Adaptive high learning rate probabilistic disruption predictors from scratch for the next generation of tokamaks

Abstract

The development of accurate real-time disruption predictors is a pre-requisite to any mitigation action. Present theoretical models of disruptions do not reliably cope with the disruption issues. This article deals with data-driven predictors and a review of existing machine learning techniques, from both physics and engineering points of view, is provided. All these methods need large training datasets to develop successful predictors. However, ITER or DEMO cannot wait for hundreds of disruptions to have a reliable predictor. So far, the attempts to extrapolate predictors between different tokamaks have not shown satisfactory results. In addition, it is not clear how valid this approach can be between present devices and ITER/DEMO, due to the differences in their respective scales and possibly underlying physics. Therefore, this article analyses the requirements to create adaptive predictors from scratch to learn from the data of an individual machine from the beginning of operation. A particular algorithm based on probabilistic classifiers has been developed and it has been applied to the database of the three first ITER-like wall campaigns of JET (1036 non-disruptive and 201 disruptive discharges). The predictions start from the first disruption and only 12 re-trainings have been necessary as a consequence of missing 12 disruptions only. Almost 10 000 different predictors have been developed (they differ in their features) and after the chronological analysis of the 1237 discharges, the predictors recognize 94% of all disruptions with an average warning time (AWT) of 654 ms. This percentage corresponds to the sum of tardy detections (11%), valid alarms (76%) and premature alarms (7%). The false alarm rate is 4%. If only valid alarms are considered, the AWT is 244 ms and the standard deviation is 205 ms. The average probability interval about the reliability and accuracy of all the individual predictions is 0.811 ± 0.189.

Feature selection for disruption prediction from scratch in JET by using genetic algorithms and probabilistic predictors

“…The list of signals is explained in Table 1, which is the same as used in the work [5]. As previously, it consists of 14 features belongs to 7 plasma signals.…”

“…The quality of each chromosome is estimated by a classifier. A probabilistic classifier based on Venn predictors [5] is used. Venn predictors make prediction directly from data by transduction (without generating any rules), instead of repeatedly training classifiers to generate models.…”

“…It happens when a new machine starts its operation (great relevance thinking of ITER). Continuing with this paradigm, recently, a high learning rate probabilistic predictor (based on Venn machines) has been developed at JET [5] under the 'from scratch' approach. Venn predictors are used as probabilistic classifiers instead of using bare classifiers like SVM.…”

“…They have the advantage of providing a confidence level for each individual prediction. Moreover, in the specific implementation carried out in [5], it has been possible to reduce the input sample space by means of grouping the data to use a nearest centroid taxonomy (NCT). To concentrate as much as possible the information of both plasma behaviors (the disruptive and non-disruptive activities), the Venn predictor uses the NCT that is fully described in Ref.…”

“…To concentrate as much as possible the information of both plasma behaviors (the disruptive and non-disruptive activities), the Venn predictor uses the NCT that is fully described in Ref. [5]. This results in a faster predictor when testing new discharges.…”