Real-time prediction of high-density EAST disruptions using random forest

Abstract

Abstract A real-time disruption predictor using random forest was developed for high-density disruptions and used in the plasma control system (PCS) of the EAST tokamak for the first time. The disruption predictor via random forest (DPRF) ran in piggyback mode and was actively exploited in dedicated experiments during the 2019–2020 experimental campaign to test its real-time predictive capabilities in oncoming high-density disruptions. During dedicated experiments, the mitigation system was triggered by a preset alarm provided by DPRF and neon gas was injected into the plasma to successfully mitigate disruption damage. DPRF’s average computing time of ∼250 μs is also an extremely relevant result, considering that the algorithm provides not only the probability of an impending disruption, i.e. the disruptivity, but also the so-called feature contributions, i.e. explainability estimates to interpret in real time the drivers of the disruptivity. DPRF was trained with a dataset of disruptions in which the electron density reached at least 80% of the Greenwald density limit, using the zero-dimensional signal routinely available to the EAST PCS. Through offline analysis, an optimal warning threshold on the DPRF disruptivity signal was found, which allows for a successful alarm rate of 92% and a false alarm rate of 9.9%. By analyzing the false alarm causes, we find that a fraction (∼15%) of the misclassifications are due to sudden transitions of plasma confinement from H- to L-mode, which often occur during high-density discharges in EAST. By analyzing DPRF feature contributions, it emerges that the loop voltage signal is that main cause of such false alarms: plasma signals more apt to characterize the confinement back-transition should be included to avoid false alarms.