AI-Based Segmentation of Power System Faults using Rms Values : A GRU-Attention Framework for Time-Series Fault Segmentation in Electrical Grids

PQD such as voltage sags, swells, and interruptions present significant risks to the stability and efficiency of modern power systems. Traditional approaches for segmenting these events often rely on rule-based thresholding or raw waveform analysis, which can be both inaccurate and computationally expensive. This thesis proposes a deep learning-based approach incorporating GRU to segment PQD in power systems by leveraging features extracted from rms voltage and current signals. A comprehensive pipeline is developed, incorporating half-cycle rms and derivative based feature extraction, sequence windowing, and supervised learning. Multiple deep neural architectures are evaluated to detect and localize the start and end times of disturbances within time-series data. The dataset used here contains signals which belong to PQD with a sampling rate of 4000 Hz. This means the PQ monitors collected 4000 data points in a second. Each signal is composed of multichannel three-phase time-series data, with annotations indicating the start and end times of PQD. The methodology is designed to support various input scenarios, including single-phase and multi-phase signals, enabling flexible and scalable deployment for real-world grid monitoring systems. This work contributes to advancing intelligent fault segmentation systems and supports ongoing efforts in smart grid automation and reliability improvement.