In this paper, a neural predictor has been built using plasma discharges selected from two years of ASDEX Upgrade experiments, from July 2002 to July 2004. In order to test the real-time prediction capability of the system, its performance has been evaluated using discharges coming from different experimental campaigns, from June 2005 to July 2007. All disruptions that occurred in the chosen experimental campaigns were included with the exception of those occurring in the ramp-up phase, in the ramp-down phase (if the disruption does not happen in the first 100 ms), those caused by massive gas injection and disruptions following vertical displacement events. The large majority of selected disruptions are of the cooling edge type and typically preceded by the growth of tearing modes, degradation of the thermal confinement and enhanced plasma radiation. A very small percentage of them happen at large beta after a short precursor phase. For each discharge, seven plasma diagnostic signals have been selected from numerous signals available in real-time. During the training procedure, a self-organizing map has been used to reduce the database size in order to improve the training of the neural network. Moreover, an optimization procedure has been performed to discriminate between safe and pre-disruptive phases. The prediction success rate has been further improved, performing an adaptive training of the network whenever a missed alarm is triggered by the predictor.
Disruption prediction at JET through deep convolutional neural networks using spatiotemporal information from plasma profiles