In tokamak discharge experiments, the plasma position prediction model’s research is to understand the law of plasma motion and verify the correctness of the plasma position controller design. Although Maxwell equations can completely describe plasma movement, obtaining an accurate physical model for predicting plasma behavior is still challenging. This paper describes a deep neural network model that can accurately predict the HL-2A plasma position. That is a hybrid neural network model based on a long short-term memory network. We introduce the topology, training parameter setting, and prediction result analysis of this model in detail. The test results show that a trained deep neural network model has high prediction accuracy for plasma vertical and horizontal displacements.
28 May 2020 in Plasma Physics and Controlled Fusion
The modeling and control of the plasma equilibrium response is still one of the more important research areas in tokamak discharge experiments. Although theoretically, first principles can predict the plasma instability, how to build a physical model for accurate prediction is still a challenging problem. Therefore, a deep learning method is proposed to model the plasma vertical displacement system in the HL-2A tokamak experiment, whose method expands the modeling strategy for tokamak plasma control systems. Through the training of a large number of high-dimensional experimental data, the obtained deep neural network model in this paper has a higher precision prediction ability. Additionally, to illustrate the significance of the predictive model in controller design, a data-driven adaptive control algorithm is proposed to replace the traditional proportional-integral-derivative control algorithm for controlling the vertical displacement of plasma. The simulation results showed that the proposed algorithm had less adjustable parameters, strong self-adaptability, and effective control for the HL-2A plasma vertical displacement.
This paper introduces a self-consistent field-null optimization algorithm of a poloidal magnetic field that precisely accounts for the influence of vacuum vessel eddy currents. Building on existing poloidal field (PF) coil currents, the algorithm can refine these waveforms to achieve various target field-null configurations. Firstly, based on the TokSys toolbox, a response model, including the PF coils and vacuum vessel circuits for the HL-3 tokamak, is developed under the MATLAB® and Simulink™ framework. The resistivity parameters of the model are calibrated using experimental data obtained from single-coil discharge tests. Subsequently, an iterative method was employed to simultaneously solve the dynamic field-null optimization problem within a specified spatial region and precisely account for the effect of passive eddy currents. Typically, within a large area can be obtained with this iterative scheme, which can be stably sustained for over 15 milliseconds to ensure the robustness of breakdown. Finally, a low-pass filtered PID controller is applied to the model to achieve precise control of the PF coils currents, confirming the feasibility of implementing the proposed algorithm in real experiments.
In the discharge experiments of the HL-2A Tokamak device, plasma position prediction plays a crucial role in aspects such as the closed - loop control of HL-2A and verifying the correctness of the plasma controller design. In previous studies, methods based on traditional equations and LSTM have made significant progress in plasma position prediction. However, plasma position prediction is of extremely high complexity, and these methods have difficulty in achieving further improvement on this issue. To address the problem that the performance of traditional methods is hard to enhance further, we propose a Non - Stationary - GLU Transformer (NSGT) model integrating multi - source data processing. In this model, we select data from historical discharge experiments as the input of the NSGT model, and use the improved Transformer structure to model the global features of the multi - source fused data. The experimental results show that the NSGT model has a significant reduction in the root - mean - square error of prediction compared with the LSTM method, effectively supporting the real - time adjustment of the closed - loop control system. This method not only improves the prediction accuracy but also provides important technical support for achieving more stable and safe operation of the Tokamak device.
1 February 2026 in Plasma Physics and Controlled Fusion
For enhancing plasma parameters, stable plasma operation is a critical aspect. Precise control of the poloidal field (PF) coils promotes the exploration of plasma shape and significantly mitigates the damage caused by plasma disruptions to the device. In actual tokamak control, the adjustment of feedback control parameters and the development of new control algorithms are complex and time-consuming tasks that require substantial funding for validation and optimization. Since conventional simulation environments, though highly accurate, involve long computation times, there is an urgent need to construct a reliable RZIP (the displacement in the R direction, the displacement in the Z direction, and plasma current) response model for real-time control. This study utilizes a total of 1100 shots of historical data to develop a deep learning-based RZIP prediction model aimed at predicting RZIP data 10ms in advance. By employing a multi-task learning (MTL) approach, the model ’ s performance is enhanced to obtain a more accurate RZIP prediction model. Furthermore, by combining PID feedback control with the M -matrix (i.e. the decoupling matrix between RZIP and PF coils), a coil voltage conversion model is constructed, with RZIP as the input and PF coil control voltage as output. This research also derives more suitable PID control parameters for the current IP based on the IP data 10 ms later, leading to more reasonable PF coil control voltage data. The study predicts the RZIP using the RZIP MTL model, and the mean squared error between the predicted RZIP and the actual RZIP experimental data meets the experimental requirements. Additionally, the output voltage control for the PF coils, generated by combining PID feedback control with the M -matrix, satisfies the experimental needs.