使用機器學習進行輸電線路故障診斷

Abstract

本研究旨在探索與完善輸電線路故障診斷技術，開發基於機器學習之故障檢測、故障類型分類、故障原因識別和非同步量測校正方法。本研究採用機器學習模型Conformer來學習序列資訊。本研究主要提出二種架構之機器學習模型，一為基於多任務監督式機器學習，另一為自監督和無監督式機器學習之方法。 多任務監督式機器學習模型解決了故障檢測、故障類型分類和非同步量測校正同步。透過使用滑窗法與最小化前處理來降低計算複雜度。該模型生成兩種類型的輸出：第一種輸出處理故障檢測和分類，區分不對稱故障、對稱故障、不對稱輸電線斷線故障和無故障。第二種輸出指示故障在序列中開始的取樣點，便於測量同步。本文提出的方法將對比其他先進的方法或模型。 在故障原因識別方面，我們回顧並總結了現有文獻。基於歷史記錄的稀缺性，我們提出了一種基於模擬數據的自監督和無監督式機器學習技術的解決方案。 本論文提出之方法將透過歷史事故資料驗證。通過整合現有的輸電線路參數估測和故障定位演算法，開發出一種全面的故障診斷系統。

This research aims to explore and improve fault diagnosis techniques for transmission lines by developing machine learning-based methods for fault detection, fault type classification, fault cause identification, and asynchronous measurement synchronization. This research employs the Conformer: Convolution-augmented Transformer model to learn sequential information. Two main machine learning model architectures are proposed: one based on multi-task supervised learning, and the other on self-supervised and unsupervised learning methods. The multi-task supervised machine learning model effectively handles fault detection, fault type classification, and asynchronous measurement synchronization with minimal pre-processing of input feature sequences. The sliding window technique is utilized to reduce computational complexity. The model produces two types of outputs: the first addresses fault detection and classification, distinguishing between asymmetric faults, symmetric faults, asymmetric conductor broken faults, and no fault. The second output indicates the sample point at which the fault begins in the sequence, facilitating measurement synchronization. The evaluation between the proposed methodology and other state-of-the-art technique is present. For fault cause identification, we review and summarize existing literatures. Given the scarcity of field records, we propose a feasible solution using self-supervised and unsupervised machine learning techniques based on simulation data. Historical events are used to validate the proposed methodology in this thesis. By integrating existing transmission line parameter estimation and fault location algorithms, a comprehensive fault diagnosis system is developed.