2023 4th International Conference on High Voltage Engineering and Power Systems (ICHVEPS)

Prediction Health Index Using Machine Learning

and Its Correlation with Transformer Age on
Historical Data
2023 4th International Conference on High Voltage Engineering and Power Systems (ICHVEPS) | 979-8-3503-1867-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICHVEPS58902.2023.10257560

Gemelfour Ardiatus Sudrajad Suwarno Gemelthree Ardiatus Subekti

School of Electrical Engineering and School of Electrical Engineering and Department of Accountancy
Informatics Informatics Politeknik Negeri Madiun
Bandung Institute of Technology Bandung Institute of Technology Madiun, Indonesia
Bandung, Indonesia Bandung, Indonesia [email protected]
[email protected] [email protected]

l I

Abstract— The transformer is a vital of equipment in the which can result in social and economic losses to electricity
power transmission and distribution system. Optimum use of service providers and users of electricity services [4],[5].
transformers can increase the economic value and reliability of Some transformer failures can also be detrimental to the
electricity supply. Power apparatus is expected to operate close power system and the environment. Because transformer
to or exceed the design life of the transformer. Health index
failure can also be accompanied by fire to explode.
(HI) is used to provide a general assessment of the condition of
the transformer. In determining the health index, it can be In addition, the initial failure that occurs repeatedly in a
done using conventional methods or non-conventional long time can result in premature aging of the transformer.
methods. According to the IEEE loading guide lines, the age of Interference from heat, oxygen, and water can result in
the transformer ranges from 25-30 years. The "effective age" chemical reactions of pyrolysis, hydrolysis, and oxidation
of the transformer can be determined by the probability of which damage the chemical structure of the insulation
failure of insulation system or which can be based on HI. In system [6]. Aging of paper insulation accompanied by a
this study, transformer oil testing was used to determine the decrease in mechanical strength is an important factor in the
health index and calculate the age of the transformer. The operating life of a transformer [6],[7]. Aging of paper
parameters used are BDV, Water, Acid, IFT, DGA, and 2-
insulation will cause several aging products such as CO,
furaldhyde. In this study, 79 transformer oil sample data were
used which had known health indexes from field assessments. CO2, water, and furan compounds dissolved in transformer
The health index method uses calculation, ANN, and naïve oil [8]. The weakening of paper insulation can be seen from
Bayes methods. Parameters affecting the health index BDV (r the degree of polymerization (DP) and tensile strength. The
= 0.57), water (r = -0.55), 2FAL (r = -0.58), and acetylene (r = - "effective age" of the transformer can also be seen from the
0.25). Parameters that affect the life of the transformer are relationship between the probability of failure based on the
BDV (r = 0.45), water (r = -0.66), Acid (r = -0.53), IFT (r = - Health Index and the failure rate of failures that have been
0.49), 2FAL (r = - 0.58), and CO2 (r = -0.40). Parameters BDV, detected before [9].
water, and 2 FAL have a significant correlation in determining Health Index (HI) is a method that combines field
the health index and predicting the life of the transformer.
inspections, monitoring operations, field and laboratory
Parameters 2FAL and water are related to chemical reactions
in the insulation system which result in aging of the testing into quantitative and objective values to provide an
transformer. With the deteriorating condition of the assessment of overall asset condition[10],[11]. In addition to
transformer (HI), it further increases the possibility of initial using all field and testing data, the health index can be
disturbance to failure of the transformer. represented by the results of oil testing. Testing the quality
of oil, dissolved gas, and furan can provide an overview of
Keywords— Power transformer, Health Index, Expected life the condition of the transformer. The results of laboratory
transformer, and Machine Learning tests will be given an assessment and a weighting factor, in
I. INTRODUCTION order to quantify the value of all parameters. [12],[13]. The
results of the health index can be used as recommendations
Transformers are vital equipment in electric power for asset managers in assessing and prioritizing maintenance
transmission and distribution systems. Transformers are and repair.
expected to operate for a long time by ensuring their Apart from using the conventional method (scoring and
reliability and performance in providing a reliable electrical weighting factor), the calculation of the health index can use
system [1],[2]. Based on IEEE design standards, the lifespan the machine learning method. Machine learning learns from
of transformers is estimated to be 25-30 years [3]. The training data to look for relationships between variables that
transformer will experience initial disturbances as the determine the health index. The relationship between these
operating period of the transformer in the electrical system. variables is used to predict the new health index data.
Disturbances can take the form of electrical stress, thermal Several machine learning algorithms that can be used to
stress, mechanical stress, and chemical stress. Some predict the health index are Artificial Neural networks,
disturbances can be fatal to the condition of the transformer Decision trees, naïve bayes, support vector machines,
until the failure of the transformer operation. These random forests, and other algorithms. In this study, artificial
disturbances can cause power outages in certain areas,

979-8-3503-1867-8/23/$31.00 ©2023 IEEE 188

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.
 neural network and naïve Bayes algorithms were used. B. Health Index calculation
Neural networks imitate the workings of the human brain
and the relationship of each neuron in producing actions.
Connections between neurons can differ according to the
strength of the connections in these neurons and their
variable function. The result of the sum of the function
results and the error value will be the output of the neuron
[14]. Naïve Bayes is a statistical method that applies
probabilistic relationships to all data and makes calculations
of the relationships between data [15]. If the work of the
neural network is to find the value of the weighting factor
(to determine the strength between neurons) randomly to
optimal. While naïve Bayes uses the probability method to
look for relationships between variables. The advantage of
machine learning is that it can find relationships between
dependent and independent variables.

II. METHODOLOGY
In this study, 79 data from field inspections were used to
find the health index value and estimate the age of the Fig. 1. Block diagram of health index calculation
transformer. Then look for the correlation between variables
in determining the age of the transformer and the health
The health index calculation is used to quantify a number
index. In addition to using the (conventional) calculation
of test parameters into a value that can generalize asset
method to determine the health index, this study also uses a conditions. The calculation of the health index depends on
machine learning method with 2 algorithms, namely the assessment of the test parameters and the weighting
artificial neural network and naïve Bayes. In this study, factor of the test parameters [17]. The parameters used are
1328 transformer data were used for machine learning transformer oil testing parameters such as breakdown voltage
model training. The machine learning model will be used to (BDV), water, acidity (TAN), interfacial tension (IFT),
predict field data whose health index is known. dissolve gas (DGA), and furanic compounds. To calculate HI
with the non-historical method 12 input parameters are used
and the historical method uses 26 input parameters and one
A. Transformer Remaining Life Estimation
output parameter (HI). Calculation of the health index can be
Calculation of the remaining life of the transformer seen in the block diagram above. Scoring parameters use
using paper insulation condition parameters. The several standards such as IEC 60422, IEEE C57.104 – 2008,
degradation of paper insulation results in a higher IEEE C57.104 – 2019, and Limitations from SDMyers.
probability of transformer failure, thereby reducing the life Meanwhile, the weighting factor serves to determine the
expectancy of the transformer. Unlike transformer oil which urgency of each parameter in determining the value of the
can be replaced, paper insulation is a material that is highly health index. The value of the scoring and weighting factor
unlikely to be replaced [16]. Even if a replacement is made, of each parameter can be seen in table 1 and table 2 below.
it can reduce the performance of the transformer and the The health index calculation formula is as in the formula
replacement price is very expensive (almost the same as below.
replacing a transformer). Paper insulation consists of long
(6)
chains of cellulose, which when aging or disturbance occurs
will cause the cellulose chains to decay. The breakdown of
the cellulose chain was detected by the appearance of furan TABLE I. WEIGHTING FACTOR OF EACH PARAMETER
compounds (specifically 2furaldhyde) dissolved in the Parameter Weight Factor
transformer oil. 2furaldhyde (2FAL) compounds can be BDV 3
used to calculate the Degree of Polymerization (DP), which
Water 4
is a number that determines the strength of the cellulose Oil Quality 8
chains in paper insulation. The DP calculation can be used Acid 1
as a benchmark to determine the estimated remaining life of IFT 2
the transformer.
Chengdong Equation 2FAL 5
Dp = -285.7 x Log10 (2FAL x 0.88) + 1288.6 (1) DGA 10
SDMyers Equation
Dp = -343.8 x Log10 (total furan) + 1387.5 (2)
% life remaining = 166.1xLog10(DP)-382.2 (3) Heath index calculations are performed to create machine
Elapsed life (years) = 20.5*ln(1100/DP) (4) learning training data. 1328 data are used as training data
Remaining life (years) = 35 – Elapsed life (5) with details of 992 data having historical testing and the rest
The formula used is from several studies by Chen Dong, do not have historical data. In this study, 79 field data were
SDMyers, and Pradhan.

189

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.
 used to check the performance of the health index from Cigre 741. The results of the oil testing of the 79 data
calculation. are also known, the test results are used to recalculate the
health index and predictions with machine learning. The
TABLE II. PARAMETER HEALTH INDEX results of machine learning calculations and predictions will
Rating Condition Description Scoring be measured for their performance by comparing these
A Good < 1,2 4 results with the health index results from field assessments.
The performance/accuracy formula for machine learning
B Acceptible 1,2 – 1,5 3
calculation methods and predictions can be seen in the
C Need Cautions 1,5 – 2 2 formula below.
D Poor 2–3 1

E Very Poor >3 0

C. Predicting Health Index using Machine learning

Machine learning is an alternative method in determining
the health index. Machine learning has advantages in
finding dependent and independent relationships of input
parameters and output parameters. There are 2 machine
learning algorithms used, namely Naïve Bayes and Artificial
neural network.
Naïve Bayes is an algorithm in machine learning that
Fig. 3. Pie chart of 79 Validation Transformers Health Index
functions as a classifier. Naïve Bayes works by assuming an
assumption that is tested statistically. This assumption is Fair = 8; Good = 2; Poor = 2; Very poor = 0; Very good =
known as class conditional independence. In Naïve Bayes, 67.
all data parameters are calculated prior probability and Accuracy Equation
posterior probability. Posterior probability is the highest that
determines the distribution of classes. The probability (8)
formula can be seen in the formula below.
E. Correlation between parameter
P(h|D) = P(D|h)P(h)/ P(D) (7)
Artificial Neural Network (ANN) is a machine learning In the method of calculating the health index and
classifier. Neural networks work by imitating the workings machine learning, it is known that each parameter has an
of the human brain and nervous system. In ANN there are independent or dependent relationship to the health index
several main parts, namely input parameters, hidden layers and the age of the transformer. The correlation can be
consisting of neurons, and class (output). ANN works by explained by the magnitude of the weighting factor and the
looking for strength between neurons on input and class theory of reactions between parameters on the transformer's
(output) variables. Each input variable has a scoring and the health index and remaining life. In addition, this correlation
strength between neurons (weight) will be included in the can be tested using statistics using Pearce correlation testing
activation function. The smallest error value determines the (T-test).
class distribution. The neural network diagram can be seen (9)
in the image below.
Input Hidden Output
Layer III. RESULT AND DISCUSSION
In this study, 79 transformer oil data were used which
were measured for estimated age and health index. The 79
data transformers have been subjected to field assessments.
A. Remaining Life estimated of Transformer

TABLE III. EXPECTED LIFE ESTIMATED

Age Remaini
Count Count Count
(Years) 2FAL ng Life
(Year)
Fig. 2. Artificial Neural Network
- 16 < 100 62 > 30 58
100-
0 – 10 8 9 20 – 30 14
D. Field Assesment of Transformer 250
250-
10 – 20 9 5 10 – 20 7
In this study, 79 data transformers were identified whose 1000
health index was known. The health index is obtained from 1000-
20 – 30 36 1 5 – 10 0
1500
the results of field assessments, maintenance history, and
> 1500
fault history. The data is then assessed based on the standard > 30 10 2 <5 0
- 300

190

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.
 Estimation of the remaining life of the transformer can
be measured based on the condition of the paper insulation.
Paper insulation is the main protector when an internal
failure occurs in the transformer, its function is to prevent a
short circuit from occurring in the transformer. The 2FAL
compound can be used as a measure of the degradation of
paper insulation in transformer oil.
Transformer aging can be caused by operational or
premature aging reactions. Premature aging reactions
consist of oxidation, hydrolysis and pyrolysis reactions. The
pyrolysis reaction is a heating reaction but without oxygen.
This reaction arises from thermal stress and electrical stress
on the transformer. High temperature conditions can cause Fig. 4. Comparasion of health index model
the cellulose chains on paper insulation to decay and give
rise to furan compounds (one of which is 2FAL). The The accuracy of the calculation of the health index can
collapse of 2FAL can be used as a measure in determining be seen in Table IV, where the accuracy of the non-
the degree of polymerization of paper insulation. The DP historical method and the historical method is 84.8%. This
value determines the estimated remaining life of the indicates that the delta and gas rate parameters in the
transformer. historical method do not have a strong correlation with the
B. Health Index calculation health index.
The comparation of health index model using
In this research, a health index was calculated to create
conventional (calculation) and non conventional (naïve
machine learning training data. As well as calculations used
bayes and neural network) can be seen in figure 4. In this
to recalculate the health index from field data and calculate
study, non-conventional methods were used to determine the
the level of accuracy. The calculation of the health index
health index. One of the methods used is machine learning.
depends on the value of the weighting factor, the weighting
The machine learning algorithms used are naïve Bayes and
factor defines the significance value of an input parameter to
neural networks. The accuracy of the machine learning
the output parameter (health index). Some parameters that
method is obtained from a comparison between the
have high weighting factor values are breakdown voltage,
prediction results and the results of field inspections. In
water, acetylene, and 2 furfural.
table V it is known that the accuracy value of the neural
Accuracy is obtained from a comparison of the results of
network with the non-historical method is the highest. The
the field assessment (real data) with the results of
historical method has a lower level of accuracy, but has a
calculating the health index. The results of the field
variety of output classes. Machine learning is very
assessment can be seen in the figure 3 (pie chart 1).
dependent on the balance of the output class dataset. The
TABLE IV. HEALTH INDEX CALCULATION
non-historical method data set has a better balance than the
historical method. This causes the accuracy of the non-
Health
Non Historical Historical historical method to be higher than the historical method.
Index Data Data Data Data
training real training real D. Correlation each parameter for remaining life and
Very Good 878 71 865 72 health index
Good 191 6 91 5 From the calculation of the estimated remaining life of
Fair 99 2 32 2 the transformer and the health index, there are several
Poor 59 0 4 0 parameters (Test) that have a high correlation with the age
Very Poor 101 0 0 0 and health index. The remaining life of the transformer is
Accuracy 84,8% 84,8% affected by parameters 2 furfural, water, interfacial tension,
acidity, breakdown voltage, and carbon dioxide. The
C. Health Index calculation pearson correlation (R) value and significance (P-Value) can
be seen in table VI. The age of the transformer is determined
TABLE V. HEALTH INDEX PREDICTION
by the condition of the insulation paper. Compound 2
Non Historical Historical furfural is one of the compounds that arise due to thermal
Health
Index Naïve Neural Naïve Neural faults and pyrolysis reactions that occur in paper insulation.
bayes Network bayes Network Pyrolysis reaction occurs due to excessive heating on
Very Good 54 69 61 67 paper insulation. Excessive heating that occurs continuously
Good 25 10 4 9 or repeatedly can reduce the condition of paper insulation.
Fair 0 0 0 2 In addition, one of the products of the pyrolysis reaction is
Poor 0 0 14 1 carbon monoxide, carbon dioxide, and water. Insulation
Very Poor 0 0 0 0 power can also be weakened due to hydrolysis and oxidation
Accuracy 68,3% 82,3% 68,3% 77,2% reactions. The results of the hydrolysis and oxidation
reactions are water content, polar compounds, and acidic
compounds. These polar compounds become contaminants
in the oil causing low IFT and BDV values.

191

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.
 TABLE VI. CORRELATION EACH PARAMETER TO REMAINING LIFE hydrolysis, and oxidation. This reaction produces polar
Test R P- Test R P-Value compounds, acidic compounds, water, and carbon dioxide.
Value The breakdown voltage, water, and 2 furfural parameters
BDV 0,45 0,000 H2 0,12 0,000 also affect the health index. In addition, the acetylene
Water -0,66 0,000 CH4 0,23 0,000 parameter also indicates a poor health index.
Acid -0,53 0,000 C2H6 0,21 0,004 The health index and remaining life of the transformer
IFT 0,49 0,006 C2H4 0,13 0,000 are used to predict transformer reliability. Transformer
2FAL -0,76 0,005 C2H2 0,08 0,000 reliability can be used as a strategy for maintenance, repair,
CO 0,05 0,000 operation and investment.
CO2 0,40 0,000

TABLE VII. CORRELATION EACH PARAMETER TO HEALTH INDEX REFERENCES

Test R P- Test R P-Value [1] Li, Y., et al.: “Aging assessment of power transformer using multi‐
parameters”. Int. J. Electr. Eng. Inf. 5(1) (2013)
Value
BDV 0,57 0,000 H2 -0,28 0,007 [2] Li, S., Li, J.: “Condition monitoring and diagnosis of power
equipment: Review and prospective”. High Volt. Instit. Eng. Technol.
Water -0,55 0,000 CH4 -0,20 0,000 2(2), 82–91 (2017)
Acid -0,12 0,000 C2H6 -0,23 0,000 [3] “IEEE Guide for Loading Mineral-Oil-Immersed Transformers”,
IFT 0,12 0,000 C2H4 -0,08 0,018 IEEE Std. C57.91-1995.
2FAL -0,58 0,002 C2H2 -0,25 0,003 [4] Tang, S.; Hale, C.; Thaker, H. Reliability modeling of power
CO -0,19 0,000 transformers with maintenance outage. Syst. Sci. Control Eng. 2014,
CO2 -0,40 0,000 2, 316–324.
[5] Sang, Z.; Mao, C.; Lu, J.; Wang, D. Analysis and Simulation of Fault
Table VII shows that BDV, 2FAL, water content, and Characteristics of Power Switch Failures in Distribution Electronic
Power Transformers. Energies 2013, 6, 4246–4268.
CO2 have a high degree of correlation with the health index.
[6] "CIGRE 227 - Life Management Techniques for Power Transformer,"
In addition, these parameters have a weighting factor value CIGRE, 2003.
in determining the calculation of the health index. The [7] "CIGRE 323 - Ageing of Cellulose in Mineral-Oil Insulated
health index is influenced by parameters 2 furfural, Transformers," CIGRE, 2007.
breakdown voltage, water content, acetylene and carbon [8] Kanumuri, D.; Sharma, V.; and Rahi, O. P., “Analysis Using Various
dioxide. From field inspection data, it is known that Approaches for Residual Life Estimation of Power Transformers”,
International Journal on Electrical Engineering and Informatics, Vol.
transformers that have a low health index are undergoing 11, no 2, June 2019.
maintenance or repairs due to leaks, purification, and [9] Jahromi, A.N.; Piercy, R.; Stephen; Cress; Service, J. R. R.; and Fan,
planning for transformer replacement. Leakage and W, “An Approach to Power Transformer Asset Management Using
purification are caused by a high water content value, thus Health Index”, IEEE Transactions on Dielectrics and Electrical
causing a low breakdown voltage value. The high content of Insulation, vol. 25., no. 2. 2009.
acetylene indicates the presence of an electrical fault in the [10] Thor Hjartarson, Shawn Otal, "Predicting Future Asset Condition
Based on Current Health Index and Maintenance Level" IEEE 11th
transformer. Parameter 2 furfural indicates a weakening of International Conference on Transmission & Distribution
the paper insulation. The weakening of paper insulation Construction, Operation and Live-Line Maintenance, 2006, ESMO,
causes a high water value and the possibility of a fault is Oct. 2006.
higher. [11] Nick Dominelli, "Equipment Health Rating of Power Transformers",
IEEE International Symposium on Elecmcal Insulation, Indianapolis,
IV. CONCLUSION IN USA, 19-22 September 2004, pp. 163 -168.
[12] S. Li and J. Li, ``Condition monitoring and diagnosis of power
The transformer is an equipment that supports the equipment: Review and prospective,'' High Voltage, vol. 2, no. 2, pp.
reliability of the electricity system. Transformer failure can 82-91, Jun. 2017.
result in material and non-material losses. This study aims to [13] I. G. N. S. Hernanda, A. C. Mulyana, D. A. Asfani, I. M. Y. Negara,
determine the parameters of transformer oil in influencing and D. Fahmi, ``Application of health index method for transformer
condition assessment,'' in Proc. IEEE Region Conf. (TENCON), Oct.
the remaining life and health index of transformers. In this 2014, pp. 1-6.
study, 79 transformer oil data were used from field [14] J. Faiz and M. Soleimani, “Assessment of Computational Intelligence
assessment results to determine the estimated remaining life and Conventional Dissolved Gas Analysis Methods Transformer
and health index. As well as 1328 transformer oil data used Fault DIagnosis”, IEEE Transactions on Dielectrics and Electrical
Insulation, Vol. 25, no. 5, 2018.
as machine learning training data.
In determining the health index, calculations and [15] Xiao. Y; Pan. W; Guo. X; Sheng Bi. S; Feng. D;and Lin. S, “Fault
Diagnosis of Traction Transformer Based on Bayesian Network“,
predictions are used using machine learning. The accuracy Journal MDPI Energies, 2020.
of calculating the health index is 84.8%. The accuracy of [16] Al Hamdani, M. F.; Prasojo, R.A.; Suwarno; and Siada, A. A.;
machine learning with the naïve Bayes algorithm is 68.3% “Power Transformer Degradation Condition and Insulation Index
and the neural network algorithm is 82%. The calculation of Estimation Based on Historical Oil Data”, 2 nd. Int. Conf. on High
Voltage Engineering and Power Systems, 2019.
the health index is very dependent on the weighting factor.
[17] Prasojo, R.A.; Suwarno; and Siada, A. A.; “Dealing With Data
Meanwhile, machine learning depends on the balance of Uncertainty for Transformer Insulation System Health Index”, IEEE
training data. Access, 2019.
The age of the transformer can be determined from the [18] IEC, IEC “60422 Mineral insulating oils in electrical equipment
condition of the insulation paper. Compound 2 furfural is a Supervision and maintenance guidance.” IEC Fluids for
compound that arises due to a fault in the transformer. The Electrotechnical Applications Technical Committee, (2013).
reactions that occur in premature aging are pyrolysis, [19] WG. A2.49, “Condition assessment of power transformers,” CIGRE,
2019.

192

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.
 [20] SD. Myers, JJ. Kelly, RH. Parrish, “A guide to transformer [24] Wilson, G.: ‘Asset management of transformer oil’, presented at the
maintenance,” Transformer Maintenance Institute, S D Myers Inc, Euro TechCon, Stratford, UK, December 2015
Jun. 1981. [25] R. A. Prasojo, K. Diwyacitta, Suwarno, and H. Gumilang,
[21] IEEE C57.104 2008 - IEEE Guide for the Interpretation of Gases “Transformer paper expected life estimation using ANFIS based on
Generated in Oil Immersed Transformers oil characteristics and dissolved gases (Case study: Indonesian
[22] IEEE C57.104 2019 - IEEE Guide for the Interpretation of Gases transformers),” Energies, vol. 10, no. 8, 2017.
Generated in Oil Immersed Transformers
[23] D. Rediansyah, R. A. Prasojo, Suwarno, and A. Abu-Siada, “Artificial
Intelligence-Based Power Transformer Health Index for Handling
Data Uncertainty,” IEEE Access, vol. 9, pp. 150637–150648, 2021.

193

Authorized licensed use limited to: King Mongkut's University of Technology North Bangkok. Downloaded on December 03,2024 at 10:40:21 UTC from IEEE Xplore. Restrictions apply.