Speaker
Tatsuya Yokoyama
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P2.1079.pdf
Disruption Prediction with Sparse Modeling by Exhaustive Search
Tatsuya Yokoyama1 , Takamitsu Sueyoshi1 , Yuya Miyoshi2 , Ryoji Hiwatari2 ,
Yasuhiko Igarashi3 , Masato Okada1 , and Yuichi Ogawa1
1 Graduate School of Frontier Science, The University of Tokyo, Kashiwa, Japan
2 Rokkasho Fusion Institute, QST, Rokkasho, Japan
3 Japan Science and Technology Agency, PRESTO, Saitama, Japan
Plasma disruption is one of crucial phenomena in a tokamak fusion reactor. To realize nuclear
fusion reactor, it is necessary to elucidate and control it. However, its physical mechanism is not
clearly identified yet, so there are some studies trying to predict occurrence of disruptions based
on experimental data.
In this research, we constructed disruption predictor using a support vector machine(SVM)
based on the large experimental data in JT-60U and feature extraction by sparse modeling was
carried out. The concept of sparse modeling exploits the inherent sparseness that is common to
all high-dimensional data and enables us to efficiently extract the maximum amount of informa-
tion from data. For the sparse modeling, we used exhaustive search with SVM, assuming that
the optimal combination of explanatory variables is K-sparse [1].
We have obtained some results showing that feature extraction can contribute to improvement
of disruption prediction performance and understanding of the physical background of disrup-
tion. As a variable before narrowing down, we chose 17 parameters from physical knowledge.
We selected normalized beta and plasma internal inductance because we use results of high-
beta experiment. We also selected safety factor 95% of minor radius, and these parameters are
obtained from MHD equilibrium calculation. We use not only those parameters’ value, but also
time derivative value. In our results, 6 parameters including mode lock amplitude and its time
derivative are extracted as the optimal combination of parameters.
We will try to specify dangerous parameter area where disruption is likely to occur using
sparse modeling.
References
[1] Y. Igarashi, H. Takenaka, Y. Nakanishi-Ohno, M. Uemura, S. Ikeda, and M. Okada. Exhaustive search for
sparse variable selection in linear regression, 2017.
