Experimental evaluation of ensemble learning models for partial discharges detection in transformer insulation mediums

The reliability and stability of power systems largely depend on the integrity of power transformer insulation systems. Early detection of partial discharge (PD) activity plays a crucial role in preventing catastrophic failures, as PDs are localized dielectric breakdowns that indicate insulation degradation and may develop into severe faults if left undetected. Most conventional diagnostic techniques rely on single-sensor measurements, which may not fully capture the characteristics of PD events under varying operating conditions. This paper proposes a multi-signal ensemble learning framework for the accurate detection, identification, and classification of PDs in transformer insulation systems. Current, voltage, and acoustic PD signals were collected, filtered, and analyzed using time-domain feature extraction. Two ensemble classifiers, Random Forest (RF) and Gradient Boosting (GB), were employed to classify PD patterns occurring in air, mineral oil, and kraft insulation paper. The results show that RF achieved the highest classification accuracy of 96.51% using acoustic signals, outperforming GB, which achieved 95.87%. Compared with the best directly comparable previously published acoustic-sensor-based method, which reported an accuracy of 93%, the proposed RF approach achieved an improvement of 3.51 percentage points, equivalent to a relative improvement of 3.77%. Under the experimental conditions investigated in this study, these findings demonstrate the strong diagnostic capability of acoustic sensing combined with ensemble learning and highlight its potential for reliable and high-precision monitoring of transformer insulation health.