Data-driven simulation model for tokamak magnet coils on HL-3

Abstract

Magnet coils are essential components in tokamaks, playing a critical role in maintaining plasma equilibrium and ensuring efficient fusion reactions. Accurate and real-time control of these coils is vital for both anomaly detection and controller design. While traditional physics-based models offer precise representations of the coupling between magnetic fields and plasma dynamics, their high computational cost limits their practical use in real-time control applications. To address this, we propose a data-driven simulation model based on a modified WaveNet architecture to predict the real-time responses of the Poloidal Field (PF) and Central Solenoid (CS) coils in the HL-3 tokamak. The model demonstrates good single-step prediction performance, with the Mean Absolute Error (MAE) between the true values and predicted values being only 7 A, averaged across all coils. This result indicates the model's capability for real-time coil fault detection. Furthermore, in autoregressive predictions over 200 $\text{ms}$, our model achieved a cosine similarity exceeding 0.98, demonstrating its suitability for integration with Model Predictive Control (MPC) frameworks. This study presents a viable alternative to traditional physics-based modeling, effectively supporting the control needs of tokamak coils in various scenarios.