Advances in Adaptive Learning Methods for Disruption Prediction in JET

The development of disruption predictors based on machine learning methods requires the use of a training dataset from past discharges. In order to cover as much as possible the parameter space of the predictor, a large number of discharges (the more the better) from several operation scenarios is necessary. This approach is no longer valid for ITER or DEMO. Disruption predictors have to learn in an adaptive way from scratch, i.e. without information of past discharges. The paper shows an adaptive approach of a disruption predictor for mitigation purposes based on the nearest centroid method. Only the mode lock signal normalised to the plasma current (ML/Ip) is used. The predictor parameter space is a two-dimensional space, whose dimensions are the amplitudes of the ML/Ip signal of two consecutive samples. The adaptive predictor has been developed and tested with the JET database. A set of 1510 discharges (113 disruptive and 1397 non-disruptive) has been used. The training process start with only one disruptive sample and five non-disruptive ones to compute the disruptive and non-disruptive centroids. A new retraining is accomplished when an alarm is missed or when a tardy detection is produced. In the analysis of 1510 discharges, only 5 re-trainings have been performed (1 missed alarm and 4 tardy detections). The global fraction of detected disruptions is 99.11%, which corresponds to a success rate (predictions with positive warning time) 95.54% and a tardy detection rate of 3.57%. The false alarm rate has been 1.58% and the average warning time 247 ms. Therefore, adaptive predictors for ITER or DEMO can be valid candidates to address the problem of reliable disruption prediction from scratch.