Condition assessment of power transformer under data unavailability and hyperparameter insensitivity

Abstract

Power transformers are vital assets in transmission grids, and accurate condition assessment is essential for power system reliability and cost-efficient maintenance. The health index (HI) is widely used for transformer condition assessment, integrating diagnostic information from dissolved gas analysis (DGA), oil quality (OQ) assessments, and insulation aging indicators such as the degree of polymerization (DP). However, accurate health index estimation is often challenged by measurement uncertainties, laboratory limitations, and incomplete datasets. This paper proposes a novel hybrid machine learning (ML) framework based on stacking to enhance health index prediction under both complete data and missing data conditions. This study employs Random Forest Regression (RFR) and Gradient Boosting Regression (GBR) as base learners, and the meta-learner combines their outputs to improve predictive accuracy. This framework is validated using data from 885 transformers, and 15 data-unavailability scenarios are investigated to assess the robustness and stability of the models. Results show that the hybrid model outperforms conventional methods and maintains high predictive accuracy even with substantial missing data. Sensitivity analysis confirms the robustness of the proposed model against hyperparameter variations and shows that these changes have a negligible effect on the model’s performance and accuracy. It can be observed that the model’s reliability under varying operational conditions and data uncertainties, making it a robust tool for data-driven decision-making in maintenance planning and asset management of power transformers.