Real-Time Fault Risk Forecasting on Transmission Lines Using Hybrid Deep-Learning Models and Generative Data Augmentation

This paper presents a probabilistic fault-risk forecasting framework that integrates time-series generative modelling with deep sequence learning to enhance early identification of transmission line fault conditions using multivariate Phasor Measurement Unit (PMU) data.

The approach addresses data scarcity and class imbalance while providing a graded PMUderived risk signal to complement traditional protection and support earlier operator actions particularly in unconventional systems. Historical PMU measurements of three-phase voltages and currents were used to construct multivariate time-series for normal and fault operating conditions. . A preprocessing pipeline was applied that includes feature scaling using a train-only min–max normalization and segmentation into fixed-length overlapping time windows. To address class imbalance, a generative adversarial network specifically designed for time-series data was employed for data augmentation. A controlled scarcity stress-test evaluates performance under rare-event conditions.

The augmented dataset was used to train a hybrid deep-learning architecture combining convolutional neural networks (CNNs) for multichannel feature extraction and long shortterm memory (LSTM) networks for temporal dependency modeling. The forecasting model predicts multivariate PMU trajectories. Deviations between predicted and observed trajectories are quantified using a multivariate distance-based metric, from which a probabilistic fault-risk score supporting early warning and situational awareness.

Experimental results show that, on the segmented PMU dataset, the proposed framework achieves near-perfect window-level discrimination between normal and faulted regimes .

Under a controlled scarcity stress-test, detection performance degrades sharply when trained only on limited historical fault data (recall score = 0.019, F1-score = 0.03, ROC AUC = 0.019). When synthetic fault windows generated by the proposed augmentation framework are incorporated during training, detection performance is restored (recall score = 1.000, F1score = 1.000, ROC AUC = 1.000). These results demonstrate improved sensitivity to faultrelated PMU patterns under data-scarce conditions.

By integrating generative augmentation with deep probabilistic forecasting, this work contributes a scalable approach to enhancing transmission system resilience in low-inertia power systems. The proposed framework is particularly relevant for emerging transmission systems with unconventional sources, where reduced inertia and converter-dominated dynamics can accelerate disturbance propagation and reduce available operator response time.

Because the methodology is data-driven and based on multivariate PMU forecasting, it provides a flexible monitoring approach that can be extended to evolving grid conditions without relying on explicit modeling of specific generation technologies. The framework is extendable to other corridors and operating regimes.