Table 1 Performance comparisons between models in J-TEXT.

From: Disruption prediction for future tokamaks using parameter-based transfer learning

Model

SAR

FAR

BA

NN

77.27%

9.09%

84.09%

SVM

90.91%

12.73%

89.09%

FFE

93.64%

9.09%

92.28%

  1. The models are trained with the same diagnostics and discharges, and tested on the same discharges on J-TEXT. Alarms raised earlier than 300 ms are considered as early alarms, while alarms raised later than 10 ms are considered as tardy alarms. The disruption predictor based on Fusion Feature Extractor (FFE) reached the best performance among all models. The predictor based on manual feature extraction as input (the Support Vector Machine (SVM) model) reaches a similar Success Alarm Rate (SAR) but a worse False Alarm Rate (FAR). The traditional deep neural network (NN) model performs the worst with a much lower SAR.
Back to article page