Disruption Prediction on JET during the ILW Experimental Campaigns. (1st April 2016)

• Record Type: Journal Article
• Authors: Moreno, R.; Vega, J.; Dormido-Canto, S.; Pereira, A.; Murari, A.
• Is Part Of: Fusion science and technology. Volume 69:Number 2(2015)
• Page Start: 485
• Page End: 494
• Publication Date: 2016-04-01
• Subjects: Fusion reactors -- Periodicals; Nuclear fusion -- Periodicals; Fusion reactors; Nuclear fusion; Periodicals; 621.48405
• DOI: 10.13182/FST15-167
• Languages: English
• ISSNs: 1536-1055

Abstract

Abstract: The Advanced Predictor of Disruptions (APODIS) has been working in the JET real-time network since the beginning of the ITER-like wall (ILW) campaigns. APODIS is a data-driven system based on a multilayer structure of Support Vector Machines (SVM) classifiers. APODIS was trained with JET data corresponding to carbon wall discharges between April 2006 and October 2009, without any retraining in spite of its use with metallic wall discharges. This paper has two main parts. First, APODIS disruption prediction capabilities are evaluated during the ILW run period from July 2013 to October 2014. The results obtained from these experimental campaigns (more than 1059 nondisruptive discharges and 390 nonintentional disruptions) showed 2.46% false alarms and 85.38% success rate. Taking into account that ITER (International Thermonuclear Experimental Reactor) will work with a similar wall to the current ILW at JET, the purpose of the second part of this study is to compare predictors trained with data from JET ILW campaigns. The high computational cost that APODIS training requires and the great performance of SVM have motivated the development of a one-layer SVM predictor. Therefore, an APODIS version and a simpler one-layer predictor have been compared. They have been trained with data between September 2011 and July 2012 (1036 nondisruptive discharges and 201 nonintentional disruptions) and tested with experimental data in the period July 2013 to October 2014 (1051 nondisruptive discharges and 390 nonintentional disruptions). The one-layer predictor shows slightly better results than the APODIS structure.