You are on page 1of 11

Electric Power Systems Research 163 (2018) 441–451

Contents lists available at ScienceDirect

Electric Power Systems Research


journal homepage: www.elsevier.com/locate/epsr

A binary coded brain storm optimization for fault section diagnosis of power T
systems

Guojiang Xionga, , Dongyuan Shib, Jing Zhanga, Yao Zhangc
a
Guizhou Key Laboratory of Intelligent Technology in Power System, College of Electrical Engineering, Guizhou University, Guiyang 550025, China
b
State Key Laboratory of Advanced Electromagnetic Engineering and Technology, Huazhong University of Science and Technology, Wuhan 430074, China
c
Guizhou Electric Power Grid Dispatching and Control Center, Guiyang 550002, China

A R T I C LE I N FO A B S T R A C T

Keywords: Fault section diagnosis (FSD) of power systems plays an important role in power system operation. In order to
Binary coded brain storm optimization quickly and accurately diagnose the fault section or sections after the occurrence of an event, a novel variant of
Fault section diagnosis brain storm optimization (BSO) in objective space algorithm, referred to as BCBSO (binary coded BSO), is
Power system proposed in this paper. The FSD problem is transformed into a 0–1 integer programming problem. The difference
Logical operation
between the reported alarms and the expected states of protective relays and circuit breakers is used as the
objective function. In BCBSO, each population individual is directly encoded as a binary vector and thereby the
transcoding process can be avoided when solving the 0–1 integer programming problem. In addition, logical
operations instead of floating operations are employed for binary strings, making the evolutionary process more
convenient. In order to verify the performance of BCBSO, three test systems, i.e., the typical 4-substation power
system, IEEE 118-bus system, and a practical power grid in Jilin province of China with different fault scenarios
including single fault and multiple faults with failed and/or malfunctioned protective devices are employed. Six
popular metaheuristic methods including ABC, BBO, DE, GA, PSO, and BSO are utilized to validate the effec-
tiveness of BCBSO. The experimental results comprehensively demonstrate the superiority of BCBSO in terms of
successful rate, diagnosis error, robustness, computation efficiency, convergence speed, and statistics. In addi-
tion, the effect of population size is investigated as well.

1. Introduction functioned operation of PRs and/or CBs, or even a simultaneous


concurrence of them, will make the task tougher. In such a context, it
During the operation of a power system, fault is entirely in- is highly essential to turn to advanced diagnosis techniques such as
evitable. When a fault event occurs, the well-designed relay protec- expert systems (ESs) [1–3], artificial neural networks (ANNs) [4–7],
tion system will operate to isolate the fault section or sections from Petri nets (PNs) [8–10], Bayesian networks (BNs) [11,12], Cau-
the sound part of the power system. After eliminating the fault event, se–Effect networks (CE-Nets) [13], spiking neural P systems (SN P
how to quickly and accurately identify the fault section or sections is systems) [14–16], and optimization based method. These methods
eminently important for recovering the operation and which is pre- possess their own exclusive advantages and disadvantages. ESs based
cisely the whole point of the fault section diagnosis (FSD) of power method firstly constructs a rule-based knowledge base and then uses a
systems. reasoning engine to diagnose the fault section or sections. It can make
The FSD of power systems can be achieved by retrieving the states full use of experts’ empirical knowledge, but the acquisition and
of a set of protective relays (PRs) and circuit breakers (CBs) from the maintenance of knowledge database is burdensome. ANNs based
supervisory control and data acquisition (SCADA) systems. This task method needs to extract numerous representative fault samples to
is manageable if the fault scenario is simple. However, a complicated train the ANNs. It has good fault tolerance and strong learning ability.
scenario accompanied by multiple fault sections, failed and/or mal- But it also faces the same situation that the acquisition and main-


Corresponding author at: College of Electrical Engineering, Guizhou University, Room #401, Jiaxiu South Road, Huaxi District, Guiyang City, Guizhou Province
550025, China.
E-mail addresses: gjxiongee@foxmail.com, gjxiongee@163.com (G. Xiong).

https://doi.org/10.1016/j.epsr.2018.07.009
Received 29 October 2017; Received in revised form 16 May 2018; Accepted 10 July 2018
0378-7796/ © 2018 Elsevier B.V. All rights reserved.
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

tenance of comprehensive training data is onerous. With regard to the Two functionalities, i.e. capability learning and capacity developing
PNs, BNs, CE-Nets, and SN P systems based methods, their diagnosis [31] make BSO have a good equilibrium between the exploration and
principles are methodologically similar. They are graphical-based exploitation. This fascinating feature soon makes BSO attract many
reasoning methods which firstly build a causal model to explicitly attentions and be applied in various research fields [32–37]. However,
express the cause and effect connections between a fault event and to the best of our knowledge, BSO has not been applied to the FSD
the corresponding protective devices and then employ respective problem of power systems.
reasoning methods to diagnose the fault section or sections. They do The original BSO utilizes the k-means clustering algorithm to cluster
not need to extract representative fault samples and the diagnostic the population. It is implemented recursively in each iteration and thus
procedures are transparent. However, how to establish a sound causal is time-consuming. To reduce the computational burden, a variant of
model and to improve the fault tolerance ability requires further BSO algorithm named BSO in objective space (BSO-OS) algorithm [38]
studies. is proposed. BSO-OS replaces the k-means clustering algorithm with a
The optimization based method, methodologically different from simple grouping method. This grouping method is just based on the
those aforementioned methods, creatively formulates the FSD as a individuals’ fitness values and can make the algorithm more efficient
0–1 integer programming problem and then utilizes optimization al- and easier to implement. Hence the BSO-OS algorithm is extended for
gorithms to solve it. This method possesses some technological traits the FSD problem in this paper. The contributions of this work are as
such as rigorous mathematical logic, strong theoretical foundation, follows:
easy to implement, and fast response time. In consequence, it is very
promising to be applied in practice [17]. Up to now, many optimi- (1) A binary coded BSO-OS algorithm, referred to as BCBSO, is pro-
zation algorithms have been successfully applied to the FSD problem. posed. In BCBSO, each idea is directly encoded as a binary vector
Among them, metaheuristic methods have gained considerable at- and thereby the transcoding process can be avoided when solving
tention in recent years mainly due to that they have no strict re- the 0–1 integer programming problem.
quirements on the problem formulation and can avoid the influences (2) Logical operations instead of floating operations are employed for
of the initial condition sensitivity and gradient information. For ex- binary strings, making the evolutionary process more convenient.
ample, Lin et al. [17] firstly proposed an improved objective function (3) BCBSO is applied to the typical 4-substation power system, IEEE
and then hybridized genetic algorithm (GA) with tabu search (TS) to 118-bus system, and a practical power grid in Jilin province of
solve it, which can improve the diagnosis accuracy. Bedekar et al. China. It is comprehensively validated through various severe fault
[18] presented a continuous GA with less storage to solve a Hebb’s scenarios by different performance criteria. The comparison results
learning law based objective function. The objective function is with other popular algorithms consistently demonstrate that BCBSO
simple but it is based on a number of extracted representative fault can be used as a competitive alternative for the FSD problem of
samples. Zhang et al. [19] utilized redundancy and temporal in- power systems.
formation of PRs and CBs to construct an improved analytic model
and then solved it by the TS algorithm with the purpose of improving The remainder of this paper is organized as follows. Section 2 briefly
the correctness and efficiency. Leão et al. [20] used the parsimonious introduces the objective function of FSD problem of power systems. The
set covering theory to formulate an improved objective function proposed algorithm, BCBSO, is fully elaborated in Section 3. In Section
which was optimized by an immune algorithm. Utilizing the same 4, experimental results and comparisons are presented. Finally, Section
objective function, Escoto and Leão [21] designed an adaptive GA to 5 is devoted to conclusions and future work.
further improve the diagnosis efficiency. He et al. [22] proposed an
improved model which takes the failure of PRs and CBs into account
and then employed a binary particle swarm optimization (PSO) to 2. Problem formulation
minimized it. Zhao et al. [23] firstly introduced the stochastic time
domain simulation to generate an improved objective function which When a power system suffers from a fault event, the well-configured
was then solved by a well-designed history driven differential evo- relay protection system will quickly and accurately detect the fault and
lution (DE). Huang et al. applied different optimization algorithms activate the corresponding protective relays (PRs) to trip off the circuit
including honey-bee mating optimization (HBMO) [24], biogeo- breakers (CBs) to isolate it. To know the detailed cooperative re-
graphy-based optimization (BBO) [25], and artificial bee colony lationships among the main PR, primary backup PR, and secondary
(ABC) [26] to the FSD problem. Sobhy et al. [27] employed two test backup PR, one can refer to Ref. [39]. Those reported alarms, i.e., the
cases to verify the effectiveness of ABC in solving the FSD problem operated PRs and the tripped CBs will be stored in the SCADA systems.
once again. Abdelaziz et al. [28] introduced adaptive reduction factor Theoretically, the fault section or sections can be diagnosed by using
and heuristic workers to enhance the local search capability of the reported alarms. Therefore, the mathematical model of FSD is to
HBMO. These methods have verified their own merits in the FSD adopt a backward reasoning method to build a fault hypothesis, in
problem, yet much remains to be explored with regard to consistently which the operating behavior of the reported alarms can be explained
improving the diagnosis results with higher performance according to as logically as possible. Namely, when a fault event occurs under such a
the famous No Free Lunch theorem [29]. fault hypothesis, the expected states of PRs and CBs should be in
Brain storm optimization (BSO), proposed in 2011 [30], is an effi- agreement with the reported alarms as much as possible. The mathe-
cient and versatile swarm intelligence algorithm. BSO is inspired by the matical model can be expressed as the following 0–1 integer pro-
human brainstorming process, in which a diverse group of people with gramming problem:
different backgrounds and expertise gather together to come up with
minF (S )
new ideas to solve a problem. In BSO, each population individual is
represented as an idea and all ideas are clustered into several groups by ⎧ Si=0 or 1
k-means clustering algorithm. Then the ideas are updated based on one s.t . S ∈ ZD

or two ideas in clusters by neighborhood searching and combination. ⎩i = 1, 2, …, D (1)

442
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

where S = [S1, S2,…SD] is the decision variable vector. D is the number the BSO in objective space algorithm [38] was proposed. In which, the
of candidate fault sections. Si denotes the fault state of the ith candidate k-means clustering algorithm is replaced by a simple grouping method
fault section (fault = 1, sound = 0). The objective function can be which is just based on the individuals’ fitness values. The grouping
formulated as follows [40]: method makes the BSO in objective space algorithm more efficient and
easier to implement. Therefore, the BSO in objective space algorithm is
min F (S ) = wm ∑ rk m − rk*m 1 − rk p rk*p − ∑⊕ rk s rk*s employed in this paper. For convenience, the BSO in objective space
+ wp ∑ rk p − rk*p 1 − ∑⊕ rk s rk*s + ws ∑ rk s − rk*s algorithm is referred to as BSO later in this paper unless otherwise
specified.
+ wmal ∑ r jmal − r j*mal + wdir ∑ dirn − dirn* In BSO, each population individual is called an idea denoted as
+ wc ∑ Ci − Ci* 1 − rimal ri*mal (2) Xi = [x i1, x i2 , …, x iD], where i = 1,2, …, N , N is the population size and D
is the problem dimension. Each idea is a real-coded vector and is in-
where Ci and Ci* denote the actual and expected states of the ith CB, itialized as:
respectively. rk and rk* denote the actual and expected states of the kth
PR, respectively. m, p, s, mal, and dir denote the main PR, primary Xid = ld + rand(0,1) × (ud − ld ) (9)
backup PR, secondary backup PR, breaker failure protection, and di- where rand(0,1) is a uniformly distributed random real number in (0,1).
rectional element, respectively. The operator ∑⊕ represents the ld and ud are the lower bound and upper bound of the dth dimension,
“CONTINUOUS XOR” operator meaning that if the elements contained
respectively.
in ∑⊕ are not all zero, the summation result is 1, otherwise the result is
BSO mainly consists of three operators: grouping operator, replacing
0. wm , wp , ws , wmal , wdir , and wc are the weight factors and set to be 0.9,
operator, and creating operator. In the grouping operator, a simple
0.8, 0.7, 0.7, 0.85, and 0.95, respectively.
grouping method just based on the fitness values is implemented to
The expected state of main PR can be formulated as follows:
cluster the N ideas into two groups: the top perce percentage as elitists
rk*m = Sk (3) and the remaining as normals.
In the replacing operator, one randomly selected dimension of each
The expected state of primary backup PR can be formulated as
idea is occasionally replaced by a newly generated random value with a
follows:
probability of prep . This operator, on one hand, can effectively improve
rk*p = Sk ⊗ ( ¬ rk m ) (4) the population diversity and thereby enhance the exploration ability.
On the other hand, it is able to control the randomness by properly
where ⊗ are ¬ denote the logical AND and NOT operators, respectively. setting the value of prep to keep the exploitation ability.
The expected state of secondary backup PR can be formulated as In the creating operator, BSO firstly makes use of a probability pe to
follows: determine whether to generate a candidate idea Vi (i = 1,2, …, N ) based
on “elitists” or “normals”, and then employs a probability pone to de-
rk*s = Sk ⊗ ( ¬ rk m ) ⊗ ( ¬ rk p) ⊕ ∑⊕,Sj∈B (rks) (Sj ⊗ ∏ ( ¬ Cv ) )
Cv ∈ L (r k s, Sj ) termine whether based on one or two selected ideas. The generating
strategy is as follows:
(5)
where ⊕ denotes the logical OR operator. B (rks) denotes the set of Vi = Hi + ξi N (μ, σ ) (10)
sections that protected by rks . L (rk s, Sj ) denotes the set of CBs that
contained in the electrical transmission path from rks to Sj . Xi , one idea
Hi = ⎧
The expected state of breaker failure protection can be formulated ⎨ wi Xi + (1 − wi ) Xj , two ideas (11)

as follows:
where i = 1,2, …, N , j = 1,2, …, N . N (μ, σ ) is the Gaussian random
* = ( ¬ Cj ) ⊗
r jmal ∑⊕,rq∈Z (Cj) rq (6)
value with mean μ and variance σ . w is a weight factor within the range
(0,1). ξ is an adjusting factor which can be calculated as:
where Z (Cj ) denotes the set of PRs that can trip Cj .
ξ = logsig((gmax /2 − g )/ K ) × rand(0,1) (12)
The expected state of directional element can be formulated as
follows: where logsig(⋅) is a logarithmic sigmoid transfer function. g and gmax are
dirn = Sn (7) the current iteration and the maximum number of iterations, respec-
tively. K is for adjusting the logsig(⋅) function’s slope.
The expected state of breaker failure protection can be formulated
After generating the candidate idea Vi , the parent idea Xi will be
as follows:
replaced by Vi if Vi has a better fitness value.
Ci* = ∑⊕,rq∈Z (Ci) (rq* ⊗ rq)
(8)
3.2. Binary coded brain storm optimization

3. Binary coded brain storm optimization It should be noted that in the original BSO and the BSO in ob-
jective space algorithm, each idea is directly encoded by floating
3.1. Brain storm optimization point for the global continuous optimization problems. When im-
plemented in optimization problems with discrete binary or integer
Brain storm optimization (BSO) [30] is a simple yet promising search spaces, BSOs usually transform each idea into the target en-
metaheuristic inspired by human being’s behavior of brainstorming. In coding version with the aid of sigmoid transfer function [41] in each
the original BSO, the k-means clustering algorithm is utilized to cluster iteration. In this paper, in order to directly adopt BSO to solve the FSD
the population into several groups. It is implemented recursively in problem of power systems, a binary coded BSO in objective space
each iteration and thus is time-consuming. To simplify the grouping algorithm referred to as BCBSO is proposed. The main procedure of
process and improve the implementation efficiency, a BSO variant, i.e., BCBSO is shown in Algorithm 1.

443
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Algorithm 1. The main procedure of BCBSO. binary number with value 0 or 1. ξ is determined as follows:

0, if logsig((gmax /2 − g )/ K ) < rand(0,1)


ξ=⎧

⎩ 1, else (17)

The flowchart of FSD of power systems using the proposed BCBSO is


depicted in Fig. 1. When a fault event occurs, the outage area can be
immediately determined through the monitoring picture by the dis-
patchers. Because all fault section or sections are undoubtedly con-
tained in the outage area, the candidate fault section or sections thus
can be easily identified via both static data and real-time data derived
from the SCADA systems. After determining the candidate fault section
or sections, the corresponding objective function can be constructed
swimmingly. Finally, BCBSO is utilized to solve the objective function
to obtain the diagnosis result and ultimately terminate the FSD process.

4. Case studies

In order to illustrate and validate the feasibility and effectiveness of


the proposed FSD method, experiments are executed on a 3.7-GHz Intel
(R) Core(TM) computer with 8.0-GB RAM under MATLAB 2010b. For
comparison, some well-known optimization methods including ABC,
BBO, DE, GA, PSO, and BSO are employed and they are all encoded in
floating scheme. For them, the sigmoid transfer function [41] is used to
transform the population into the binary version.

4.1. Parameter settings for simulation

In this work, the maximum number of fitness function evaluations


(Max_NFFEs) is used as the stopping condition and its value is set to be
500 for all fault scenarios. The population size N is set to be 10 for all
methods unless a change is mentioned. 100 independent runs are
conducted to evaluate the optimization performance. The parameter
settings for the optimization methods are listed in Table 1.

4.2. Performance criteria

The following performance criteria are used to compare the per-


formance of each optimization algorithm.

(1) Successful rate: A successful run of a method means that the method
can correctly diagnose the fault section or sections within the
Max_NFFEs. It is equal to the number of successful runs divided by
the total number of runs.
(2) Diagnosis error: The diagnosis error calculated as F(X) − F(X*) is
BCBSO adopts a binary vector to represent an idea. In the in- recorded when the Max_NFFEs is reached, where X and X* are the
itialization phase, a random number rid (i = 1,2, …, N , d = 1,2, …, D ) diagnosis result and the practical fault section or sections, respec-
generated within the range (0,1) is used to initialize Xid : tively. The mean and standard deviation of the diagnosis error va-
lues of 100 runs are recorded.
0, if rid < 0.5
Xid = ⎧ (3) Computation time: The computation time of each run is recorded

⎩1, else (13) when the Max_NFFEs is reached. The mean computation time value
of 100 runs is recorded.
The following strategy (14) is used to update Xid in the replacing
(4) Convergence graphs: The convergence graphs show the mean error
operator.
performance of 100 runs.
Xid ⟵ ( ¬ Xid ) (14) (5) Statistical ranking: The rankings of all involved algorithms on all
fault scenarios are obtained by using the Friedman test.
In the creating phase, the generating strategy is modified as follows:

Vi = Hi ⊕ ξi (15) 4.3. Test system 1

Xi , one idea The typical 4-substation power system [17,42] shown in Fig. 2 is
Hi = ⎧ firstly employed for study. This test system has been widely used in a
⎨ (wi ⊗ Xi ) ⊙ (( ¬ wi ) ⊗ Xj ), two idea (16)

number of reported papers. It consists of 28 sections, 84 PRs, and 40
where ⊙ denotes the logical XOR operator, respectively. w is a random CBs. 28 sections consist of 12 buses (A1, A2, …, A4, B1, B2, …, B8),

444
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Fig. 1. Flowchart of applying BCBSO in solving the FSD problem of power systems.

eight transmission lines (L1, L2, …, L8), and eight transformers (T1, T2,
Table 1 …, T8). The detailed protective configuration can be found in Ref. [42].
Parameter settings of different methods. Some typical fault scenarios listed in Table 2 are employed for de-
Method Parameters settings monstration. All of these scenarios are severe fault cases. For example,
scenario 1 is a single fault case with one CB (CB6 fails and the fault is
ABC limit = 100 isolated by the corresponding secondary backup PRs) failed to be
BBO I = E = 1.0, mmax = 0.01
tripped off. Scenario 3 is a multiple faults case with failed PR (L1Rm)
DE F = 0.5, CR = 0.9
GA pcrossover = 0.9, pmutation = 0.01
and information lost (main PR L2Sm). Scenario 5 is also a multiple
PSO w = 0.9, c1 = c2 = 2.0 faults case but simultaneously coupled with more PRs (T8m, L5Rm, and
BSO perce = 30, prep = 0.2, pe = 0.5, pone = 0.7, K = 25 L7Sm) and CBs (CB31 and CB40) failed to work properly. Consequently,
BCBSO perce = 30, prep = 0.2, pe = 0.5, pone = 0.7, K = 25 these fault scenarios present high demands for the FSD methods.

Fig. 2. Typical 4-substation power system.

445
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Table 2

0.00E + 00(0.00E + 00)


0.00E + 00(0.00E + 00)
0.00E + 00(0.00E + 00)
0.00E + 00(0.00E + 00)
0.00E + 00(0.00E + 00)
0.00E + 00(0.00E + 00)
Typical fault scenarios of test system 1. (R and S denote the power receiving and
sending ends of a transmission line, respectively).

Mean (std dev)


Scenario Activated protective devices Faulted section
or sections

BCBSO
1 B1m, L2Rs, L4Rs, CB4, CB5, CB7, CB9, CB12, CB27 B1
2 B1m, L1Sp, L1Rm, CB4, CB5, CB6, CB7, CB9, CB11 B1, L1
3 B1m, B2m, L1Sm, L1Rp, L2Rm, L2Rp, CB4, CB5, CB6, B1, B2, L1, L2
CB7, CB8, CB9, CB10, CB11, CB12
4 T3p, L7Sp, L7Rp, CB14, CB16, CB29, CB39 T3, L7

0.00E + 00(0.00E + 00)

0.00E + 00(0.00E + 00)


3.50E − 02(2.11E − 01)
5.60E − 02(2.26E − 01)
6.00E − 03(2.95E − 02)

4.03E − 01(6.41E − 01)


5 T7m, T8p, B7m, B8m, L5Sm, L5Rp, L6Ss, L7Sp, L7Rm, L5, L7, B7, B8, T7,
L8Ss, CB19, CB20, CB29, CB30, CB32, CB33, CB34, T8
CB35, CB36, CB37, CB39

Mean (std dev)


6 L1Sm, L1Rp, L2Sp, L2Rp, L7Sp, L7Rm, L8Sm, L8Rm, CB7, L1, L2, L7, L8
CB8, CB11, CB12, CB29, CB30, CB39, CB40

BSO
Table 3
Successful rates for all fault scenarios of test system 1.

0.00E + 00(0.00E + 00)


4.30E − 02(2.54E − 01)
3.70E − 02(2.03E − 01)
3.40E − 02(2.81E − 01)

2.63E − 01(5.29E − 01)


1.76E − 01(7.71E − 01)
Scenario ABC BBO DE GA PSO BSO BCBSO

1 98 99 100 99 97 97 100

Mean (std dev)


2 98 99 98 98 95 89 100
3 99 97 97 98 95 96 100
4 100 100 100 100 100 100 100
5 58 90 88 86 79 68 100

PSO
6 97 100 99 100 95 100 100

0.00E + 00(0.00E + 00)

0.00E + 00(0.00E + 00)


4.3.1. Simulation results and comparison

1.70E − 02(1.70E − 01)


3.00E − 03(2.11E − 02)
3.00E − 03(2.11E − 02)

1.83E − 01(4.67E − 01)


(1) Successful rate
Mean (std dev)

The successful rates of different methods for all fault scenarios are
presented in Table 3. The best results are highlighted in boldface. From
Mean and standard deviation (std dev) of the diagnosis error values for all fault scenarios of test system 1.

the results, it is observed that only the proposed BCBSO can consistently
GA

diagnose all fault scenarios accurately, while other methods cannot


entirely diagnose them correctly in all runs. Although BSO is able to
0.00E + 00(0.00E + 00)

0.00E + 00(0.00E + 00)

diagnose the fault scenarios 4 and 6 in all runs, it fails on the other fault
1.80E − 02(1.27E − 01)
4.50E − 03(2.57E − 02)

1.67E − 01(4.82E − 01)


3.50E − 02(3.50E − 01)

scenarios especially on the scenario 5 which is the severest fault case


with 8 candidate fault sections. For the scenario 4, although it is a
Mean (std dev)

multiple faults case with failed main PRs, the faults are successfully
isolated by the corresponding primary backup PRs and thus the outage
areas are not extended. Under such a circumstance, the decision vari-
ables are just two in number and the global optimum can be easily
DE

obtained. While for other fault scenarios, the fault situations are more
severe and thus not every method can diagnose them correctly in all
0.00E + 00(0.00E + 00)

0.00E + 00(0.00E + 00)

runs. In conclusion, BCBSO performs significantly well in terms of the


9.00E − 03(9.00E − 02)
9.00E − 03(9.00E − 02)
5.60E − 02(3.84E − 01)

1.39E − 01(4.23E − 01)

successful rate compared with other methods. It can indeed enhance the
performance of BSO dramatically.
Mean (std dev)

(2) Diagnosis error


BBO

The successful rate is used to reflect the diagnosis accuracy, while


the diagnosis precision can be evaluated from the diagnosis error cri-
The best results are highlighted in boldface.

terion. Table 4 summarizes the diagnosis error values of different


0.00E + 00(0.00E + 00)
1.80E − 02(1.27E − 01)
1.80E − 02(1.27E − 01)
1.50E − 03(1.50E − 02)

4.70E − 01(6.20E − 01)


1.05E − 01(6.00E − 01)

methods for all fault scenarios. According to Table 4, BCBSO is able to


provide the better diagnosis error values of the best-so-far diagnosis
solutions than its competitors. The comparison confirms the same
Mean (std dev)

conclusion that BCBSO can markedly improve BSO’s searching ability


from another perspective.
ABC

(3) Robustness

Since metaheuristics adopt random values to initialize the popula-


Scenario

tion and possess randomization procedures, so randomness is an in-


Table 4

herent characteristic of them. As a consequence, it may not be appro-


1
2
3
4
5
6

priate to synthetically evaluate their performance just through a single

446
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Table 5 Table 6
Mean of the computation time (ms) for all fault scenarios. Ranking of different methods according to the Friedman test on all fault sce-
narios.
Scenario ABC BBO DE GA PSO BSO BCBSO
Method Friedman ranking Final ranking
1 16.37 11.64 9.41 15.84 10.40 15.25 14.95
2 16.32 11.59 8.76 15.10 9.66 14.82 14.44 ABC 4.67 5
3 15.99 10.64 8.82 15.20 9.75 14.92 14.39 BBO 3.08 2
4 16.01 9.15 8.44 14.92 9.60 14.66 14.19 DE 3.67 4
5 17.21 14.98 10.22 16.70 11.25 16.44 16.16 GA 3.50 3
6 15.88 10.69 8.69 15.14 9.70 14.94 14.47 PSO 5.92 7
BSO 5.33 6
The best results are highlighted in boldface. BCBSO 1.83 1

run. Their robustness, i.e. stability and consistency can be judged The best results are highlighted in boldface.
through a number of independent runs with different initial popula-
tions. The standard deviation of the diagnosis error values of 100 runs converge faster than BCBSO in the beginning stage, but they are quickly
for all fault scenarios presented in Table 4 clearly indicates that trapped into local search and thus suffer from prematurity. In short,
BCBSO’s recorded values are all zero and they significantly outperform BCBSO is capable of enhancing BSO’s global searching ability for
those of other methods. This observation means that BCBSO possesses avoiding stagnation, and thus creating a better balance between the
remarkable strong robustness and it is able to obtain the correct diag- exploration and exploitation.
nosis result easily just with only one trial.
(6) Statistical analysis
(4) Computation efficiency
The significance difference among different methods can be mea-
The FSD of power systems has high requirement on the diagnosis sured by the statistical analysis. In this paper, the Friedman test is
speed. The mean computation time required for different fault scenarios conducted to obtain the rankings of all involved methods on all fault
are tabulated in Table 5. The result demonstrates that DE needs a scenarios. The test result is summarized in Table 6. It is clear that
minimum of computation time while ABC is computationally the most BCBSO obtains the best ranking, followed by BBO, GA, DE, ABC, BSO,
expensive. Although BCBSO needs more computation time than DE, and PSO. The result further verifies the abovementioned conclusion
BBO, and PSO, it beats BSO and the efficiency can still satisfy the that BCBSO significantly outperforms other methods on all fault sce-
practical demand of application. The comparison manifests that BCBSO narios. In addition, it is observed that BCBSO can highly move up the
has the capability of improving the implementation efficiency of BSO. ranking based on that of BSO. The improvement is due to the binary
encoding technique and logical operations.
(5) Convergence property

Convergence speed is another important criterion for measuring the 4.3.2. Effect of population size
performance of an optimization method. Fig. 3 displays the con- The population size N is an important parameter that affects the
vergence curves of all methods. It can be seen that although BCBSO optimization performance of metaheuristic methods. In this subsection,
cannot provide the fastest convergence speed on all fault scenarios in the sensitivity of involved methods to the variability of population size
the beginning stage, it is able to consistently converge to the optimal is investigated. All the parameter settings are the same as those men-
solutions during the whole evolutionary process. Other methods may tioned in subsection 4.1 except the population size which is set from 10

Fig. 3. Convergence graphs of different methods for test system 1. (a) Scenario 1. (b) Scenario 2. (c) Scenario 3. (d) Scenario 4. (e) Scenario 5. (f) Scenario 6.

447
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Table 7 Table 8
Effect of population size N on the successful rates. Alarm information of test system 2.
N ABC BBO DE GA PSO BSO BCBSO Sequence no. Alarm information

10 58 90 88 86 79 68 100 1 Main protective relay B85m


12 63 91 92 84 88 60 100 2 Secondary backup protective relay L3Rs
14 60 96 98 83 80 73 100 3 Secondary backup protective relay L4Rs
16 67 98 97 93 81 73 100 4 Circuit breaker CB3
18 56 96 98 93 88 82 100 5 Circuit breaker CB4
20 60 95 96 90 90 76 100 6 Circuit breaker CB6
7 Circuit breaker CB8
8 Circuit breaker CB15
to 20 with interval 2. The experiment is conducted on the fault scenario
5 due to that it is the most complex scenario and thus the effect of
population size can be observed clearly. The successful rate results are Table 9
tabulated in Table 7. It is shown that the population size indeed affects Experimental results of test system 2.
the search performance to some extent. Overall, a too small or too large Method Successful rate Diagnosis error Computation Rankinga
value of population size will deteriorate the performance. The reason time (ms)
might be that a small population size can increase the number of
iterations and speed up the convergence at the very beginning evolu- ABC 94 1.67E − 01(6.79E − 01) 16.02 5
BBO 99 2.50E − 02(2.50E − 01) 10.75 2
tionary stage, but the population diversity of possible movements is DE 98 5.80E − 02(4.12E − 01) 8.77 3
poor and thereby the premature is easily caused. Whereas a large po- GA 93 1.75E − 01(6.41E − 01) 15.34 7
pulation size can increase the diversity but the iteration number will be PSO 96 1.00E − 01(4.92E − 01) 9.81 4
decreased significantly, and the probability to find the correct search BSO 94 1.50E − 01(5.97E − 01) 15.08 5
BCBSO 100 0.00E + 00(0.00E + 00) 14.39 1
direction is also decreased. Actually, there is no agreement among re-
searchers on the setting of the population size so far [43] and a tem- The best results are highlighted in boldface.
perate value of population size is recommended. In addition, we ob- a
The ranking result is achieved based on the successful rate.
serve that BCBSO is able to diagnose the fault scenario correctly with
one hundred percent successful rate while other methods have the 15 lines: L1–L15), 103 PRs, and 30 CBs. Line L7 is in maintenance and
possibility of misdiagnosis. It means that BCBSO is highly robust to the out of service. The fault scenario is derived from Ref. [44] and is de-
variability of population size while other methods are more susceptible tailed as follows. During the operation, a fault occurred on bus B85 and
to the effect of population size. the main PR B85m operated to send a signal to trip off the circuit
breakers CB4, CB5, CB7, and CB15. CB4 and CB15 succeeded while CB5
4.4. Test system 2 and CB7 failed. Then the corresponding secondary backup PRs L3Rs
and L4Rs operated to successfully trip CB6 and CB8 off, respectively,
In this subsection, the IEEE 118-bus system is employed to further and thereby the fault was isolated. The observed alarm information are
verify the proposed FSD method. Because the proposed FSD method is listed in Table 8.
related with the outage area rather than the entire system, therefore it The experimental results are summarized in Table 9. It can be seen
has nothing to do with the scale of the test system. Here we only pro- that only BCBSO can correctly diagnose the fault scenario in all runs.
vide a sub-network of the IEEE 118-bus system as Fig. 4 depicts. This Although the successful rates of other methods are over 90%, they
sub-network consists of 28 sections (13 buses: B77, B82–B92, B96, and misdiagnose in some runs. The diagnosis error results also support the

Fig. 4. Sub-network of the IEEE 118-bus system.

448
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

Table 10
Alarm information of test system 3.
Sequence no. Alarm information

1 Main protective relay B7m


2 Main protective relay L11Sm
3 Secondary backup protective relay L1Rs
4 Secondary backup protective relay L3Rs
5 Secondary backup protective relay L4Ss
6 Secondary backup protective relay L10Ss
7 Circuit breaker 1504
8 Circuit breaker 3104
9 Circuit breaker 8204
10 Circuit breaker 4802
11 Circuit breaker 4803
12 Circuit breaker 4811
13 Circuit breaker 3106
14 Circuit breaker 3107

Fig. 5. Convergence curves of different methods for test system 2.

aforementioned statement from another perspective. BCBSO is not the


most parsimonious in computation time but it is also fast. In addition, Table 11
Experimental results of test system 3.
the convergence curves plotted in Fig. 5 indicate that BBO is the fastest
method in the early stage, but it is surpassed by BCBSO after about 130 Method Successful rate Diagnosis error Computation Rankinga
function evaluations. time (ms)

ABC 64 6.29E − 01(8.58E − 01) 16.77 7


4.5. Test system 3 BBO 91 1.53E − 01(4.89E − 01) 12.31 2
DE 88 2.37E − 01(6.80E − 01) 9.69 3
GA 85 3.05E − 01(7.73E − 01) 16.02 4
Besides the benchmark test cases, we also use an actual fault sce- PSO 82 3.39E − 01(7.60E − 01) 10.58 5
nario [45] happened in the Siping power grid in Jilin province of China BSO 77 4.26E − 01(8.13E − 01) 15.42 6
to validate the proposed FSD method. The related power system and the BCBSO 100 0.00E + 00(0.00E + 00) 14.73 1
observed alarm information are presented in Fig. 6 and Table 10, re-
The best results are highlighted in boldface.
spectively. The fault scenario is detailed as follows. The bus B2 suffered a
The ranking result is achieved based on the successful rate.
from a fault but the main PR B2m was out of service due to a certain
reason. Therefore, the corresponding secondary backup PRs L1Rs, L3Rs,
and L4Ss operated to successfully tripped the circuit breakers 8204, the circuit breaker 3106 off. Compared to the previous fault scenario,
1504, and 3104 off, respectively. At the same time, another fault oc- this scenario has more candidate fault sections.
curred on bus B7. The main PR B7m operated and the circuit breakers The experimental results are tabulated in Table 11. It is obvious that
4802, 4803, and 4811 were successfully tripped off while 4801 failed to all the successful rates of the involved methods except BCBSO drops
be tripped off. As a result, the corresponding secondary backup PR considerably whereas BCBSO is still able to diagnose the scenario cor-
L10Ss operated to trip the circuit breaker 3107 off successfully. Influ- rectly in all runs, which means once again that the proposed method
enced by the fault current, the main PR L11Sm malfunctioned to trip

Fig. 6. Siping power grid in Jilin province of China.

449
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

References

[1] T. Minakawa, Y. Ichikawa, M. Kunugi, K. Shimada, N. Wada, M. Utsunomiya,


Development and implementation of a power system fault diagnosis expert system,
IEEE Trans. Power Syst. 10 (2) (1995) 932–940.
[2] E. Vazquez M, O.L. Chacon M, H.J. Altuve F, An on-line expert system for fault
section diagnosis in power systems, IEEE Trans. Power Syst. 12 (1) (1997) 357–362.
[3] Y.M. Park, G.W. Kim, J.M. Sohn, A logic based expert system (LBES) for fault di-
agnosis of power system, IEEE Trans. Power Syst. 12 (1) (1997) 363–369.
[4] T. Bi, Y. Zheng, F. Wen, Y. Ni, C.M. Shen, F.F. Wu, Q. Yang, On-line fault section
estimation in power systems with radial basis function neural network, Int. J.
Electr. Power Energy Syst. 24 (4) (2002) 321–328.
[5] G. Cardoso, J.G. Rolim, H.H. Zurn, Identifying the primary fault section after
contingencies in bulk power systems, IEEE Trans. Power Deliv. 23 (3) (2008)
1335–1342.
[6] G. Xiong, D. Shi, J. Chen, L. Zhu, X. Duan, Divisional fault diagnosis of large-scale
power systems based on radial basis function neural network and fuzzy integral,
Electr. Power Syst. Res. 105 (2013) 9–19.
[7] W.A. Dos Santos Fonseca, U.H. Bezerra, M.V.A. Nunes, F.G.N. Barros,
J.A.P. Moutinho, Simultaneous fault section estimation and protective device
failure detection using percentage values of the protective devices alarms, IEEE
Fig. 7. Convergence curves of different methods for test system 3.
Trans. Power Syst. 28 (1) (2013) 170–180.
[8] V. Calderaro, C.N. Hadjicostis, A. Piccolo, P. Siano, Failure identification in smart
grids based on petri net modeling, IEEE Trans. Ind. Electron. 58 (10) (2011)
performs well and is highly robust against different fault scenarios. In
4613–4623.
addition, it can diagnose the scenario within 15 ms and the computa- [9] X. Luo, M. Kezunovic, Implementing fuzzy reasoning petri-nets for fault section
tion efficiency completely meets the practical demand of application. estimation, IEEE Trans. Power Deliv. 23 (2) (2008) 676–685.
The convergence curves illustrated in Fig. 7 show that BBO indeed [10] Y. Zhang, Y. Zhang, F. Wen, C.Y. Chung, C. Tseng, X. Zhang, F. Zeng, Y. Yuan, A
fuzzy petri net based approach for fault diagnosis in power systems considering
converges the fastest while BCBSO is relatively slow in the early stage. temporal constraints, Int. J. Electr. Power Energy Syst. 78 (2016) 215–224.
However, all other methods are overtaken by BCBSO after half of the [11] Y. Zhu, L. Huo, J. Lu, Bayesian networks-based approach for power systems fault
evolutionary process. BCBSO is capable of successively converging to- diagnosis, IEEE Trans. Power Deliv. 21 (2) (2006) 634–639.
[12] Q. Shi, S. Liang, W. Fei, Y. Shi, R. Shi, Study on Bayesian network parameters
ward the optimal solutions whereas other methods easily stagnate and learning of power system component fault diagnosis based on particle swarm op-
thus obtain poor solutions. timization, Int. J. Smart Grid Clean Energy 2 (1) (2013) 132–137.
[13] W.H. Chen, S.H. Tsai, H.I. Lin, Fault section estimation for power networks using
logic cause-effect models, IEEE Trans. Power Deliv. 26 (2) (2011) 963–971.
[14] T. Wang, G. Zhang, J. Zhao, Z. He, J. Wang, M.J. Pérez-Jiménez, Fault diagnosis of
5. Conclusions and future work electric power systems based on fuzzy reasoning spiking neural P systems, IEEE
Trans. Power Syst. 30 (3) (2015) 1182–1194.
[15] G. Xiong, D. Shi, L. Zhu, X. Duan, A new approach to fault diagnosis of power
In this paper, a novel binary variant of BSO in objective space al- systems using fuzzy reasoning spiking neural P systems, Math. Probl. Eng. 2013
gorithm, referred to as BCBSO, is proposed to solve the FSD problem of (2013) 13. Article ID. 815352.
power systems. In BCBSO, on one hand, the binary encoding technique [16] H. Peng, J. Wang, J. Ming, P. Shi, M.J. Perez-Jimenez, W. Yu, C. Tao, Fault diag-
nosis of power systems using intuitionistic fuzzy spiking neural P systems, IEEE
is directly utilized to code the population and thereby the transcoding
Trans. Smart Grid 99 (2017) 1–8.
process is avoided, making the implementation more explicit. On the [17] X. Lin, S. Ke, Z. Li, H. Weng, X. Han, A fault diagnosis method of power systems
other hand, the using of logical operations instead of floating operations based on improved objective function and genetic algorithm-tabu search, IEEE
for binary strings is able to make the evolutionary process more con- Trans. Power Deliv. 25 (3) (2010) 1268–1274.
[18] P.P. Bedekar, S.R. Bhide, V.S. Kale, Fault section estimation in power system using
venient. The experimental results demonstrate that, the proposed Hebb’s rule and continuous genetic algorithm, Int. J. Electr. Power Energy Syst. 33
BCBSO, strictly following the implementation process of BSO in ob- (3) (2011) 457–465.
jective space algorithm without any improvement by other evolu- [19] Y. Zhang, C.Y. Chung, F. Wen, J. Zhong, An analytic model for fault diagnosis in
power systems utilizing redundancy and temporal information of alarm messages,
tionary operators or algorithms, is able to quickly diagnose various IEEE Trans. Power Syst. 31 (6) (2016) 4877–4886.
severe fault scenarios accurately and precisely. Additionally, the com- [20] F.B. Leão, R.A.F. Pereira, J.R.S. Mantovani, Fault section estimation in electric
parisons also indicate that BCBSO can obtain more robust diagnosis power systems using an optimization immune algorithm, Electr. Power Syst. Res. 80
(11) (2010) 1341–1352.
results and provide faster convergence speed with competitive com- [21] E.F. Escoto, F.B. Leão, Fault section estimation in power systems using an adaptive
putation efficiency. Conclusively, BCBSO can be used as an efficient and genetic algorithm, Proceedings of IEEE Power and Energy Society General Meeting,
reliable alternative for the FSD of power systems. Boston, USA, 2016, pp. 1–5.
[22] Z. He, H.D. Chiang, C. Li, Q. Zeng, Fault-section estimation in power systems based
BSO is relatively young yet promising evolutionary algorithm. In on improved optimization model and binary particle swarm optimization,
future work, it will be extended and enhanced to solve other con- Proceedings of IEEE Power and Energy Society General Meeting, Calgary, Canada,
strained power system optimization problems, such as non-convex 2009, pp. 1–8.
[23] J. Zhao, Y. Xu, F. Luo, Z.Y. Dong, Y. Peng, Power system fault diagnosis based on
economic dispatch and optimal power flow.
history driven differential evolution and stochastic time domain simulation, Inf. Sci.
275 (2014) 13–29.
[24] S.J. Huang, X.Z. Liu, W.F. Su, T.C. Ou, Application of enhanced honey-bee mating
Acknowledgments optimization algorithm to fault section estimation in power systems, IEEE Trans.
Power Deliv. 28 (3) (2013) 1944–1951.
[25] S.J. Huang, X.Z. Liu, Fault section estimation in distribution systems using bio-
The authors would like to thank the editor and the reviewers for geography-based optimization approaches, Int. Trans. Electr. Energy 25 (1) (2013)
their constructive comments. This work was supported by the Scientific 155–168.
[26] S.J. Huang, X.Z. Liu, Application of artificial bee colony-based optimization for fault
Research Foundation for the Introduction of Talent of Guizhou section estimation in power systems, Int. J. Electr. Power Energy Syst. 44 (1) (2013)
University under Grant No. [2017]16, the Guizhou Province Science 210–218.
and Technology Innovation Talent Team Project under Grant No. [27] M.A. Sobhy, A.Y. Abdelaziz, M. Ezzat, W. Elkhattam, A. Yadav, B. Kumar, Artificial
bee colony optimization algorithm for fault section estimation, in: S.S. Dash,
[2018] 5615, the Science and Technology Foundation of Guizhou K. Vijayakumar, B.K. Panigrahi, S. Das (Eds.), Artificial Intelligence and
Province under Grant No. [2016]1036, and the Guizhou Province Evolutionary Computations in Engineering Systems: Proceedings of ICAIECES,
Reform Foundation for Postgraduate Education under Grant No. [2016] Springer, Singapore, 2017, pp. 127–139.
[28] A.Y. Abdelaziz, W. Elkhattam, M. Ezzat, M.A. Sobhy, Fault section estimation in
02. power systems based on improved honey-bee mating optimization, Proceedings of
Eighteenth International Middle East Power Systems Conference, Cairo, Egypt,

450
G. Xiong et al. Electric Power Systems Research 163 (2018) 441–451

2017, pp. 1–7. for power electronic circuit optimization, Proceedings of the Annual Conference on
[29] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE Trans. Genetic and Evolutionary Computation, Vancouver, Canada, 2014, pp. 183–184.
Evol. Comput. 1 (1) (1997) 67–82. [38] Y. Shi, Brain storm optimization algorithm in objective space, Proceedings of IEEE
[30] Y. Shi, Brain storm optimization algorithm, Proceedings of International Conference Congress on Evolutionary Computation, Sendai, Japan, 2015, pp. 1227–1234.
in Swarm Intelligence, Springer, Heidelberg, 2011, pp. 303–309. [39] F. Wen, C.S. Chang, A tabu search approach to fault section estimation in power
[31] S. Cheng, Q. Qin, J. Chen, Y. Shi, Brain storm optimization algorithm: a review, systems, Electr. Power Syst. Res. 40 (1) (1997) 63–73.
Artif. Intell. Rev. 46 (4) (2016) 445–458. [40] G. Xiong, D. Shi, An improved analytic model for fault diagnosis of power grids and
[32] J. Chen, S. Cheng, Y. Chen, Y. Xie, Y. Shi, Enhanced Brain Storm Optimization its self-adaptive biogeography-based optimization method, Trans. China
Algorithm for Wireless Sensor Networks Deployment, Springer, 2015, pp. 373–381. Electrotech. Soc. 29 (4) (2014) 205–211.
[33] H. Duan, C. Li, Quantum-behaved brain storm optimization approach to solving [41] S. Mirjalili, A. Lewis, S-shaped versus V-shaped transfer functions for binary particle
loney’s solenoid problem, IEEE Trans. Magn. 51 (1) (2015) 1–7. swarm optimization, Swarm Evol. Comput. 9 (2013) 1–14.
[34] X. Ma, Y. Jin, Q. Dong, A generalized dynamic fuzzy neural network based on [42] F. Wen, Z. Han, Fault section estimation in power systems using a genetic algo-
singular spectrum analysis optimized by brain storm optimization for short-term rithm, Electr. Power Syst. Res. 34 (3) (1995) 165–172.
wind speed forecasting, Appl. Soft Comput. 54 (2017) 296–312. [43] Y. Wang, Z. Cai, Q. Zhang, Differential evolution with composite trial vector gen-
[35] C. Sun, H. Duan, Y. Shi, Optimal satellite formation reconfiguration based on eration strategies and control parameters, IEEE Trans. Evol. Comput. 15 (1) (2011)
closed-loop brain storm optimization, IEEE Comput. Intell. Mag. 8 (4) (2013) 55–66.
39–51. [44] H. Ren, Z. Mi, Power system fault diagnosis modeling techniques based on encoded
[36] J. Wang, R. Hou, C. Wang, L. Shen, Improved v-support vector regression model Petri nets, Proceedings of IEEE Power Engineering Society General Meeting,
based on variable selection and brain storm optimization for stock price forecasting, Montreal, USA, 2006, pp. 1111–1118.
Appl. Soft Comput. 49 (2016) 164–178. [45] T. Bi, C. Yang, S. Huang, Q. Yang, Improved petri net models based fault diagnosis
[37] G. Zhang, Z. Zhan, K. Du, W. Chen, Normalization group brain storm optimization approach for power networks, Power Syst. Technol. 29 (21) (2005) 52–56.

451

You might also like