A Probabilistic Learning Framework for Distribution Transformer Residual Useful Life Estimation

An electrical distribution system needs distribution transformers that can convert the voltage to a safe, controlled level and ensure a continuous supply to users at all times. The Residual Useful Life (RUL) of transformers is estimated to plan maintenance and manage assets. In this paper, transformer insulation ageing will initially be estimated using the thermal ageing model specified in IEC 60076-7. The insulation loss of life (LOL) is determined by the dynamics of hot-spot temperatures. The ageing calculations are then used to generate the RUL, and these R values serve as reference labels for training probabilistic machine learning models. In addition, two nonlinear regression models to interpret nonlinear correlations between the operational variables and the transformer RUL are presented: Gaussian Process Regression (GPR) and Bayesian Neural Networks (BNN). The results show very high consistency of the predictions of BNN and GPR: BNN predicts 7499.60 ± 0.17 days, and GPR predicts 7499.59 ± 0.11 days, where the uncertainty is very small. The performance of the model is measured using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and coefficient of determination (R2), which show high predictability and low estimation error. To make it more interpretable, explainable artificial intelligence techniques, including LIME and SHAP, are used to visualise local and global feature inputs. The suggested approach offers a method for predicting the uncertainty-aware RUL of the distribution transformer with high reliability and can be used to inform maintenance decisions aimed at managing the distribution transformer's assets.