Overview of disruption prediction at JET during the ILW experimental campaigns

The advanced predictor of disruptions, APODIS, has been working in the JET real time network since the beginning of the ILW campaigns. APODIS is a data driven system based on a multilayer structure of SVM classifiers. APODIS was trained with JET data corresponding to carbon wall discharges between April 2006 and October 2009. The total number of training discharges was 8407: 7648 non-disruptive discharges and 512 non intentional disruptions. This article has two main parts. Firstly, the APODIS disruption prediction capabilities in the period July 2013-October 2014 are analysed. During these experimental campaigns (over 1059 non-disruptive discharges and 390 non intentional disruptions), APODIS shows 2.46% of false alarms, 14.62% of missed alarms, 2.56% of tardy detections (alarms triggered with less than 10ms), 3.33% of premature alarms (alarms triggered with more than 1.5s) and, finally, 79.49% of valid alarms (warning times between 10ms and 1.5s). It is important to note that the average warning time of valid alarms is 262ms and the standard deviation is 293ms. As mentioned, the above results are obtained with APODIS trained with C wall data and without any retraining in spite of its use with metallic wall discharges. The purpose of the second part of this study is to compare the APODIS results with predictors trained with the JET data from the ILW campaigns. A single predictor, which has been trained with data between September 2011 and July 2012 (1036 non-disruptive discharges and 201 non intentional disruptions), has been developed. It has been tested with experimental data in the period July 2013 - October 2014 (1051 non-disruptive discharges and 390 non intentional disruptions). It shows 2.1% of false alarms, 9.23% of missed alarms, 10.51% of tardy detections, 4.1% of premature alarms and, finally, 76.16% of valid alarms. The average warning time of valid alarms is 263ms and the standard deviation is 276ms.