Unbiased and non-supervised learning methods for disruption prediction at JET

Abstract

The importance of predicting the occurrence of disruptions is going to increase significantly in the next generation of tokamak devices. The expected energy content of ITER plasmas, for example, is such that disruptions could have a significant detrimental impact on various parts of the device, ranging from erosion of plasma facing components to structural damage. Early detection of disruptions is therefore needed with evermore increasing urgency. In this paper, the results of a series of methods to predict disruptions at JET are reported. The main objective of the investigation consists of trying to determine how early before a disruption it is possible to perform acceptable predictions on the basis of the raw data, keeping to a minimum the number of 'ad hoc' hypotheses. Therefore, the chosen learning techniques have the common characteristic of requiring a minimum number of assumptions. Classification and Regression Trees (CART) is a supervised but, on the other hand, a completely unbiased and nonlinear method, since it simply constructs the best classification tree by working directly on the input data. A series of unsupervised techniques, mainly K-means and hierarchical, have also been tested, to investigate to what extent they can autonomously distinguish between disruptive and non-disruptive groups of discharges. All these independent methods indicate that, in general, prediction with a success rate above 80% can be achieved not earlier than 180 ms before the disruption. The agreement between various completely independent methods increases the confidence in the results, which are also confirmed by a visual inspection of the data performed with pseudo Grand Tour algorithms.