A shapelet-based neural network for binary and multi-class disruption prediction for prevention at JET

Abstract

Disruptions in tokamaks remain, to this day, an unsolved issue on the path toward fusion power plants. Such events should be avoided or mitigated, requiring adequate detection of the disruption causes. However, due to the complex nature of disruption causes, this is, in general, challenging. Despite recent progress designing disruption prediction systems via data-driven methods, many questions remain open—such as disruption-type identification or the transfer of such methods between different tokamaks. We propose a shapelet-based neural network for binary and multi-class disruption identification that can be used for post-disruption analysis or real-time prediction. The performance is compared to two approaches from the literature, retrained on our data: support-vector machines similar to the advanced predictor of disruption, and a recurrent neural network. For the present study, we compiled a dataset of discharges from the Joint European Torus tokamak containing stable discharges and seven disruption types. Due to the availability of such classification, all shots predate the installation of the ITER-like wall. Using the binary and multi-class classification results, we report on the performance of the three models and discuss the advantages of our method. Our model ranks first and second on binary and multi-class tasks, respectively. The shapelets' contribution to the results is evaluated by conducting an ablation study. We show that shapelets with normalized Euclidean distance are enough for binary classification, but multi-class predictions require the absolute value of the signals. The good results obtained from locally normalized signals are promising for future cross-tokamak studies.