Disruption prediction at ASDEX Upgrade using neural networks

n this paper, a Multi Layer Perceptron is trained to act as disruptions predictor at ASDEX Upgrade. In particular, an optimization procedure is performed to identify a time instant that discriminate between disruptive and safe phases of disruptive discharges. The neural predictor has been trained, validated and tested using 149 disruptive pulses, selected from two years of ASDEX Upgrade experiments from 2002 to 2004. Non disruptive pulses has not been used to design the predictor, because the disruptive discharges at ASDEX Upgrade present a safe phase sufficiently long to well represent also the behavior of safe pulses. In order to limit the neural network size, for each disruptive shot, seven plasma diagnostic signals have been selected from numerous signals available in real time. A Self Organizing Map has been used to reduce the shot dimensionality in order to improve the training of the Multi Layer Perceptron, greatly increasing the prediction capability of the system. The results are quite good, with a prediction success rate greater than 90%.