Abstract
The reliability and safety of power transformers are essential for the stable operation of
electrical power systems. Partial Discharge (PD) is a key indicator of insulation degradation
within transformers, and early detection of PD can prevent catastrophic failures and prolong
the life of these critical assets. PD occurs as localized electrical discharges within insulation
material that, over time, can deteriorate insulation and lead to significant damage. Accurate
localization of PD sources is crucial for diagnosing and mitigating faults before they escalate.
Traditional methods for PD localization such as Dissolved Gas Analysis (DGA), ultrasonic
techniques, and machine learning algorithms like Support Vector Machines (SVM), K-Nearest
Neighbors (K-NN), and Random Forest—have been extensively utilized. These methods
analyze PD signals and estimate their origin based on time delays and signal characteristics.
However, they often struggle with challenges in complex environments, where noise, signal
attenuation, and reflections can hinder localization accuracy. Furthermore, traditional machine
learning approaches require extensive feature engineering and may not generalize well across
different transformer types and operational conditions.
Recent advancements in deep learning, particularly Convolutional Neural Networks (CNNs),
provide a promising alternative. CNNs are capable of automatically learning hierarchical
features from raw data, making them highly suitable for complex pattern recognition tasks. In
PD localization, CNNs can analyze spatial and temporal features of PD signals, potentially
improving accuracy and robustness. Additionally, Time Reverse Modelling (TRM) is a signal
processing technique that reconstructs the original signal by reversing it in time, enhancing
resolution and localization accuracy particularly in noisy environments where traditional
methods falter. This research proposes an integrated approach that combines CNNs with TRM
to address limitations of conventional PD localization methods. By leveraging CNNs’ featurelearning
capabilities and TRM’s signal reconstruction benefits, the approach aims to achieve
higher accuracy, efficiency, and scalability in PD localization. The study begins with a
comprehensive review of existing literature on PD localization, examining the limitations of
traditional methods and recent advances in CNNs and TRM. The methodology involves
designing a CNN-TRM framework tailored to analyze PD signals, capturing spatial and
temporal features, and applying TRM for signal enhancement. The model is evaluated against
traditional methods (SVM) on metrics such as localization accuracy, computational efficiency,
and robustness to noise.
Page 6 of 126
Through case studies, the CNN-TRM approach is tested under varied conditions, including
different noise levels, transformer types, and environmental factors. Results demonstrate that
this integrated approach significantly outperforms traditional methods, particularly in
challenging scenarios where noise and signal distortions are prevalent, marking a step forward
in reliable, real-world PD localization for power transformers.