Predicting the Oil Quality of Power Transformers Using Machine Learning

Abstract:
The reliable prediction of the oil quality of power transformers is crucial for an efficient and safe energy supply. In this contribution, a modified nonlinear autoregressive network with exogenous inputs (NARXnet) approach for oil quality prediction is presented that achieves high accuracy. By investigating the time dependence of the measurement data, it is shown that an overabundance of measurements as input variables can degrade the prediction performance. A sensitivity analysis confirms the physical plausibility of the results and underlines the comprehensibility of the machine learning model. The comparison with traditional Markov chain methods shows that machine learning models are superior, when applied to single transformers. These findings provide valuable contributions to optimizing maintenance strategies and extending the service life of transformers.