Plasma Disruption Prediction Based On Decision Tree And Disruption Mitigation For J-TEXT And HL-2A

Disruption which leads the discharge to terminate suddenly is a cataclysmic event in the operation of tokamak.The heat fluxes,electromagnetic loads and runaway electrons caused by the disruption will seriously damage the safety of the device structure.Therefore,it is an important research to accurately predict the disruption combined with the disruption mitigation.At present,various machine learning algorithms have greatly improved the performance of predictors,but the cross-machine performance of unfamiliar new devices is relatively poor.In order to improve fracture prediction,using the two device databases from J-TEXT and HL-2A to study cross-machine disruption prediction can accumulate research experience and provide reference for extrapolating the disruption prediction model to other devices in the future.In this paper,a new method is proposed to explore the cross-machine disruption prediction by using the interpretable algorithm gradient boosting decision tree.A high-performance predictor is built successfully,and the key factors affecting the cross-machine disruption prediction are found.This paper firstly introduces the methodology of the disruption prediction model and the machine learning algorithm.The algorithm based on gradient boosting decision tree is mainly used to build the model.At the same time,the model analysis tools and evaluation indexes are given.Secondly,the independent signals of the two devices were selected for model training and testing,and two high-performance disruption prediction models were built.The true positive rate and false positive rate of the independent model on J-TEXT are 91.9%and 10.1%,while the true positive rate and false positive rate on HL-2A are 83.2% and 11.1%.Thirdly,this paper analyzes the model in terms of the importance of the input feature and the contribution of them to the disruption.The relative feature importance ranking of disruption prediction model can reflect the physical mechanism of rupture to a certain extent.Those features associated with typical disruption discharges on the device rank higher than others.The relative feature importance ranking of different devices is different.Through the analysis of feature contribution,it is found that this is because there are certain differences in the disruption types of different disruption databases.Next,the paper trains the two models by selecting the common signals of the two devices.However,compared with the independent model in terms of prediction accuracy and warning time,these models can not be compared with the previous independent model.This is because there are too few common signals to train an efficient and accurate model.The results show that the difference of disruption types and common signals in different device databases is an important factor affecting the cross-machine disruption prediction.In this paper,the real-time disruption prediction system is further used to help the disruption mitigation system to complete the new experiment of real-time disruption mitigation,which realizes the early warning and effective mitigation of plasma disruption.Disruption prediction needs real-time testing and combines with disruption mitigation.Massive gas injection(MGI)is the main disruption mitigation method for J-TEXT.Density limit disruption prediction system has been established on J-TEXT.A density limit disruption prediction and mitigation experiment was designed and completed in the autumn experiment of 2020.By comparing the discharge conditions of the reference group and the experimental group,it is found that the density limit disruption prediction system can accurately predict the density limit rupture and the warning time is relatively sufficient.At the same time,it can quickly and effectively combine with the MGI disruption mitigation system to quickly shut off the plasma discharge when disruption approches,the thermal energy dissipation is effectively controlled and the damage caused by disruption is mitigated.In this paper J-TEXT and HL–2A device using the gradient boosting decision tree algorithm to find the key factors affecting cross-machine disruption prediction,further on JTEXT using density limit disruption prediction system and MGI to realize the early warning and effective disruption mitigation,it provides a reference for the research on disruption prediction and mitigation of large-scale devices in the future.