Table 2 The results of the cross-tokamak disruption prediction experiments using different strategies and models.

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

Case no.

SAR

FAR

BA

1

86.11%

30.00%

78.06%

2

83.33%

40.56%

71.39%

3

78.89%

48.89%

65.00%

4

69.44%

48.89%

60.28%

5

49.44%

30.56%

59.44%

  1. All experiments use the same test set from EAST (360 discharges, 180 disruptive). The five cases are FFE-based disruption predictors applying the same set of hyper-parameters, which are chosen according to the source model mentioned above from J-TEXT. Case 1 is trained with all discharges currently available from EAST (1896 discharges, 355 disruptive). Case 2 is the best practice using parameter-based transfer learning, which uses the source model trained with J-TEXT data and is further tuned with 20 discharges from EAST (10 disruptive). Case 3 is trained with the mix of the training set from J-TEXT (494 discharges, 189 disruptive) and 20 discharges from EAST (10 disruptive). Case 4 is trained with 20 discharges from EAST (10 disruptive). Case 5 directly applies the source model trained with J-TEXT data to predict disruptions in EAST. SAR success alarm rate, FAR false alarm rate, BA balanced accuracy are used to evaluate the performance of the cases.
Back to article page