Research On Real-time Evaluation And Prediction Of EAST Plasma Equilibrium Parameters Based On Neural Network

Tokamak is currently one of the most promising devices for nuclear fusion energy.Effective control of plasma equilibrium and vertical displacement instability is an important issue for the safe operation of the tokamak device.Aiming robust and fast control in future for tokamak plasma,the plasma equilibrium parameter estimation and vertical displacement prediction of the tokamak device are studied based on the neural network method.This article first applies a basic feedforward neural network to the EAST discharge database which generated by Tokamak simulation code(TSC),trains and tests the plasma equilibrium parameter estimation model.And the effects of missing signal and input noise on the robustness of the model are discussed separately.For the optimized neural network evaluation model,the average error of the dimensional equilibrium parameters(plasma current center position,major radius,minor radius,X point position)is within 1 millimeter under the configuration of the divertor and limiter,and the average error of the dimensionless parameters(elongation ratio,poloidal pressure ratio,inductance,safety factor)is estimated to be within one percent.For the hypothesized missing measurement signal and input noise,the model is verified to be robust by adding random noise and zeroing the missing signal for retraining verification.It takes about 1 millisecond to implement the model on a normal desktop with MATLAB software,which preliminarily verified the feasibility of this approach.In the identification of the position of the plasma current center,this article first uses simulation data to verify the feasibility of the neural network method,and then uses experimental data to train the model.The trained basic feed-forward neural network model was still verified to accurately identify the plasma current center;at the same time,reducing the inputs and exploring the impact of different input groups on the performance of the model,only using the flux loop diagnosis or the poloidal field coil current as inputs are not sufficient under the model.Considering that the poloidal field coil currents are important control variables for the change of the plasma current center position,this paper discusses the application of the NARX neural network model to establish the mapping relationship between the control quantity and the response quantity;compared with the basic feedforward network model,the accuracy of the current center position inference using only the control quantity is significantly improved.However,in the current work,training data is only obtained from the same type of discharge data,and the scope of application of the model is limited.This article finally screened and organized the 2018 EAST experimental database,covering more than a thousand shots,and applied long short-term memory neural networks to the prediction of plasma vertical position;at the same time,a GPU is used to accelerate the trained model in parallel.In this study,the first five time slices of electromagnetic measurement data are used to predict the vertical position of the plasma at the next moment,and accuracy and robustness of the model with input noise and signals absence are discussed.After continuous training and testing,continuous performance improvement,the final model parameters are fixed,and the initial speed of the CPU implementation is on the order of 1 millisecond;through algorithms and hardware acceleration,the model is finally implemented using the GPU,and the prediction time can be up to 50 microseconds.It is an order of magnitude higher than the current EAST control speed,which provides a new method for the fast control of plasma vertical displacement in the future.