Position Prediction And Control Based On Deep Learning For HL-2A And HL-2M Plasma

Nuclear fusion is polymerizing two lighter nuclei into a heavier nucleus and releasing a large amount of energy.The fusion of deuterium and tritium,the isotope of hydrogen,is the most typical and easy fusion reaction.Nuclear fusion does not produce long-term and high-level nuclear radiation,nuclear waste,greenhouse gases and does not pollute the environment caused like nuclear fission.Thus,scientists have considered Controlled nuclear fusion as one of the most promising research fields to solve permanent clean energy in the future.With the start of the International Thermonuclear Experimental Reactor(ITER)Program,the tokamak experiment device has gradually become the mainstream research object of scientists worldwide.The tokamak is a transliteration of an acronym of four separate Russian words: toroidal,kamera,magnet,and kotushka.The development of tokamak experiment research made researchers find that it is necessary to design and build larger-scale experimental equipment,innovate the tokamak modeling method and optimize the controller in order to obtain plasma with higher parameters and higher control precision,such as advanced configuration plasma with higher temperature,higher current,higher density,and larger elongation ratio.The research work in this thesis focuses on the deep learning modeling and plasma control of HL-2A and HL-2M.The main research contents and innovations are as follows.(1)The thesis analyzes the traditional modeling method of plasma equilibrium response from the HL-2A tokamak.Then,according to the mathematical model of singleinput single-output(SISO)system,design the PID controller of each independent subsystem of HL-2A plasma(plasma current feedback control system,plasma vertical displacement feedback control system,and plasma horizontal displacement feedback control system).By analyzing the plasma control system of HL-2A tokamak,the section clarifies the relationship between the inputs and outputs in each subsystem control loop.In the HL-2A plasma controller optimal design,this thesis proposes an adaptive PID control algorithm based on radial basis function neural network for the HL-2A network models.The new control algorithm obtains a good control effect in the simulation study and enriches the design method of the tokamak plasma controller.(2)For obtaining a higher precision plasma prediction model,this thesis proposes a deep learning modeling method based on a long short-term memory network(LSTM)according to some historical experimental data of HL-2A.The model is applied to predict the tokamak plasma position for the first time.In the prediction analysis of this network model,the thesis evaluates its prediction performance from multiple dimensions,such as the root mean square error,the goodness of fit,correlation,and compares it with traditional modeling methods.The results show that the prediction accuracy of the deep neural network model designed in this thesis has high accuracy in predicting the plasma current,vertical displacement,and horizontal displacement in the discharge experiment.(3)In HL-2M,because the system inputs are interrelated and coupled,the system presents the characteristics of a multi-input multi-output system(MIMO),the controlled quantities(plasma current,the plasma’s vertical displacement,and the plasma’s horizontal displacement)are no longer controlled separately.Therefore,it is necessary to study the deep learning modeling of the multi-input multi-output system based on a long short-term memory network for the HL-2M.A plasma configuration simulation platform based on MATLAB developed by the Southwestern Institute of Physics is used to simulate the HL-2M discharge experiments and generate 5500 shot experimental data.These experimental data are applied to build a training set,a verification set,and a test set.And the thesis applies the long short-term memory network to the deep learning modeling of the MIMO tokamak plasma system.The trained deep neural network model has good prediction performance for the HL-2M plasma current and horizontal displacement simulated experimental data.(4)The first discharge experiment of the HL-2M device requires producing a limiter plasma.According to the characteristics of the MIMO system of HL-2M plasma controlled plant,an improved scheme of adaptive PID control algorithm based on radial basis function neural network is proposed to be suitable for the MIMO system in this thesis.The controller parameters obtained by the algorithm provide a useful reference for the HL-2M discharge experiments in December 2020.Finally,the thesis analyzes the HL-2M control system structure and makes the experimental plan in detail.After hundreds of experiments,the HL-2M has obtained the output response with continuous plasma current successfully.