A Machine Learning Predictor for Vertical Displacement Events on EAST

Abstract

To obtain higher discharge parameters in experimental advanced superconducting tokamak (EAST) experiments, an elongated plasma configuration is applied. As a result, vertical displacement events (VDEs) are easy to occur, especially for plasma with high vertical instability growth rate. Studying the prediction and avoidance of VDEs is of great importance for the protection of the plasma-facing and structural components of EAST tokamak. Data-driven methods based on supervised learning are widely used in disruption prediction. Labels as the key of the supervised learning are difficult to accurately divide. In this article, we first propose a labeling method based on the Jensen-Shannon (JS) divergence, enabling a specific analysis and evaluating the precursor onset time for each discharge. By comparing prediction accuracy and warning time prior to disruption using different algorithms with dataset collected from EAST experiment of 2021–2023, it is found that random forest (RF) model works best and achieves a success VDE alarm rate of 99.2% with a false alarm rate of 2.1% for nondisruptive discharges. The results show that models trained with dataset collected with class split point found by the JS divergence method from each discharge outperforms models trained with the dataset collected from each discharge with class split point of fixed time before disruption.