Abstract
<jats:p xml:lang="en">In transmission systems, power transformers are key high-value components, and their consistent operation is fundamental for maintaining grid stability and achieving cost-effective performance. The Transformer Health Index (THI) integrates key diagnostic parameters including dissolved gas analysis, water content, oil quality indicators, and power factor providing essential insights for asset condition assessment and investment planning. In this study, THI prediction is conducted using the Random Forest algorithm, recognized in literature for its high predictive accuracy for transformer applications, in combination with data preprocessing and filtering techniques applied to transformer dataset. For the first time, to the best of our knowledge in the THI prediction literature, various wavelet families are systematically compared at the preprocessing stage to examine their influence on predictive accuracy. The results show that the Symlet-2 configuration consistently outperformed other families in both filtered and non-filtered datasets, while Coiflet-3 and Coiflet-5 achieved higher efficiency through dimensionality reduction but with an accuracy decrease of approximately 0.09–0.10 in R2 compared to Symlet-2. The findings demonstrate that the choice of wavelet family in the preprocessing phase directly impacts feature selection outcomes and model performance, offering valuable guidance for the development of high-accuracy transformer condition assessment frameworks.</jats:p>