Neural-net predictor for beta limit disruptions in JT-60U

Yoshino, Ryuji

Prediction of major disruptions observed at the -limit for tokamak plasmas has been investigated in JT-60U with developing neural networks. A sub-neural network is trained to output a value of the limit every 2 ms. The target limit is artificially set by the operator in the first step training and is modified in the second step training using the output limit from the trained network. To improve the prediction performance further, the difference between the estimated limit and the measured and the other 11 parameters are inputted to a main neural network to calculate the stability level. Major disruptions have been predicted with a prediction success rate of 80% at 10 ms prior to the disruption while the false alarm rate is lower than 4%. This 80% is much higher than about 10% previously obtained. A prediction success rate of 90% has been also obtained with a false alarm rate of 12% at 10 ms prior to the disruption. This 12% is about a half of previously obtained one.