Scholarlink

<p>Creeping discharges along oil–pressboard interfaces are a major degradation mechanism that limits the reliability and service life of power transformer insulation systems. While previous studies have primarily focused on the detection and classification of these discharges, the prediction of their propagation limits remains insufficiently explored. This study addresses this gap by formulating the prediction of creeping discharge stopping length as a supervised regression problem based on experimentally measured discharge current signals. The model is based on current signals generated during discharge propagation over liquid–solid interfaces under AC voltage. Five insulating liquids were investigated: synthetic ester, methyl oleate, olive oil, rapeseed oil, and mineral oil. Four features, mean, root mean square, time-domain energy, and wavelet-based frequency energy, were extracted and used as inputs to two machine learning models: multilayer perceptron and support vector regression. The models were evaluated on independent test datasets using <i>R</i><sup>2</sup>, MAE, and RMSE metrics. The results demonstrate that the multilayer perceptron and support vector regression models both achieved high predictive accuracy, as evidenced by their high <i>R</i><sup>2</sup> values and low mean absolute error (MAE), root-mean-square error (RMSE), and mean squared error (MSE). For SVR, the results reaching <i>R</i><sup>2</sup> = 0.94 (MAE = 1.08, RMSE = 1.35), while MLP achieved comparable accuracy with <i>R</i><sup>2</sup> = 0.93 (MAE = 0.76, RMSE = 1.00). The novelty of this work lies in using regression-based learning to predict the length of discharge directly. Future work will focus on advanced models and real-time, non-intrusive monitoring for practical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative prediction of creeping discharge stopping length: a multilayer perceptron and support vector regression approach

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Mohammed Adaika,
  • Abderrahim Reffas,
  • Hicham Talhaoui,
  • Tarek Bouguettaya,
  • Abderazak Bennia,
  • Oualid Aissa,
  • Abderrahmane Beroual,
  • Hocine Moulai

摘要

Creeping discharges along oil–pressboard interfaces are a major degradation mechanism that limits the reliability and service life of power transformer insulation systems. While previous studies have primarily focused on the detection and classification of these discharges, the prediction of their propagation limits remains insufficiently explored. This study addresses this gap by formulating the prediction of creeping discharge stopping length as a supervised regression problem based on experimentally measured discharge current signals. The model is based on current signals generated during discharge propagation over liquid–solid interfaces under AC voltage. Five insulating liquids were investigated: synthetic ester, methyl oleate, olive oil, rapeseed oil, and mineral oil. Four features, mean, root mean square, time-domain energy, and wavelet-based frequency energy, were extracted and used as inputs to two machine learning models: multilayer perceptron and support vector regression. The models were evaluated on independent test datasets using R2, MAE, and RMSE metrics. The results demonstrate that the multilayer perceptron and support vector regression models both achieved high predictive accuracy, as evidenced by their high R2 values and low mean absolute error (MAE), root-mean-square error (RMSE), and mean squared error (MSE). For SVR, the results reaching R2 = 0.94 (MAE = 1.08, RMSE = 1.35), while MLP achieved comparable accuracy with R2 = 0.93 (MAE = 0.76, RMSE = 1.00). The novelty of this work lies in using regression-based learning to predict the length of discharge directly. Future work will focus on advanced models and real-time, non-intrusive monitoring for practical applications.