Partial discharge localization in power transformer tanks using machine learning methods.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Partial discharges (PD) in power transformers pose significant risks to grid stability.
• Accurate localization of PD is crucial for effective maintenance and fault prevention.
• Existing localization methods often face challenges with complex transformer geometries and signal noise.

Purpose of the Study:

• To compare the effectiveness of various machine learning (ML) and deep learning (DL) methods for 3D PD localization in power transformers.
• To evaluate the performance of different ML/DL models based on input signals, methodologies, and accuracy metrics.
• To identify the most suitable ML/DL approach for precise PD source identification.

Main Methods:

• Investigated Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) for PD localization.
• Utilized single-sensor electric field measurements as input data.
• Analyzed methods based on input signal characteristics, core algorithms, correlation coefficients, and Root Mean Square Error (RMSE).
• Conducted multiple case studies with varying sensor positions, signal frequencies, and transformer tank sizes.

Main Results:

• Convolutional Neural Networks (CNN) significantly outperformed other ML/DL methods in PD localization accuracy.
• The study quantified performance differences using correlation coefficients and RMSE.
• Model performance varied based on specific case study attributes like sensor placement and signal frequency.

Conclusions:

• CNN models offer a superior solution for accurate 3D partial discharge localization in power transformers.
• The findings provide valuable insights for developing advanced diagnostic tools for power equipment.
• Further research can explore hybrid models and real-time implementation for enhanced grid reliability.