Implementing deep learning-based disruption prediction in a drifting data environment of new tokamak: HL-3

Abstract A deep learning-based disruption prediction algorithm has been implemented on a new tokamak, HL-3. An Area Under receiver-operator characteristic Curve of 0.940 has been realized offline over a test campaign involving 72 disruptive and 240 non-disruptive shots, despite the limited training data available from the initial two campaigns. In addition to the well-documented challenge of insufficient training data, a previously unanticipated issue is addressed that the data distribution of a new device is continuously drifting. The plasma scans across a broad parameter space, bringing a drifting distribution of disruption causes and diagnostic data. This problem is often overlooked in previous implementations on steadily operating tokamaks, necessitating greater attention in future tokamaks like ITER. To address these challenges, innovative modules including predict-first neural network, data augmentation, and pseudo data placeholders are developed and implemented, which promotes the accuracy by up to 20%. A series of advantages are also brought by the modules, including the robustness in handling missing input channels, and the interpretability to identify which parameter of plasma is under abnormal condition. The results demonstrate that, with dedicated data collection and algorithm implementation, the issues of limited data and drifting distribution can be overcome, and further, the deep learning-based algorithm can reliably provide disruption alarms on a new tokamak.