Enhancing transformer health index prediction using dissolved gas analysis data through integration of lightGBM and robust EM algorithms

Abstract

The dissolved gas analysis (DGA) data play a crucial role in evaluating the transformer health index (HI). In recent years, data-driven approaches have attracted significant research interest for the HI prediction with various health condition data. However, the DGA data collection is prone to missing or erroneous data due to sensors or data transfer issues. Consequently, handling missing data requires careful attention for accurate HI computation. In this paper, a novel data-driven hybrid approach is proposed that leverages the Light Gradient Boosting Machine (LightGBM) as a regression method and the Robust Expectation-Maximization (robust-EM) as a missing data imputation technique to predict the HI of transformers using DGA data. The proposed method is evaluated through five case studies with the percentage of missing data at 0%, 5%, 10%, 15%, and 20%. The proposed method has been compared with seven benchmark methods through six evaluation metrics, showing superior performance. The proposed method is also analyzed with and without robust-EM, and 22% – 71% performance improvements across various case studies and performance metrics have been achieved with robust-EM.