Speaker
Description
High-precision and rapid evolution prediction of magnetic measurement signals are essential for tokamak plasma configuration control and the safe operation of discharges. Traditional physics-based models typically suffer from high computational complexity and long execution times, making it difficult to satisfy real-time requirements. Meanwhile, existing data-driven methods often encounter issues with insufficient reliability when performing long-term extrapolation. To address these challenges, this paper proposes a real-time autoregressive magnetic measurement evolution method based on the Transformer architecture. The model takes historical magnetic measurement data and real-time control commands as inputs, leveraging the self-attention mechanism of the Transformer to capture long-range temporal dependencies. At each time step, the model predicts the magnetic measurements for the next step and recursively updates these predictions as subsequent inputs, thereby achieving continuous temporal evolution. Experimental results demonstrate that the proposed single-step autoregressive approach accurately tracks the evolution trajectory of magnetic signals during the discharge process, achieving an accuracy of no less than 92% within a 1000 ms duration. Furthermore, its single-step inference structure is well-suited for low-latency hardware deployment, fulfilling the rigorous demands of real-time control systems. This research provides a high-precision, high-efficiency digital environment for closed-loop control simulators and the development of advanced controllers based on reinforcement learning for magnetic confinement fusion devices.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | No |