High-beta disruption prediction study on HL-2A with instance-based transfer learning

Yu Zhong,Wei Zheng,Zhongyong Chen,Wei Yan,F Xia,Liming Yu,Fengming Xue,Chengshuo Shen,Xin.Kun Ai,Zongyu Yang,Yinlong Yu,Zisen Nie,Yong Hua Ding,Yunfeng Liang,Zhipeng Chen

DOI: https://doi.org/10.1088/1741-4326/ad59b5

IF: 3.3

2024-06-20

Nuclear Fusion

Abstract:Unmitigated disruptions pose a much serious threat when large scale tokamak is operating in high performance regime. Machine learning based disruption predictors can exhibit impressive performance. However, their effectiveness is based on a substantial amount of training data. In future reactors, taking a substantial amount of disruption shot in high-performance regimes, without risking damage to the machine, is highly improbable. Using machine learning to develop disruption predictors on data from low performance regime and transfer to high performance regime is an effective solution for a large reactor sized tokamak like ITER and beyond. In this study, a number of models are trained by creating different sub datasets of HL-2A tokamak experiment. SHapley Additive exPlanations (SHAP) analysis is executed on the models, revealing there are different even contradicting patterns between different performance regimes. Thus, just mixing data among different performance regimes will not yield optimal results. Based on this analysis, we proposed an instance-based transfer learning technique which trains the model using a dataset generated with an optimized strategy. The strategy involves instance and feature selection based on the physics behind differences in high and low-performance discharges, as revealed by SHAP model analysis. Tradaboost technique significantly improved the model performance from 76% TPR (true positive rate) to 95% at same 23% FPR (false positive rate) with a few high-performance operation data.

physics, fluids & plasmas