subsets would be
and .

B. Optimization Algorithm

The optimization algorithm consists is finding in which


branches one should install RCS in order to maximize the }
objective function of (2). The projected problem of (3) consists
in solving the problem in . To compute for each , we
evaluate all the possible combinations of RCS and make IV. NUMERICAL RESULTS
. We then make use of
Property 1 to infer on the convexity of in and derive the The methodology presented in Section III is illustrated with
algorithm stopping criteria. To infer on the convexity of in an application example build over a real 15 kV urban distribu-
take the following steps: tion network. Fig. 3 depicts the geographical view of an under-
ground distribution network. The network cables are Aluminum
— Define ; cored with size between 50 (98,680 Mils) and 240
— Note that is a monotone increasing function of ; (473 600 Mils).
— Note that is a sublinear function of (Property 1); From all possible RCS locations, eleven normally open
— As is linear in , is convex in . points were considered. These have been selected to be the
The convexity of assures that there is a single relative max- most promising sites as their feeders had a history of high
imum in , and that can be used to reduce the optimization ef- failure rates and difficulty of site access. When operations
fort. Convexity allows to find the optimum by defining any practice and historical data are not available the sites to be con-
increasing direction (e.g., increasing ) and stopping just before sidered can be selected based on individual RCS benefits. The
decreases. The following algorithm can be used to find sub- individual benefits of installing each RCS have been calculated
problem solutions of (2). and are shown in Table I. These are shown in percentage of the
Algorithm 2 total benefit of installing all eleven RCS devices.
Given a subset of dependent RCS, the following loop To illustrate the projection of the maximization problem in
can be used to find the optimum solution . into a maximization problem in [see (3)], we also show in
CARVALHO et al.: A DECOMPOSITION APPROACH TO OPTIMAL REMOTE CONTROLLED SWITCH ALLOCATION IN DISTRIBUTION SYSTEMS 1035

TABLE I
INDIVIDUAL RCS BENEFITS

TABLE II
COMBINED RCS MAXIMUM BENEFITS

Fig. 4. The figure shows three functions of the number of RCS devices, k . The
functions are (i) the maximum benefit B (shown with 3); (ii) the investment
+
cost c 1 kxk (shown with ); and (iii) the function F as the difference of the
previous two functions (shown with o). Values are shown in percentage of the
maximum benefit. The optimal solution is found for k =5 and the following
branch selection fb; c; d; g; ig.

Table II the combined maximum benefits for each value to-


gether with the best combination of RCS devices. Fig. 4 com-
pletes the illustration by showing the combined benefits together
with the corresponding investment and the tradeoff costs evolu-
tion with the number of devices, .
To illustrate the advantages of the algorithms proposed con-
sider the following. To be able to find (exhaustively) the best
combination of RCS branches one would have to evaluate
solutions by running reliability analysis 2047 times
to access the ENS cost differences. Supposing that we do so,
exhaustive search would yield an optimal solution with
. This solution involves the selection of the set of branches
and corresponds to the following array:
(5)
Fig. 5. Subproblem objective function F for each of the four independent
Now observe that the partitioning of S with Algorithm 1 leads S S S S
subspaces , , and , respectively. The optimal solution is found for
to four independent subsets, namely: , , k =1 S in and the selection of branch b; k =2 S in and the selection of
, . The partitioning process in- =1 S
branches fc; dg; k in and the selection of branch g ; k=1 S in and
the selection of branch i. Values are shown in percentage of the total maximum
volves solely evaluations. The resulting benefit (as in Fig. 4).
subproblem objective functions are shown in Fig. 5.
Finding the best combination of RCS by exhaustive search in
, , and now involves By comparing (5) with (6) note that the equality
evaluations instead of the previous 2047. The holds. The set of branches
optimal solutions for each subset , , and correspond is shown in entry number five of Table II and its corresponding
to the following arrays: cost/benefit tradeoff presented in Fig. 4 for .
The partitioning of S can be used to reduce problem com-
(6) plexity in a great deal. Algorithm 1 alone reduces the necessary
1036 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005

number of evaluations by 97% (from 2047 to [2] “Bibliography of distribution automation,” IEEE Trans. Power App.
evaluations). Syst., vol. 103, pp. 1176–1182, Jun. 1984.
[3] G. Levitin, S. Mazal-Tov, and D. Elmakis, “Optimal sectionalizer alloca-
Using Algorithm 2 after the partitioning of S (with Algorithm tion in electric distribution systems by genetic algorithm,” Elect. Power
1) will again reduce the already very small number of necessary Sys. Res., pp. 97–102, 1994.
evaluations.1 Fig. 5 reports the subproblem objective functions [4] J.-H. Teng and C.-N. Lu, “Feeder-switch relocation for customer inter-
ruption cost minimization,” IEEE Trans. Power Del., vol. 17, no. 1, pp.
for each of the four independent subspaces , , and 254–259, Jan. 2002.
. Remark that: (i) to find the optimum in we need just [5] G. Celli and F. Pilo, “Optimal sectionalizing switches allocation in
ten evaluations instead of fifteen, , evaluations; and (ii) distribution networks,” IEEE Trans. Power Del., vol. 14, no. 3, pp.
1167–1172, Jul. 1999.
to find the optimum in we need six evaluations instead of [6] L. A. Jorge, E. Quaresma, F. Mira, L. A. F. M. Ferreira, P. M. S. Carvalho,
seven, . Together, the two algorithms allow an important S. N. C. Grave, F. F. C. Silva, and L. M. F. Barruncho, “Optimal distri-
reduction on the necessary number of evaluations—a reduction bution planning: meeting new regulation criteria for reliability,” in Int.
Conf. Exhibition on Electricity Distribution, CIRED’2003, Barcelona,
from 2047 evaluations to fifty-four (32 22) evaluations. Spain.
[7] D. Bertsekas, Dynamic Programming and Optimal Control, 2nd
ed. Belmont, MA: Athena Scientific, 2000, vol. I.
V. CONCLUSION
RCS devices allow improving the fault isolation and recon-
figuration time and this way increasing the system quality of
service. Because the investments in such devices can be costly, P. M. S. Carvalho received the electrical and computer engineering degree, the
M.Sc. degree, and the Ph.D. degree from the Instituto Superior Técnico (IST),
their allocation needs to be optimized. In the paper we propose Lisbon, Portugal, in 1992, 1994, and 1999, respectively.
an approach to determine the optimal number of RCS devices Since 1992, he has been with IST, Department of Electrical Engineering and
and their optimal location in order to maximize the tradeoff be- Computers, where currently he is an Assistant Professor.
tween ENS costs and investment costs.
The paper presents two problem properties that may be used
to design (i) a polynomial time algorithm for space partitioning L. A. F. M. Ferreira received the electrical engineering degree from the In-
into independent subspaces; and (ii) a convex optimization al- stituto Superior Técnico (IST), Lisbon, Portugal in 1977 and the M.S.E.E. and
gorithm to solve the optimization problem in each subspace. We Ph.D. degrees from the Georgia Institute of Technology, Atlanta, in 1983 and
1986, respectively.
proposed an algorithm for each purpose and illustrate the overall From 1986 to 1989, he was with the Pacific Gas and Electric Company, San
solution framework with an application example taken over a Francisco, California, where he was a major developer of the Hydro-Thermal
real urban distribution network. Optimization program. Since 1989, he has been with IST, Department of Elec-
trical Engineering and Computers, where he is currently an Associate Professor.

REFERENCES
[1] R. Billinton and S. Jonnavithula, “Optimal switching device placement
in radial distribution systems,” IEEE Trans. Power Del., vol. 11, no. 3, A. J. Cerejo da Silva received the electrical and computer engineering degree
pp. 1646–1651, Jul. 1996. from the Instituto Superior Técnico (IST), Lisbon, Portugal, in 1996.
From 1995 to 1998, he was with the Planning Department of Electicidade
1Note that for larger subsets Algorithm 2 reduces significantly the number do Sul (SLE) at Electricidade De Portugal and with Delphi Packard in Castelo
of necessary evaluations. E.g., for the set S it reduces the number necessary Branco. Since 1998 he has been with Instituto Politécnico de Castelo Branco,
evaluations by 27% C = 1486 instead of 2047). Department of Electrical Engineering and Telecommunications.

You might also like