Prediction of Transformer Oil Insulation Strength Based on Physical Quantities Using Ensemble Machine Learning Model

Abstract

Existing machine learning models for predicting transformer oil breakdown voltage (BDV) often depend on chemical properties, which require laboratory tests, thus making them time-consuming and costly. Furthermore, these models are typically developed using standalone methods, which may fail to achieve satisfactory prediction accuracy. This study introduces an interpretable ensemble prediction model for transformer oil BDV, leveraging physical properties instead of chemical properties. The ensemble model integrates four standalone models: multilinear regression (MLR), multilayer perceptron (MLP) neural network, support vector machine (SVM), and long short-term memory (LSTM) neural network, to enhance prediction accuracy. A Local Interpretable Model-Agnostic Explanation framework was used to provide insights into the model's predictions. The model's performance was assessed using key metrices, including the correlation coefficient (R), determination coefficient (R²), root mean squared error (RMSE), and mean absolute error (MAE). Results revealed that the ensemble model outperformed individual models, achieving R = 0.9796, R² = 0.9596 during training and R = 0.9781, R² = 0.9567 during testing, along with RMSE = 1.7678, MAE = 1.2972 during training, and RMSE = 1.7347, MAE = 1.2985 during testing. The proposed model reliably predicted transformer oil BDV using physical properties, eliminating complex experiments while offering transparency to enhance applicability in the power industry