Speaker
Description
Precise control of plasma density is critical for high-performance steady-state operation in fusion devices. The candidate fueling method for density control for future tokamak will be pellet injection. However, pellet injection introduces rapid, highly non-linear density perturbations that challenge the latency and accuracy limitations of traditional feedback systems. This study presents a real-time density evolution prediction and controlling architecture for the EAST Tokamak. The core challenge addressed is the requirement to process high-dimensional multi-physics diagnostic data—including magnetic equilibrium parameters, radiation profiles, and Dα signals—and generating reliable future density estimation within the 10ms control cycle required for active feedback.
To achieve this, we deploy a lightweight, attention-based LSTM model optimized for low-latency inference. The system integrates a rolling-window mechanism that synchronizes data acquisition, pre-processing, and model inference of 100Hz. This design ensures that the prediction of density evolution, particularly the sharp rise and decay following pellet ablation, is computed and fed into a Model Predictive Control (MPC) solver within the allocated time budget. Off-line test results with EAST experiment data demonstrate that the system successfully captures complex non-linear dynamics with negligible computational lag. By overcoming the trade-off between model complexity and real-time responsiveness, this framework enables precise, automated trajectory tracking and oscillation suppression during high-frequency pellet injection scenarios.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | Yes |