Disruptions are a potential showstopper on the route to developing a tomakak fusion reactor. Since their consequences can be more severe the larger the devices, in the next generation of machines they will have to be carefully managed from the beginning of operation. On the other hand, in new devices coming on line, the diagnostic coverage is typically limited and there will be no opportunity to collect many examples for the training of traditional machine learning classifiers. It is therefore important to develop predictors that can ideally operate satisfactorily without training and with minimal diagnostic information. A technique capable of satisfying these requirements is described in the present work. It is based on converting the time series of macroscopic basic signals, such as the plasma current or the locked ML, into a string of symbols, before quantifying the complexity of the resulting sequences with permutation entropy. The application to a large dataset of discharges of JET with a metallic wall has provided very interesting results. In addition to good statistical performances, the warning times are sufficient not only for mitigation but also for the prevention of most disruptive events. The transfer of the technique to JET with a carbon wall has also been quite encouraging and therefore it is proposed to deploy the approach in new machines such as JT-60SA and DTT.
This paper presents a technique to predict disruptions in fusion reactors, even with limited data and diagnostic information. It converts time series data into symbol sequences and uses permutation entropy to quantify their complexity, enabling disruption prediction without extensive training.