Evolutionary Algorithm for Protection Relay Setting Coordination

K.K.Li, C. W. So
Hong Kong Polytechnic University
Abstract- The protection relay setting coordination manages
the protection relay operations to clear a system f a d t in several
steps of contingence. Relays which are missoordinated will trip
out unnecessary circuits resulting in electric supply interruption.
The Time Coordination Method (TCM) which formulates the
coordination of relay settings into a set of constraint equations
and objective function is developed to manage the relay settings.
The protection system coordination is a highly constrained
optimization problem and conventional methods fail in
searching for the global optimum. This paper presents the
application of Evolutionary Algorithm (EA) in optimizing the
protection relay setting coordination in comparison with other
intelligent methods. The result shows that Evolutionary
Algorithm is an effective tool to search the optimum protection
setting with maximum constraint satisfactions.

I
I

Initialization
Generation

Objective Value
Evaluation

I.

INTRODUCTION
The protection relay setting coordination manages the
protection relay operations to clear a system fault in several
steps of contingence. Relays which are mis-coordinated will
trip out unnecessary circuits resulting in electric supply
interruption. The Time Coordination Method (TCM) [l] is
developed to manage the relay settings. It formulates the
coordination of relay settings into a set of constraint
equations and objective function, which are optimized by the
Evolutionary Algorithm (EA). EA is a novel technique for
solving highly constrained discrete optimization problems [2]
such as protection relay coordination. This problem is
difficult to be solved by conventional optimization technique
such as linear programming or steeper descend gradient
search [2]. This paper presents the application of
Evolutionary Algorithm on the protection relay setting
Coordination. The results show that EA effectively searches
for the optimum protection relay settings with maximum
constraint satisfactions.
ALGORITHM
11. EVOLUTIONARY
Evolutionary Algorithm (EA) is one branch of the
Evolutionary Computation. It can search for the optimum
solution for a highly constrained problem. The flow chart for
EA is shown in Fig 1.

4 Yes
End of EA
Fig. 1
Evolutionary Algorithm Processes Flowing
Diagram
A. Initialization
The initialization process of EA is similar to all
Evolutionary Computational Methods such as Genetic
Algorithm and Evolutionary Programming. It provides the
starting points for the EA to search for the optimized solution.
The greater number of points to start, the higher is the chance
to search for the global optimum solution. The initialization
of the TCM generates a set of relay settings and formulated a
column vectorX,as shown in equation (1).

Where

kjis the j setting in relay n.

X*=

0-7803-6338-8/00/$10.00(~)2000IEEE

Note
For example, if RI is Inverse
Definite Multiple Time Lag
(IDMTL) Overcurrent (OC)
Relay, RIslis the Current Setting
Multiplier (CSM) and RI, is the
Time Multiplier (TM) of R,.

(1)

813

The dimension of X,,is the summation of all protection

relay settings in the TCM to be processed. Typically, EA
requires pure random initialization. It can broaden the search
area and increase the chance of searching out the global
optimum solution. Unfortunately, protection setting
coordination is a highly constrained problem. The pure
random generated relay settings very often fail due to
constraint violations [3]. For example, a random generated
relay settings may not satisfy the operation time margin
between upstream and downstream relays [l] under fault
conditions. Any insufficient relay operation time margin may
cause unnecessary system supply interruption, which is
classified as a constraint violation case. Those initialized
relay settings with constraint violations will be discarded.
Another set of relay settings will be generated and it will be
tested against the constraint violations as before.
The successful rate of a pure random initialized
protection relay settings without any constraint violation may
be calculated in the equation (2).

NJU)

N=gm

(21

Where
n is the number of relays in the power system,
m is the number of settings in relay n,
N,(ij) is the number of setting steps of relay i setting
j which satisfy the constraint violations,
N,&) is the number of settable steps of relay i
settingj,
N is the successfil rate of the protection relay
settings without constraint violations.

For example, if the total number of relays is 10. Each

relay has two settings with 100 steps in each setting range. If
the chance to satisfy the constraints is only 10% in each
setting range, thus N = (1O h OO)20 = 1x 1
From equation 2, if the number of relays is increasing, the
successful rate of the initialized relay settings without
constraint violations will decrease and approaches to zero. To
maximize the successful rate, the setting pusher technique is
developed [l] to push the random generated protection
settings from unfeasible solution region to feasible solution
region.
The processing of EA introduce the continuous
improvement to relay settings in which some constraint
violation cases are corrected to within the constraint violation
limits. A small number of constraint violations is thus
allowed at initialization stage. In the TCM, the maximum
number of constraint violations is defmed. It counts the
number of constraint violations for the initialized relay
setting during constraint checking. If the checked number of
constraint violations of the initialized relay settings is greater
than the pre-defined value, it will be discarded. Otherwise, it

814

will be put into the eligible pool for TCM process. The
number of constraint violations will be reflected on the
objective value. The initialization process will be terminated
after the pre-defined number relay settings are initialized.

B. Generation
The EA is responsible for the generation of new relay
settings. It is carried out by mutation, which is different from
genetic algorithm [4] and evolutionary programming [5]. For
generation n of the k relay settingsX f i ] , the n + l generation
of k relay settings X,,+,F/ is generated by equation (3).

X,,FI = X f i I +
ofiI=(

onsin
FI = JP

d w x~fiIfl(OY1)

(3)

- ~.)(xnrki)
+Y

where
/7

is the scale factor for EA mutation.

y is the offset for EP mutation.

@(XJk-n is the objective value of the relay settings XJk-1.
N(0,l) is the Gaussian normal distribution noise.
PmJk] is a mutation enabling matrix.
a,@] is a step matrix.
The step matrix 0 Jk] is calculated before mutation
process. This is generated from the objective value @(X,@J
of the protection setting XJk] and each entry
,[RI is
independent of the others.
The mutation enabling matrix Pmfi] is designed to
decrease the number of relay settings alternation in each
mutation process. It is found that a larger number of relay
setting altemation will result in a larger number of constraint
violations. For the Genetic Algorithm, the single point
crossover operator [6] may provide smooth relay setting
alteration and introduce smaller number of constraint
violations, but the speed of searching for the optimum relay
setting is slow. If multi-point crossover operator is applied,
the relay setting altemations in each generation is greater and
will caused larger number of constraint violations.
Consequently, the Genetic Algorithm fails in the Time
Coordination Method application.
At the end of generation, the new generated relay settings
and the old relay settings will undergo a selection process to
select the better relay settings for the next generation. The
EA uses stochastic selection via a tournament [5]. Each new
generated relay settings face competition against a preselected number of opponents and receives a ''win" if it is at
least as good as its opponent in each encounter. Selection
then eliminates those sets of relay settings with the least
wills.

C. Objective Value Evaluation

The objective value evaluation is taking the key-roll in
the TCM. It generates all system constraints according to the
system configurations, fault types and fault locations [ 11. The
constraint checking is playing the important part in the
objective value evaluation. It checks the relay settings
satisfaction in all constraints and counts the number of
constraint violations. The number of constraint violations is a
punishment to the relay settings as it is reflected in the
objective value. The larger number of constraint violations
scores higher objective value resulting in less chances to
survive in the next generation.

D. Termination
The termination of the EA process is similar to the other
evolutionary computation methods, such as evolutionary
programming, by applying the fixed number of generations.
As EA introduce continuous improvement process, the
occurrence of the global optimum solution cannot be
predicted. Unlike Genetic Algorithm which generates
offspring mainly by crossover operator, EA generates relay
settings by mutation. It can get rid of the pre-mutual
dominance which is the solution trapped in the local
optimum. For some other optimization algorithms, the
termination is by monitoring the difference of the objective
values between two consecutive generations approaching to
the pre-defined value. This technique fails in the TCM
because the local optimum relay settings always last for
several number of generations which satisfies the termination
criteria but it is not the best optimum solution.
111.

L67

Line L571

B
-us
B7
0 IDMTL Phase Fault / Earth Fault
Overcurrent Relay
Fig. 2
Typical distribution network

Table 1
Circuit Parameters

SIMULATION

The control parameter of EA are as follows:

Number of generations - EA termination criteria.
Population size - The number of sets of relay settings in
each generation.
Mutation Probability - To generate the mutation
enabling matrix h J k ] .

Note : All values are per-unit @U) at IOOMVAbase and

all Lines are working at 11kV.

The TCM also has a set of control parameters to be set

and are described in [11.
To demonstrate the effectiveness of EA in Protection
Setting Coordination, a typical distribution network with 8
circuits and each circuit is protected by a IDMTL Phase Fault
/ Earth Fault Overcurrent Relay as shown in Fig. 2 for study.
The circuit parameters are listed in Table 1.
Systems
Parameters
EA
SurvivalSize
EA
Offsety

EA
EA

ScaleFactor p
Mutation Factor

Values
10
0
0.9
0.1

815

Population
No. of
Objective
Generations Value
Size
500
0.000670
30
500
0.000650
50
500
0.000730
I00

Simulation
Time per
Generation for
Pentium II
350MHz
0.912 sec
I.728 sec
3.563 sec

The optimum solution among these simulation cases

occurs in case 2 with the smallest objective value. The relay
settings are shown in Table 4. The optimum relay settings
can protect the system with fastest fault clearance time,
maximum operation time margin and minimum number of
constraint violations for all system conditions [13.
Table 4
Optimum Protection Relay Settings
Line

Phase Fault
CSM
I TM

Earth Fault
CSM I
TM

a
&
]
. The mutation enabling matrix Pm&] is controlled by
the Mutation Factor 0.1. The PmJk] is generate in each EA
generation by comparing the Mutation Factor and random
numbers.
The larger population size also allow more sets of
protection settings survives in. each generation and the
divergent effect is reflected on the maximum objective values
in Fig 3, 4, 5 and 6. The divergent effect should be limited
and specific to the problem. In case 2, the divergent effect is
the minimum.
V. CONCLUSION
The Evolutionary Algorithm is successfully applied in the
Time Coordination Method for protection setting
coordination. The results show that the population size and
the number of generations should be pre-determhate by
several trials. The number of relays forms the problem
domain and imposes the divergent effect, which can be
suppressed by the selection of the correct population size.
The future work would be the development on a method to
find out the right population size and the number of
generations automatically.
VI. ACKNOWLEDGEMENTS
The authors would like to thank the Hong Kong
Polytechnic University for supporting the research and
publishing this work.

VII.
The population size is the number of sets of relay settings
in each generation to be processed. Obviously, a larger
population size would use more computation power. Thus,
case 1 is the fastest and the case 3 is the slowest. To examine
the EA performance, all trails are recorded as shown in Fig. 3,
4, 5 and 6. Fig.3 shows the best, average and m a x i
objective values recorded in each generation for the first 100
generations in case 1. From 21 to 93 generations, it is
found that the best objective values are improved
significantly. Beyond 93 generations, the improvements
becomes less significant. When all individuals are improved,
the better relay settings is prepared by EA and stored in
several sets of relay settings. Eventually, the new best relay
settings are generated. This improvement is carrying on for
the first 300 generations as shown in Fig 4. In Fig 4 , s and 6,
the improvement becomes minimum, and the average and the
best objective value becomes almost constant for the last 200
generations. Beyond 450 generations, the trend of
improvement for both average and best objective values
becomes flat. Typical effect also occurs in several other trials
on the case. Therefore, 500 generations is selected to be the
tetmination criteria.
The Survival Size is controlled the tournament size and
10 is recommended by D.B. Fogel [SI. The Offset and Scale
Factor is set to 0 and 0.9 and they control the step matrix

816

REFERENCE

[11 C W So, K K Li, Time Coordination Method for Power

[2]

[3]

[4]
[5]

[6]

System Protection by Evolutionary Algorithm, 1999

IEEE Industry Application society Annual Meeting,
Phoenix Arizona, U.S.A., 3-7 October 1999, Session 53,
paper no 53.4.
R Salomon, Evolutionary Algorithms and Gradient
Search
Similarities and Differences,
IEEE
Transactions on Evolutionary Computation, Volume 2,
July 1998, pp 45-55.
C.W. So, K.K. Li, K.T. Lai, K.Y. Fung, Overcurrent
Relay Grading Coordination Using Genetic Algorithm,
IEE APSCOM-97 International Conference, Hong
Kong, November 11-14,1997, Vol. 1, pp. 283-287.
D.E. Goldberg, Genetic Algorithm in Search,
optimization and Machine Learning, Addison-Wesley,
Reading MA, 1989
D.B. Fogel, An analysis of evolutionary programming,
Proc. of the First Cod. on Evolutionary Programming,
Evolutionary Programming Society, La Jolla, CA, 1992,
pp 43-5 1.
C.W. So, K.K. Li, K.T. Lai, ICY. Fung, Application of
Genetic
Algorithm
for
Overcurrent
Relay
Coordination,IEE 6 Intemational Conference on
Developments in Power System Protection, Nottingham,
UK,March 1997, pp. 66-69

Fig 3
EA performance case lin the first 100 generations

Fig 5
EA performance for case 2

817