The 20th International Symposium on High Voltage Engineering, Buenos Aires, Argentina, August 27 – September 01, 2017

METHODOLOGY FOR RELIABILITY ANALYSIS OF POWER

TRANSFORMERS BASED ON FAILURE DATA
1* 2 3 4
A. B. Oliveira Neto , E. G. Costa , V. S. Moraes , T. V. Ferreira
1
Post-Graduate Program in Electrical Engineering (PPgEE/COPELE), Federal University
of Campina Grande (UFCG), Aprígio Veloso St n 882, Brazil
2
Professor, Electrical Engineering Department, UFCG, Aprígio Veloso St n 882, Brazil
3
Graduation Student, Electrical Engineering Department, UFCG, Brazil
4
Professor, Electrical Engineering Department (DEL), Federal University of Sergipe,
Marechal Rondon Av., São Cristóvão, Brazil
*Email: <[email protected]>

Abstract: The global electrical sector is constantly evolving due to the increasing energy
demand since the industrial revolution of the 19th century. This evolution brings new
challenges for companies that generate, transmit and distribute energy. In order to meet
these demands, research and technologies are developed aiming at obtaining systems
and equipment for the electrical sector with reduced costs, high efficiency and reliability.
Reliability and availability estimates are necessary assist in decision-taking and estimate
costs related to operation and maintenance of the power system and its equipment,
among which the power transformer stands out. Power transformers are essential
equipment in the electric power system: they convert voltage levels to interconnect
generation, transmission and distribution systems. This work presents a methodology to
analyze the reliability of electrical power system equipment with emphasis on power
transformers. Therefore, a database is collected concerning the times and failure modes
of these equipment. The collected samples are evaluated by parametric probabilistic
models largely used in reliability analysis. The Anderson-Darling quality of fit criterion is
used to determine which model best fits the study sample. When the probabilistic
representation model is found, the maximum verisimilitude and least squares methods
are used to estimate the input parameters of the model. Thus, reliability curves, failure
rate, risk function and mean time to failure for the equipment were found and analyzed.
Besides, it was elaborated the survey of fixed costs related to preventive and corrective
maintenance. The curves of preventive, corrective and total were found, as a function of
time, using a numerical method for maintenance. An estimate of the optimal time to
perform the preventive maintenance was also implemented. Additionally, an economic
analysis was elaborated as a function of a maintenance policy. Therefore, the developed
methodology allows to analyze the reliability of the power transformers under different
aspects and assists decision-taking regarding operation and maintenance.

reliability of the system, as well as a decrease in

1 INTRODUCTION
peak demand and a higher revenue return [3].
The increasing level of market requirement
In this context, transformers are vital for the
demands a system that assures the fulfillment of
electrical system and electric power companies.
the needs as well as quality of the product or
Failures on transformers may result in power
service. Prolonged and effective operation of the
interruption, which not only reduce utility revenue,
productive systems of goods and services is a vital
but also cause loss of customer confidence,
requirement [1]. In services such as energy
whether industrial, commercial or residential [4]. It
production, transmission and distribution, sudden
is therefore imperative that energy companies
failures caused by random factors must be
conduct studies to evaluate the reliability of
understood and counterbalanced, avoiding
transformers.
damages of an economic nature.
The choice of probabilistic criteria and the
Reliability assessment is fundamental for planning
calculation of probability indices constitute a
and operating the electrical system [2]. Thus,
relevant subsidy for the actions of the electric
system topological information, failure modes and
system planning [5]. Inserted in this context, the
historical data of components, such as
operation and maintenance (O&M) encompass
transformers, aid operational planning and
decisions and actions related to the control and
development of advanced methodologies for fault
maintenance of goods and equipment.
detection, isolation and service restoration. This
results in reduction losses, increase quality and
In view of the exposed scenario, it is intended, in ( )
( ) . (5)
this work to elaborate and validate a methodology ( )

for evaluating the reliability of power transformers.

In addition, this paper verifies which probabilistic 2.2 Parametric Probabilistic Models
model best fits the samples of the failure times of
these equipment and what is the best method to In the following, the main distributions used in the
estimate the parameters of the model. Thus, based context of the reliability analysis will be briefly
on the choice of a probabilistic model, it is possible described.
to elaborate a maintenance plan through the 2.2.1 Exponential Distribution
reliability analysis, stipulating the functions of
corrective, preventive and total maintenance over The exponential distribution plays an important role
time, helping decision making and optimizing O&M. in reliability studies because it is always the only
continuous distribution with a constant risk function
2 THEORETICAL BACKGROUND [8]. Their reliability representations are described
according to the expressions from (6) to (9).
2.1 Reliability
( ) , (6)
Reliability is the probability that an item will ( ) , (7)
satisfactorily perform its required function, under ( ) , (8)
specific conditions in a predetermined period of [ ] . (9)
time [6, 7]. The main reliability representations are
described below, from topic 1 to topic 5. Where: = Failure Rate

2.1.1 Reliability Function 2.2.2 Weibull Distribution

The reliability function, ( ), can be expressed as: The Weibull distribution is an expression widely
used in maintenance engineering. It is appropriate
to time-to-failure modeling, presenting constant,
( ) ( ) (1)
strictly increasing and strictly decreasing risk
functions [9]. Their reliability representations are
Where: = Continuous random variable described according to the expressions from (10)
corresponding to the equipment failure time to (13).
= Final time of the period during which
the component is observed
( )
= Probability of occurrence of the event ( ) , (10)
( ) ( )
2.1.2 Cumulative Distribution Function ( ) , (11)
( ) ( )
The cumulative distribution function, ( ), ( ) , (12)
corresponds to the probability that the component
( ). (13)
fails before t. ( ) is expressed by (2):

( ) ( ) (2) Where: = Form parameter

= Scale parameter
2.1.3 Probability Density Function = Gamma function

The probability density function, ( ), describes the 2.2.3 Normal Distribution

shape of the fault time distribution. ( ) is defined
by (3): A population that fits the normal distribution has
symmetrically arranged variations around the
( ) ( ) mean. One of the reasons for its application comes
 ( )  (3) from the fact that when a value is subject to many
variations that add up, regardless of how these
2.1.4 Mean Time To Failure (MTTF) variations are distributed, the result of the
composite distribution is normally distributed
The MTTF corresponds to the mean, or expected
[9, 10]. The reliability representations of this
value, of the probability distribution and is defined
distribution are given by expressions (14) to (16).
by (4):

  ( ) ( ), (14)
∫ ( ) ∫ ( ) . (4)
( ) ( )
* + , (15)
√
2.1.5 Risk Function ( )
The risk function h(t), also known as the failure ( ) . (16)
( )
rate, is interpreted as the amount of risk associated
with a unit at time t, defined by (5): Where: = Mean of failure times
 = Standard deviation of failure times The application of the MLS, after linearization of
= Cumulative distribution of standard probability distribution, consists of minimizing the
normal sum of the squares of the deviations between the
ξ= Value of standard normal distribution observed value and the estimated value ̂ [11].
density function For elements of the sample, we have the
expression defined in (23):
2.2.4 Lognormal Distribution
The lognormal distribution is more versatile than ∑ . (23)
the normal distribution, since it has a more varied
form, which allows a better adjustment of the In this situation, the values of and that
population [10]. Their reliability representations are minimize (23) are obtained by calculating the
described in accordance with expressions (17) to partial derivatives of and setting equal to zero.
(20). Finally, the system solution allows estimating the
parameters of the linear regression model. In
( ) addition, it is possible to estimate the correlation
( ) ( ), (17)
( )
coefficient that measures the adhesion of the
( ) , * + -, (18) model to the analyzed data.
√
[( ) ]
( ) , (19)
[( ) ] 2.3.2 Method of Maximum Likelihood
( ) Estimation (MMLE)
. (20)
The function expressed in (24) informs about the
Where: = Mean of natural logarithms of failure possibility or likelihood of the random variables
times ,... , assuming the values ,... , . This
= Standard deviation of natural possibility is given by the probability density
logarithms of failure times
function. For the case of discrete variables, the
( ) = Value of the standardized normal
likelihood is a probability value.
distribution function
= Value of the standardized normal
distribution density function ( ) ∏ ( ). (24)

2.3 Estimation of the Model Parameters Where: ( ) = Likelihood function associated with
the sample
The parameters are estimated from the sample ( ) = Probability distribution
observations, allowing the determination of the = Unknown parameter
models and analysis. There are some estimation
methods known in the statistical literature, among 2.4 Choice of Probabilistic Model
which the following are highlighted.
The choice of model to be used is an extremely
2.3.1 Method of Least Squares (MLS) important step in parametric analysis of reliability
data. In this context, the Anderson-Darling test is
The MLS aims to find the regression function that widely used for non-normality [12].
minimizes the sum of the distances between the
adjusted function and the observed points [11]. To 2.4.1 Anderson-Darling Fit Quality Test
do so, an adjustment is made to a regression line,
as expressed by (21) and thus calculates the The Anderson-Darling fit quality test (AD) is based
distance between the estimated and observed on a probability graph. When the data fits well into
value, as defined in (22). the hypothetical distribution, the probability graph
will be close to a straight line. The AD test statistic
̂ , (21) measures the total distance between the actual
̂. (22) and assumed distributions using a weighted
square distance [12]. Thus, Anderson and Darling
Where: ̂ = Estimated value of dependent variable [13] proposed the expression given by (25):
= Observed value of independent
[ ( ) ( )]
variable ∫ ( ). (25)
( )( ( ))
= Estimated coefficient by means of
experimental points of the sample
Where: = Quadratic statistical value
= Coefficients of the estimated model by
( ) = Continuous cumulative distribution
means of experimental points of the sample
function
= Distance between observed value and
( ) = Cumulative empirical distribution
estimated value by the regression line
function
= Observed value of dependent variable
= Random sample
2.5 Numerical Method for Maintenance The main potential failure mode detected was
Analysis classified as design or fabrication error.

In order to define the best maintenance strategy, it In Table 2 can be seen the survey of average
is necessary to take into account the probability of operating costs related to power transformers
failure of the component and its reliability over under study [15].
time [14]. The functions of corrective, preventive
and total maintenance as a function of time are Table 2: Survey of average operating costs
given by equations (24) to (26).
Replacement
Run to
[ ( )] policy up to
( ) . (26) failure
age 60 years
∫ ( )
[ ( )] Replacement policy ($) – 3,310,000.00
( ) . (27) Replace after major failure ($) 4,560,000.00 1,270,000.00
∫ ( )
( ) ( ) ( ). (28) Maintenance ($) 1,830,000.00 1,390,000.00
Losses ($) 5,440,000.00 4,310,000.00
Major failures - Direct costs ($) 590,000.00 160,000.00
Where: = Fixed cost for corrective maintenance
Larger faults – Largest parallel
= Fixed cost for preventive maintenance fault ($)
11,830,000.00 3,290,000.00
( ) = Cost of corrective maintenance as Major failures - Minor miss in
a function of time t deviation ($)
2,130,000.00 570,000.00
( ) = Cost of preventive maintenance as Total ($) 26,380,000.00 14,300,000.00
a function of time t
( ) = Cost of total maintenance as a
3.2 Methods
function of time t
The methodology adopted in this work follows the
3 MATERIAL AND METHODS
sequence of steps described below:
3.1 Material
Step 1 – Collect a set of samples of the number of
failures occurred in observation time intervals of the
The material used in this work corresponds to a equipment under analysis.
database, referring to the failure times (FT) of a
single-phase supply power transformer bank,
Step 2 – Apply the exponential, Weibull, normal
according to Table 1 [15].
and lognormal probabilistic models to the samples
of the failure times, considering right censoring.
Table 1: Failure data for a single-phase power
transformer bank
Step 3 – Evaluate the fit of probabilistic models
through the Anderson-Darling fit quality test. In this
Voltage
FT test, an admission number (AD = A) is obtained.
Failure Cause Censorship Frequency Class
(years) The lower value of AD, the more adjusted is the
(kV)
Design or
14 Failed 1 220/33 distribution to evaluated data.
Manufacturing error
Design or
16 Failed 1 110/33
Manufacturing error Step 4 – Choose the distribution to be applied, and
Design or employ the MLS or MMLE to estimate the
18 Failed 1 110/34
Manufacturing error
Design or parameters. The method with lowest AD will be
27 Failed 1 50/11
Manufacturing error used to estimate parameters of the model to be
Design or
Manufacturing error
28 Failed 1 66/11 chosen. In the case of lower AD of the same value
Design or for the two methods used, besides the graphical
32 Failed 2 220/33
Manufacturing error visualization of suitability of the models to samples,
Design or
Manufacturing error
34 Failed 3 220/33 the MLS is chosen as the criterion of choice, since
Design or it has the correlation calculation of the sample to
39 Failed 1 220/33
Manufacturing error the model. In this case, the one with the highest
Design or
Manufacturing error
42 Failed 1 110/50 correlation is chosen.
Design or
43 Failed 1 110/11
Manufacturing error Step 5 – Elaborate the construction of the reliability
Design or
Manufacturing error
44 Failed 1 50/11 representations, from the estimated parameters,
Design or
60 Censored 25 –
such as: probability density function; survival
Manufacturing error
function and failure rate function. In addition,
estimate model statistics, such as MTTF.
The data of Table 1 have a right censorship, since
after the observation time, 25 transformers are in Step 6 – Elaborate a policy of substitution of the
operation in the system, with 60 years or more. transformers using the estimated parameters, the
numerical method for maintenance and the transformers is the Weibull distribution with
estimation of the operational costs. estimation of the parameters made by the MLS.

To assist the construction and analysis of the steps Table 3: AD values for the distributions used in the
described was used computational tool samples of failure times of the transformers using
Minitab 17®. MLS and MMLE

4 RESULTS MLS MMLE

Distribution
AD Coefficient AD Coefficient
Next, the reliability representations are shown and Weibull 154.727 154.775
discussed for the samples in Table 1. Lognormal 154.730 154.751
Exponencial 154.891 154.862
Initially, the presented distributions were plotted Normal 154.728 154.817
and analyzed by the Anderson-Darling fit quality
test, considering MLS and MMLE, as can be seen Then, the parameters of form and scale were
in Figure 1 and Figure 2, respectively. estimated from Weibull distribution by MLS, with
= 2.55 and θ = 60.96, as well as estimated
MTTF = 54.11 years. In Figure 3 can be seen as
functions: probability density, survival and failure
rate.

Figure 1: Anderson-Darling fit quality test applied

to the distributions used in the analysis of samples
of failure times of the transformers, considering the
MLS.
Figure 3: Representations of the reliability
functions for the Weibull distribution by MLS.

After selecting the model of representation and

estimating its parameters, considering the survey
of the costs seen in Table 2 and using the
numerical method for maintenance, the
representation of the transformers substitution over
time was elaborated, according to Figure 4.

Figure 2: Anderson-Darling fit quality test applied

to the distributions used in the analysis of samples
of failure times of the transformers, considering the
MMLE.

The AD values of the models to the samples can

be seen in Table 3. It can be noted that the lowest
AD value for both methods employed was Figure 4: Graph of replacement cost over time.
AD = 154.727. Therefore, following the criterion
presented in the methodology, the model chosen Within the limitations of this analysis, and in
for representing samples of fault times of the particular the difficulties of estimating the failures,
there is evidence that the preventive replacement
of the single-phase transformer bank is estimated Distribution Transformers”, Colloquium on
at around 56 years. The average minimum total Humanities, Science and Engineering
cost per unit for this preventive replacement is Research (CHUSER), IEEE. Penang,
$ 449,584.14 per year. Malaysian. 5-6 Dec., 2011.
[5] Empresa de Pesquisa Energética (EPE),
The expected behavior of transformers at 56 years “Estudos Associados ao Plano Decenal de
of operation was analyzed, in which these Energia PDE 2007/2016: Transmissão de
equipment will have: reliability of 55.92% and Energia Elétrica – Análise dos índices de
probability of failure 44.08%. confiabilidade do SIN”, 29 Junho, 2007.
[6] L. Leemis, “Reability: Probabilistic Models and
5 CONCLUSION Statistical Method”, Prentice-Hall, New York,
1995.
This work presented a methodology to analyze the [7] E. A. Elsayed, “Reliability Engenieering”,
reliability of power transformers based on failure Adssion Wesley, Massachustts, 1996.
time data. Its contribution is to demonstrate the [8] F. Fogliatto, J. Ribeiro, “Confiabilidade e
application of probabilistic models for the analysis Manutenção Industrial”, Elsevier, Rio de
of the reliability of high voltage equipment and the Janeiro, 2009.
use of methods to stipulate the most appropriate [9] J. R. B. Lafraia, “Manual de Confiabilidade,
model to the samples and to estimate the Mantenabilidade e Disponibilidade”,
parameters of this model. Qualitymark, Rio de Janeiro, 2001.
[10] E. R. Carnaúba and M. A. Sellitto, “Análise de
The reliability, probability density and failure rate Confiabilidade e Evolução de Uma Máquina de
curves were obtained and analyzed, in addition to Envase de Leite UHT ao Longo da Curva da
stipulating the MTTF for the equipment studied. Banheira”, Revista Liberato, Vol. 14, No. 22,
pp 113-238, 2013.
In fact, this research developed a substitution [11] N. Naghettini and E. J. A. Pinto, “Hidrologia
policy focused on the optimal choice of Estatística”, CPRM, Belo Horizonte, 2007.
replacement time of power transformers. p.402.
Therefore, the study assists decision making [12] P. D. Doanem and L. E. Seward, “Estatística
regarding of planning the O&M of the equipment Aplicada à Administração e Economia”, 4°ed.
analyzed. Bookman, 2014.
[13] T. W. Anderson and D. A. Darling, "A test of
In future studies, it is intended to develop an goodness-of-fit", Journal of the American
optimum model of choice of probabilistic models Statistical Association, Vol. 49, pp. 765-769,
and methods of estimation and sensitivity with 1954.
application in different equipment or sets of [14] J. D. Campbell, and A. K. S. Jardine,
equipment, optimizing O&M of the electric power “Maintenance Exellence: Optimizing
system. Equipment Life - Cycle Decisions”, New York,
Marcel Dekker, Inc., 2001.
ACKNOWLEDGMENTS [15] Transpower New Zealand Limited. ACS Power
Transformers-Fleet Strategy. November 2013.
Authors acknowledge the support of UFCG URL: https://www.transpower.co.nz/node/1095
Post-Graduate Program in Electrical Engineering 1/fleet-strategies, Last acessed: April 03, 2017.
(COPELE) and Coordination of Improvement of
Higher Education Personnel (CAPES) for the
financial support.

REFERENCES

[1] V. F. Campos, Controle de qualdide total

(TQC), Ed. Edg, 1992.
[2] N. Balijepalli, S. S. Venkata and R. D. Christie,
“Predicting Distribution System Performance
Against Regulatory Reliability Standards”,
IEEE TPD, Vol. 19, No. 1, pp. 350-356, 2014.
[3] P. C. Sekhar, R. A. Deshpande and V. Sankar,
“Evaluation and Improvement of Reliability
Indices of Electrical Power Distribution
System”, IEEE Power Systems Conference
(NPSC), IEEE. 19-21 Dec., 2016.
[4] M. I. Ridwan, M. R. Samsudin and Y. Z. Y.
Ghazali, “Reliability Analysis of Premature
Failed 11/0.433 kV Hermetically Sealed