Machine Learning For Predictive Maintenance of Transformer

e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science

Volume:03/Issue:04/April-2021 Impact Factor- 5.354 www.irjmets.com
MACHINE LEARNING FOR PREDICTIVE MAINTENANCE OF TRANSFORMER
Mehak Asad*1, Fareeha*2
*1Electrical Energy System Engineering, US Pakistan Center for Advance Studies in Energy
University of Engineering and Technology, Peshawar, Pakistan
*2Electrical Energy System Engineering, US Pakistan Center for Advance Studies in Energy
University of Engineering and Technology, Peshawar, Pakistan
ABSTRACT
The research paper presents a methodology to carry out predictive maintenance of power transformers. The
core concept of the methodology is to monitor some of the parameters and predict the condition of
transformer. This helps to prevent catastrophic failure and unnecessary testing and maintenance thus reducing
the overall operating cost and time required for maintenance. The common methods of predictive maintenance
of transformer include dissolve gas analysis, artificial neural network, support vector machine, multi-class least
square support vector machine. The paper is based on methodology for predictive maintenance of power
transformers by using Linear Regression and Principal Component Analysis (PCA). The electrical parameters
used for conducting this research includes phase voltage’s magnitude, phase voltage angles, phase current’s
magnitude, phase current’s angles, active power, frequency and reactive power. Two different approaches are
used for this study. First approach makes use of Phase A voltage’s magnitude and Phase A current’s magnitude
to train the linear regression model. The applied model is trained and then tested to predict the value of Phase
A voltage’s magnitude for the given values of Phase A current’s magnitude. Where as in second approach, PCA is
applied on the whole data set and its output was given to the linear regression model as input. The model is
trained and tested to predict Phase A voltage’s magnitude for the given principal components (input). To obtain
the result, the data collecting device i.e. PMU was installed at Sheikh Muhammadi Grid with an 11KV
distribution transformer. The device records the values of parameter (phase voltage, phase current, reactive
power, active power and frequency) at time interval of 5 minutes.
Keywords: Predictive maintenance, Transformer, Machine learning, Linear Regression, Principal Component
Analysis (PCA), Dissolved gas analysis (DGA), Support vector machine (SVM)
I. INTRODUCTION
As transformer is an essential component, they are quite expensive alot of time is required for maintenance and
repairing if it gets damaged. The main problems that arise within power transformers are because of different
faults occurring in them. Due to variable loading and weather conditions, these devices are under mechanical,
thermal stress and electrical stress. These changing conditions can result into different type of faults giving rise
to abnormal conditions and leading to cascading failure under worst case. These faults are broadly divided into
internal and external faults.
External faults arise due to events happening outside this static device and hence can’t be prevented. Whereas,
the internal faults that arises due to events occurring inside this device can be prevented by testing and
maintenance. The possibly occurring internal faults arise due to either contamination in transformer oil, or due
to insulation failure or overheating. These conditions don’t cause the failure instantly but the conditions
buildup with the passage of time and becomes severe resulting in cascading failure. Hence predictive
maintenance can be used to monitor these conditions and take the appropriate steps and actions needed to
reduce the risk of failure.
There are different ways for maintenance of transformer. Few of them are predictive maintenance, preventive
maintenance, and corrective maintenance. Predictive maintenance can be done by a number of methods used in
the literature review. These include predictive maintenance by vibration analysis, thermal analysis, insulating
oil analysis, partial discharge analysis and frequency response analysis. Besides the different method used for
these analysis, dissolve gas analysis, artificial neural network (ANN), support vector machine, multi-class least
square support vector machine are a few to mention.
Study has been carried out to predict and classify the faults present in transformer by using SVM as the core
technique. DGA is used for condition assessment. This method is to detect and predict faults in oil-filled
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

[414]
e-ISSN: 2582-5208
transformer by analyzing the concentration or amount of specific gases that are dissolved in the insulation oil of
the transformer. Its is used to determine the ratio and rate of gasses. Along with SVM, K-nearest neighbor is
applied to fill the missing values of the DGA data set. In this particular study, Euclidan distance and city block
distance metric were used. Further its output is used as input to SVM [10].
Studies are conducted to diagnose and detect the faults using DGA data to training two step ANN. ANN is
trained to capture the relation between the gas composition and incipient fault conditions. ANN detects these
relationships and uses them as a base for interpreting the amount of gases dissolved in the oil. For this study,
two step ANN is used, one to classify main fault types and the second used to find the involvement of cellulose.
The results of this approach are promising even with limited amount of data. However to achieve higher
accuracy and to learn complex relations, more data is required [11].
To detect incipient faults in power transformer, two neural network structures (N1 and N2) were proposed
which were trained using the data of dissolved gas. Neural networks were having 5 input node and 2 and 6
output nodes. Different gases were given as inputs to the two networks. N1 and N2 were trained using the error
back propagation method and the achieved accuracy was 93.17% and 89% respectively. The scheme was used
to detect the faults (based on DGA) which remained unidentified when conventional methods were used [13].
Zhou et al. [14] used visualization and NN techniques together for faults prediction in power transformer. Oil-
dissolved gas data was used for this research. The results obtained show that the proposed methodology works
effectively for the condition monitoring of the power transformers.
Besides other methods, for condition monitoring of power transformers key gas method is used and results
were verified by the Ratios and Total Combustible Gas method. This approach uses the DGA method for
analyzing the historical test results. Two transformer models of 115/22kV and 230/115/22kV were used for
this study [16].
Insulating oil analysis is a powerful and important method to monitor conditions of transformer. For oil
analysis, DGA is one of a very powerful test for detecting the abnormal conditions on the basis of different gas
ratios present in the oil [17][18].Beside different gases that are dissolved in the oil, presence of moisture can
also accelerate the insulation aging process and results in different type of failure. Partial discharge can also be
measured to detect the upcoming faults in transformer [19].
Multi-class least square support vector machines (LS-SVM)-based classifier were used to diagnose faults in
transformer. The particle swarm optimization algorithm was applied to select the optimal feature for the multi-
class LS-SVM classifiers [21].
The objective of the research is to propose a reliable solution for detecting any abnormal conditions by
developing predictive maintenance algorithms to reduce the overall maintenance cost. The approach used here
for the predictive maintenance is based upon linear regression and principal component analysis. To predict
healthy and unhealthy conditions based on the past data machine learning algorithms are used.
II. METHODOLOGY
The proposed methodology for predictive maintenance of power transformer is to use some machine learning
algorithms. Machine learning is one of the applications of Artificial Intelligence (AI) that allows the system to
learn from the past data and using that past data provide some information about the future or make some
predictions. It main idea is to develop some computer program that uses the past data and learns the
underlying logic in the data by itself. It learn from the past data and use the same knowledge to make
predictions.
a) Linear Regression
To model relationship between dependent and independent variable(s), linear regression is used. It helps to
understand the relation between the variables by fitting a line to the data set. The fitted line is defined by an
intercept (wₒ) and slope (w₁) where wₒ is the regression coefficient or weight on the feature x. Weight on the
feature tells about the amount of variation in y by changing x. The equation of a simple of linear regression
model is given as:
𝑦 = wₒ + w₁𝑥

[415]
e-ISSN: 2582-5208
b) Principal Component Analysis
PCA is a statistical procedure and an unsupervised machine learning algorithm that is widely used to reduce the
dimensionality of the data set by transforming a set of large features into a smaller set by keeping important
information contained in larger set. Reducing the dimensionality helps in speeding up a machine learning
algorithm because it transforms a large set of correlated variable into linearly uncorrelated variables. These
components (Principal components) are new variable constructed by the combinations of initial variable. The
initial variables are correlated but the new components i.e. principal components are uncorrelated containing
all the information of initial data. The basic idea behind is that a 5-dimensional data can produce 5 principal
components. The first component has maximum possible information; the second component has remaining
information and so on.
III. RESULTS AND DISCUSSION
To obtain the results, first the available parameters were plotted with respect to time to visualize the trends in
the data set. After data visualization, two different approaches were used. In first approach, linear regression
mathematical model is applied on Va magnitude and Ia magnitude. The obtained results are explained and
discussed below in this section. In second approach, PCA was applied on the data set and the first two
components of PCA were taken as input for linear regression model to predict Va magnitude (taken as y).
a) Data Visualization
VA_MAG VS TIME
Figure 1: Phase A voltage’s magnitude with respect to time

VAB_MAG VS TIME
Figure 2: Phase AB voltage’s magnitude with respect to time

[416]
e-ISSN: 2582-5208
FREQUENCY VS TIME
Figure 3: Frequency with respect to time

MVAR VS TIME
Figure 4: Reactive Power with respect to time

MWH VS TIME
Figure 5: Active Power with respect to time

b) First Approach-Linear Regression
In this approach, the selected parameters were used as input to train the linear regression mathematical model.
The parameter VA_MAG was taken as output (y) and IAX_MAG was taken as input (x).
[417]
e-ISSN: 2582-5208
• The parameters chosen for this regression models are magnitude of phase A voltage (VA_MAG) and
magnitude of phase A current (IA_MAG). The relation among the two variables is visualized by scatter plot
shown in figure 1.
Figure 6: VA_MAG VS IAX_MAG

• The data set is split up into 80% training set and 20% test set.
• The model is trained on training set. The algorithm used to train the mathematical model is same as used in
the first model (i.e. Newton method). The results obtained after training the model are shown in figure 2.
Figure 7: Linear regression mathematical model on transformer data set

• This model was evaluated on test data. Following are the results that were obtained.
• The prediction made by the model along with the original data was plot on a graph shown in figure 3. The
original data is indicated by green dots while the predicted values are show by red line.
Figure 8: Visualizing the actual and predicted data

[418]
e-ISSN: 2582-5208
• The table below shows of linear regression mathematical model i.e. the slope and the intercept of the line that
fits the data best giving the minimum value of RSS over all possible combinations.
Figure 9: Model coefficients

In above table, the value of the intercept indicates the point at which the line crosses the y axis which is 6587.25
for this case. The coefficient of IAX_MAG indicates the variation in phase A voltage’s magnitude with respect to
phase A current’s magnitude is -0.697.
The equation of the line is
y= 6587.25 -0.967x
The variable loading conditions results in variation in voltage can be seen on the plot (as for same value of
current we have different values of voltage). When load increases, voltage drops. But to keep the voltage at same
level, different voltage control methods are used. At grid, tap changers keep the voltage at a constant value.
During faulty conditions, the current’s threshold value for this particular setup is 653A. Above this value the
relays connected with the system send trip signal. The limit is shown with an orange dashed line on the plot. Any
point lying outside the limit indicates the presence of fault. But as there was no fault on the grid since the data is
recorded, therefore there is no point that lies outside the orange line.
c) Second Approach- Linear Regression and Principal Component Analysis
In this approach, PCA was applied first and then the principal components were given as input to the linear
regression model.
• PCA was applied on the data set because it contains 28 features. The first two principal components were
chosen to be taken as input for linear regression model. The figure 11 below shows the graph between
principal component 1 and 2.
Figure 10: PC1 VS PC2

• The results below shows the variance that is explained but PC1 and PC2 i.e. 42.1% and 22.74% respectively.
As these two components contain the maximum information so these two are used as input to the linear
regression model.
• In this approach, the first two principal components were given as input to train linear regression
mathematical model. VA_MAG was taken as output for the model. The obtained results are shown below.
• To use the model used in this approach, it should be trained first on the available dataset. After training the
model, the model is tested on the same data set to check how accurately it predicts the observations. So for

[419]
e-ISSN: 2582-5208
training and testing the model, the available data set is split into two subsets with 80% data in the training set
and the remaining 20% for the test set.
• After data splitting, the model is trained on training set. The algorithm used for model training is Newton
method which is chosen automatically by the model used here. The obtained results after training the model
are shown in figure 11.
Figure 11: Linear Regression on Principal Components

• This model was evaluated on test data. The obtained results are shown below.
This shows that max error is reduced by using this approach as compared to just applying linear regression
algorithm.
• The table below shows the coefficients of linear regression model i.e. the slope and the intercept of the line
that fits the data best giving the minimum value of RSS over all possible combinations.
Figure 11: Model coefficients

In above table, the value of the intercept shows the point where the line crosses the y axis which is 6445.50 for
this case. The variation in coefficient of PC1 is -55.47 per volts of phase A voltage’s magnitude. Whereas the PC2
coefficients show that the variation is 4.38 per volts of phase A voltage’s magnitude.
IV. CONCLUSION
Power systems is a complex network and power transformers serves as the backbone of the power system so, its
proper functionality is of core importance for the reliable and secure operation of the system. Being essential
components, they are quite expensive and require a lot of time for maintenance and repairing if it gets damaged.
The main problems that arise in this static device (power transformers) are different faults occurring in them.
Due to variable loading and weather conditions, these devices are under mechanical, thermal stress and
electrical stress. These changing conditions can result into numerous faults giving rise to abnormal conditions
and leading to cascading failure under worst case. There are a number of ways for the maintenance of
transformers. These include predictive, corrective and preventive maintenance.
The research approach used here for the predictive maintenance is based on linear regression mathematical
model and principal component analysis. The machine learning algorithms are used to predict healthy and
unhealthy conditions based on the past data.

[420]
e-ISSN: 2582-5208
To get the results, first the available parameters were plotted with respect to time to visualize the trends in the
data set. After data visualization, two different approaches were used. In first approach, linear regression
algorithm is applied on Va magnitude and Ia magnitude. The obtained results are explained in results section.
V. FUTURE WORK
The data that is currently used for this research work includes parameters like phase voltage’s magnitude, phase
voltage’s phase angle, phase current’s magnitude, phase current’s phase angle active and reactive power,
frequency. However if in future the collected data include temperature, concentration of different dissolved
gases in transformer oil, mechanical and thermal parameters then these can be utilized for different analysis like
vibration analysis, thermal analysis, Insulating oil analysis, partial discharge analysis and frequency response
analysis. More the number of parameters in the data set better are the results. More parameters can help to use
other predictive maintenance techniques including Artificial Neural Network, Support Vector Machine, Particle
Swarm Optimization, Dissolved gas analysis and many other. By using these advance techniques, more accuracy
can be achieved that will be helpful for analysis resulting in better approach for predictive maintenance of
transformers.
VI. REFERENCES
[1] /@yufengg. (2017, September 07). The 7 Steps of Machine Learning. Retrieved from
https://towardsdatascience.com/the-7-steps-of-machine-learning-2877d7e5548e
[2] Catalog. (n.d.). Retrieved from https://www.vizimax.com/knowledge-center/131-english/phasor-
measurement-unit
[3] /@zakaria.jaadi. (2019, May 24). A step by step explanation of Principal Component Analysis. Retrieved
from https://towardsdatascience.com/a-step-by-step-explanation-of-principal-component-analysis-
b836fb9c97e2
[4] Principal component analysis. (n.d.). Retrieved from
https://www.sciencedirect.com/topics/engineering/principal-component-analysis
[5] Yildirim, Tülin, M., Kurt, & Bülent. (2018, March 14). Aircraft Gas Turbine Engine Health Monitoring
System by Real Flight Data. Retrieved from https://www.hindawi.com/journals/ijae/2018/9570873/
[6] William H. Bartley, “Failure analysis of Transformer”, The Hartford Steam Boiler Inspection & Insurance,
2003.
[7] Transformer Failures, Causes & Impact. (n.d.). Retrieved from
http://iieng.org/images/proceedings_pdf/8693E0215039.pdf
[8] M. F. Cabanas, F. Pedrayes, M. G. Melero, C. H. Rojas G. A. Orcajo, J. M. Cano & J. G. Norniella, “INSULATION
FAULT DIAGNOSIS IN HIGH VOLTAGE POWER TRANSFORMERS BY MEANS OF LEAKAGE FLUX
ANALYSIS”, Progress In Electromagnetics Research, Vol. 114, 211{234, 2011.
[9] Mohsen Akbari, P. Khazaee & Sabetghadam, P. Karimifard “Failure Modes and Effects Analysis (FMEA) for
Power Transformers” Power Transmission & Distribution Research Center Niroo Research Institute (NRI)
Tehran, Iran. 28th international power system conference, iran, November2013.
[10] Sahri, Z., & Yusof, R. (2015). Fault diagnosis of power transformer using optimally selected DGA features
and SVM. 2015 10th Asian Control Conference (ASCC). doi:10.1109/ascc.2015.7360340
[11] Zhang, Y., Ding, X., Liu, Y., & Griffin, P. (1996). An artificial neural network approach to transformer fault
diagnosis. IEEE Transactions on Power Delivery,11(4), 1836-1841. doi:10.1109/61.544265
[12] Kabir, F., Foggo, B., & Yu, N. (2018). Data Driven Predictive Maintenance of Distribution
Transformers. 2018 China International Conference on Electricity Distribution (CICED).
doi:10.1109/ciced.2018.8592417
[13] Nagpal, T., & Brar, Y. S. (2014). Neural network based transformer incipient fault detection. 2014
International Conference on Advances in Electrical Engineering (ICAEE).
doi:10.1109/icaee.2014.6838535

[421]
e-ISSN: 2582-5208
[14] Zhou, Z., Jiang, Y., Yin, X., & Chen, S. (2002). The Application of Visualization and Neural Network
Techniques in a Power Transformer Condition Monitoring System. Developments in Applied Artificial
Intelligence Lecture Notes in Computer Science,325-334. doi:10.1007/3-540-48035-8_32
[15] Morais, D., & Rolim, J. (2006). A hybrid tool for detection of incipient faults in transformers based on the
dissolved gas analysis of insulating oil. 2006 IEEE Power Engineering Society General Meeting.
doi:10.1109/pes.2006.1709217
[16] Suwanasri, T., Chaidee, E., & Adsoongnoen, C. (2008). Failure statistics and power transformer condition
evaluation by dissolved gas analysis technique. 2008 International Conference on Condition Monitoring
and Diagnosis. doi:10.1109/cmd.2008.4580333
[17] Faria, H. D., Costa, J. G., & Olivas, J. L. (2015). A review of monitoring methods for predictive maintenance
of electric power transformers based on dissolved gas analysis. Renewable and Sustainable Energy
Reviews,46, 201-209. doi:10.1016/j.rser.2015.02.052
[18] Lewand, L. R. (n.d.). Condition Assessment of Transformers — Analysis of Oil Data and Its
Quality(Publication).
[19] Ciulavu, Cristina & Helerea, Elena. (2008). Power transformer incipient faults monitoring. Ann Univ
Craiova-Electr Eng ser. 32.
[20] Abu-Elanien, A. E., & Salama, M. M. (2007). Survey on the Transformer Condition Monitoring. 2007 Large
Engineering Systems Conference on Power Engineering. doi:10.1109/lescpe.2007.4437376
[21] Zheng, H., Liao, R., Grzybowski, S., & Yang, L. (2011). Fault diagnosis of power transformers using multi-
class least square support vector machines classifiers with particle swarm optimisation. IET Electric
Power Applications,5(9), 691. doi:10.1049/iet-epa.2010.0298
[22] Yadaiah, N., & Ravi, N. (2012). Statistical method for fault detection in synchronous generators. 2012
International Conference on Computer Communication and Informatics. doi:10.1109/iccci.2012.6158896
[23] Xayyasith, S., Promwungkwa, A., & Ngamsanroaj, K. (2018). Application of Machine Learning for Predictive
Maintenance Cooling System in Nam Ngum-1 Hydropower Plant. 2018 16th International Conference on
ICT and Knowledge Engineering (ICT&KE). doi:10.1109/ictke.2018.8612435
[24] Åsnes, A & Willersrud, Anders & Kretz, F & Imsland, Lars. (2018). Predictive maintenance and life cycle
estimation for hydro power plants with real-time analytics.
[25] Betti, Alessandro & Luisa Lo Trovato, Maria & Salvatore Leonardi, Fabio & Leotta, Giuseppe & Ruffini,
Fabrizio & Lanzetta, Ciro. (2019). Predictive Maintenance in Photovoltaic Plants with a Big Data Approach.
10.4229/EUPVSEC20172017-6DP.2.4.
[26] Kankar, P., Sharma, S. C., & Harsha, S. (2011). Fault diagnosis of ball bearings using machine learning
methods. Expert Systems with Applications,38(3), 1876-1886. doi:10.1016/j.eswa.2010.07.119
[27] Susto, G. A., Schirru, A., Pampuri, S., Mcloone, S., & Beghi, A. (2015). Machine Learning for Predictive
Maintenance: A Multiple Classifier Approach. IEEE Transactions on Industrial Informatics,11(3), 812-820.
doi:10.1109/tii.2014.2349359
[28] Li, H., Parikh, D., He, Q., Qian, B., Li, Z., Fang, D., & Hampapur, A. (2014). Improving rail network velocity: A
machine learning approach to predictive maintenance. Transportation Research Part C: Emerging
Technologies,45, 17-26. doi:10.1016/j.trc.2014.04.013
[29] Yang, Z., Tang, W., Shintemirov, A., & Wu, Q. (2009). Association Rule Mining-Based Dissolved Gas Analysis
for Fault Diagnosis of Power Transformers. IEEE Transactions on Systems, Man, and Cybernetics, Part C
(Applications and Reviews),39(6), 597-610. doi:10.1109/tsmcc.2009.2021989

[422]

Machine Learning For Predictive Maintenance of Transformer

Uploaded by

Copyright:

Available Formats

Machine Learning For Predictive Maintenance of Transformer

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning For Predictive Maintenance of Transformer

Uploaded by

Copyright:

Available Formats

e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Figure 1: Phase A voltage’s magnitude with respect to time

Figure 2: Phase AB voltage’s magnitude with respect to time

Figure 3: Frequency with respect to time

Figure 4: Reactive Power with respect to time

Figure 5: Active Power with respect to time

Figure 6: VA_MAG VS IAX_MAG

Figure 7: Linear regression mathematical model on transformer data set

Figure 8: Visualizing the actual and predicted data

Figure 9: Model coefficients

Figure 10: PC1 VS PC2

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Figure 11: Linear Regression on Principal Components

Figure 11: Model coefficients

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

You might also like