Unsupervised Clustering of Disruption Alarms in JET

Dealing with disruptions is a main research subject in nuclear fusion. In case they are detected with enough anticipation, mitigation actions (to minimize their possible damage to the plasma facing components) or even avoidance strategies (to redirect the ongoing pulse into a safer space) can be carried out. Several data-driven models, that attain high prediction rates, have been developed in the last 15 years to tackle this problem. These systems are designed to trigger alarms once they detect an early precursor of the phenomenon. However, not all the disruptions evolve in the same manner or imply a similar risk. Executing effective control actions demands not only a trigger but also some information about the reason of the alarm activation. The work developed in the paper is basedon analysing the alarms of a previously developed JET machine-learning disruption predictor that uses five variables selected with genetic algorithms. They are the Locked mode amplitude and the time derivative of the stored diamagnetic energy signals. Also, it uses time derivatives of the plasma internal inductance, the plasma elongation and the plasma vertical centroid position. The values of these five variables, at the time the alarms are activated, are input to a k-means non-supervised clustering. With this methodology, 4 main disruptive clusters have been identified. Also, a non-disruptive cluster was created using the values of the 5 parameters in non-disruptive pulses. Each disruptive cluster is characterized by a specific mean warning time: 1.14s,590ms, 251ms and 204ms. The developed clustering has allowed analysing the evolution of the discharges. As it was expected, most of the pulses (disruptive and non-disruptive) start as non-disruptive (the 5 parameters, at the beginning of the plasma current flat-top, are closest to the non-disruptive cluster). As the discharges approach a disruption, the membership of the parameters moves towards the disruptive clusters. Then, the clusters represent different operational spaces (safe and disruptive) and a path from a non-disruptive to a disruptive state. The space the discharges belong to can be a significant factor to develop efficient control actions. This methodology can be directly applied in real-time. The clusters membership can be followed on-line to estimate how far away the shot is to the disruptive zone. In case an alarm is triggered, the closest disruptive cluster (with its specific characteristics) can be automatically identified. Then, a pre-defined control action targeting the singular features of the detected cluster can be developed. As a final observation of the potential of this technique, it should be noticed that these categorized disruptive alarms can be used in future works to train cluster-oriented disruption predictors.