Network Reconfiguration for Improving Reliability in Distribution Systems*

Richard E. Brown (SM)

Abstract-Feeder reconfiguration i s widely used to halance feeder loading, improve voltage regulation and reduce system losses. Reconfiguration is accomplished by closing normally-open switches and opening normally-closed switches. This paper extends the use of feeder reconfiguration to improve reliability. A predictive reliability model is used to compute reliability indices for the distribution system and a novel algorithm, an annealed local search, is used to adjust switch positions until an optimal solution is identified. Properties of the annealed local search are explored on a hypothetical test system and seven actual distribution systems with different reliability characteristics. Keywords-Feeder, Distribution System, Reconfiguration, Reliability, Optimization, Annealed Local Search
goritbm is then applied to seven actual distribution systems with different reliability characteristics. 11. PREDICTIVE RELIABILITY ASSESSMENT Reliability assessment models are needed to allow feeder configurations to be optimized for reliability. These models must he able to predict reliability at each customer location based on system topology and component reliability data. The first predictive model was produced by EPRI in 1978 [41 and was not widely adopted. Several utilities developed their own distribution reliability assessment capabilities in the early and mid 1990s 15-71, Commercial packages became available in the late 1990s. This paper uses an analytical simulation to predict distribution system reliability. Analytical simulation (1) models each system contingency, (2) computes the reliability impact of each contingency of each customer, and (3) weights the impact based on the expected frequency of each contingency. This method accurately models complex system behavior and dynamically enumerates each possible system state. The analytical simulation used in this paper produces the following results for each customer:
- Expected number of momentary interruptions per year - Expected number of sustained interruptions per year - Expected annual interruption hours per year

1. INTRODUCTION

RIMARY distribution feeders carry power from distribution substations to distribution transformers. On radial systems, each customer is connected to a spccific feeder for a given configuration. A feeder fault causes a protection device to open and results in interruptions for customers downstream of the device. Normally-closed switches are installed on feeders to provide sectionalizing capability and improve reliability. These switches allow fault isolation and restoration of certain customcrs during line repair as well as switching flexibility for maintenance. Different feeders are connected through normally-open tie switches, allowing certain intcrrupted customers to he temporarily transferred to adjacent feeders during maintenance and repair. A distribution system can be reconfigured by changing the location of normally-open switches. This changes the allocation of customers and the flow of power for the affected feeders. Feeder reconfiguration methods have been developed to minimize losses [I-21 and operational costs 131. This paper uses the feeder reconfiguration concept to improve system reliability by using an annealed local search that takes advantage of the radial nature of many distribution systems. Properties of the annealed local search are explored on a hypothetical test system. The al-

The specific reliability assessment algorithm used has been designed to analyze radially operated distribution systems, and is described in detail in [E]. A radially operated system may he highly interconnected, but at any particular time, a component will have a unique electrical path hack to a distribution substation.
A. Radial Tree Structures

Dr. R. E. Brown is with ABS Power Syslms 81 940 Main campus Drive, Raleigh. NC. 21606 (e-mail: nchanl.e.brown~us.abb.cam). * This paper is based on a panel prcsentation given by the author at the [ E W E S 2001 Transmission and Distribution Conference and Exposition in Atlanta. CA. Oct. 2001

In a radial system, the direction of power flow is well defined. Figure 1 shows an example of a radial tree structure consisting of a component (A) with a parent (B)and two children (C, D). Power will flow into Component A from a unique component that is upstream (in the direction of the substation). The upstream component is referred to as the parent of Component k There may also he a set of components just downstream of Component A. Each of these components is referred to as a child of Component A. The parenuchild structure is important since its modification is the goal of the feeder reconfiguration process. A radial system is modeled by representing components as objects and parenuchild relationships as pointers between objects. Each component has a pointer to its parent, and a list of pointers to its children. This data structure

0-7803-7989-6/03/$17.00 02003 IEEE

2419

allows for the efficient implementation of power flow and reliability algorithms using tree navigation techniques. Upstream components such as protection devices and sources can he found by following parent pointers. Downstream components such as interrupted customers can be found using depth-first and breadth-first search techniques.

of the feeder reconfiguration is to this function suhject to equipment overloads and voltage violations:
Problem Formulation minimize: w,SAIDI + uvSAIH + K~MAIFI s.t.: No equipmcnt overloads No voltage drop violations wK: Reliability Index Weight

(4)

, . . . . . . . . . . .~

. . . . . . . .

First Child

The reliability index weights are chosen to reflect utility targets, performance based rate structures or relative cost to customers. are implemented as penalty fac~ ~ . Constraints ~ tors added to the optimization function. The penalty factor formulation for loading violations and voltage violations is:

List of Pointers j to Children j

Component

Pfi,. = E m a x [
i

kVAi -Rated -kVA. Rated - kVAi ' ,o]

(5)

Next Child

Pf," = c im a x [

Min -Voltagei - Voltagei D] Miri _Voltagei

(6)

Figure 1. Radial T r e e S~NC~UE

B. Power NOWConsiderations
When examining possible system reconfigurations, it is important not to violate equipment loading and voltage drop criteria. Excessive equipment loading may result in accelerated loss of life and unsafe line clearances. Line voltage drops may violate regulatory obligations and may damage customer equipment. A power flow for each system configuration (using a robust current-voltage iteration algorithm) is performed to identify voltage and capacity violations.

pfr,, = Penalty Factor due to loading violations pf,, = Penalty Factor due to voltage violations

The penalty factors dcfined in Equations 5 and 6 can be implemented simply and efficiently in the optimization algorithm.

1V. ANNEALED LOCAL SEARCH

Feeder reconfiguration using existing switch locations is a discrete optimization problem. Simulated annealing and genetic algorithms are powerful discrete optimization methods that have hcen successfully applied to reliahility optimization problems [ l0,l I]. Both randomly traverse the scarch space and slowly converge on a solution. They are problematic to apply to radial structure reconfiguration problems for two reasons: (1) most of the generated switch position combinations will not represent feasible solutions, and ( 2 ) generating a new radial tree structure for each combination of switch positions is computationally intensive.

111. PROBLEM FORMULATION

The novel algorithm presented in this paper uses feeder reconfiguration to optimize distribution system reliability. Predictive reliability assessment provides detailed customer reliability information for each customer and allows customized reliability mctrics to be defined. These values can be used to examine the expected reliability for a particular customer, or can be aggregated for a specific area (such as a feeder, a substation, or an entire system) into reliahiljty indices. The most commonly used indices correspond to momentary and sustained intemptions [9] and are defined as: MAIFI =
SAIFl =
# of Customer Momentary Intemptions It of Customen Served # of Customer Sustained Intemptions

peryr (1) peryr (2) hr/yr

A local search avoids the problems of simulated annealing and genetic algorithms by making small changes in the radial structure by closing a normally open switch and opening a nearby upstream switch. This process, referred to hereafter as a tie switch shift, allows incremental changes to the tree structure as opposed to complete reformulation of the structure for each switch position

change.
To illustrate tie switch shift, consider the simple system in Figure 2. It consists of a normally closed switch (SI), a normally open switch (s?), and six line segments (~1.~6). The radial tree structure is shown by solid arrows (pointers to parents) and dotted anows (pointers to children). P is defined as the set of components o n the path from SI to S2 and D is defined as the set of components directly connected to P. In Figure 2, PE(LI.LZ.L~) and D E ( ~ A . L ~ . L ~ ) .

# of Customers Served

Sum of Customer Intemption Durations # of customers Served

(3)

The optimization function for feeder reconfiguration is defined as a weighted sum of reliability indices. The goal

2420

Consider a tie switch shift from S2 to SI. Prior to the shift, power flows from SI to S2. After the shift, a reversal occurs and power flows from S2 to SI. This requires the following changes to the radial tree structure:

IO. Iffbrr, has changed since the last change in r, go to step

3
11. End The annealed local search probabilistically allows system reconfigurations that result in worse reliability andor violate constraints. Occasionally accepting worse solutions allows a greater area of the search space to he explored and increased the likelihood of a near-optimal solution being discovered 1121. Local optimality is assured since the annealed local search reduces to integer programming at a zero temperaturc.
.................
L4 L5 L6

Radial Tree Structure Changes After a Tie Switch Shift 1. For PI and yz in P and a pointer P I + P ~ ,reverse thc connection to become P?+PI. 2. For PI in P, DI in D and a child pointer PI+DI, alter the pointer to hecome P?+)DI where P? is the new parent of P I . 3. For P I in P, D I in D and a parent pointer connecting ~ 1 3 ~ change 1 , the pointer to become D I ~ P Z where PZ is the new parent of PI.
After performing a tie switch shift. the system shown in Figure 2 is reconfigurcd to become the system shown in Figure 3. Power now flows in the reverse direction and terminates at the new normally open point. The tie switch shift is well suited for the optimization technique of integer programming. Integer programming tests tie switch shifts and keep changes if the optimization criterion improves. The process continues until a local optimum is reached. Unfortunately, integer programming has difficulty identifying solutions that differ substantially from thc initial condition. A new approach, referred to as a local annealed scarch, overcomes the limitations of integer programming; allowing a greater search area to be explored by prohabilistically accepting certain solutions with worse objective functions. The primary variable that drives a local annealed search is temperature, r e [O,l]. The temperaturc of a system represents the probability that a worse solution will be accepted. A temperature of 1 gcneralizes to an undirected random search and a temperature of zero reduces to integer programming. A local annealed search is characterized by two parameters: initial temperature ruand anncaling ratc p. The initial temperature determines how freely the algorithm will initially traverse the solution space, and the annealing rate determines the rate that the temperature is reduced. When the temperature keeps dropping without the solution changing, the process is frozen and the algorithm terminates. The annealed local search algorithm is outlined below.

............... :

.................

14

L5

L6

.................

............... 7

Pointerto Child Pointer to Parent

Figure 3. Radial Tree Structure After Tie Switch Shift

V.TESTSYSTEM RESULTS The characteristics of an annealed local search have been examined on an eight feeder test system. Each feeder is a multi-branch layout [I31 and the total system consists of 576 load points, 1865 components, 208 fused laterals, and 200 switches. The test system is shown in Figure 4. Figure 4 represents the most reliable configuration for this system if all feeders are topologically identical and all loads are equal. The optimal system configuration changes if the loading patterns are altered sufficiently. To create a system for optimization, each feeder was assigned a random average loading level and each load point assigned a random demand ranging from zero to twice the average

Annealed Local Search I . Set r = 5,. rus [O,I] 2. Compute the objective function fbpc, 3. Select a tie switch, shift the tie switch in Direction I , and compute the objective functionf,,,, 4. Let r be a uniform random number: r E [O,l] 5. Ifffex, >fbes, or I > r, set&,, =he%, and go to stcp 8. 6. Shift the tie switch back to its previous position 7. Repeat steps 3-7 for Direction 2 8. Repeat steps 3-8 for all untested tie switches 9. Set r = p r

2421

demand. The randomization resulted in LWO of the feeders being overloaded and provides an opportunity for optimization. The impact of annealing rate on solution quality was assessed by performing ten trials for annealing rates between 0 and I .O. The initial temperature was set equal to the annealing rate so that fast annealing rates tended to explore solutions similar to the initial conditions, and slow annealing rates tended to explore a wide range of solutions. Results for optimizing SAID1 (w, = 1, w2 = 0, and w.? = 0 in Eq. 4) are shown in Figure 5.

equal to unity. The plot shows an exponential relationship as the annealing rate approaches unity. Based on Figure 5, an annealing rate of 0.99 was selected to provide reasonable assurance of a high quality solution with a single trial. Similar trials were performed to determine the minimum required initial temperature, rh that consistently produces a quality solution (a low initial temperature being dcsirable to reduce execution time). Ten trials were performed for initial temperatures between 0.1 and 0.99. SoluLion quality versus ro is shown in Figure 7 and execution time versus rois shown in Figure 8. Solution quality and solution variance do no( significantly improve for initial temperatures greater than 0.3. Execution time increases linearly for increasing starting temperatures.

-"
.IC

40

30
U

2:
20

N ._ -

z
jqO,Swit&

10

' ,

..... ,

0
0
0.1

0.2

0.3

0.4

- @Main
~

0.5 0 . 6 Annealing Wle

0.7

0.8

0.9

- - - Feeder Boundary
Figure 4. Eight Feeder Test System

IgLaeral

Figure 6. Execution Tim Venus Annealing Rate

5
B

......................................
1.4

..............

S1.41

..................
1.3

.................................... 1.2

.......................

1.2

I
0
0.1

I
0.2 0 . 3
0.4

0 . 5

0.6

0.7

0 . 8

0 . 9

1
0
0.1

I
0.2

Starting Temperature

0 . 3

0.4

0.5

0 . 6

0 . 7

0 . 8

0 . 9

Annealing w

Figure 7. Solution Quality Venus Staning Tempxatwe

Figure 5. Solution Quality Venus Annealing Rate

Figure 5 shows that average solution quality improves as the annealing rate increases and shows that the variance of solution quality remains high until very slow annealing rates are used. The disadvantage of high annealing rates is increasing execution time. A normalized plot of the of average execution time versus annealing rate is shown in Figure 6. The plot is normalized such that the execution time for an annealing rate of zero (integer programming) is

2422

._ E ._ E"
L

"-

8
U
Y

I
0

(Annealing Fate = 0.99)

30 -

I
0.1

0.2

0 . 3

0 . 6 0.7 Starling Temperatwe

0.4

0.5

0.8

0.9

Figure 8. ExecutionTim Versus Starling Temperature

When applying the annealed local search to actual systems, optimization parameters must he selected a priori. Based on results shown on Figures 5-8, an annealing rate of 0.99 and a starting temperature of 0.3 were selected for the analyses performed in the next section. These values are shown to rcsults in high quality solutions for a single trial while avoiding unnecessary computation time. VI. ACTUALSYSTEM RESULTS The annealed local search methodology has been applied to seven actual distribution systems in order to test the ability of the algorithm to optimize reliability through feeder reconfiguration. Brief descriptions of these topologically diverse systems are now provided.
Servs a cormneTcial and hotel distrin in a large southeastem coastal city. I t i s served by an a& underpound system and i s moderately loaded. Serves an industrial and c o m r c i d dislfict in a large Urban 2: midwestem city. I1 is served by an aging underground system and i s heavily loaded. This system serves a newly developing suburban area in the Suburban 1: Southeast. There are several moderately s i z d industrial customers served by this system Serves B quickly growing suburban mea in the Southeast Suburban2: that i s primarily residential. This system serves B nual area in the Southeast. The infraRural 1: stmcture i s lightly loaded in most areas. but i s heavily loaded in an area that serves several industrial customen. This system serves a mal area in the Southeast. Load Rural 2 density i s low and m s t of the circuit exposure i s single phase overhead c~n~tmction. Imtil~lionaI: This system serves a large institutional campus in a coastal area ofthe Southeast. All buildings are served from a single substation and the site is vrvd through undergound feders and pad-mounted eqlllpmnt.
Urban 1:

quality solutions. Key characteristics of the test systems are summarized in Table I . For each test system, the algorithm is instructed to optimize SAID1 (corresponding to weights of w, = I , w2 = 0 and w = 0 in Eq. 4). The results of the reliability optimization for the seven test systems are summarized in Table 2. The annealed local search is able to find configurations with higher reliability for each of the seven test systems without violating voltage and loading constraints. This improvement ranges from less than 1% to nearly 15%. In general, higher density systems result in greater reliability improvement. This is true for load density (MVA per circuit mile), switch density (switches pcr circuit mile) and switches per MVA of load. With thc exception of the institutional system, the relationship appears to be nonlinear for the first two measures and linear with respect to switches per MVA of load. The sample set is small and further investigation is required determine whither the relationship is evidenced in a broader array of systems.
Table 1. Exawle SystemChracteristics

These seven test systems are diverse in terms of topol-

The results seen on these seven test systems are encouraging. In an era deregulation and increasing customer reliability needs, utilities are under pressure to improve reliability without increasing cost. For certain types of systems, the annealed local search is able to improve reliability by nearly 15% without any capital spending. Assuming a conservative capital reliability improvement cost of $IO/kVA-hr, reconfiguration optimization results in a reliability improvement that could cost nearly $580.000 for the "Urban I" system. VII. CONCLUSIONS This paper has presented a new algorithm to optimize the location of normally-open points on distribution systems for reliability improvement. The algorithm uses an annealed local search to overcome difficulties associated with simulated annealing, genetic algorithms and integer

ogy, load density, construction type and customer type.

The number of normally-open and normally-closed switches also varies considerably, resulting in varying solution space sizes (solution space grows as the number of switches grows). Larger solution spaces require more time to explore, but incrcase the likelihood of finding higher

2423

programming optimiLation methods. The concept of a tie switch shift is used to maintain the radial structure of the system as normally open points are adjusted to improve system reliability. The algorithm enforces voltage and loading constraints by using penalty functions to ensure that reliability improvements arc not detrimental to other system requircmcnts. The algorithm has been tested on seven actual distribution systems with diverse characteristics. In each case, the algorithm was able to offer improved system reliability through reconfiguration of the normally open points. Improvements range from 1% to 15%, and increase as the number of switches per MVA of load increase. Feeder reconfiguration can improve reliability for little or no capital expenditure. As such, feeder reconfiguration for reliability optimization presents electric utilities with an opportunity to become more competitive by providing higher levels of reliability at a lower cost.

IX. BIOGRAPHY
Richard E . Brom received his B.S.E.E. M.S.E.E.. and Ph.D. in elrctricnl rnginecring from the Universily of Wnshingmn in 1991. 1993. and 1996, respectively. He worked 8 s a consulting engineer from 199I lo 1993. joined ABB in 1996 and is currently the Director or Engineering at ABB Power System. Dr. Brown spcivlizes in the arras of distribution system, reliability asiessmnt, design optimization and computer applications to power systemanalysis. He is a senior " b e r of IEEE and is il rcgiistcred prolessional engineer in the state ol North Carolina.

VIII.REFERENCES
N. Rajakoric. 'Application Asscssmnts of Distribution Network Minimum Loss Reconfiguration.' IEEE Tronrrrctium 081 PowrDeIivery. Vol. 12. No. 4, Oct. 1997. pp. 1786-1792. R. Takeski, D. Rajicic, 'Distribution Network Reconfiguration for Energy Loss Rcductian,' IEEE Tran.suciionr on Power S w r m s . Vol. 12, No. I . Feb. 1 9 9 7 . p ~198-406. . Q. Zhou. D. S h i m h a m d i , W. H. E. Liu, 'Distribution Reconfiguration for Operation Cast Rcduction.' IEEE Tronrodnar on PowrSwre,nr, Vol. 12. No. 2, May. 1997. pp. 730-735. EPRl Kepon EL-2018. Dn.rlopment qf Dirtriburion Reliubiliry mrd Rkk Analysis Modelr. Aug. 1981. ( R e p n prepared by Westinghause
V. Borozan,

COW.)
S . R. Gilligm. 'A Method for Estimatingthe Reliability of Distribution Circuits.' IEEE Truraacrio,n on Power Delivery, Vol. 7, No. 2. April 1992. pp. 694-698. G. Kjolle and Kjcll Sand. 'RELRAD - An Analylical Approach for Distribution System Reliability Asiessment,' IEEE Tmrmzcriurrs on P o s e r Delivery, Vol. 7, No. 2, April 1992. pp. 809-814. R.E. Brown. S . Guptu, S.S. Vekata, R.D. Christie, and K. Fletcher, 'Distribution System Reliability Assessmnt Using Hierarchical Markov Modeling. IEEE Transacrioris on Power Delivery. Vol. 11, N a . 4 . Oct., 1996, pp. 1929.1934. R. E. Brown. Electric Power Dbwiburion Reliability. Marcel Dekker, 2002. IEEE Working Group on System Design, Triol U.w Guidefor Power Disrriburinn Reliobiliry Indices,Repon P1366. EEE, 1998. [IO] R.E. Brawn, S . Gupta, S.S. Vmkata. R.D. Christie, and R. Fletcher, 'AutomaLed Prinnry Distribution System Design: Reliability and Cost Optimization,' IEEE Tmnsarrions on Power D e l i v e v . Vol. 12, No. 2. April 1997, pp. 1017-1022. [ I I ] R. E. Brown, "The Impact 01 Heuristic Initialization on Distribution System Reliability Optimization:' lnternoriond Journal qf engineer^ ing Intelligenr Sysrems lor Electrical Engineering and Cnmmunicaiinnr. Vol. 8. No. I , March 2ooO. pp. 45-52. [I21 U. E. Brown. "Distribution Reliability Asscssmnt and Recontigumtion Optimization," 2WI IEEWPLS Tmansmii.$ionand Distribution
Cmfereace arid Exposirion, Atlanta, CA, Oct. 2001 [ 131 H.L. Willis. Power Disrribulion Reference Book. Marcel-Dekker,

1997.

2424