Transmission lines are an important component of the power system and act as a conduit through which electrical energy is transferred from the power generation station to the wider distribution network. With the recent decades of global growth in electricity demand, maintaining stable and reliable transmission lines has become increasingly essential. However, transmission lines tend to be attributed to environmental conditions such as lightning, wind, and storms, as well as various types of faults caused by aging of equipment, insulation failures, or external mechanical effects. If these failures are detected and not addressed quickly, they can lead to massive power outages, infrastructure damage, and disruption of grid stability. Therefore, accurate and timely fault identification is an important requirement for the safe operation of modern power systems.
Due to the dynamic nature of the power system and the interaction between its numerous components, the process of detecting and classifying faults in electrical transmission lines is inherently complicated. The power lines operate within a vast, interconnected electrical network, including generators, transformers, circuit breakers and protective relays. Because these elements interact continuously with each other, it is essential that the fault detection system not only identifies the occurrence of a fault, but also accurately classifies it according to its type and origin. Transmission system failures can manifest in a variety of ways, from single-line to underground faults, line to line to row faults, double-line to underground faults, or symmetrical three-phase faults. Each type has its own electrical properties and meaning, requiring a coordinated detection and protection strategy to minimize equipment damage and restore system stability.
In modern power systems, fault detection forms the fundamental element of a protection mechanism. When a failure occurs, the protection scheme must quickly and reliably detect events, determine the nature and location of the failure, and trigger appropriate control actions. Traditional fault detection approaches rely heavily on hardware-based solutions such as electromechanical relays and threshold-based mechanisms, and despite their simplicity, they lack adaptability and fail to meet the accuracy and speed requirements of today's dynamic grids. Due to the increased complexity of power systems, such traditional approaches demonstrate the limitations of response times, feature generalizations, and adaptability to various failure scenarios, especially in systems with renewable energy sources and distributed architectures. Recent developments in sensor technology and data collection systems have resulted in intelligent monitoring devices being installed throughout the power supply network. Such devices continuously monitor a variety of electrical parameters such as voltage, current, frequency, and phase angles to provide a large amount of high-resolution data. The use of this data through advanced computational methods is a major factor in creating more accurate and adaptive fault detection models. In particular, model-based methods are increasingly being used to overcome the flexibility of traditional rule-based systems.
Machine learning (ML) and pattern recognition algorithms have proven to be powerful candidates for fault detection and diagnosis. By learning from past or simulated failure data, these systems can automatically develop complex patterns of association between system measurements and failure conditions. For example, we use support vector machines (SVMs), artificial neural networks (ANNs), and decision tree algorithms to successfully classify failures in three-phase transmission lines. These models can identify nonlinear patterns and enhance generalization through monitored learning, improving performance under varying system conditions.
Nevertheless, despite classic ML models that provide superior flexibility and prediction accuracy compared to previous threshold-based methods, they tend to rely on hand-designed features and cannot explain the temporal evolution of electrical signals. This is particularly difficult in cases of time-changing obstacles in subtle but beneficial patterns of electrical parameters. Furthermore, ML models may not be as robust as necessary for real-time use, especially when noise and behavior variations are present. These challenges motivate the use of deep learning techniques, allowing the user to automatically learn relevant functional representations from raw input data without the need for cumbersome preprocessing.
Several deep learning models have been proposed for electrical failure diagnosis, including convolutional neural networks (CNNS), long-term memory (LSTM) networks, and hybrid models. These models can learn complex spatial and temporal relationships within the data. Although effective, some models are difficult to balance model complexity, interpretability, and computational costs. Furthermore, most models do not have a mechanism to concentrate on the most beneficial areas of input data, resulting in less efficiency and vulnerability to unrelated information.
To overcome these drawbacks, in this paper, we propose a attention-glu-based fault classifier (AGFC-NET), a deep learning model that integrates the convolutional layers, attention mechanisms, and the advantages of gated recurrence units (Grus), as shown in Figure 1. The attention mechanism receives these features. These features selectively pay attention to the most appropriate signal sections to minimize noise effects and maximize interpretability. The GRU layer is used later to detect backward dependencies, allowing the model to understand how electrical parameters change over time during fault conditions. This blend helps AGFC-NET achieve improved classification performance and maintain computational efficiency and noise robustness. The dataset employed in this study consists of approximately 12,000 annotated samples representing both the faulty and normal states of an emulated three-phase transmission system. Each sample includes six features: line voltage (VA, VB, VC) and line current (IA, IB, IC) captured under various fault types and conditions. The defects range across all combinations of possible phase failures, covering differences in fault location and impedance, providing a wide, comprehensive training set. This annotated dataset is ideal for monitored learning as the model can learn the identification features that distinguish between different types of faults and normal operations.
The main goal of this study is to create a powerful, accurate and scalable model that can perform real-time fault detection and classification on power lines. AGFC-NET aims to address the problems of traditional machine learning and deep learning models by incorporating attention-based learning and a GRU architecture. The long-term objective is to contribute to the development of intelligent fault detection systems that are not only capable of recognizing complex fault patterns, but also suitable for implementation within modern power grids that require real-time monitoring and decision-making.
