Machine Learning Prediction for Experimental Quench Voltage of HTS Coils Using a Combined Framework With Time-Series Generative Adversarial Network and Long Short-Term Memory Techniques

Fast and reliable quench detection is critical for protecting high temperature superconducting coils, magnets, and devices, yet experimental datasets remain limited due to cryogenic constraints and risk of conductor damage. This limitation is one of the key challenges in developing machine learning (ML) techniques for predicting quench through regression-based forecasting of voltage signals. To address this, we propose a framework that combines Time-series Generative Adversarial Networks (TimeGAN) for artificial sequence generation with Long Short-Term Memory (LSTM) predictors. From 36 measured voltage sequences, TimeGAN generated 90 artificial sequences to expand the dataset and mitigate overfitting. Augmenting the LSTM improved quench prediction accuracy to ∼3% mean relative error. This study demonstrates the first application of TimeGAN to superconducting quench signals and highlights the benefit of ML-driven augmentation for improving quench prediction robustness.