Real-Time Fault Diagnosis in UHV Power Systems Using Bayesian-Optimized 1D CNN on Raw Time-Series Signals

Hani Albalawi^1,2, Rab Nawaz³, Abdul Wadood^1,2,*, Anwar Ul Haq³, Shahbaz Khan^1,2, Bakht Muhammad Khan¹, Aadel Mohammed Alatwi^1,2
1 Zero Emission Technologies Innovation Center, University of Tabuk, Tabuk, Saudi Arabia
2 Electrical Engineering Department, Faculty of Engineering, University of Tabuk, Tabuk, Saudi Arabia
3 Department of Electrical Engineering, Mirpur University of Science and Technology, Mirpur, Azad Jammu and Kashmir, Pakistan
* Corresponding Author: Abdul Wadood. Email: email
(This article belongs to the Special Issue: Advanced Artificial Intelligence and Machine Learning Methods Applied to Energy Systems, 2nd Edition)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.084732

Received 28 April 2026; Accepted 24 August 2026; Published online 14 September 2026

Download PDF

Abstract

Ensuring the reliability of Ultra-High Voltage (UHV) power systems remains a critical challenge, as series compensation improves stability while introducing complex fault dynamics. Although machine learning and deep learning methods have advanced fault diagnosis, existing approaches often depend on computationally intensive preprocessing and struggle with data scarcity, limiting real-time applicability. This study proposes a streamlined One-Dimensional Convolutional Neural Network (1D CNN) optimized via Bayesian learning for efficient and robust fault classification in a 735 kV, 32-bus UHV system. The model operates directly on raw time-series signals, eliminating the need for domain-specific transformations while preserving the natural characteristics of the data. Quantitatively, the proposed framework achieves 99.91% test accuracy using combined voltage-current signals, with a narrow standard deviation of

\pm 0.06 %

across multiple evaluations, and correctly classifies 12,998 out of 13,010 test instances with only twelve misclassifications across eleven fault categories. For resource-constrained applications, voltage-only signals deliver 99.72% accuracy while reducing training time by 38.6% (680 s) and memory usage by 61.8% (1272 MB), whereas current-only signals achieve 99.90% accuracy with the fastest inference speed of 2995 samples per second. The model demonstrates strong dependability under noisy conditions, maintaining 97.03% test accuracy even at 10 dB SNR, and exhibits robustness against downsampling with a Sampling Index of 0.0235. The Bayesian optimization process identifies optimal hyperparameters including a learning rate of 0.002611, dropout rate of 0.182, and batch size of 19. From an economic perspective, the elimination of expensive preprocessing, noise cancellation, and high-frequency sampling (typically 0.2–10 MHz) reduces computational resource requirements and operational costs, enabling practical deployment in real-time monitoring systems. These findings demonstrate the practical potential of the proposed approach for cost-effective, scalable fault diagnosis in large-scale power grids.