An adaptive disruption predictor based on FDI approach for next generation Tokamaks

Disruptions have the potential to create serious damage to large reactor-scale. Thus, disruption detection is essential to allow proper mitigation actions to be triggered when preventing disruptions is not feasible. Several contributions have been proposed using neural network models with good performances in different tokamaks. In particular, Support Vector Machines in JET [1, 2] and Multi-layer Perceptrons both in JET [3] and ASDEX Upgrade [4]. The main drawback of these methods is the need of a set of disruptions to implement the predictive model. As largely known, ITER cannot wait for hundreds of disruptions to develop a successful disruption predictor. Hence, a prediction system starting from only few safe discharges will be required. Thus, the proposed approaches are not directly applicable. To this purpose, the disruption prediction can be formalized as a fault detection and isolation (FDI) problem, where the safe pulses are assumed as the normal operating condition and the disruptions are assumed as status of fault [5]. The main advantage of the proposed FDI methods is that the model can be developed without any information about disruptions. In this work, in view of ITER, an adaptive disruption predictor based on FDI approach is developed. In particular, an autoregressive model is trained to represent the normal operating conditions (NOC) described by few safe shots. Then, the model is progressively updated as soon as a new safe configuration is performed. The dynamic structure of each pulse is estimated through the fitting of the NOC model, the discrepancy between the outputs provided by the NOC model and the actual measurements (residual) is an indication of the plasma disruptivity. The prediction performance is evaluated for JET using a set of safe and disrupted discharges in terms of correct predictions, missed and false alarms. Preliminary results show the suitability of the proposed method.