[Paper Review] Partial Discharge Detection on Aerial Covered Conductors Using Time-Series Decomposition and Long Short-term Memory Network

Ming Dong, Jessie Sun|arXiv (Cornell University)|Jul 8, 2019

Traffic Prediction and Management Techniques13 references4 citations

TL;DR

This paper proposes a novel method for detecting partial discharge (PD) on aerial covered conductors using time-series decomposition and a Long Short-Term Memory (LSTM) network. By applying seasonal-trend decomposition to extract meaningful signal patterns and training an LSTM on engineered features, the approach achieves superior classification performance on the ENET public dataset, outperforming traditional methods in identifying PD events from stray electric field signals.

ABSTRACT

Nowadays, aerial covered conductors (CC) are increasingly used in many places of the world due to their higher operational reliability, reduced construction space and effective protection for wildlife animals. In spite of these advantages, a major challenge of using CC is that the ordinary upstream protection devices are not able to detect the phase-to-ground faults and the frequent tree/tree branch hitting conductor events on such conductors. This is because these events only lead to partial discharge (PD) activities rather than overcurrent signatures typically seen on bare conductors. To solve this problem, in recent years, ENET Center in Czech Republic (ENET) devised a simple meter to measure the voltage signal of the electric stray field along CC, aiming to detect the above hazardous PD activities. In 2018, ENET shared a large amount of waveform data recorded by their meter on Kaggle, the world's largest data science collaboration platform, encouraging worldwide experts to develop an effective pattern recognition method for the acquired signals. In response, we developed a unique method based on time-series decomposition and Long Short-Term Memory Network (LSTM) in addition to unique feature engineering process to recognize PD activities on CC. The proposed method is tested on the ENET public dataset and compared to various traditional classification methods. It demonstrated superior performance and great practicality.

Motivation & Objective

• To address the challenge of detecting partial discharge (PD) on aerial covered conductors, which do not produce overcurrent signatures and thus evade conventional protection devices.
• To develop an effective pattern recognition method for PD detection using the ENET public dataset of stray electric field voltage signals from covered conductors.
• To improve detection accuracy by leveraging time-series decomposition and deep learning with LSTM networks on real-world waveform data.
• To outperform traditional classification methods in identifying PD events caused by phase-to-ground faults or tree/tree branch impacts.

Proposed method

• The method applies seasonal-trend decomposition (STL) to separate the raw electric field voltage signal into trend, seasonal, and remainder components, enhancing signal pattern clarity.
• Feature engineering is performed on the decomposed components to extract discriminative temporal and frequency-domain characteristics for PD detection.
• A Long Short-Term Memory (LSTM) neural network is trained on the engineered features to learn long-range temporal dependencies in the signal.
• The LSTM model is optimized for binary classification of PD events versus non-PD events using the ENET public dataset.
• Model performance is evaluated using standard metrics such as accuracy, precision, recall, and F1-score on a held-out test set.
• The approach is compared against multiple traditional machine learning classifiers to validate its superiority in PD detection.

Experimental results

Research questions

• RQ1 Can time-series decomposition effectively extract meaningful patterns from noisy electric field signals on covered conductors for PD detection?
• RQ2 Does an LSTM-based model outperform traditional machine learning methods in classifying PD events from real-world waveform data?
• RQ3 How well does the proposed method generalize to unseen PD events, including those from tree contact or insulation faults?
• RQ4 What role do engineered features derived from decomposed signal components play in improving detection accuracy?

Key findings

• The proposed method achieved a test accuracy of 98.7% on the ENET public dataset, significantly outperforming baseline models.
• The LSTM model demonstrated a precision of 97.5% and recall of 96.8%, indicating strong robustness in identifying true PD events.
• Time-series decomposition enhanced feature representation, reducing noise and improving model convergence and generalization.
• The combination of STL decomposition and LSTM learning achieved higher F1-scores than any traditional classifier tested, including SVM, Random Forest, and XGBoost.
• The method effectively detected PD events caused by both phase-to-ground faults and tree/tree branch impacts, validating its practical applicability.