Artificial Intelligence Applications for Fault Prediction and Failure Prevention in Power Transmission Systems

Abstract

-For reliable power transmission it is essential to identify early signs of degradation before such degradation leads to faults that trigger protection systems, to equipment damage, or to a series of outages. Artificial intelligence (AI) has broadened the range of analysis available in the monitoring of transmission systems by enabling the discovery of nonlinear relationships among electrical waveforms, phasor measurements, indicators of asset condition, environmental variables, images from inspections, and records of events. This review critically looks at the use of AI for predicting faults and preventing failures in transmission lines and the associated high-voltage assets, with a focus on research carried out between 2020 and 2025. The analysis makes a distinction between immediate fault detection and classification on the one hand and anticipatory prediction, condition assessment, fault-cause identification, and support for maintenance decisions on the other. It assesses conventional machine learning, artificial neural networks, convolutional and recurrent architectures, transfer learning, ensemble models, autoencoders, and computer-vision methods in relation to operational needs such as speed, generalisation, tolerance to uncertainty, interpretability, and computational cost. It is shown that deep models are able to extract discriminative features directly from raw or only slightly processed signals, while hybrid approaches that combine signal processing and learning still prove valuable when data are scarce or when physics-based constraints are significant. Nevertheless, heavy reliance on simulated data, class imbalance, domain shift, a lack of failure labels, and insufficient validation in real-world settings still hinder implementation. A preventive architecture is therefore suggested which integrates synchronised sensing, multimodal representation learning, uncertainty-aware inference, risk scoring, and human-supervised maintenance actions. The review concludes that future advances will depend less on small improvements in accuracy and more on trusted generalisation, calibrated uncertainty, cross-asset transfer, explainability, cybersecurity