트랜스포머 모델을 이용한 전력선의 부분 방전 탐지

Abstract

Power line maintenance has become a major issue as the demand for electricity rise every year. Damaged power line casued by poor maintenance leads to partial discharge, causing blackouts and fire. In this research, we have conducted partial discharge detection based on the power line data from ENET Centre of VSB Technical University in Ostrava, Czech Republic. Power line data are electrical signal data, which can be easily converted to sequence list. Recurrent Neural Network (RNN) model shows its best performance on analyzing sequence list, so we ran our first analysis using RNN. Yet RNN model had a problem that it took a long time

to calculate because of its high calculation cost. As a supplement for this problem, we used a Transformer model built with Attention Mechanism to run our last analysis. Using F1-Score and MCC (Matthews Correlation Coefficient) values as comparison standards, we were able to confirm that Transformer model has drastically reduced the time required for calculation and has shown higher performance compared to eariler RNN model.|전력의 소비량이 매년 증가함에 따라 전력선의 유지와 관리는 중요한 문제로 대두되고 있다. 전력선의 관리가 올바르게 되지 않아 발생하는 손상은 부분 방전으로 이어지며, 잦은 부분 방전은 정전과 화재 등을 유발한다. 본 연구에서는 체코 Ostrava 에 위치한 VSB Technical University 의 ENET Centre 에서 제공한 전력선 자료에 대하여 부분 방전 탐지를 진행하였다. 전력선 자료는 전기 신호 데이터로 순차열의 구조로 바꾸기 용이하며 이러한 순차열 구조에 뛰어난 성능을 보이는 순환신경망 모델을 사용하여 1 차 평가를 진행하였다. 하지만 순환신경망 모델은 계산 비용이 커서 계산 시간이

오래 걸리는 문제가 있어, 이러한 단점을 보완할 수 있는 어텐션 메커니즘으로 구축한 트랜스포머 모델을 사용하여 최종 분석을 진행하였다. 기존 순환신경망 모델과 F1-Score, MCC (Matthews Correlation Coefficient) 값을 통해 성능을 비교하였으며, 트랜스포머 모델은 계산 시간을 비약적으로 줄이며 성능 역시 기존 대비 향상된 것을 확인할 수 있었다.; Power line maintenance has become a major issue as the demand for electricity rise every year. Damaged power line casued by poor maintenance leads to partial discharge, causing blackouts and fire. In this research, we have conducted partial discharge detection based on the power line data from ENET Centre of VSB Technical University in Ostrava, Czech Republic. Power line data are electrical signal data, which can be easily converted to sequence list. Recurrent Neural Network (RNN) model shows its best performance on analyzing sequence list, so we ran our first analysis using RNN. Yet RNN model had a problem that it took a long time

to calculate because of its high calculation cost. As a supplement for this problem, we used a Transformer model built with Attention Mechanism to run our last analysis. Using F1-Score and MCC (Matthews Correlation Coefficient) values as comparison standards, we were able to confirm that Transformer model has drastically reduced the time required for calculation and has shown higher performance compared to eariler RNN model.