Disruption prediction by support vector machine and neural network with exhaustive search

Abstract

A disruption is an event in which the plasma current suddenly shuts down in a tokamak reactor. Establishing methods to predict, mitigate, and avoid disruptions may be indispensable for realizing a tokamak reactor. In the present study, we have used the large dataset of high-beta experiments at JT-60U to develop a method for predicting the occurrence of disruptions. The method is based on sparse modeling that exploits the inherent sparseness common to all high-dimensional data, and it enables us to extract the maximum amount of information from the data efficiently. To carry out the sparse modeling, we have used exhaustive searches with a support vector machine and a neural network. In this research, we repeated the training and evaluation of the predictor while changing the combination of plasma parameters. As a result of the exhaustive search, we found |Bn=1r| and d|Bn=1r|/dt to be the dominant parameters for disruption predictions. This is not surprising, because MHD instabilities are considered to be the direct triggers of disruption. In addition, we have succeeded in identifying several important parameters that may also be strongly related to disruptions, i.e., βN, βP, q95, σ, fGW, and frad.