1

Determination of Transformer Health Condition Using Artificial Neural Networks

Ahmed E. B. Abu-Elanien, Student Member, IEEE, M. M. A. Salama, Fellow, IEEE, and Malak Ibrahim Department of Electrical and Computer Engineering

University of Waterloo Waterloo, ON, Canada

Abstract—This paper presents a method to estimate a transformer health condition based on diagnostic tests. A feed forward artificial neural network (FFANN) is used to find the health index of the transformer. The health index is used to find the health condition of the transformer. The training of the FFANN is done using real measurements of 59 working transformers. The testing of the trained neural network performance is done using real data for 29 working transformers. The performance evaluation of the trained FFANN shows that the trained neural network is reliable in finding the health condition of any working transformer.

Keywords-component; Asset management; condition monitoring; dissolved gas analysis ; health condition; health index; furans analysis; transformer.

I. INTRODUCTION In today's competitive energy market, power utilities wish

to have sustainable operation of all assets with minimum cost of the existing equipment over their life span. The current trend is to utilize existing equipment at ever-higher capacity levels in order to defer the capital investment in new facilities or in refurbishment of the existing facilities [1].

Power transformers have the single highest value of the equipment installed in high-voltage substations, comprising up to 60% of total investment [2]. They also occupy large portion of the investments in the distribution level. Deployment of new transformers is costly and may not be justified in terms of the utility bottom line. [1]. Most power utilities have a growing demand for extending the lifetime of their fleets of transformers by improving their financial and technical performances [2]. One important aspect that allows power utilities to enhance the financial and technical performances of power transformer is the knowledge of transformer health condition. Health condition is important to improve reliability and assist in determining appropriate asset management decisions benefits, enable operators to develop effective maintenance and replacement strategies based on the condition of the transformer, identify transformers that could benefit from life extension measures [3].

The purpose of asset condition assessment is to detect and quantify long-term degradation and to provide a means of quantifying remaining asset life [4]. Assets that are at or near end-of-life or at high risk of failure can be identified to save

require major capital expenditures to either refurbish or replace the assets.

Insufficient research work is presented to evaluate the health condition of a transformer. In [2, 5], clear steps are given to calculated the health index of a power transformer. However, the methods presented in [2, 5] ignored important measurement in assessing the health index, which is the amount of total solids in the transformer oil. Amount of total solids is a good indication about the condition of the insulation system of the transformer. Further, they give the measurements of total furans less weight than other less important factors such as the winding resistance and oil quality during the calculation of the health index. It is known that the amount of furans is an important indication to the degree of polymerization of the solid insulation. The degradation of solid insulation (paper) can be considered the primary reason for a transformer end of life [6-10]. The work done in [1] includes all parameters affect the health of the transformer; however, the paper did not reveal the method used in the calculation of the health index. The total furan in the transformer oil is ignored in the method presented in [11]. Further, no case studies are given to illustrate the method.

The main aim of this paper is to use the artificial neural network to evaluate the health index of a transformer using real measurements from working transformers. The calculated health index is used to evaluate the health condition of the transformer. In order to find the health index, the author used real data from the field. The diagnostic tests data are taken from distribution systems of an industrial installation in the Middle East. This company commissioned a specialized in Asset Management Health Assessment (AMHA) at UK to derive and populate a health index for the distribution transformers located on their electric system. To facilitate this, AMHA undertook a sampling and test program for 90 oil filled transformers. The provided data are the water content, acidity, break down voltage (BDV), hydrogen content (H2), methane content (CH4), ethylene content (C2H4), Acetylene content (C2H2), furans content, loss factor, and total solids in oil for each transformer. The health indices of all transformers calculated by AMHA are also given.

Measuring moisture, acidity, solid contamination and breakdown strength of the oil gives a good indication of the overall condition of the oil and internal components. The quality of the oil is also critical in preventing premature ageing

978-1-61284-922-5/11/$26.00 ©2011 Crown

2

of the transformer and extending service life. Moreover, Furfuraldehyde (furans) analysis gives a highly accurate indication of the condition of paper insulation. The furans content is correlated to the degree of polymerisation of the paper [12, 13]. When furans levels reach specific values, we know that the insulation has effectively broken down and the probability of failure is very high. Furthermore, analysing the levels of different dissolved gases in the oil identifies electrical discharge, arcing and thermal activity within the transformer.

The method of calculating the heath index is considered as a secret and it is not provided in the final report presented to the industrial installation. The values of the AMHA health index ranges between zero and ten. A value of zero of the health index means a brand new transformer. A value of ten of the health index means the transformer should be replaced. In between zero and ten of the health index, the condition of the transformer can be good, moderate, or bad. The condition of the transformer is considered to be good for health indices from zero to 4, the condition of the transformer is considered to be moderate for health indices from 4 to 7, and the condition of the transformer is considered to be bad for health indices from 7 to 10. In this paper FFANN is used to find the relation between the 11 data categories and the health index of each transformer with the aid of the health indices given by AMHA.

II. METHOD CONCEPTS Artificial intelligence is the study of systems that act in a

way that to any observer would appear to be intelligent [14]. One of the most important artificial intelligence branches is the Artificial Neural Network (ANN). ANN is an information processing paradigm that is inspired by the biological nervous systems in human brain [14]. The key element of this paradigm is the novel structure of the information processing system. ANN has a remarkable ability to derive meaning from complicated or imprecise data. They can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer techniques. A trained neural network can be thought of as an "expert" in the category of information it has been given to analyze. This expert can then be used to provide projections given new situations of interest and answer "what if" questions.

The artificial neural network (ANN) is used to estimate the health index of a transformer. The available data for the 88 transformers are used to train a feed forward artificial neural network (FFANN) to be able to find the health index of a transformer.

The given data should be preprocessed first before using them. The main aim here is to adjust the given data to be ready for analysis. The elements that contain irregular data should be removed. These irregular data include too high attributes, negative or corrupted attributes. Further, all data are normalized according to their respective maximum. In the regard of irregular data, elements 76 and 31 of the available data seem to be irregular and should be eliminated. This is because the amount of total furans of element 76 is too high

Fig. 1. FFANN configuration.

and the loss angle of element 31 is too high. These very high inputs may affect the remaining values of the same category when we perform the normalization. The next step in the pre-processing stage is the normalization of all data categories by dividing all elements of each data category by their respective maximum. The given health indices are normalized too by dividing all entries of health indices by their maximum.

The available measurements for the 88 transformers are divided into two sets: training set and testing set. The training set consists of the data for 59 transformers selected randomly from the available data for the 88 transformers (67% of the available data). The testing set consists of the data for the remaining 29 transformers (33% of the available data).

A four-layer FFANN is used to estimate the health index of the transformers. The FFANN is shown in Fig. 1. The neural network consists of one input layer, one output layer, and two hidden layers. The input layer consists of 11 neurons; the inputs to the input layer are the data of the 88 transformers. The output layer consists of one neuron representing the health index for the transformer under study. The output layer neuron has output in the range of zero to one: zero represents a brand new transformer and one represents a transformer in a very bad condition and it should be replaced. With respect to the hidden layers, it is customary that the selection of the number of neurons in the hidden layers is selected by trial and error. The authors used the same approach in the proposed algorithm. The number of hidden layers is found to be two with four neurons in the first hidden layer and two neurons in the second hidden layer.

A comparison between the given health index for the 59 transformers used in the training of FFANN and the FFANN output health index for the training set is shown in Fig. 2. It is clear from Fig. 2 that the output of the FFANN is very close to the original health index of the training set of transformers. The trained FFANN is tested using the testing set which includes 29 transformers. The given health index of the testing set and the output of the FFANN for the testing set are shown

3

0 10 20 30 40 50 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Transformer number

Healt

h ind

ex

FFANN output health index for the training set

Original health index for the training set

Fig. 2. The given health index for the 59 transformers used in the training of FFANN and the FFANN output health index for the same training set.

0 5 10 15 20 25 300.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Transformer number

Healt

h ind

ex Original health index for the testing set

FFANN output health index for the testing set

Fig. 3. The given health index for the 29 transformers used in the testing of FFANN and the FFANN output health index for the same testing set.

in Fig. 3. Fig. 3 shows that the estimated health indices for the testing set are numerically close to the given health indices of the same set. Table I shows the given health indices for the testing set and the corresponding output of the FFANN.

III. DISCUSSION The condition of the transformer is the main goal of the

investigation. In this regard, the trained FFANN can classify the conditions of the transformers with excellent accuracy. The estimated and target health indices of the testing set of transformers shown in Table II are converted into health conditions (good, moderate, or bad) according to the aforementioned criteria. Table III depicts the accuracy of the FFANN classifier based on the health condition criteria.

The FFANN classifier is able to classify correctly 28 cases of the 29 test cases with accuracy of 96.55%. Moreover, with a quick look at Table I, we can found that the incorrectly classified case is case 5. The calculated and given health indices of case 5 are 0.45 and 0.334 respectively. The given

health index indicates that the transformer is at the end of the good condition zone and the calculated health index indicates that the transformer is in the beginning of the moderate condition. This means that the two outputs are not far from each other. Analyzing the transformers’ data with bad conditions, which are the most important ones, we can find that the furans content of all of them is more than 5 ppm. The furans content of five of them are more than 6.5 ppm, which means that the degree of polymerization is less than 200 units [12]. Moreover, the majority of the transformers with bad conditions have high total solids contamination and acidity level, and medium-high water content in the oil samples. Furthermore, the loss factor of all of them is higher than 0.15%, which is considered above the safe limit. This data analysis proves that the evaluation of the transformers 3, 8, 20, 21, 22, 23, and 24 as transformers in bad condition was correct.

4

Table I Given health indices for the testing set and the corresponding output of the FFANN

Transformer number

Given health index

FFANN output health index

Transformer number

Given health index

FFANN output health index

1 0.226 0.224 16 0.226 0.199 2 0.322 0.261 17 0.387 0.259 3 0.772 0.813 18 0.29 0.269 4 0.377 0.398 19 0.316 0.337 5 0.334 0.45 20 1 0.94 6 0.29 0.19 21 0.931 0.931 7 0.29 0.367 22 1 0.939 8 0.701 0.81 23 0.916 0.937 9 0.381 0.129 24 0.732 0.8156 10 0.102 0.323 25 0.354 0.271 11 0.274 0.189 26 0.45 0.535 12 0.316 0.245 27 0.414 0.432 13 0.29 0.264 28 0.291 0.339 14 0.316 0.241 29 0.414 0.4467 15 0.29 0.214

Table II The output transformer condition of the FFANN and the target conditions of the transformer testing set.

Transformer number Given health condition FFANN output

health condition Transformer

number Given

health condition FFANN output health condition

1 good good 16 good good 2 good good 17 good good 3 bad bad 18 good good 4 good good 19 good good 5 good moderate 20 bad bad 6 good good 21 bad bad 7 good good 22 bad bad 8 bad bad 23 bad bad 9 good good 24 bad bad 10 good good 25 good good 11 good good 26 moderate moderate 12 good good 27 moderate moderate 13 good good 28 good good 14 good good 29 moderate moderate 15 good good

Table III Accuracy of the FFANN classifier with respect to the

transformer condition good moderate bad total

good 18 1 0 19

moderate 0 3 0 3

bad 0 0 7 7

The trained FFANN can be used to find the health index of

any transformer. In order to find the health index of any transformer, the 11 aforementioned measurements for any transformer are fed into the trained FFANN after normalization using the same standards. The output of the FFANN is the health index of the transformer under study. The health condition of the transformer can be evaluated using the health index by applying the conversion rules stated in the introduction.

IV. CONCLUSION The usage of artificial neural network in the evaluation of a

transformer health index is presented in this paper. Real

5

Measurements of 90 working transformers are used for the purpose. The provided measurements include water content, acidity, break down voltage, hydrogen content, methane content, ethylene content, Acetylene content, furans content, loss factor, and total solids in oil for each transformer. The given health index evaluated by AMHA is used to train the FFANN. The health index calculated by the neural network is used to evaluate the health condition of 33% testing set of the available data. The evaluated health conditions of the testing set show similar results as the results of AMHA except for one transformer.

REFERENCES [1] N. Dominelli, “Equipment health rating of power transformers,” in Proc.

2004 IEEE International Symposium on Electrical Insulation, 19-22 Sept. 2004.

[2] A. Jahromi, R. Piercy, S. Cress, J. Service and W. Fan, "An approach to power transformer asset management using health index," IEEE Electrical Insulation Magazine, vol. 25, pp. 20-34, 2009.

[3] D. Hughes, G. Dennis, J. Walker and C. and Williamson, "Condition based risk management (CBRM)- enabling asset condition information to be central to corporate decision making," in Proc. of the First World Congress on Engineering Asset Management (WCEAM), Gold coast, queensland, Australia,11-14 July, 2006.

[4] T. Hjartarson and S. Otal, "Predicting future asset condition based on current health index and maintenance level," in proc. IEEE 11th

International Conference on Transmission & Distribution Construction, Operation and Live-Line Maintenance, ESMO, 2006.

[5] A. Naderian, S. Cress, R. Piercy, F. Wang and J. Service, "An approach to determine the health index of power transformers," in proc. Conference Record of the 2008 IEEE International Symposium on Electrical Insulation, ISEI , 2008, pp. 192-196.

[6] H. Herman, M. J. Shenton, G. C. Stevens and R. J. Heywood, "A new approach to condition assessment and lifetime prediction of paper and oil used as transformer insulation," in proc. IEEE 7th International Conference on Solid Dielectrics, ICSD '01., 2001, pp. 473-476.

[7] A. J. Kachler and I. Hohlein, "Aging of cellulose at transformer service temperatures. Part 1: Influence of type of oil and air on the degree of polymerization of pressboard, dissolved gases, and furanic compounds in oil," IEEE Electrical Insulation Magazine, vol. 21, pp. 15-21, 2005.

[8] M. K. Pradhan and T. S. Ramu, "Criteria for estimation of end of life of power and station transformers in service," in proc. Annual Report Conference on Electrical Insulation and Dielectric Phenomena, CEIDP '04., 2004, pp. 220-223.

[9] J. P. Van Bolhuis, E. Gulski and J. J. Smit, "Monitoring and diagnostic of transformer solid insulation," IEEE Transactions on Power Delivery, vol. 17, pp. 528-536, 2002.

[10] A. M. Emsley, X. Xiao, R. J. Heywood and M. Ali, "Degradation of cellulosic insulation in power transformers. Part 3: effects of oxygen and water on ageing in oil," IEE Proceedings Science, Measurement and Technology, vol. 147, pp. 115-119, 2000.

[11] S. Miletic, "Asset condition assessment by health index benchmarking," in proc. 20th International Conference on Electrical Distribution, CIRED, Prague, Czech Republic, 2009 .

[12] A. E. B. Abu-Elanien and M. M. A. Salama, "Asset management techniques for transformers," Electr. Power Syst. Res., vol. 80, pp. 456-464, 2010.

[13] I. Chendong, "Monitoring paper insulation aging by measuring Furfural Contents in oil," in proc. 7Th Int. Symp. on High Volt. Eng., 1991 .

[14] B. Coppin, Artificial Intelligence Illuminated. Sudbury, Massachusetts, USA: Jones and Bartlett Publishers, 2004.