NEURAL NETWORKS APPLICATION TO FAULT DETECTION

IN ELECTRICAL SUBSTATIONS

Luiz Biondi Neto, Pedro Henrique Gouvêa Coelho

Electronics and Telecommunications Department, State University of Rio de Janeiro - UERJ

Rua São Francisco Xavier, nº 524, Bl. A, Sala 5036, Maracanã, 20550-013, Rio de Janeiro, RJ, Brazil

Alexandre Mendonça Lopes, Marcelo Nestor da Silva

AMPLA Energia e Serviços S.A., Praça Leoni Ramos, nº 1, São Domingos, Niterói, RJ, Brazil

David Targueta

Pro-Energy Engenharia LTDA, Rua cinco de Julho nº322, sala 902, Icaraí, Niterói, RJ, Brazil

Keywords: Fault detection in substations, Alarm Processing, Neural Networks, Decision Making Support.

Abstract: This paper proposes an application of neural networks to fault detection in electrical substations, particularly

to the Parada Angélica Electrical Substation, part of the AMPLA Energy System provider in Rio de Janeiro,

Brazil. For research purposes, that substation was modeled in a bay oriented fashion instead of component

oriented. Moreover, the modeling process assumed a substation division in five sectors or set of bays

comprising components and protection equipments. These five sectors are: 11 feed bays, 2 capacitor bank

bays, 2 general/secundary bays, 2 line bays and 2 backward bays. Electrical power engineer experts mapped

291 faults into 134 alarms. The employed neural networks, also bay oriented, were trained using the

Levenberg-Marquardt method, and the AMPLA experts validated training patterns, for each bay. The test

patterns were directly obtained from the SCADA (Supervisory Control And Data Acquisition) digital

system signal, suitably decoded were supplied by AMPLA engineers. The resulting maximum percentage

error obtained by the fault detection neural networks was within 1.5 % which indicates the success of the

used neural networks to the fault detection problem. It should be stressed that the human experts should be

the only ones responsible for the decision task and for returning the substation safely into normal operation

after a fault occurrence. The role of the neural networks fault detectors are to support the decision making

task done by the experts.

1 INTRODUCTION

Electrical substations evolved rapidly in time not

only in their conception but also in their protection

equipment that are now fully electronic instead of

electromechanical, in the digital substation age. The

Standard IEC (International Electromechanical

Commission) 61850 and the possibility of using high

speed and reliable Ethernet LAN networks brought

new developments to the area. Such developments

include sharing information among several IEDs

(Intelligent Electronic Devices) as well as the

capability of providing these information to several

Electrical Energy Companies users or industrial

heavy consumers (Cascaes et alli., 2007) As a

natural consequence, the current supervision,

automation, and control systems are gradually being

adapted to that new reality that is present in the

Brazilian electrical sector. Nowadays, the main

facility to aid the operator in the supervision system

system comprised by multiple alarms having the

purpose of making the operator aware of the

problems that is afflicting the electrical sector

(Chan, 1990). The operator must detect the fault

based on the set of fired alarms and proceed to the

corrective action towards a quick recover of the

system and to normal operation conditions. So all

the responsibility to return the system to normal

operation lies on the shoulders of the operator which

can suffer a lot of stress and pressure. Due to

484

Biondi Neto L., Henrique Gouvêa Coelho P., Mendonça Lopes A., Nestor da Silva M. and Targueta D. (2008).

NEURAL NETWORKS APPLICATION TO FAULT DETECTION IN ELECTRICAL SUBSTATIONS.

In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 484-487

DOI: 10.5220/0001697204840487

 SciTePress

operator fatigue and inexperience, and an excessive

number of simultaneous fired alarms, usually a

significant number of wrong diagnostics occur

which seriously affects security and efficiency in

electrical systems (Kwang-Ho et alli., 1993).

Modern techniques (Biondi et alli., 2007) such as

Neural Networks can help in the solution of the

problem, leaving the operator focused on the

corrective action in which his or her participation is

very important. The investigated substation is part of

the AMPLA system and is called Parada Angélica

(PAR) which diagram is shown in Figure 1.

The susbstation comprises 19 BAY’s as indicated in

Table 1 and shown in Figure 1.

Table 1: BAY's Distribution in the substation.

Type of BAY Quantity

Feeder 11

Capacitor Bank 02

General/Secondary 02

Line 02

Backward 02

TOTAL 19

 Feeder BAY’s are the following: PAR 5,

PAR 22, PAR 17, PAR 11, PAR 14, PAR

8, PAR 6, PAR 15, PAR 9, PAR 7, and

PAR 10.

 Capacitor Bank BAY’s are: BCO 3 and

BCO 4.

 General/Secundary BAY’s are: General

T3 and General T4.

 Line BAY’s are: LI ALC/ADR # 3 and LI

ALC/ADR # 1.

 Backward BAY’s are: all other BAY’s.

diferindo apenas, relativamente às linhas LI

#3 e LI #1.

The substation in study has two (138/13.8 KV)

three-phase transformers (TRAFO), and a high and a

low Bus (BUS), subdivided in two, connected by a

link disconnector.

Figure 1: Parada Angélica Substation Diagram.

NEURAL NETWORKS APPLICATION TO FAULT DETECTION IN ELECTRICAL SUBSTATIONS

485

2 DATA PREPROCESSING

The AMPLA data base concerning the mapping of

alarms into faults characterizes, in a reliable way,

the Parada Angélica substation operation in respect

of alarms processing. The dimension of the matrix

which maps alarms into faults, for each bay, is

shown in Table 2, for 291 faults.

Table 2: Mapping Dimension.

BAY Alarms X Faults

Feeder 16 x 20

Capacitor Bank 19 x 20

General/Secondary 19 x 62

Line 13 x 32

Backward 67 x 157

Table 3 shows the mapping of alarms into faults,

only for the line Bay. The mappings concerning the

other neural networks were not shown here due to

the oversize of such Tables. As seen in Table 3, each

fault is defined by the set of fired alarms and that

behavior is trained by 5 neural networks, one for

each substation Bay.

Table 3: Line BAY Mapping.

A 0

0 1 1 0 0 0 0 0 0 0 0 0 0

1 1 1 0 0 0 0 1 0 0 0 0 0

1 1 1 0 0 0 0 1 1 0 0 0 0

1 1 1 0 0 0 0 1 0 1 0 0 0

1 1 1 0 0 0 0 1 0 0 1 0 0

1 1 1 0 0 0 0 1 0 0 0 1 0

0 1 1 0 0 0 0 0 0 0 0 0 1

1 1 1 0 0 0 0 1 0 0 0 0 1

0 0 0 1 1 0 0 0 0 0 0 0 0

F10

1 0 0 1 1 0 0 1 0 0 0 0 0

F11

1 0 0 1 1 0 0 1 1 0 0 0 0

F12

1 0 0 1 1 0 0 1 0 1 0 0 0

F13

1 0 0 1 1 0 0 1 0 0 1 0 0

F14

1 0 0 1 1 0 0 1 0 0 0 1 0

F15

0 0 0 1 1 0 0 0 0 0 0 0 1

F16

1 0 0 1 1 0 0 1 0 0 0 0 1

F17

0 0 1 0 0 1 0 0 0 0 0 0 0

F18

1 0 1 0 0 1 0 1 0 0 0 0 0

F19

1 0 1 0 0 1 0 1 1 0 0 0 0

F20

1 0 1 0 0 1 0 1 0 1 0 0 0

F21

1 0 1 0 0 1 0 1 0 0 1 0 0

F22

1 0 1 0 0 1 0 1 0 0 0 1 0

F23

0 0 1 0 0 1 0 0 0 0 0 0 1

F24

1 0 1 0 0 1 0 1 0 0 0 0 1

F25

0 0 0 0 1 0 1 0 0 0 0 0 0

F26

1 0 0 0 1 0 1 1 0 0 0 0 0

F27

1 0 0 0 1 0 1 1 1 0 0 0 0

F28

1 0 0 0 1 0 1 1 0 1 0 0 0

F29

1 0 0 0 1 0 1 1 0 0 1 0 0

F30

1 0 0 0 1 0 1 1 0 0 0 1 0

F31

0 0 0 0 1 0 1 0 0 0 0 0 1

F32

1 0 0 0 1 0 1 1 0 0 0 0 1

The Levenberg-Marquardt backpropagation

algorithm was used for the neural networks training.

3 MODELING, TRAINING AND

RESULTS

The neural network for the feeder BAY had two

hidden layers with 45 and 35 neurons and was

trained with the 16 input alarms and had 20 output

neurons. (16-45-35-20).

The training 16 x 20 matrix was validated by the

correlation statistics technique in order to avoid the

mapping of the same set of alarms into different

faults. The neural network general model that is the

basis for all neural networks used in this paper is

shown in Figure 2 for the feeder BAY. The best

neural network architecture was the (16-45-35-20)

one, that is the neural network with 16 input

neurons, 45 and 35 neurons in the hidden layers and

20 output neurons.

Figure 2: Neural Net Model - feeder BAY.

The cross-validation technique entitled to train,

validate, and test all the neural networks during

training. The training results for training and tests

and, the percentage error curve are shown in Figure

3 and, Figure 4 respectively.

0 2 4 6 8 10 12

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Epoch

Squared Error

Training

Tes t

Figure 3: Test and Training - feeder BAY.

0,2

0,4

0,6

0,8

1,2

1,4

1 2 3 4 5 6 7 8 9 1011121314151617181920

TESTES

ERRO

Figure 4: Error Percentage - feeder BAY.

eede

-BAY

A-1

A-16

F-1

F-20

ICEIS 2008 - International Conference on Enterprise Information Systems

486

The neural network for the Capacitor bank BAY was

trained for 19 input alarms and had two hidden

layers with 46 and 33 neurons, and an output layer

with 20 neurons resulting in a (19-46-33-20)

architecture. The neural network for the

General/Secundary BAY was trained for 19 input

alarms and had two hidden layers with 78 and 44

neurons, and an output layer with 62 neurons

resulting in a (19-78-44-62) architecture. The neural

network for the line BAY was trained for 13 input

alarms and had two hidden layers with 65 and 46

neurons, and an output layer with 32 neurons

resulting in a (13-65-46-32) architecture.

Finally, the neural network for the backward BAY

was trained for 67 input alarms and had two hidden

layers with 410 and 220 neurons, and an output layer

with 157 neurons resulting in a (67-410-220-157)

architecture. For all neural networks used, the

activation functions were of the sigmoid type and

the training algorithm was the Levenberg-Marquardt

method. Over 5000 tests proposed by the AMPLA

experts were carried out and validated the mapped

faults in Parada Angélica substation and the results

were within the quality standards required by the

AMPLA energy provider group. Table 4 shows the

maximum percentage error found for each BAY.

Table 4: Maximum Percentage Error.

BAY Maximum

Percentage Error

Feeder 1.2610 %

Capacitor Bank 0.6450 %

General/Secondary 0.4679 %

Line 0.8385 %

Backward 1.4493 %

4 CONCLUSIONS

The results indicate the success of BAY oriented

application of fault detection in electrical

substations. Table 4 shows that the highest

percentage error found in all BAYs is less than

1.5%. According to AMPLA experts, the application

produces faster and more reliable responses

compared to traditional procedures of fault detection

which are completely dependent on human beings

empirical analysis. Due to the good results so far, the

authors of this paper are pursuing an ongoing action

for a man-machine interface to be imbedded in an

expert system in order to explain each occurred fault

and also in the concept modeling of a knowledge

base for that project.

REFERENCES

Kwang-Ho Kim and Jong-Keun Park, 1993. Application

of hierarchical neural networks to faults diagnosis of

power systems, Electrical Power & Energy Systems,

Vol 15 No. 2 , pp 65-70.

Biondi Neto, L. ; Ferreira, R. A. B., Targueta, David ;

Mello, João Carlos Correia Baptista Soares, 2007.

Intelligent System for Detection Fault in Electrical

Systems, III European-Latin-American Workshop on

Engineering Systems, 2007, Curicó, 2007.

Cascaes Pereira, A.; Cáceres, David; Biondi Neto, L.;

Ordacgi Filho, J. M.; Correia, J. R.; Pellizzoni, R.,

2007. Rede de IEDs de Proteção como obter o máximo

benefício para proteção e automação de Subestações,

XII Encontro Regional Ibero-americano do CIGRÉ -

ERIAC, 2007, Foz do Iguaçu. XII ERIAC.

NEURAL NETWORKS APPLICATION TO FAULT DETECTION IN ELECTRICAL SUBSTATIONS

487