Time series regression network using exogenous variables for missing effective ion charge reconstruction in EAST tokamak

Authors:

Abstract
The effective ion charge ( Z eff ) is a key parameter for characterizing plasma impurity levels and is important for investigating impurity transport and assessing discharge conditions in tokamaks. However, diagnostic system failures can lead to missing or incomplete Z eff signals. Accurate reconstruction remains challenging because Z eff is related to multiple plasma parameters and exhibits temporal dependencies. In this study, a time series regression network was developed to reconstruct missing Z eff values from exogenous plasma parameters in Experimental Advanced Superconducting Tokamak (EAST). The model employed a sequential CNN-BiGRU-Transformer encoder architecture, in which the CNN extracted local temporal features, the BiGRU captured bidirectional temporal information, and the Transformer encoder modelled dependencies across different timescales through multi-head self-attention. The model was trained and validated using a dataset comprising 1566 EAST discharges. The results show that the average similarity between the reconstruction and experimental data is 90.3%, while the average coverage probability is 91.4%, indicating that most reconstruction values fall within the standard deviation error range of the experimental results. Overall, this study presents a feasible offline signal reconstruction method to maintaining the integrity of Z eff signals in EAST and provides a useful reference for reconstructing missing diagnostic data in future fusion devices.
Funding Information
  • Excellent Youth Project of the Education Department of Anhui Province (Contract Nos. 2023AH030112)
  • National Key Research and Development Program of China (Project No. 2024YFE03000200)

This publication has 63 references indexed in Scilit: