Comparison between evolutionary programming

and a genetic algorithm for fault-section estimation

L.L.Lai
A.G. Sichanie
B.J.Gwyn

Indexing terms Power systems, Computational intelligence, Evolutionary algorithm, Evolutionary programming, Genetic algorithm

response time for complex systems can be long, The

Abstract: The paper proposes an application of application of ANNs to FSE is an active research area,
evolutionary programming (EP) to fault-section however, the correctness of the estimation result cannot
estimation in power systems. Several techniques be proved theoretically, and as a result the system is
have been employed to solve this problem so far. questionable. As the FSE objective function is usually a
A genetic algorithm (GA) has been reported to be high-order polynomial, the GA optimisation method
one of these techniques. In order to measure the has been employed to deal with such a problem [6].
efficiency of EP and make comparisons, a GA Evolutionary programming is an optimisation algo-
has also been used to solve the same problem. rithm based on the mechanics of natural selections,
Different parameters which affect the EP mutation, competition and evolution The process of
convergence have been investigated. Two object- evolution inevitably leads to the optimisation of
oriented software codes have been developed to ‘behaviour’ within the context of a given criterion. EP
implement the algorithms. A sample power does not require crossover operation and it has a
system is used to examine the algorithms. It shorter run time when compared with GAS [7].
shows that EP is superior to the GA for the type Another approach to fault-section estimation is to
of coding strategy and evolution as defined for use model-based reasoning [8], for which a large invest-
the GA. ment has to be made into protection models and expert
knowledge Many modern protection relays are diffi-
cult to model as the exact algorithm is difficult to
obtain, and for new systems there is little or no opera-
Introduction tional experience. In these cases, an EP or GA
approach could be complementary to model-based rea-
To enhance service reliability and to reduce power out- soning. This is the first time evolutionary programming
age time, rapid restoration of the power system is has been used for FSE
required. As a first step to fault restoration, the fault
section should be estimated quickly and accurately. 2 Problem formulation
The fault-section estimation (FSE) identifies fault com-
ponents in a power system by using information on the To simplify the explanation of the proposed approach,
operation of protective relays and circuit breakers. a model is developed. It is assumed that the modelled
However, this task can be difficult, especially for cases power system consists of the following components and
where the relay or circuit breaker fails to operate and parameters. a circuit breaker (CB), a relay (R), a bus-
for multiple faults. bar (BB), a transformer (T), a transmission line (L) and
Several papers have reported surveys on evolutionary a section z(A,)
algorithms’ (EA) applications in power systems [l, 21 There are three types of relays; a main protective
Few methods have so far been employed to solve the relay (MPR), a primary backup protective relay
FSE problem. They include, expert systems [3] and (PBPR) and a secondary backup protective relay
other computational intelligence techniques (CIT), such (SBPR).
as, artificial neural networks (ANNs) [4, 51 and genetic The operation and description of the protective
algorithms [6]. Among these methods, expert systems relays are as follows:
are based on production rules and involve a great Each busbar has one MPR. It is used to initiate the cir-
number of rules describing the complex protection-sys- cuit breakers to disconnect the fault in the busbar.
tem behaviour. This will result in various problems. Each transformer has three relays, namely, MPR,
The maintenance of such a complex rule-based ltnowl- PBPR and SBPR. The MPR is used to initiate the two
edge base is very difficult and time consuming. The circuit breakers at its ends. The PBPR is to initiate the
0IEE, 1998 circuit breakers when a fault is on one of its neighbour-
IEE Proceedings onlrne no 19982192 ing elements and the main protective relay fails to
Paper first received 22nd October 1997 and In rewsed form 18th March operate. The purpose of the SBPR is to protect the
1998 transformer in case of a fault occurring on one of its
The authors are wth the Energy Systems Group, Clty Umversity, Lon- neighbouring elements and the main protective relay of
don, UT( the faulted element fails to operate.
616 IEE Proc -Gener Trnnsm Distrib , Vol 145, No 5, September 1998
Each transmission line has two sets of MPR, PBPR and ( 1 - 2(T,-PR1M))TZ(1
- T,-MAIN)
SBPR; one for the sending end and one for the receiv- (5)
ing end. The MPR of each end is to actuate the circuit - L,S-MAIN)
(1- 2(L,S-PRIA4))LZ(1 (6)
breaker of that end when there is fault on the line. The
PBPR in each end is used to protect the line in case a ( 1 - 2(L,R_PRIM))L2(1
- L,R-MAIN) (7)
fault has occurred while the main relay fails to operate. Where PRIM and MAIN show the primary back-up
The SBPR is used to protect the transmission line in and main protective relays, respectively, for each com-
case a fault occurs on one of its neighbouring elements ponent. T, and L, are the sections.
but the main protective relay fails to operate.
SBPR of any transformer:
3 Mathematical model and fitness function (1 - 2(T,-SEC))x { 1 - [1- A,(1 - A,-MAIN)]
To use the EP technique, the problem has to be mod- x [l - Ak(1 - A k - M A I N ) ]} x (CBiICB,)
elled mathematically. The EP technique is based on the
assumption that a fitness landscape can be character- (8)
ised in terms of variables and there is one or a set of Where Tl-SEC is the SBPR of transformer TI,AI is one
optimum solutions. A mathematical model has been of the sections protected by this relay, AJ_MAIN is the
introduced in [6] and is used to formulate a 0-1 integer main relay of this section, A, is the other section which
programming problem which is then solved by the GA. is protected by this relay and A,MAIN is its main
Although EP [9] and GAS [lo, 111 are evolution algo- relay. CBl and CB, are neighbouring circuit breakers of
rithms, they have a different data structure. EP has the transformer.
more flexibility for selection of data type. It operates SBPR of any transmission line:
directly on an individual type, e.g. the floating-point. A For sending end:
form of the same mathematical model used in [6] is
adapted in this paper. This modified version can be ( I - 2 ( L 2 S - S E C ) ) {-
1 [I - A,(CBk)]} (9)
operated in conjunction with EP characteristics in the For receiving end:
forms of a [min, max] floating-point programming
problem. l [l - A 3 ( C B k ) ] } (10)
(1 - 2 ( L , R _ S E C ) ) { -
The fitness function is one of the main elements of an Where L,S-SEC and L,R-SEC are the secondary send-
EP algorithm. The fitness function evaluates each indi- ing and receiving end relays, respectively. CBk is the
vidual and returns a value indicating how fit that indi- corresponding circuit breaker of the corresponding sec-
vidual is, to be considered as a solution to the problem. tion A].
Each component of the system corresponds to a The whole fitness function is the summation of the
mathematical term which appropriately affects the final corresponding terms of each protective relay and cir-
value of the fitness function. The status of busbars and cuit breaker. Each configuration of circuit breakers and
relays are presented by two values, 0 for non opera- relays produces an input pattern. A list of possible
tional and 1 for operational conditions. Each section of fault sections forms the system output.
the power system model is considered as an individual
for EP. This individual varies over a predefined range, 4 Evolutionary programming
A,, < Ai < A,,,. In this paper we considered A,,, = 1
and A,, = 0. Any values in between could also repre- EP is a computational intelligence method in which an
sent the probability for a fault to occur in the section optimisation algorithm is the main engine for the proc-
concerned. ess of three steps, namely, natural selection, mutation
The combination of 0-1 integers from relays and and competition. According to the problem, each step
busbars status and floating-point values of individuals could be modified and configured in order to achieve
created by EP, will result in an appropriate value to the optimum result. Each possible solution to the prob-
show the fitness of that particular individual. The final lem is called an individual. The mathematical form of
output of the fitness function will then be deducted the ith individual is
from a large positive constant number in order to
secure a positive fitness. p z = [ A i ] , k = 1 , 2 ,...,m (11)
By considering the above points, the corresponding
where m is the maximum number of parameters in any
mathematical terms for each relay are directly related
possible solution and A,,, < A," < A,,,. In order to use
to the protection system, explained in Section 2, and
EP, the mathematical model should be capable of deal-
are listed as follows:
ing with the data type and structure of individuals.
MPR of any busbar, transformer, sending and receiving
The EP algorithm has four main stages:
ends of transmission line:
Initialisation: The initial population consists of individ-
(1- 2(Ai_MAIN))Ai (1) uals (sections) and is created randomly. The fitness
score J1 of each p , is obtained by a fitness function.
( 1 - 2(Ti_MAIN))Ti (2)
Statistics: The maximum fitness, minimum fitness, sum
(1- 2 ( L i S _ M A I N ) ) L i (3) of fitness and average fitness of this generation are cal-
culated.
(1- 2 ( L i R _ M A I N ) ) L i (4) Mutation; Each p l is mutated in order to generate a
Where MAIN shows the main protective relay of each new population.
component and Ai, Tiand Liare the sections. Competition: Each individual p 1 in the combined popu-
PBPR of any transformers, sending and receiving ends of lation has to compete with some other individuals to
each transmission line: get its chance to be transcribed to the next generation.
IEE Proc-Gener. Transm. Distvib., Vol. 145, No. 5, September 1998 617
 The number of generations is determined by test or tion probability determines how often mutation occurs.
by monitoring the convergence of the system, e.g. the A mutation probabilty of 0.1 is used in our tests.
difference between maximum and minimum fitness. The elitist generation method is used in order to :
transfer the best individual through the generations and
5 Genetic algorithms therefore guarantee the convergence.
Genetic algorithms are the most popular and widely 6 Genetic algorithms versus evolutionary
used of all the evolutionary algorithms. They have been programming
widely applied to solve complex non-linear optimisa-
tion problems in a number of engineering disciplines. Considering both algorithms, it can be concluded that
They operate on populations of strings (chromosomes) the EP run needs less computational time than the GA
that encode the parameter set of the problem to be run. The difference in computational speed can be
solved over some finite alphabet. In the selected case related to the characteristic of each algorithm. GA
study there are 10 parameters in the problem. Each individuals are represented in a binary form, therefore
parameter is encoded into an 8-bit binary string (gene)
producing an 80 (8 x 10) bit chromosome. As the range crossover and mutation are binary operators. Execu-
of each parameter is from 0 to 1, this gives a step size tion of these operators takes a longer time in compari-
(precision) of 0.0039 which is quite acceptable for the son to the simple EP mutation operator, especially
selected case. Each encoding represents an individual in when the number of system-free parameters is high and
the GA population. The population is initialised to chromosomes are long and/or population size is large.
random individuals (random chromosomes) at the start The following tests can demonstrate the efficiency of
of the GA run. The GA searches the space of possible each algorithm for this specific problem. In the follow-
chromosomes (hamming space) for better individuals. ing Section the results with both EP and the GA are
The search is guided by 'fitness' values returned by the presented. The presented graphs which show the varia-
modelled protection strategy explained earlier. The tion of the maximum fitness during the generations,
same fitness function as used in EP is adopted. This can be used to coinpare the behaviour of EP and the
gives a measure of how good each individual is in GA.
terms of solving the problem and hence determining its
chance of' appearing in future generations.
Two types of rules are used by genetic algorithms in
their search for highly fit individuals; selection rules
and combination rules.
Selection: The selection rule is used to determine the
TI I - - 1
BB3 I
T3 I

individuals that will have a representation in the next

generation of the GA. Two inajor selection mechanisms
are commonly adopted in a GA search; roulette-wheel
I
I I
selection and tournament selection. In roulette-wheel T4
selection, the probability of being selected is propor- Fig. 1 Power network
tional to an individual's fitness value. In tournament
selection, a fraction of the individuals in the population
are randomly selected into a sub-population and com-
petition carried out to select the fittest individuals in
102
_.
10,
100-
c
each sub-population. Both selection mechanisms have
5 99-
._
been used in this paper to produce the results of the
2
+

tests. In each case, the best results have been selected 98-
and presented. '3 97-

g5
Combination: Combination rules are used to intro- E 96-
duced new individuals into the current population or to
create a new population based on the current popula- 94
1 15 29 43 57 71 85 99 113127141 155169183 197
tion. The most commonly used genetic operators are generations
crossover and mutation which are used here. The com- Fig.2 Maximum fitness against number of generations, case I (populu
bination rules act on individuals that have been previ- tion 50)
GA
ously selected by the selection mechanism. A
~

~ EP
reproduction process takes place between the selected
individuals in the current population to produce off-
spring that become individuals in the next generation.
Crossover is a combination rule that produce off-
spring in the hamming interpolation of the parents. In
this paper the uniform crossover is adopted. A crosso-
ver mask is generated initially for all chromosomes.
For each mask position, a real value is generated at
random and compared to the crossover threshold. If
the value is less than the threshold, then the mask at
the position is set, otherwise the mask is cleared. A 1 15 29 43 57 71 85 99 113127141155169183 197
crossover probability of 0.85 is used for production of' generations
results. Fig.3 MaximumJitness against numbev of generations, case I (populu-
tion 10)
Mutation is the randoin occasional alteration of the __ GA
information contained in the chromosome. The muta- ~ EP

618 IEE Pvoc.-Gene?. Tvansm. Distuib., Vol. 145, No. 5, Septembev 1998
7 Case studies

95""""""""'"'
1 15 29 43 57 71 85 99 113 127 141 155 I69183 197
generations
Fig.6 Maximum fitness against number of generations, case 3 (popula-
tion 50)
~ GA
~ EP

102-

97 .

97
96 ' . ' . . . . . . . " . . . . . ' .
1 15 29 43 57 71 85 99113 127 141 155169183197
generations
Fig.7 Maximum jitness against number of generations, case 3 (popula-
tion 10)
~ GA
~ EP

R1 to R28 are BBI-MAIN, BB2_MAIN, BB,-MAIN,

BB4-MAIN, T I M A I N , TZMAIN, T3-MAIN,
Td-MAIN, LI R-MAIN, L,S-MAIN, L&MAIN,
9
96 7 L,S-MAIN, Ti-PRIM, T,-PRIM,
~ T3PRIM,
15 29 43 57 71 85 99 113 I27 I41 155 169183197
1 Td-PRIM, Ti-SEC, Tz-SEC, T3_SEC, T4-SEC,
generations L I R P R I M , LIS-PRIM, L,R-PRIM, L,S-PRIM,
Fig.5 Maximum$tness against number of generations, case 2 (popula- L,R-SEC, L,S-SEC, L,R-SEC, and L2S-SEC, respec-
tion IO)
~ GA tively.
~ EP
Case 1: For case 1 , the operated relays and circuit
breakers are given in Table 1. The number of genera-
As defined in problem formulation, Section 2, A, and tions, population size and estimated fault sections with
R define the section and relay, respectively. The sec- EP and the GA are also presented in the same Table.
tions, protective relays and their assignments are as fol- Figs. 2 and 3 show the maximum fitness during genera-
lows: tions.
A1 to A10 are BB,, BB,, BB3, BB4, Ti, T,, T3, T4, L , In this case, a possible explanation could be as fol-
and L,, respectively. lows:

Table 1: Status of relays, circuit breakers and results

No of Reported actuated Results: Fault Results: Fault

Size Of Reported actuated relays
generations population circuit breakers section EP section GA

200 10 and 50 L,R-SEC, L,R-SEC, T,-SEC CB8, CB,,, CB,, BB3 BB3

Table 2: Status of relays, circuit breakers and results

No of Size of Reported actuated Results: Fault Results: Fault

Reported actuated relays
generations population circuit breakers section EP section GA
200 10 and 50 T,-SEC, T,-SEC, L,S-MAIN, CB,, CB,, CB8 BBZ, Li BBZr Li
L,R MAIN. L,S PRIM

Table 3: Status of relays, circuit breakers and results

No of Size of Reported actuated Results: Fault Results: Fault

Reported actuated relays
generations population circuit breakers section EP section GA
200 10 and 50 TS
,EC, TLPRIM, BB,MAIN CBq3,CB,,, CB,, BB,, T4 BB4, T4

IEE Proc-Gener. Transm. Distrib., Vol. 145,No. 5, September 1998 619

 BB3-MAIN, LISPSEC and L,S-SEC fail-to-trip. FSE problem with an even higher speed and efficiency.
CBlo is actuated, so T3-SEC may operate but the sig- Fault-section estimation in a control centre can be
nal has not been received by the control centre. CB9 difficult if the information from the energy manage-
operates but no signal has been received by the control ment system is corrupted or missing as well as if there
centre also. CB, may be tripped by L2R-SEC. is a malfunction or failure to operate of the protection
Case 2: For case 2, the operated relays and circuit relays. This paper demonstrates that EP and GA
breakers are given in Table 2. The number of genera- approaches can be used for fault-section estimation
tions, population size and estimated fault sections are when the information available is inaccurate.
also presented in the same Table. Figs. 4 and 5 show
the maximum fitness for both EP and GA generations. 9 Acknowledgments
Case 3: For case 3, the operated relays and circuit The authors would like to thank the National Grid
breakers are given in Table 3. The number of genera- Company plc, UK in providing both the financial and
tions, population size and estimated fault sections are technical support. Special thanks to H C Subasinghe
also presented in the same Table. Figs. 6 and 7 show for completing the paper.
the maximum fitness for both EP and GA systems dur-
ing the generations.
10 References
This computer simulation on average will give an
optimum result iii about 100 to 150 generations. Two 1 SRINIVASAN, D , WEN, F , CHANG, C S and LIEW, A C ~

different population sizes are simulated at 10 and 50. A survey of applications of evolutionary computing to power
system’ Proceedings of the international conference on Intellzgent
EP is proven to be more efficient for smaller popula- systems applications to power systems, Jan IFeb 1996, pp 3541
tion sizes and needs less generations compared to the 2 MIRANDA, V I SRINIVASAN, D., and PROENCA, L M,
GA. This can also be concluded from the above dia- Evolutionary Computation in power systems’ Proceedings of the
12th Powev systems computation conference, Ger
grams. All the simulation results are performed on a pp 2540
Pentium 166MHz PC. The CPU processing time on 3 FUKUI, C , and KAWAKAMI, J ‘An expert system for fault
section estimation usrng information from protective relays and
average for EP is about 13 times faster than for the circuit breakers’, I E E E Tvans Powev Deliv , 1986, 1, (4), pp 83-
GA. Some discussions on the use of GAS to real-life 90
large scale power systems have been presented in [la], 4 LAI, L L , NDEH-CHE, F , CHU, K H , RAJROOP, P , and
WANG, X.F ’ ‘Design of neural networks with genetic algo-
and it can be forseen that EAs could have a good rithms for fault section estimation’ Proceedings of the 29th Uni-
potential for practical applications. versities Power engineering conference, GalGay, Ireland, 1994,
Vol 2, pp 596-599
5 YANG, H T , CHANG, W Y , and HUANG, C L ‘A new neu-
8 Conclusions

An EP approach has been developed for solving the Trans Power Dehv , 1994, 9, (1), pp 220-230
6
FSE problem including malfunctions of protective
relays and/or circuit breakers and multiple fault cases.
A comparison is made with the GA approach at the ference on Power system teclznology, 1994 (International Aca-
demic Publishers, Beijing, China), pp 1051-1055
same time. Two different population sizes are tested for 7 FOGEL, D B ‘A comparison of evolutionary programming and
each case In general, EP is shown to have a faster genetic algorithms on selected constrainted optimisation prob-
computational speed than the GA and on average, it is lems’, Szmulatron, 1995, 64, (6), pp 397-404
8 WHIPP, P , and LEWIS, T ‘Model based reasonmg and simula-
13 times faster. The final results were almost the same tion’ in VADERA, S (Ed ) ‘Expert system apphcations’ (Sigma
The convergence speed (the required number of genera- Press, 1989)
9 LAI, L L ‘Intelligent system apphcations in power engineering
tions to get an optimum result) is a very important fac- evolutionary programming and neural networks’ (John Wiley and
tor in real-time applications. Therefore, according to Sons, Inc , 1998)
the tests carried out, EP has a benefit over the GA for 10 HOJJAT, A , and SHIH-LIN, H ‘Machine learnmg. neural net-
works, genetic algorithms and fuzzy systems’ (John Wiley and
the coding strategy and type of evolution as defined for Sons, Inc , 1995)
the GA However, as both EP and the GA are evolu- 11 LAI, L L , MA, J T , YOKOYAMA, R , and ZHAO, M
tionary algorithms their efficiencies are largely depend- ‘Improved genetic algorithms for optimal power flow under both
normal and contmgent operation states’, Int J Electr Power
ent on the complexity of the problem which might Energy Syst , 1997, 19, (5) -292
differ from case to case EAs are also ideal for parallel 12 WEN, F., and CHANG, new approach to fault section
estimation in power systems based on the set covering theory and
processing computer systems or hardware. Therefore, a refined genetic algorithm’ Proceedings of the 12th Power sys-
with this kind of equipment, it is possible to solve the tems computation conference, Germany, Aug. 1996, pp. 358-365

620 IEE Proc -Gene? Tvansm Distrih , To1 145, No 5, September I998