Download as pdf or txt
Download as pdf or txt
You are on page 1of 29

Tutorial from WG A2:62 International

Transformer Reliability Survey

Dr Dan Martin, on behalf of WG A2:62 chaired by Prof. Stefan Tenbohlen


ETEL transformers and CIGRE AU.A2 panel member representative
CIGRE NZ.A2 panel convener

Auckland, 18th May 2021


Learning Objectives

 Examples of the use of statistics to make engineering decisions.


 Apply Weibull distribution to model probability of failure.
 Understand bathtub curve and relevance of different sections.
 Misuse of statistics to make decisions.
 Randomness applied to catastrophic failures such as fires and
explosions.
Historical perspectives

 First CIGRE survey published 1983.


 Australian & NZ 1996 by Petersen.
 Follow up CIGRE survey 2015.
 Inspired Australian and NZ survey,
published in 2018.
Use of statistics in failures
What can we do?

 AER regards the use of failure statistics to


calculate risk costs as following good
practice when planning replacement.
In this article spend limits on existing plant is discussed. Repairs or refurbishment
are classed as an investment. The probability of failure is used to calculate the
chance of this investment being written off early if the asset fails soon after.
 Calculate the number of spares which should be available.
 Model works by calculating the number of failures expected in a given time
frame (e.g. how long it takes to procure a replacement unit).
 Closer look at differences in useful life, and whether the model should be
adjusted for an older fleet if the probably of failure is increasing.
Misuse of statistics
It is not that useful to apply fleet failure statistics to one unit.

Why?

Lets suppose out of 100 units of the same age one will probably† fail this year.

An average, each of the 100 units will have a 1% chance of failure. Not that useful
to determine which transformer of the 100 actually fails.

† Disclaimer – failures are statistically independent events. Even if there is a 1%


annual failure rate, in one year there might be zero or more than one failures.
Assumptions using statistics

 Future predictions of failure rate are dependent on historic


events
 Any changes to the asset, utilization, maintenance or operational
stresses may impact failure rate!
 Still the failure distribution in any given year provides a snap-
shot in time. Fast forward ten years and we may find that
changes in manufacture or operation have worked through.
 On one hand it is beneficial to have a large number of years of
data to analyse for failure statistics, on the other there might be
differences in management from the beginning to the end of
the period.
Bath tub is a frequently applied technique
 Abstract model of three different
distributions.

 Infant mortality: minimised through quality


design, manufacture and testing to
remove items with defects.

 Useful life: failures which occur randomly,


not related to time.

 Wearout: failure rate increasing.

 The beta parameter refers to the shape


parameter of the Weibull distribution (one
of its variables).
Weibull statistical distribution

For a sample of 6,000 power transformers looked at


the failure rate, which was:

𝑛𝑛 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓(𝑡𝑡)
𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 =
𝑛𝑛 𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢 (𝑡𝑡)

When failures are plotted as ln(age), they line up


which shows that the use of this distribution is valid.
The line gradient is the beta parameter. Individual
three sections of the bathtub curve is not always
apparent (e.g. if random PoF is so low).
What is wear-out?
Failure rate starts to increase due to x-axis (age).
Not that every transformer ‘starts to wear out’.
Why would its start differ between utilities?

 Quality of transformers and subcomponents.


 Maintenance.
 Usage and network faults.

A utility may put more focus on assets older than the start of wear-out
to check condition.
Data #1 - Transmission

 A relatively small number of


failures.
 There could be two groups
emerging, one near random
and the other perhaps
wearout.
 However, uncertain as need
more data points.
Data #2 from Australia

 Transmission – absence of early failures –

Probability of failure (log scale)


20 years.
 Failure wearout seems to start at 20 years.

D. Martin, J. Marks, T. Saha, O. Krause, and N. Mahmoudi, “Investigation into


Modelling Power Transformer Failure and Retirement Statistics”, IEEE
Transactions on Power Delivery, vol. 33, no. 4, Aug. 2018.
Data #3 - Transmission

 When failures are plotted on these


log scales, if they line up on a
straight line the use of the Weibull
distribution is valid.
 Gradient (beta) is nearly 1,
indicating nearly random.
 Only two early failures, and no
wear out (yet).
 Careful of extrapolating Probability
of failure because end of useful life
not observed!
MTBF

 The MTBF (mean time between failures) is commonly used


metric for reliability

Mathematically is 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 = 1/λ where λ is the failure rate.

MTBF expressed in units of time.


What is useful life and why is it important?

 Across a population, the


failure rate starts to increase
due to degradation or usage.
 Check condition regularly
after the end of the useful life
– detect units which are more
likely to fail.
Analysis of useful life

Difficult to see where the useful


life ends for each utility.

D. Martin, T. K. Saha, G. Buckley, S. Chinnarajan and T. MacArthur, “Analyzing


Differences in Useful Life of Power Transformers across Utilities for Better
Strategic Spares Management”, IEEE PESGM, Aug. 2018.
Modelling random events - Poisson distribution

 Early application: modelling deaths of Prussian army officers


from falling off horses (Bortkiewicz 1898).

 Calculate number of incidents that would be expected to


occur by random for a population. If more incidents occur
than what is expected from randomness, could indicate
problems with training/horses.

 Assumption that incidents are random. When might this not


be so?

 Modern application:
Failure of equipment due to earthquakes, fires, surges, where failure
has not been made more likely due to condition or age.
Randomness

 Poisson distribution is used


to model randomness, to
investigate the useful life.
 Using statistics, we can
check the observed number
of failures which would have
occurred by random.
 Up to 20 years observations
fit that expected by
randomness.
2018 Fires and explosions in Australia and New
Zealand investigated

 92,119 service years from 6,500


power transformers from 2000 1
1
onwards, focussing on when the fault 1 Bushing
resulted in a fire or explosion 1 7 OLTC
Link box
2 Unknown
Cable Termination
Causes of fires or explosions in Australian & New Zealand Insulation
2
power transformers over 92,119 service years. Surge arrestor
Winding
Note equal number of OLTC and bushing faults. 7

But:
1 fires per 1,300 service years from CIGRE 2015 vs 1 fire per
4,200 service years for ANZ (fewer fires in ANZ)
Fire Statistics – Australia & New Zealand
≤ 66 kV 110 & 132 kV ≥ 220 kV
≤ 66 kV 110 & 132 kV ≥ 220 kV

0.004 TNSP retirements DNSP retirements

4 20
Catastrophic failure rate

0.003

3 15

Retirements
Failures
0.002
2 10

0.001 1 5

0 0
0
2000 2002 2004 2006 2008 2010 2012 2014
Year
Age (years)

 Probability of fire or explosion seems related to age, but the


yearly number of fires has fallen from about 2008 onwards.
Poisson distribution - randomness

≤ 66 kV 110 & 132 kV ≥ 220 kV

0.9
 On average, about 1 fire per
0.8
year.

Probability of occurance
0.7
 However, in any given year there 0.6
might be either no fires, or more 0.5
than 1. 0.4
 Asset manager should therefore 0.3
consider extremes. 0.2
0.1
0
0 1 2 3
Number of fires each year from 2000 to 2016
Fires and explosions before and after 2008
Bushing
1
1
OLTC
1
6 Link box 1 1
Bushing
2 Unknown
OLTC
Cable
Termination Winding
2 Insulation
0
Surge arrestor

6 Winding
1
 After 2008 there have been fewer fires – especially bushings and OLTCs
 Two possible reasons:
 Vacuum interrupters becoming prevalent
 Industry more aware of bushings requiring management, or newer technologies being used.

 Shows that industry is taking note of statistical surveys and updating their asset
management practices!
Summarising remarks

 Statistical models extremely beneficial to make decisions on


risk and spares management.

 Usually, one organisation does not incur enough failures for a


statistically significant analysis. Consequently, an international
survey is beneficial.
Current work of WG A2.62

 Conducting another worldwide survey of transformers ≥100 kV.


 There have been calls for data over the past year.
 The more data the better our outcome.
Progress to data

Data from 18
countries.

Respondees own
10,000 power
transformers.

300 failures and 700


retirements.
The end
Thanks for listening

Dr Dan Martin
Dan.Martin@eteltransformers.co.nz
Further reading
 D. Martin, C. Beckett, J. Brown, and S. Neilsen, “Analysis and Mitigation of Australian and New Zealand Power Transformer Failures Resulting in
Fires and Explosions”, IEEE Electrical Insulation Magazine, vol. 35, no. 6, pp. 7-14, 2019.

 D. Martin and N. Watson, “Statistical analysis of Australian and New Zealand Power Transformer Catastrophic Fires”, CMD conference, Perth,
Australia, Sept. 2018.

 D. Martin, T. K. Saha, G. Buckley, S. Chinnarajan and T. MacArthur, “Analyzing Differences in Useful Life of Power Transformers across Utilities for
Better Strategic Spares Management”, IEEE PES General Meeting, Aug. 2018. (Selected as one of the best papers of the conference.)

 D. Martin, J. Marks, T. Saha, O. Krause, and N. Mahmoudi, “Investigation into Modelling Power Transformer Failure and Retirement Statistics”,
IEEE Transactions on Power Delivery, vol. 33, no. 4, Aug. 2018.

 D. Martin, J. Marks, and T. Saha, “Survey of Australian Power Transformer Failure Modes and Retirements”, IEEE Electrical Insulation Magazine,
vol. 33, no. 5, pp. 16-22, 2017.

 D. Martin, J. Marks, T. Saha, O. Krause, G. Russell, and A. Alibegovic-Memisevic, “On the Development of Power Transformer Failure Models: an
Australian Case Study”, IEEE PES General Meeting, USA, July 2017. (Selected as one of the best papers of the conference.)

 D. Martin, J. Marks, and T. Saha, “2016 Survey on Australian Power Transformer Failure and Retirement”, Asia Pacific TechCon, Apr. 2017, Sydney,
Australia.

 J. Marks, D. Martin, T. Saha, O. Krause, A. Alibegovic-Memisevic, G. Russell, G. Buckley, S. Chinnarajan, M. Gibson, and T. MacArthur, “An
analysis of Australian power transformer failure modes, and comparison with international Surveys”, 2016 Australasian Universities Power
Engineering Conference (AUPEC), pp. 1 – 6, Australia, 2016.
Copyright © 2021

This tutorial has been prepared based upon


the work of CIGRE and its Working Groups.
If it is used in total or in part, proper
reference and credit should be given to
CIGRE.

Copyright &
Disclaimer notice Disclaimer notice

“CIGRE gives no warranty or assurance


about the contents of this publication, nor
does it accept any responsibility, as to the
accuracy or exhaustiveness of the
information. All implied warranties and
conditions are excluded to the maximum
extent permitted by law”.

You might also like