Transformer Condition Assessment & Survival Model

Abstract

Managing electrical distribution assets entails making the decision on how to proceed with an ageing asset. The three options that asset managers have are run to fail, perform maintenance or repairs and lastly to replace. This decision making process is becoming incrementally more important due to the threat of replacement waves. This a phenomenon that comes forth from the fact that a large portion of the assets were installed around the 1970's, the so called installation wave, combined with the fact that these assets have a life expectancy of 40 to 60 years. It is up to the asset managers to accurately asses the condition of these assets and prioritize investments due to budget restrictions. Asset management has been evolving from time based maintenance (TBM) to condition based maintenance (CBM) and risk based maintenance (RCM). To aid in the decision making process for the latter two methods, accurate condition assessment and failure probability methods are needed. At Stedin, one of the three largest network operators in The Netherlands, the asset managers require a better condition assessment method for transformers, as the one being used now lacks certain capabilities; trend analyses and proper prioritization is difficult and time consuming. Furthermore, there lies the question whether the condition of the transformer can be used to improve the failure or survival probability model. The assumption is that the condition indicators of an individual asset can be used to adjust the population based probability models, which are then comprehensive and more accurate. In this thesis, an improvement to the transformer condition assessment is implemented and the effect of the condition indicators on the survival probability of the transformer is studied. Several condition assessment methods were reviewed, with the chosen method being the so called health index (HI). The results from this method is accurately reflected in the transformers which are to be replaced by Stedin. However, it has become clear that using one single number for decision making is not recommended, as the subsystems that are in moderate or bad condition might be masked by those that aren't. This can be compared to the analogy that a chain is only as strong as its weakest link, and thus the condition of that link should not be masked by the condition of the rest. For the survival model, machine learning and classic statistical methods were reviewed. The chosen method was Cox's Proportional Hazard Model, which has the advantage of being applicable to situations in which the underlying probability distribution of the events is unknown. Due to missing failure data, the survival model was used to model the effect of the condition indicators on the probability of corrective maintenances. Further constraints in the data quality lead to the confidence bands of the model's parameters to be relatively large. However, the univariate models do prove the significance of the effect of these condition indicators. The conclusion is that there is a relationship between the condition and the observed failures, but that proper documentation of failures is necessary to increase the accuracy of the model.