Disruption prediction and model analysis using LightGBM on J-TEXT and HL-2A

• Published in: Plasma physics and controlled fusion 2021-07, Vol.63 (7), p.75008
• Main Authors: Zhong, Y, Zheng, W, Chen, Z Y, Xia, F, Yu, L M, Wu, Q Q, Ai, X K, Shen, C S, Yang, Z Y, Yan, W, Ding, Y H, Liang, Y F, Chen, Z P, Tong, R H, Bai, W, Fang, J G, Li, F
• Format: Article
• Language: English
• Subjects: disruption prediction; feature importance; LightGBM; machine learning
• ISSN: 0741-3335; 1361-6587
• DOI: 10.1088/1361-6587/abfa74

Using machine learning (ML) techniques to develop disruption predictors is an effective way to avoid or mitigate the disruption in a large-scale tokamak. The recent ML-based disruption predictors have made great progress regarding accuracy, but most of them have not achieved acceptable cross-machine performance. Before we develop a cross-machine predictor, it is very important to investigate the method of developing a cross-tokamak ML-based disruption prediction model. To ascertain the elements which impact the model’s performance and achieve a deep understanding of the predictor, multiple models are trained using data from two different tokamaks, J-TEXT and HL-2A, based on an implementation of the gradient-boosted decision trees algorithm called LightGBM, which can provide detailed information about the model and input features. The predictor models are not only built and tested for performance, but also analyzed from a feature importance perspective as well as for model performance variation. The relative feature importance ranking of two tokamaks is caused by differences in disruption types between different tokamaks. The result of two models with seven inputs showed that common diagnostics is very important in building a cross-machine predictor. This provided a strategy for selecting diagnostics and shots data for developing cross-machine predictors.