The application of Machine Learning (ML) techniques to predict disruptions has shown potential to considerably improve the detection rates and the warning times in JET [1] and other tokamaks. However, ML predictors learn from past events, which implicate an already stored database to develop them. Therefore, a significant problem arises at the time of developing MLbased systems for ITER. In this work and to tackle this problem, a Genetic Algorithms - optimized (GAs) predictor based on a previous study [1] was trained using, initially, only AUG data and tested with a wide database of JET. This "smaller to larger" tokamak approach is meant as a test for future extrapolation of this technique to ITER. The outcomes of the direct application of the cross - predictor derived in a 44.47% of false alarms and more than a 65% of premature alarms, which indicates the need of some information about the target device to achieve reasonable performance. Then, in a second approach, a new predictor was trained with AUG database plus one disruptive and one non - disruptive pulse of JET. The procedure was repeated twice to ensure the robustness of the results, since the GAs evolves towards different solutions in every run. In both runs the final cross - predictions (over the chronologically first 500 shots after the t raining) reached an average of a ~94% of total detected disruptions (~90% of them with anticipation times higher to 10 ms). The false alarms in that period were, on average, of 12,95%. An analysis of the ageing effect (assessing the decrease in the model's accuracy with time after the last training) has been performed for a database that contains more than 5000 discharges. Finally, and to provide a solution to the predictor's ageing, an adaptive approach, based on the system's retraining each time a disrup tion is missed, has been introduced and applied.
AUG - JET cross - tokamak disruption predictor
Murari A;
2017
Abstract
The application of Machine Learning (ML) techniques to predict disruptions has shown potential to considerably improve the detection rates and the warning times in JET [1] and other tokamaks. However, ML predictors learn from past events, which implicate an already stored database to develop them. Therefore, a significant problem arises at the time of developing MLbased systems for ITER. In this work and to tackle this problem, a Genetic Algorithms - optimized (GAs) predictor based on a previous study [1] was trained using, initially, only AUG data and tested with a wide database of JET. This "smaller to larger" tokamak approach is meant as a test for future extrapolation of this technique to ITER. The outcomes of the direct application of the cross - predictor derived in a 44.47% of false alarms and more than a 65% of premature alarms, which indicates the need of some information about the target device to achieve reasonable performance. Then, in a second approach, a new predictor was trained with AUG database plus one disruptive and one non - disruptive pulse of JET. The procedure was repeated twice to ensure the robustness of the results, since the GAs evolves towards different solutions in every run. In both runs the final cross - predictions (over the chronologically first 500 shots after the t raining) reached an average of a ~94% of total detected disruptions (~90% of them with anticipation times higher to 10 ms). The false alarms in that period were, on average, of 12,95%. An analysis of the ageing effect (assessing the decrease in the model's accuracy with time after the last training) has been performed for a database that contains more than 5000 discharges. Finally, and to provide a solution to the predictor's ageing, an adaptive approach, based on the system's retraining each time a disrup tion is missed, has been introduced and applied.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
