The development of a disruption predictor using a data-driven solution is an effective way to avoid or mitigate tokamak device disruptions. The black-box nature of the model itself determines the agnostic nature of its decision base and becomes a key factor limiting the further development of disruption predictors. To identify the factors affecting the model performance and achieve better understanding of its predictions, this study proposes an interpretable method for disruption prediction. Based on the physical characteristics of the disruption, 2094 disruption shots and 4858 non-disruption shots from 2022 to 2024 were screened as training shots, and then the disruption prediction model was trained using the eXtreme Gradient Boosting (XGBoost) algorithm from training samples consisting of 16 diagnostic signals, such as plasma current, density, and radiation. The experimental results show that the XGBoost can accurately predict the disruption shot 30 ms before the disruption (96.7% true positive rate), whereas the false positive rate of the non-disruption shot is 6.58%. Through the analyses of Shapley Additive exPlanations interpretability, the degree of relative importance of each signal was determined while the shot disruption type was based on the signals close to the disruption. The results demonstrate the feasibility of EAST plasma disruption prediction using the method proposed in this study, which holds referential significance for the disruption prediction for future fusion devices.
This paper presents an interpretable method for predicting disruptions in the EAST tokamak using machine learning. The researchers developed a disruption prediction model based on the XGBoost algorithm and used Shapley Additive Explanations (SHAP) to identify the key factors affecting the model's predictions. The model can accurately predict disruptions 30 ms before they occur with a high true positive rate and low false positive rate.