Application Of Machine Learning In The Prediction Of No-wall Beta Limit In HL-2M Device

The HL-2M device operates in the high_mode,which are related to physical parameters corresponding to the safety factorand P-profile,and geometric parameters such as the plasma triangularity,elongation.Based on HL-2M simulation results,a database was built,which contains different plasma boundary and equilibrium profiles and the corresponding beta limit.Using this database,five ensemble models were trained to predict the beta limit for HL-2M tokamak.The Ada Boost algorithm was adopted to optimize models with better performance according to our preset evaluation indicators like MSE.The importance of each parameter to_limit was obtained through the visualization.The optimal model in the thesis was the Gradient Boosting Regressor enhanced by the Adaboost algorithm,which can predict the no-wall beta limit_<sup>9-)within 4%of the relative error range,the mean square error of prediction was about 0.02,and the model’s goodness of fit was about 0.93.Thesis was specific for the HL-2M device,which provides the opportunity to study the ideal-wall beta limit_<sup>74)-)prediction by the Machine Learning model.With the well-trained model,the_<sup>74)-)was predicted within 7%of the error range,the mean square error was about 0.15,and the goodness of fit was 0.68.As the above models were all tree models,The Support Vector Regression（SVR）was trained and calculated its prediction error,as a comparison with the tree models to prove that simple models can be mutually verified.The well-trained model also showed the weights of the critical parameters for pressure limit prediction.For no-wall beta limit,pressure peaking factor（）,the total current（_）,the internal inductance（7）)were the critical parameters.Moreover,was more relevant to the ideal-wall beta limit,which had a coefficient factor of 0.7250.Furthermore,a decision tree was plotted to validate with the visualization results of the ensemble model by observing its root nodes.The well-trained model can be used to monitor the no-wall beta limit in real-time and to forecast the possibility of disruption caused by pressure-driven instabilities during discharges in experiments.