Increased warning times in JET APODIS disruption predictor by using confidence qualifiers

JET APODIS disruption predictor follows a two - layer architecture of Support Vector Machines (SVM) classifiers. The 1 st layer contains three independent SVM classifiers that opera te in parallel. As a discharge is in execution, the classifiers use the three most recent time windows of 32 ms to make predictions about the disruptive or non - disruptive plasma behaviour. Of course, the three predictions are not necessarily the same. Ther efore, the 2 nd layer is used to combine the above three outputs into a single one. The output of the 2 nd layer determines whether or not to trigger an alarm. This paper shows a modification of the APODIS 2 nd layer classifier to obtain larger warning times, which is essential to put into operation earlier mitigation actions. The modification of the 2 nd layer classifier is based on qualifying every 32 ms how good both predictions (disruptive and nondisruptive) are. This qualification is carried out in the fra mework of conformal predictions. Given a set of examples { ? ? = ( ? ? , ? ? ) , ? = 1 , ... , ? } , where ? ? ? R ? are feature vectors and ? ? ? ? = { ? ? , ... , ? ? } are labels, conformal predictors determine the label of a new feature vector x n+1 by assuming for it all possible labels and predicting the la bel that makes x n+1 the most conformal to the initial set of n examples. The conformity measure is determined by computing the p - values ( 1 / ( ? + 1 ) <= ? - ????? <= 1 ) for all possible labels (V. Vovk et al. Proc. 16 th Int. Conf. on Machine Learning. (1999). Sa n Francisco, CA), where the minimum value means the least conformal label and the maximum value means the most conformal one. In the case of the APODIS, only two labels are possible ? = { ?????????? , ??? - ?????????? } . To increase the war ning time, an alarm is triggered when simultaneously both assumptions (disruptive and non - disruptive) are qualified with the minimum p - values. In other words, an alarm is triggered when the statistical confidences in both labels are minimum. This has been applied to 789 JET discharges in the range 82429 - 83793 (81 unintentional disruptions predicted by APODIS and 708 non - disruptive discharges). All disruptions are predicted and 94% of them have a warning time >10 ms (minimum time in JET to trigger the disrup tion mitigation valve). The average warning time of the classical APODIS predictor is 428 ms with a standar deviation of 1166 ms. With the new second layer classifier, the average warning time is 606 ms with a standard deviation of 1874 ms. The false alarm rate increases from 1% to 5%.