Review of disruption predictors in nuclear fusion: Classical, from scratch and anomaly detection approaches

J. Vega, R. Moreno, A. Pereira, G.A. Rattá, A. Murari, S. Dormido-Canto, S. Esquembri, E. Barrera, M. Ruiz

Research output: Contribution to conferencePaper

Abstract

Disruption predictors are implemented as data-driven models obtained from machine learning techniques. These data-driven models are deduced from a training process with thousands of discharges (both disruptive and non-disruptive). ITER or DEMO, the next step devices cannot afford to wait for hundreds of disruptions to start predicting. A novelty approach for disruption prediction is to avoid the use of past discharges for learning purposes. The objective is to learn in every discharge how a safe plasma evolution is and to trigger an alarm when anomalies in the data flow appear. Of course, these anomalies have to be signatures of the phenomenon's precursors. By applying these ideas in JET to a dataset of more than 1700 non-disruptive shots and more than 550 that ended in a disruption, the success rate is about 90% and the false alarm rate is slightly above 7%. On average, the alarm is triggered with an anticipation time above 200 ms.
Original languageEnglish
DOIs
Publication statusPublished - 21 Dec 2016
Externally publishedYes
Event42nd Conference of the Industrial Electronics Society, IECON 2016 - Florence, Italy
Duration: 21 Dec 2016 → …

Conference

Conference42nd Conference of the Industrial Electronics Society, IECON 2016
CountryItaly
CityFlorence
Period21/12/16 → …

    Fingerprint

All Science Journal Classification (ASJC) codes

  • Control and Systems Engineering
  • Electrical and Electronic Engineering

Cite this

Vega, J., Moreno, R., Pereira, A., Rattá, G. A., Murari, A., Dormido-Canto, S., ... Ruiz, M. (2016). Review of disruption predictors in nuclear fusion: Classical, from scratch and anomaly detection approaches. Paper presented at 42nd Conference of the Industrial Electronics Society, IECON 2016, Florence, Italy. https://doi.org/10.1109/IECON.2016.7794130