SMuK 2021 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 13: Magnetic Confinement IV & Helmholtz Graduate School IV
P 13.2: Talk
Thursday, September 2, 2021, 14:30–14:55, H4
Multi-class disruption prediction at JET using a shapelet based neural-network. — •Victor Artigues1, Frank Jenko1, and JET Contributors2 — 1Max Planck Institute for Plasma Physics, Boltzmannstr. 2, 85748 Garching, Germany — 2See the author list of `Overview of JET results for optimising ITER operation' by J. Mailloux et al. to be published in Nuclear Fusion Special issue: Overview and Summary Papers from the 28th Fusion Energy Conference (Nice, France, 10-15 May 2021)
Disruptions, the very fast, uncontrolled, termination of plasma experiments in tokamaks, remain to this day an unsolved issue on the path towards fusion-based power plants. Due to the their complex nature, disruptions have been very hard to investigate with physics-based approaches. In recent years, progress has been made with data-driven methods to build disruption detection systems, but many questions remain open such as disruption type identification, or transfer between tokamaks.
We propose a Shapelet based neural-network for the task of multi-class disruption prediction, and compare it to two approaches from the literature, trained on our data: stacked Support-Vector Machines (SVM), and a Long Short-Term Memory (LSTM) neural-network. Two datasets of discharges from the Joint European Torus (JET) tokamak, have been compiled. One containing stable discharges and 7 different disruption types, before the installation of the ITER-Like Wall (ILW). The second, with fewer shots and binary classification, from the more recent C36 campaign with ILW. Using the binary and multi-class classification results on the different datasets, we report on the performance of the three models and discuss the advantages of our method.