Speaker
Description
Traditional first-principle-based tokamak plasma magnetic response models, while pos-
sessing clear physical significance, often exhibit significant deviations when compared
with actual experimental data, limiting our deep understanding of plasma dynamics pro-
cesses. This study develops an autoregressive neural network model based on improved
WaveNet architecture that directly learns plasma behavior patterns from HL-3 device
experimental data, achieving high-precision reproduction of real plasma responses.
The model takes experimental measurements of poloidal field coil currents and cen-
tral solenoid coil currents as inputs, predicting key parameters such as plasma current,
position, minor radius, elongation, and triangularity reconstructed by EFITNN[1]. Ad-
dressing the specific requirements of experimental data validation, we designed an in-
novative dual loss function mechanism that combines teacher forcing and autoregressive
training approaches, effectively mitigating exposure bias issues and enabling the model to
better handle long-term sequence autoregressive predictions. The loss function employs
a weighted combination of MSE, MAE, and MTE, providing differentiated optimization
gradients for prediction errors of different magnitudes, significantly improving the model’s
fitting capability for both regular variations and sudden events in experimental data.
To simulate measurement noise and uncertainties in real experimental environments,
we introduced random walk noise layers and Gaussian noise layers during model training.
This design enables the neural network to adapt to various perturbations present in actual
experimental data and effectively suppress cumulative errors in long-period predictions.
Validation based on experimental data from HL-3 device shots #2000 #6698 demon-
strates that the model can reproduce single-step experimental measurement results with
99.5% accuracy, while maintaining 97.5% accuracy in autoregressive mode. More impor-
tantly, the high consistency between model predictions and real experimental data indi-
cates that the neural network successfully captures the intrinsic physical laws of plasma
magnetic response, providing a new perspective for understanding tokamak plasma dy-
namic behavior.
Through comparative analysis with traditional physics models, we found that data-
driven methods have significant advantages in reproducing experimental observations,
particularly excelling in handling nonlinear coupling effects and long-term evolution pro-
cesses. This modeling approach based on real experimental data validation not only
provides effective tools for current fusion device physics research but also establishes a
foundation for plasma behavior prediction and analysis in future large-scale devices such
as ITER.
| Speaker's email address | lijiyuan@swip.ac.cn |
|---|---|
| Speaker's Affiliation | Southwestern Institute of Physics |
| Member State or International Organizations | China |
