Speaker
Takeo Nishitani
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P4.1082.pdf
Initial Results of a Machine Learning-based Real Time Disruption
Predictor on DIII-D
C. Rea1, R.S. Granetz1, N. Eidietis2, K. Erickson3, M.D. Boyer3, R. Johnson2
1
MIT Plasma Science & Fusion Center, Cambridge, MA, US
2
General Atomics, San Diego, CA, US
3
Princeton Plasma Physics Laboratory, Princeton, NJ, US
A disruption prediction algorithm, developed using machine learning, runs in real time in the
DIII-D plasma control system (PCS), and accurately predicts impending major disruptions
with several hundred milliseconds warning time, while also having a very low rate of false
alarms. The algorithm is based on the Random Forests machine learning method, and has
been developed starting from an extensive database of more than 10000 DIII-D discharges,
both disruptive and non-disruptive ones. The algorithm uses 9 plasma parameters that are
derived from several real time diagnostic signals and real time EFIT equilibrium
reconstructions, which are provided by the PCS on a cycle time of 250 µs. Most of the
parameters are dimensionless (e.g. li, βp, …) or cast in a dimensionless form (e.g. n/nG,
Bpn=1/Btor …), which facilitates multi-machine analyses. The prediction algorithm was trained
on all types of major disruptions occurring during the flattop phase, without differentiation
by cause, and the initial results do indeed show good success at recognizing multiple types
of major disruptions during the flattop, and even during the rampdown phase of discharges.
A reliable prediction warning time of several hundred milliseconds allows, at least
conceptually, for the possibility of actively avoiding an impending disruption, if the specific
cause(s) of the disruption can be identified, and if control ‘knobs’ exist to modify the
identified cause(s). However, although Artificial Intelligence (AI) methods can accurately
make predictions, it is not well-understood how to determine which input features are
responsible for the output prediction. Determining how to do this is currently a high-priority
topic of AI research, which we are now pursuing in order to effectively close the disruption
avoidance control loop.
This work was supported by the U.S. Department of Energy under DE-FC02-04ER54698,
DE-SC0014264 and DE-AC02-09CH11466.
DISCLAIMER: This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States
Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for
the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately
owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not
necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions
of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.
