Abstract
Polarization depolarization current (PDC) analysis serves as a pivotal time-domain spectroscopy technique for assessing transformer insulation conditions. However, the low magnitude of the PDC is susceptible to noise corruption during in-situ measurements, posing challenges in accurate insulation modeling. This paper introduces an innovative hybrid multistage technique that integrates empirical mode decomposition (EMD) and encoder-decoder long short-term memory (en-de-LSTM) networks to estimate critical insulation-sensitive parameters from noise-affected polarization current (ipol) data, utilizing optimized Debye parameters. Comparative evaluation of the proposed approach with linear, Conv, and LSTM models establishes the suitability of the proposed model in capturing temporal dependencies of noise-affected PDC data. Furthermore, the Debye model determines the transfer function zero closest to the origin, a critical insulation-sensitive parameter in transformers’ oil-paper insulation (OPI) system. Moth flame optimization (MFO) finetunes parameter estimation within the Debye model, augmenting convergence to achieve precise characterization of the intricate dynamics inherent in the complex system. A correlation between moisture percentage, transfer function zero, and dissipation factor is also established to enhance the insulation diagnosis approach. Results demonstrate the consistently superior performance of the proposed noise-affected PDC investigation approach across the testing period, evidenced by the lowest logcosh and mean absolute error values. The proposed EMD-en-de-LSTM model with MFO demonstrates superior performance, achieving a significantly lower error of 1.10% compared to existing ipol denoising methodologies such as Stockwell Transform for insulation-sensitive parameter estimation. This substantiates the superiority of the hybrid approach, which utilizes adaptive non-parametric signal processing and deep learning based on EMD-en-de-LSTM networks and MFO.

This publication has 27 references indexed in Scilit: