München 2019 – wissenschaftliches Programm
P 18.93: Poster
Donnerstag, 21. März 2019, 16:30–18:30, Foyer Audimax
Analysis of disruption prediction methods on a per disruption-cause basis. — •Victor Artigues and Frank Jenko — Max Planck Institute for Plasma Physics, Boltzmannstr.2, 85748 Garching, Germany
The main approach in disruption prediction research using machine learning methods is to compile a database made of disruptive shots and safe shots, with little regard for the cause of disruption. The disruptive shots are all combined under one label. Multiple causes of disruption have been identified in a study on a large number of JET shots. On the one hand, splitting the databases with the different causes of disruption can ease the learning process and give a better understanding of the prediction and its link to physics. On the other hand, it is well known that reducing the size of the datasets will be detrimental to the prediction.
As a first step towards a cause-by-cause disruption prediction system, we analyzed the performances of state-of-the-art disruption prediction methods when trained on datasets separating the different causes. Our study is conducted on the shots from the ASDEX-Upgrade Tokamak, using a Support Vector Machine (SVM) model such as the one used at JET and a Long-Short Term Memory (LSTM) artificial neural network. We compare the ease of prediction regarding the different types and discuss future work such as data augmentation to deal with the smaller datasets.