A Fuzzy System for Detection of Incipient Fault in Power Transformers ...

0 downloads 0 Views 666KB Size Report
automatically detecting incipient fault in power transformers through analysis of dissolved gases in transformer insulating oil by introducing the IEC 599 standard ...
A Fuzzy System for Detection of Incipient Fault in Power Transformers Based on Gas-in-Oil Analysis Ronaldo R. B. de Aquino, Milde M. S. Lira, Taciana Filgueiras, Heldemarcio Ferreira, Otoni Nóbrega Neto, Agnaldo M. S. Silva, Viviane K. Asfora

Abstract— Dissolved gas analysis (DGA) is one of the most useful techniques do detect the incipient faults of power transformer. However, the identification of the faulted location by the traditional method is not always an easy task due to the variability of gas data and operational natures. This work aims to develop an intelligent system of preventive maintenance for automatically detecting incipient fault in power transformers through analysis of dissolved gases in transformer insulating oil by introducing the IEC 599 standard (International Electrotechnical Commission). The data used to model the system are taken from chromatographic analysis of CELPE (Electrical Company from Pernambuco). The results show a system with high accuracy when compared with other papers approaching the same problem. Moreover, the results also proved the ability of multiple incipient faults detection.

B

I. INTRODUCTION

ECAUSE of the great increase of the electric sector where the electric utilities earn more according to their service availability, the operation continuity or the availability of operating several equipments at any moment became of great importance. Besides, it is also of great importance to anticipate a fault, because except the largest financial losses for the company in repairing the failed equipment, the fault occurrence of the equipment also punishes the company through high fines and denigrates its image in front of the society. Therefore, it is verified that the decision of the specialist maintenance engineering for Manuscript received January 22, 2010. The authors would like to thank to CNPQ, CAPES and CELPE-Brazil for financial research support. Ronaldo R. B. de Aquino is with the Federal University of Pernambuco (UFPE), Acadêmico Hélio Ramos s/n, Cidade Universitária, Cep: 50.740530 – Recife – PE – Brazil (corresponding author: 55-81-21268986; e-mail: [email protected]). Milde M. S. Lira is with the Federal University of Pernambuco, (e-mail: [email protected]). Taciana Filgueiras is with the CELPE, Av. Joao de Barros 111, Boa Vista, Cep: 50.050-902, Recife – PE, Brasil (e-mail: [email protected]). Heldemarcio Ferreira is with the CELPE, (e-mail: [email protected]). Otoni N. Neto is with Federal University of Pernambuco, (e-mail: [email protected]). Agnaldo M. S. Silva is with the Federal University of Pernambuco, (email: [email protected]). Viviane K. Asfora is with the Federal University of Pernambuco, (email: [email protected]).

978-1-4244-8126-2/10/$26.00 ©2010 IEEE

maintaining or not operating equipments is of considerable importance and then may be very well settled. In this regard, it is verified the great importance of using artificial intelligence techniques as Artificial Neural Network (ANN) and Fuzzy Logic (FL) developed appropriately for an electrical power system. Among the most important equipments of an electric system, we can stand out the power transformers, which should work uninterruptedly in order to improve the power quality delivered. An incipient fault in a transformer should be detected so fast as possible in preventing it from future deteriorations. Thus, a correct diagnosis of incipient faults in a transformer is vital for safety and reliability of the electrical grid. A transformer in operation is subject to thermal and electrical stresses, which can destroy the insulating material, thus liberating gaseous products. The gases of insulating oil according to the chromatography analysis contain concentrations (ppm by volume) of hydrogen, oxygen, nitrogen, methane, acetylene, ethane, ethylene, carbon monoxide and carbon dioxide. The dissolved gas analysis DGA can determine the condition of the transformers by the gas concentrations dissolved in the insulating oil, gas generation rate, gas ratio and total concentrations of combustible gas in oil. Overheating, partial discharge and electric arcs are the three primary causes associated to several types of faults. Although the dissolved gas analysis in the insulating oil of the equipments has been used more than twenty years all over the world, it is still nowadays considered as the best predictive technique to diagnose internal problems in transformers. Several methods to analyze these gases are based on the DGA technique, such as, the IEC standard [1], Rogers [2], Duval [3], Dornenburg [4], among other, which diagnose the nature of the transformer deterioration. These techniques were developed through extensive observation of the relationships between the generated gases and the incipient fault. Among these methods, just the IEC / revised IEC and Rogers / modified Rogers present a “normal case”, besides others diagnoses already provided by other methods; thus being the most usual. Recent researches verify the possibility of developing a computational model using artificial intelligence techniques, such as, FL and ANN to accomplish predictive diagnosis of fault in power transformers.

The Fuzzy Logic uses the degree an object belongs to a fuzzy set, i.e., membership function value between 0 and 1, to indicate a probable diagnosis established in the IEC and IEEE standards, which are broadly used by electric utilities. In addition, this technique makes it possible a non-labeled diagnosis to be approximate to one or more labeled diagnoses by the standards. Neural networks are assigned to diagnose fault in transformers because of their great abilities to learn how to classify complex problems [5], [6]. ANNs present relevant characteristics, such as: parallel distributed processing, learning and associative memory. The database for training, validating and testing the networks, as well as developing the fuzzy system was formed from the data of chromatography analyses of dissolved gases in insulating oil of transformers at CELPE distribution substations. II. DATABASES The databases are formed by the chromatography analysis, which provide the dissolved gas ratio in transformer insulating oil of CELPE, and are composed by 4905 patterns collected with more intensity starting from 1985. Although some diagnoses are not conclusive to determine the removal of a certain equipment from the power system, they are of extreme importance because they indicate the nature or type of the problem, i.e., if the paper (cellulose) is involved or not in the incipient fault, or if it is limited to the own insulating oil. Table II shows, according to IEC 599 standard, the total amount and the total percentage of data extracted from the chromatography databases of CELPE from 1985 up to 2008, as well as the total amount of non-labeled patterns. By the IEC method, the diagnoses are obtained starting from the ratio between the gases: C2H2/C2H4, CH4/H2, and C2H4/C2H6, whose values are distributed according to the range limits of gas ratio, as depicted in Table I. Combinations of gas ratio through their range limits are associated to the IEC code presented in first column of Table II, which are consequently associated to the faults. TABLE I SHARPLY

GAS RATIO ESTABLISHED BY THE IEC CODE CODE OF DIFFERENT GAS RATIO

DEFINED RANGE OF THE GAS RATIO

C2 H 2 C2 H 4

CH 4 H2

C2 H 4 C2 H 6

0,1 > R 0,1 ≤ R < 1,0 1,0 ≤ R < 3,0 R ≥ 3,0

0 1 1 2

1 0 2 2

0 0 1 2

TABLE II NUMBER OF PATTERNS PER FAULT TYPE ACCORDING TO THE IEC CODE NUMBER PERCENTAGE DIAGNOSES (FAULT TYPE) OF IEC CODE (%) PATTERNS

Normal deterioration of insulating oil Partial discharges of low energy Thermal fault (150°C < t < 300°C) Thermal fault (t > 150°C) Partial discharges of high energy Discharges of low energy Discharges of high energy Discharges of low energy Thermal fault ( 300°C < t < 700°C)

000 001 010 020 021 022 101 102 110 202

11.64

436

8.88

237

4.83

615

12.53

208

4.24

232 41 143

4.73 0.84 2.92 0.18

9

Thermal fault (t > 700°C)

Labeled Non-labeled Total

571

99

-

2590 2315 4905

2.11 52.80 47.19 -

III. NEURAL NETWORK ARCHITECTURE A. Database Arrangement The database is of fundamental importance to generate powerful classifying models, because their outputs are strongly related to the quality and arrangement of the information used in the learning process. Thus, to provide a better result by the neural network, the databases were preprocessed. Data preprocessing is an important approach to simplify the knowledge extraction, since effective decisions by the networks are based on the quality of data. B. Input Data Preprocessing The developed neural network was provided with input vectors corresponding to the three gas ratio according to the IEC code. The input data preprocessing provides the networks a fast and more efficient learning. Therefore, the neural network inputs are: X1= C2H2/ C2H4, X2=CH4/H2 and X3= C2H4/C2H6 that were normalized according to (1).

Xi =

X i − X i min X i max − X i min

,

(1)

where, Xi = X1 , X2 or X3 ; Ximin and Ximax are the minimum and the maximum values in all 4905 patterns in the data set (labeled or non-labeled) for each gas ratio Xi, respectively. In the data from CELPE, the minimum value were zero to all gas ratio; and the maximum were X1max = 81.6, X2max = 261.25 and X3max = 330.0. C. Training, Validation and Test Set The non-labeled samples, approximately 47% of the total, were eliminated from the original database. These samples will be later assessed by the fuzzy model, whose output will be compared with the expert’s report on dissolved gas analysis. The training, validation and test sets for the ANNs contain 100 samples of each diagnosis class. The classes with

inferior amounts to 100 samples were replicated. This procedure was made necessary to avoid the ANNs tendency for choosing the classes with larger amount of samples. Thus, the network data were composed by 1000 samples, where 50% were used for training, 25% for validating and the remaining for testing. The 1000 samples were selected in a random way to compose the three subsets: training, validation and test; in order to prevent the tendency associated with the order of the sample presented to the ANN. The training set was used to train by changing the value of connection weight of the processing units; the validation set was used to avoid the overtraining, thus improving generalization; and the test set was used to compare output of different models and to plot the test set error during the training process. It is worth pointing out that the test set, in spite of being composed by samples not used in the training set; they should have the same intrinsic characteristics. D. Network Architecture and Training All of the experiments accomplished in this work created ANNs with the MLP architecture, using LevenbergMaquardt (LM) training algorithm. The ANNs have an input, a hidden and an output layer. The maximum number of iterations for all of the trainings was set to 2500 epochs. The training stopped if the early stopping implemented by MATLAB® happened 15 times consecutively, or if the maximum number of epochs is reached, or if the error gradient reaches a minimum, or still if the error goal in the training set is met. In the first layer, the ANN has 3 input nodes corresponding to the gas ratio value: C2H2/ C2H4, CH4/H2, and C2H4/C2H6; in the output layer, it has 10 nodes corresponding to the 10 diagnoses showed in Table II. Therefore, the input vector of the ANN is defined by [C2H2/C2H4 CH4/H2 C2H4/C2H6]; and the output vector is of size 10, where we assume a 1-of-10 encoding for 10 classes using output values 0 and 1, and the winner-takes-all method. This method stats that the output with the highest activation designates the class. Table III shows the 1-of-10 encoding used by the ANN and the corresponding IEC code. TABLE III OUTPUT VECTOR OF THE ANN

To determine the number of nodes in the hidden layer, some authors propose determined heuristic rules to define an initial architecture to the ANN. One of these rules states that the hidden layer must be at least a number of nodes equal to the mean between numbers of input and output nodes. The architecture with 8 hidden nodes was analyzed, since the mean was 6.5; next, we analyzed the ANN’s performance with 9 and 10 nodes. The architecture with 9 hidden nodes reached the best performance, i.e., the minimum meansquared error – MSE in the validation set. We trained the ANNs using the Levenberg-Marquardt LM algorithm [7]. This algorithm is faster than other algorithms by a factor of from 10 to 100. The main drawback of the LM algorithm is that it requires the storage of some matrices which can be quite large for certain problems. However, in version 3.0 of MATLAB®, the Reduced Memory LM algorithm was introduced, which allows for a time/memory trade off. This means that the LM algorithm can now be used in much larger problems, with perhaps only a slight increase in running time. Figure 1 depicts the final architecture of the network with 3, 9, and 10 nodes in the input, hidden, and output layers, respectively.

Wjk Wij

9 8

C2H2 C2H4

CH4 H2

Σ

Falha10

Σ Σ

7

Σ

6

Σ

5

Σ

. . .

Σ

C2H2 C2H6

4

Camada de entrada

3

Σ

Falha2

2

Σ

Falha1

1

Σ

Camada de Saída

Camada intermediária

Fig. 1. MLP network architecture selected

IEC CODE

OUTPUT ANN CODE

000

1000000000

IV. FUZZY LOGIC

001

0100000000

010

0010000000

020

0001000000

021

0000100000

022

0000010000

101

0000001000

102

0000000100

110

0000000010

202

0000000001

The Fuzzy Logic has been studied in this kind of problem by other researchers [8], [9] and showed to be very powerful. The Fuzzy Logic uses the degree to which an element belongs to a fuzzy set in order to give a probable standard diagnosis. This technique allows the results of gas ratio to approximate to a diagnosis whose limits of gas concentration are not labeled by the IEC standard. Due to the limited amount of some sample per class (gas ratio) and the unnecessary training, the Fuzzy Logic – FL was chosen to be implemented in a system for diagnosing of incipient faults in power transformers.

A considerable advantage of the FL is that its input data do not need to be huge and preprocessing. The database used in its development comprised information presented in the chromatography analysis. A. Fuzzy Model In a programming language, named MATLAB®, we developed a script to implement the fuzzy logic, in which it is possible, from the gas ratio presented in the chromatography analysis of the insulating oil to provide a diagnosis compatible with the IEC code. In the case where the range of the gas ratio is not labeled by the IEC code, the Fuzzy Inference System – FIS is able to predict multiple incipient faults, which indicates the tendency of the diagnosis in the power transformer. This diagnosis is compared with future analyses until the Diagnosis Gas Analysis – DGA passes again to be labeled, validating the tendency of the diagnosis predicted by the FIS. The limit set for each of gas ratios - C2H2/C2H4, CH4/H2, and C2H4/C2H6 - was represented by membership functions named ZERO, ONE, and TWO, respectively. These functions are curves that define how each point in the input space is mapped to a membership value (or degree of membership) between 0 and 1. Therefore, the fuzzification of the gas ratio in these membership functions assigns a membership value to each gas ratio value, turning the output of the FIS as closer as possible to the IEC code. B. Membership Function Customization Here we customize the membership functions so that the fuzzy system best models the data. Three membership functions, one for each gas ratio, were heuristically parameterized to suit us from the point of view of simplicity, convenience, and efficiency, i.e., the shape of the curve should satisfy the limits fixed by the IEC code. In the three gas ratios, we used for the lower limit the Zmembership function which has a maximum value in the beginning of the curve and decreases monotonically to a minimum in the end. The lower limit of the C2H2/C2H4 and CH4/H2 gas ratios was represented by the S – membership function and the C2H4/C2H6 gas ratio by the Sig membership function. Both S - and Sig - membership functions have a minimum value in the beginning of the curve and reach a maximum in the end. These functions were chosen because their characteristics are near to the superior limit set in the IEC code, permitting a maximum value after a specified point. In relation to the central limits, the best membership functions to C2H2/C2H4, CH4/H2 and C2H4/C2H6 gas ratios were Pi, Dsig and Gbell, respectively. These functions have maximum values in the beginning and end of the curve, and a minimum in the middle. The three membership functions chosen to represent the C2H4/C2H6 gas ratio are shown in Fig. 2.

ZERO

ONE

TWO

Zmf Gbellmf

Smf

Fig. 2. Membership functions for C2H4/C2H6 gas ratio.

While the IEC code (Table I) sets crisp values (0, 1 and 2) to the gas ratio, in the fuzzy logic approach, the gas ratio is represented as a number between 0 and 1, which indicates the probability of an example to belong to a membership function (ZERO, ONE or TWO). These probabilities are used to indicate the fault with the highest probability that could be occurring in a transformer, as shown in the following subsection. C. Fuzzy Output Set The membership functions to each probable fault state the probability of an incipient fault in the transformer, as follows: F(0) = R1_Zero + R2_Zero + R3_Zero; F(1) = R1_Zero + R2_Zero + R3_One; F(2) = R1_Zero + R2_One + R3_Zero; F(3) = R1_Zero + R2_Two + R3_Zero; F(4) = R1_Zero + R2_Two + R3_One; F(5) = R1_Zero + R2_Two + R3_Two; F(6) = maximum ((R1_One + R2_Zero + R3_One) or (R1_One + R2_Zero + R3_Two) or (R1_Two + R2_Zero + R3_One) or (R1_Two + R2_Zero + R3_Two)); F(7) = R1_One + R2_Zero + R3_Two; F(8) = R1_One + R2_One + R3_Zero; where R1, R2, and R3 denote the C2H2/C2H4, CH4/H2, and C2H4/C2H6 gas ratios, respectively; Rn_Mem is the fuzzification of the Rn gas ratio in the membership function Zero, One, and Two. For example, R1_Zero is the fuzzification of R1 gas ratio in the membership function Zero. These functions represent a value between 0 and 3 that corresponds to a fault type labeled by the IEC code. The output of the system is a vector of size 9 that corresponds to the 9 functions defined previously. The final diagnosis is represented by the function or functions of maximum fuzzy output value. In case of two functions present equal maximum value, the diagnosis will be represented by both output functions.

V. RESULTS

A. Artificial Neural Network Approach From the preprocessed data, we realized some simulations with different numbers of hidden nodes. We assessed 10 ANNs by changing initial weights. Table IV shows the classification percentage error on the test set varying the number of hidden nodes. The test set has 250 patterns composed by 25 patterns from each class. The ANN_1 with 9 hidden nodes was selected, thus it has presented the minimum mean-squared error on the validation set, besides has showed the best performance on the test set. TABLE IV CLASSIFICATION ERROR PERCENTAGE ON THE TEST SET 8 HIDDEN 9 HIDDEN 10 HIDDEN NODES NODES NODES ANN_1 22.4 2.0 24.0 ANN_2 4.4 15.6 25.6 ANN_3 5.6 12.8 22.0 ANN_4 24.0 14.0 13.2 ANN_5 4.4 4.8 6.4 ANN_6 5.2 25.2 6.0 ANN_7 18.4 13.2 33.2 ANN_8 4.4 13.6 25.2 ANN_9 22.8 15.2 15.6 ANN_10 11.2 3.6 12.8 MEAN 12.28 12.0 18.4

The results of the total error per class of the ANN_1 are presented in Table V. In this table, we can observe that the maximum classification error per class is related to the IEC code – 001, which corresponds to 8% of the total number of patterns in the test set. The total percentage of correct classification is 98%, whose value is in the range of good performance showed by certain papers that approach the same issue using ANNs [7] and [10]. TABLE V PERCENTAGE OF CORRECT CLASSIFICATION PER DIAGNOSIS BY THE ANN TOTAL PERCENTAGE IEC CODE NUMBERS OF DIAGNOSIS ERROR PATTERN

Normal

000

571

11.0 %

Thermal fault (t > 150°C)

001

436

6.9 %

Partial discharges of low energy

010

237

0%

Thermal fault ( 300°C < t < 700°C)

020

815

6.5%

Thermal fault (150°C < t < 300°C)

021

208

0.5%

Thermal (t>700°C)

022

232

3.9%

101

140

3.6%

102

143

4.9%

110

9

0%

fault

Discharges energy

of

low

Discharges energy

of

high

Partial discharges of high energy

B. Fuzzy Logic Approach From the gas ratio contained in the chromatography analysis of the insulating oil in the power transformer, the fuzzy logic yields a diagnosis compactable with the IEC code. Diagnoses, whose limits of gas concentration are nonlabeled by the IEC standard, are approximated to one or more labeled diagnosis, since simultaneous faults may occur in the transformer. Table VI depicts the amount of error and the percentage error on the total data yielded by the Fuzzy Inference System based on the IEC code. As we can see, the system was able to classify all the data correctly. A significant advantage of this system is related to its capacity for predicting incipient faults that are non-labeled by the IEC code with high accuracy. Thus, the FIS can help the maintenance staff in decision-making, even when the gas ratios are above or under the limits fixed by the IEC code, which in this case does not have any information about the stat of the transformer. TABLE VI PERCENTAGE OF CORRECT CLASSIFICATION PER DIAGNOSIS BY THE FIS TOTAL IEC PERCENTAGE NUMBERS OF DIAGNOSIS CODE ERROR PATTERN

Normal

000

571

0%

Thermal fault (t > 150°C)

001

436

0%

Partial discharges of low energy

010

237

0%

Thermal fault ( 300°C < t < 700°C)

020

815

0%

Thermal fault (150°C < t < 300°C)

021

208

0%

Thermal fault (t > 700°C)

022

232

0%

Discharges of low energy

101

140

0%

Discharges of high energy

102

143

0%

Partial discharges of high energy

110

9

0%

Table VII shows some historical diagnoses provided by the FIS to a typical case of a power transformer located in Jussaral/PE substation. It’s important to point out that in case where the IEC code gives no information about the stat of the transformer, the FIS indicates the fault with high probability, which later on is confirmed by the IEC code.

TABLE VII DIAGNOSIS BY THE IEC AND THE FIS DATE

RESULT IEC

RESULT FIS

1/9/1987

Thermal fault (150°C < t < 300°C)

Thermal fault (150°C < t < 300°C)

1/12/1988

Thermal fault (150°C < t < 300°C)

Thermal fault (150°C < t < 300°C)

20/11/1991

Non-Labeled

Thermal fault (150°C < t < 300°C)

21/1/1993

Non-Labeled

Thermal fault (150°C < t < 300°C)

25/5/1994

Non-Labeled

Partial discharges of low energy

8/7/1994

Non-Labeled

Thermal fault (150°C < t < 300°C)

31/10/1994

Non-Labeled

Thermal fault (150°C < t < 300°C)

29/10/1995

Thermal fault (t > 700°C)

Thermal fault (t > 700°C)

17/1/1996

Non-Labeled

Thermal fault (150°C < t < 300°C)

18/9/1996

Non-Labeled

Thermal fault (t > 700°C)

4/9/1997

Non-Labeled

Thermal fault (150°C < t < 300°C)

23/9/1998

Non-Labeled

Thermal fault (150°C < t < 300°C)

9/7/1999

Non-Labeled

Thermal fault (t > 700°C)

10/7/2000

Non-Labeled

Thermal fault (150°C < t < 300°C)

18/5/2001

Thermal fault (t > 700°C)

Thermal fault (t > 700°C)

1/6/2001

Thermal fault (t > 700°C)

Thermal fault (t > 700°C)

12/6/2001

Thermal fault (t > 700°C)

Thermal fault (t > 700°C)

31/10/2001

Thermal fault (t > 700°C)

Thermal fault (t > 700°C)

In spite of being very useful, the IEC code is not efficient in all case, i.e., transitions between low and high thermal fault make the IEC code not sure about its decision as we can see in Table VII in 31/10/1994 (Non-Labeled), 29/10/1995 (Thermal Fault), and 17/1/1996 (Non-Labeled). On the other hand, the FIS indicates a Thermal Fault of high or low temperature during all that time period. VI. CONCLUSION The results showed that either the ANN approach or the Fuzzy Inference System to diagnose incipient fault in a power transformer were well suited. The ANN showed low percentage error of classification 2% in the test set, while the FIS showed no error. Besides showing high performance, the FIS is able to predict incipient fault even when the fault by the IEC code is non-labeled, which is a significant advantage among others

systems that approach this same problem. In addition, the FIS developed in this work has a simple structure, whose membership functions were parameterized heuristically to suit the IEC code. In order to improve the ANN’s results and give its outputs more consistency, since they are varying considerably with the initial weights, we are going to search for a more appropriate number of hidden nodes in a higher interval, using the k-fold cross validation as showed in [11], [12]. VII. REFERENCES [1]

IEC – International Electrotechnical Commission. “Interpretation the analysis of gases in transformer and other Oil-filled impregnated Electrical Equipment in Service,” CEI – IEC – 599, First Edition, 1978. [2] IEEE – Institute of Electrical and Electronics Engineers – “Guide for the detection and determination of generated gases in oil-immersed transformers and their relation to the serviceability of the equipment,” ANSI/IEEE C57.104-1978. [3] M. Duval; 2003. “New Techniques for Dissolved Gas-in-Oil Analysis,” IEEE Electrical, Insulation Magazine, vol. 19, nº 2, pp. 615. [4] E. Dornenburg and W.Strittmatter, “Monitoring Oil Cooled Transformers by Gas Analysis,” Brown Boveri Review, vol. 61, no 5, pp. 238-247, May 1974. [5] S. Haykin, Neural Networks: A Comprehensive Foundation. Prentice Hall: NJ, 2nd ed., 1998. [6] T. B. Ludermir, A. P. Braga and A. C. P. L. F. Carvalho, Redes Neurais Artificiais: Teoria e Aplicações. Editora LTC, Rio de JaneiroRJ, Brasil, 2000. [7] M.T. Hagan and M. Menhaj, “Training Feedforward Networks with the Marquardt Algorithm,” IEEE Transactions on Neural Networks, vol. 5, no. 6, 1994. [8] C. E. Lin, J. M. Ling, and C. L. Huang, “An Expert System For Transformer Faults Diagnosis And Maintenance Using Dissolved Gas Analysis,” IEEE Trans. On Power Delivery, vol. 8, no.1, pp. 231-238 , January 1993. [9] Q. Su, C. Mi., L. L. Lai and P. Austin, “A Fuzzy Dissolved Gas Analysis Method for Diagnosis of Multiple Incipient Faults in a Transformer,” IEEE Trans. On Power Systems, vol. 15, no.2, pp. 593598, May 2000. [10] J.L.Guardado, J.L. Naredo, P. Moreno and C. R. Fuerte, “A Comparative Study of Neural Network Efficiency in Power Transformers Diagnosis Using Dissolved Gas Analysis,” IEEE Transactions on Power delivery, October, 2001. [11] M. M. S. Lira, R. R. B Aquino, A. A Ferreira, M. A. Carvalho Jr, and C.A.B.O. Lira, “Improving Disturbance Classification by Combining Multiple Artificial Neural Networks,” in IEEE World Congress on Computational Intelligent / IJCNN 2006, 2006, Vancouver, BC, Canada. IEEE Xplore, 2006. p. 3436-3442. [12] R. R. B. Aquino, A. A. Ferreira, M. M. S. Lira, G. B. Silva, O. Nóbrega Neto, J. B. Oliveira, C. F. D. Diniz, and J. Fidelis, “A Hybrid Intelligent System for Short and Mid-term Forecasting for the CELPE Distribution Utility,” in IEEE World Congress on Computational Intelligent / IJCNN 2006, 2006, Vancouver, BC, Canada. IEEE Xplore, 2006. p. 2556-2661.

Suggest Documents