Speaker
Alessandro Pau
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P2.1049.pdf
A machine learning approach towards a disruption prediction and
avoidance system: developments and perspectives
A. Pau1, A. Fanni1, B. Cannas1, S. Carcangiu1, G. Sias1, P. Sparapani1, E. Alessi2, C. Sozzi2,
M Baruzzo3, E. Joffrin4, P.J. Lomas5, A. Murari3, F. Rimini5, and JET Contributors*
EUROfusion Consortium, JET, Culham Science Centre, Abingdon, OX14 3DB, UK
1
Electrical and Electronic Eng. Dept. University of Cagliari, Italy
2
IFP-Consiglio Nazionale delle Ricerche, Milano, Italy
3
Consorzio RFX-Associazione - EURATOM ENEA per la Fusione, Padova, Italy
4
CEA, IRFM, F-13108 St Paul Les Durance, France.
5
CCFE, Culham Science Centre, OX14 3DB Abingdon, UK
* See X. Litaudon et al. Nucl. Fusion 57, 102001
Disruptive events still represent one of the main concerns for the protection of in-vessel
components of large size tokamaks, imposing several constraints on the design of the next
step experimental devices such as ITER and DEMO. This work aims to summarize the efforts
in the development of an innovative machine learning approach, based on a generative model,
towards the implementation of a disruption prediction and avoidance system.
[1]
In the proposed approach the first step is the construction of a reliable database and to the
proper selection of the discharge phases of interest for the study: the analysis, in particular,
will be mainly focused on the flat-top phase of the plasma current. In order to effectively
extract the information contained in the raw signals, a feature engineering approach has been
combined with the definition of physics-based indicators related to more structured spatial
and/or temporal information, such as the time evolution of kinetic plasma profiles, the spatial
distribution of the radiation and MHD rotating modes. In this framework, the potential of a
machine learning tool [2] built upon the Generative Topographic Mapping [3] algorithm will be
discussed emphasizing the effectiveness of the tool for the investigation of the operational
[4]
space where the relevant physics takes place . Typical patterns, describing different
processes and characterizing different types of disruption, will be compared for different
scenarios developed at JET with the ILW, extending the analysis presented in [5] to the recent
high power experimental campaign carried out in 2016. The paper will discuss how the
operational boundaries appearing in the considered parameters space are potentially modified
and how this could affect the definition of robust disruption avoidance schemes.
[1] A Pau et al 2017 “A tool to support the automatic construction of reliable disruption databases”, FED
http://dx.doi.org/10.1016/j.fusengdes.2017.10.003.
[2] A Pau 2014 “Techniques for prediction of disruptions on TOKAMAKS”, http://paduaresearch.cab.unipd.it/6664/
[3] Bishop C., Svensén M., Williams C. (1998), Neural Comp.10 215–34.
[4] B Cannas et al 2015 “Automatic disruption classification in JET with the ITER-like wall”, PPCF 57 125003
[5] A Pau et al 2017 “A first analysis of JET plasma profile-based indicators for disruption prediction and avoidance”, 27th
IEEE Symposium on Fusion Engineering, Shanghai, China, under review on IEEE TPS
