| Summary: | In electrical power utilities, there is an ever-growing need for improved asset
management. Power transformers are identi ed as one of the most critical and
high impact items of plant within an electric network. For this reason, e ective
management of transformers is required to reduce the risk to power transfer
due to unplanned outages, as well as the high consequential costs associated
with catastrophic failure.
The objectives of this work include the evaluation of e ectiveness of the current
method implemented within Eskom, of evaluating transformers based on
their condition/Health Index (HI) to develop replacement strategies, as well
as identifying possible improvements to these methods and development of a
model that can be utilized for determining the probability of failure of a power
transformer based on its HI.
There are two components of the existing model for determining failure probability:
the e ects of age and HI. Historical failure data was collected for
the period 1996 - 2014, including both severe and intermediate failures in the
Eskom Transmission network. This included failure mode, demographic information, Dissolved Gas Analysis (DGA) results, oil quality test results and
predicted Degree of Polymerization (DP). A data sample of healthy transformers
was also collected. The failure data was tted to a Weibull distribution,
and the probability of failure based on age determined. This was compared to
the existing distribution parameters and its e ectiveness evaluated. Statistical
analysis was carried out on the complete data set. Since there are multiple,
continuous predictor variables and one dichotomous output variable, a multiple
logistic regression model was tted to the data. This was done for the
existing HI, as well as for new HI parameters that were identi ed as the most
signi cant in predicting the output. The existing Weibull distribution was found to be ine ective in describing the
existing failure data for ages <10 and >50 years. The average age predicted
by this model is also unrealistically high and no practical evidence of this is
found. An alternative Weibull distribution was found that better described
the data. The logistic regression model tted to the failure data using the
existing HI parameters was found to be a poor predictor of probability of
failure. An alternative model was found enabling a more accurate prediction,
using fewer variables. Due to the large errors in measurements of the predictor
variables and in some cases, exponential tolerances, as with DP, inaccuracies
are expected within the model. The existing model is found to be ine ective
in determining the probability of failure of a power transformer. New HI
parameters, an age distribution and logistic regression model were determined,
enabling a higher accuracy in predicting failure events and can therefore be
utilized in various asset management initiatives and risk mitigation.
|
|---|
