Real-Time Estimation of Vertical Instability Growth Rate for EAST Plasma With MLP

• Published in: IEEE transactions on plasma science 2023-10, Vol.51 (10), p.3243-3249
• Main Authors: Liu, B. N., Hu, W. H., Huang, Y., Luo, Z. P., Wang, Y. H., Yuan, Q. P., Zhang, R. R., Xiao, B. J.
• Format: Article
• Language: English
• Subjects: Active control; Biological neural networks; Computational modeling; Conductors; Experimental advanced superconducting tokamak (EAST); Load distribution; Magnetic probes; Mathematical models; multilayer perceptron (MLP); Multilayer perceptrons; Neural networks; Parameter identification; Plasma; Plasma currents; Plasma equilibrium; Plasmas; Probes; real time; Real-time systems; Thermal analysis; Tokamak devices; Training; vertical displacement event (VDE); vertical instability (VI) growth rate
• ISSN: 0093-3813; 1939-9375
• DOI: 10.1109/TPS.2023.3321377

Vertical instability (VI) is one of the main challenges for fusion energy realization through advanced tokamak. The disruption caused by VI is known as vertical displacement event (VDE). VDE not only causes serious thermal load to the plasma-facing components but also generates huge mechanical load to the first wall. VI growth rate is a crucial parameter not only for VI identification but also for active control of vertical displacement. In this work, the multilayer perceptron (MLP) model is employed to estimate the plasma VI growth rate ( \gamma ) of the experimental advanced superconducting tokamak (EAST). This model shows great advantages in calculation speed compared with the conventional way of solving rigid plasma response model equations. In this model, 38 magnetic probe measurements divided by plasma current ( I_{p} ) for normalization are taken as input features. Meanwhile, the neural network was trained by taking plasma equilibrium parameters as input features for comparison. The results demonstrate a slightly lower prediction accuracy than the original model. The mean absolute error (MAE) increased from 1.02 to 1.68 \text{s}^{-1} , and the mean square error (MSE) increased from 1.89 to 10.03 \text{s}^{-2} . Finally, issues related to the interpretability of neural networks are discussed.