A multi-classifier disruption predictor based on optimized support vector machine on EAST

Abstract

Plasma disruption presents a significant risk to the operation of tokamaks. In this study, we propose a multi-classifier disruption predictor based on the support vector machine optimized using an improved particle swarm optimization algorithm. This approach is specifically designed for the EAST experiments conducted since 2020. A dataset comprising 194 disruption and 401 non-disruption shots was selected to evaluate the predictor's performance. Twelve representative diagnostic signals were selected by analyzing the statistical characteristics of the data in the plasma current flattop phase, with each signal corresponding to a sub-classifier. The performance of each signal's sub-classifier predictor was considered, leading to the development of a weighting scheme for the prediction results. The final prediction result was obtained from a linear combination of these results by their respective weights. With optimized weight schemes, the predictor effectively identified plasma diagnostic signal features, achieving a true positive rate of 93.9 % and a false positive rate of <4.99 %, indicating its ability to trigger alarms promptly. Overall, the results demonstrate the feasibility of the proposed approach for disruption prediction on EAST.