A data-driven ring voltage control method in fusion

Abstract

The present invention relates to the field of ring voltage control technology, and in particular to a ring voltage control method based on data-driven in fusion. Its technical solution includes the following method steps: selecting data for training models: taking plasma-related parameters and the target ring voltage to be controlled at the next moment as input, and outputting the LHW power at the next moment; building a database: using the discharge data of the tokamak device and the early trial discharge data of the future device as part of the database; building a neural network model: using an LSTM neural network, configured as a three-layer structure, each layer contains 35, 10, and 5 neurons respectively. The present invention combines artificial intelligence with fusion plasma control, automatically summarizes the relationship between various physical quantities through the training of neural networks, explores the potential value behind the data, and can accurately give the LHW power under the target ring voltage through the neural network, thereby achieving the purpose of precise control.

Description

Technical Field

[0001] The present invention relates to the technical field of ring voltage control, and in particular to a data-driven ring voltage control method in fusion. Background Art

[0002] Nuclear fusion is a nuclear reaction process that fuses the nuclei of two light elements into a heavier nucleus. The conditions for nuclear fusion to occur are extremely harsh, requiring ultra-high temperatures and ultra-high pressures to overcome the charge repulsion between the nuclei. In a high temperature and high pressure environment, particles are heated and excited to reach a plasma state. The temperature of this high-temperature plasma can reach hundreds of millions of degrees, and current materials cannot withstand such high temperatures. Therefore, it is proposed to use magnetic fields to confine these high-temperature plasmas. Tokamak is a device used to achieve nuclear fusion reactions, which uses a strong magnetic field to confine plasma in a ring-shaped vacuum chamber. In order to achieve fusion discharge, the operation of future fusion tokamaks must inevitably explore longer discharge times, higher discharge parameters, and safer and more stable discharge schemes, which puts higher requirements on the control of plasma in tokamaks. High-parameter long pulse stable control requires stable control of plasma parameters, such as ring voltage and density. Effective control of ring voltage can reduce volt-second consumption, thereby maintaining discharge for a longer time. At present, the control of ring voltage is mainly achieved by adjusting low-noise (LHW) power. However, the LHW-driven plasma current is affected by many factors: the wall conditions of the tokamak device, longitudinal field, electron density, electron temperature, synergistic effect of auxiliary heating, etc., which brings great challenges to the precise control of the ring voltage.

[0003] The particles in the tokamak device are heated and ionized, presenting a plasma state, and are mainly constrained by the magnetic field generated by the longitudinal field TF coil. The plasma inside the device moves in a directional manner around the inside of the device under the action of the magnetic field. In order to heat the plasma and maintain high temperature conditions, the tokamak device is equipped with a series of coils (polar field PF coils) and auxiliary heating means. Auxiliary heating means include low noise (LHW), electron cyclotron (ECRH), ion cyclotron (ICRF), etc. As mentioned earlier, precise control of the ring voltage in the tokamak device is one of the key technologies to achieve long pulse stable discharge. At present, the control of the ring voltage adopts a combination of feedforward and feedback, and the actuator is LHW. Before the tokamak discharges, the power of LHW is preset, and the feedforward control gives the preset LHW power. During the discharge process, the LHW power is adjusted in real time according to the real-time detected ring voltage signal using feedback control to control the ring voltage. The existing controller has the following disadvantages:

[0004] (1) The feedforward preset waveform is difficult to accurately match the actual discharge

[0005] The preset LHW waveform in the feedforward control is often difficult to meet a variety of different discharge conditions. In actual discharge experiments, it often needs to be adjusted according to the experimental conditions, and its robustness is poor. In addition, the adjusted feedforward waveform is often not the optimal waveform in actual discharge, which requires feedback control to have a larger margin for adjustment. In the control of the loop voltage, feedback control uses traditional PID control, and the power of LHW is limited, which also leads to limited control ability of the loop voltage. In actual discharge, if the LHW power given by the feedforward causes the loop voltage to deviate from the target by a large amount, then the overall control range will be greatly reduced, and even the entire controller will fail.

[0006] (2) Differences between modeling and actual existence of feedback controllers

[0007] As mentioned above, the LHW driven plasma current is affected by many factors and is nonlinear. The existing PID feedback controller needs to be simplified when constructing the state space equation. The simplified state space equation leads to a limited control range of the constructed PID controller, which greatly reduces the overall control capability.

[0008] In summary, this application proposes a data-driven ring voltage control method in fusion. Summary of the invention

[0009] The purpose of the present invention is to address the problem in the background technology that the LHW driven plasma current is affected by multiple factors, and to propose a data-driven ring voltage control method in fusion.

[0010] The technical solution of the present invention is a data-driven ring voltage control method in fusion, comprising the following method steps:

[0011] Select the data for training the model: take the plasma related parameters and the target ring voltage to be controlled at the next moment as input, and output the LHW power at the next moment;

[0012] Building a database: Using the discharge data of the tokamak device and the early discharge data of future devices as part of the database;

[0013] Build a neural network model: Use LSTM neural network, configured as a three-layer structure;

[0014] Model training: Build a model on the mainstream framework and use the collected data for training to obtain a controller;

[0015] Testing the model: Use a test set other than the training data to check the difference between the predicted results and the actual results;

[0016] Replace the current loop voltage control: Replace the original loop voltage control algorithm with a model that has been tested and meets the accuracy requirements, collect data in real time, and provide LHW power signals based on model prediction to accurately control the loop voltage.

[0017] Optionally, plasma-related parameters include total plasma current, longitudinal field coil current, PF coil current, average density of plasma electron strings, polar pressure ratio, plasma internal inductance, energy storage, plasma boundary, radiation power, ring voltage, plasma boundary, auxiliary heating ECRH power, auxiliary heating ICRF power, and auxiliary heating LHW power.

[0018] Optionally, the mainstream framework is either TensorFlow or Pytorch.

[0019] Optionally, in the constructed neural network model, the LSTM neural network is configured as a three-layer structure, with each layer containing 35, 10, and 5 neurons respectively.

[0020] Optionally, in the test of the model, the error is required to be less than 5%.

[0021] In summary, the present application includes at least one of the following beneficial technical effects:

[0022] The data-driven controller proposed in the present invention can avoid the complex relationship between plasmas and achieve such complex control effects only through data training. Compared with traditional controllers, it is not only easier to implement, but also has better control effects.

[0023] The present invention combines artificial intelligence with fusion plasma control, automatically summarizes the relationship between various physical quantities through neural network training, explores the potential value behind the data, and accurately gives the LHW power under the target ring voltage through the neural network, thereby achieving the purpose of precise control. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic structural diagram of a flow chart of a ring voltage control method based on a neural network of the present invention is given;

[0025] Figure 2 It is a neural network model diagram;

[0026] Figure 3 Schematic diagram of the structure of LHW power predicted based on the neural network model. DETAILED DESCRIPTION

[0027] The technical solution of the present application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians of the art without making creative work are within the scope of protection of the present application. Example

[0028] This application will develop an advanced loop voltage control method using a neural network algorithm from a data perspective. Data-driven neural networks can automatically summarize patterns from data to find potential relationships. In the LHW power control loop voltage experiment, the LHW drive current is affected by multiple factors, and the relationship is complex. It is currently difficult to summarize the patterns well, which is consistent with the application scenario of data-driven neural networks. The implementation process of the loop voltage control method based on neural networks is as follows: Figure 1 shown.

[0029] 1. Select data for training the model

[0030] Taking full account of the factors of LHW driving current control loop voltage, the plasma parameters total plasma current, longitudinal field coil current, PF coil current, average density of plasma electron string, poloidal pressure ratio, plasma internal inductance, energy storage, plasma boundary, radiation power, ring voltage, plasma boundary, auxiliary heating ECRH power, auxiliary heating ICRF power, auxiliary heating LHW power and the target ring voltage to be controlled at the next moment are taken as input, and the output is the LHW power at the next moment.

[0031] 2. Building a database

[0032] After determining the input and output of the model, we need to collect data to build a database. The discharge data of the tokamak device will be stored during operation, and we can directly select the existing data. In the early stage of the future device, we will also try some discharges, and we can also obtain these data as initial training data.

[0033] 3. Building a Neural Network Model

[0034] Since some parameters have cumulative effects over time, the long short-term memory (LSTM) neural network is considered as the training model. The model results are as follows: Figure 2 shown.

[0035] The model uses a three-layer network, each layer contains 35, 10, and 5 neurons respectively. The number of neurons and the number of layers of the neural network can be appropriately changed for different devices and the amount of data obtained.

[0036] 4. Model Training

[0037] The model can be built and trained based on the current mainstream frameworks, such as tensorflow, Pytorch, etc. Choose one of the frameworks to build the model, and then put the data into the model for training to obtain the controller.

[0038] V. Model Testing

[0039] In addition to the data used for training, you also need to prepare some data for testing the model. Input new data into the model and compare the difference between the predicted results and the actual results to determine whether the model is effective. The difference between the predicted results and the actual results is required to be no more than 5%.

[0040] 6. Replace the current loop voltage feedforward control

[0041] After completing the test in step 5, if the accuracy requirement is met, the trained model can be used to replace the traditional loop voltage control algorithm. By collecting signals in real time and providing LHW power signals in real time according to the requirements of loop voltage control, accurate control of the loop voltage can be achieved.

[0042] VII. Experimental Verification

[0043] like Figure 3 This is the LHW power predicted based on the neural network model, gun number: 120225. Based on the currently running Tokamak device, about 1800 gun discharge data are selected to train the model, and 30 gun discharge data are used to test and verify the feasibility of the invention. The trained neural network is used to predict the power of LHW. The error between the prediction result of 30 guns and the real-time LHW power is no more than 5%. One of the gun data is selected for display, and the predicted result is as follows Figure 3 shown. Figure 3 The blue solid line in the figure shows the result of the experimental discharge, and the red solid line shows the result predicted by the neural network. Figure 3 It can be found that the LHW power predicted by the neural network is basically consistent with the LHW power given by the current feedforward plus feedback control. This is enough to show that the model trained by the neural network can accurately give the LHW power. The data-driven controller can avoid the complex relationship between plasmas and achieve this complex control effect only through data training. Compared with traditional controllers, it is not only easier to implement, but also has better control effects.

[0044] The data-driven neural network controller proposed in the present invention realizes the ring voltage control method, creatively combining artificial intelligence with fusion plasma control. The relationship between various physical quantities is automatically summarized through neural network training, and the potential value behind the data is explored. The neural network can accurately give the LHW power under the target ring voltage, thereby achieving the purpose of precise control.

[0045] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A data-driven ring voltage control method in fusion, characterized in that: The method comprises the following steps: Select the data for training the model: take the plasma related parameters and the target ring voltage to be controlled at the next moment as input, and output the LHW power at the next moment; Building a database: Using the discharge data of the tokamak device and the early discharge data of future devices as part of the database; Build a neural network model: Use LSTM neural network, configured as a three-layer structure; Model training: Build a model on one of the frameworks, TensorFlow or Pytorch, and use the collected data for training to obtain a controller. Testing the model: Use a test set other than the training data to check the difference between the predicted results and the actual results; Replace the current loop voltage control: Replace the original loop voltage control algorithm with a model that has been tested and meets the accuracy requirements, collect data in real time, and provide LHW power signals based on model prediction to accurately control the loop voltage.

2. A data-driven ring voltage control method in fusion according to claim 1, characterized in that: Plasma-related parameters include total plasma current, longitudinal field coil current, PF coil current, average density of plasma electron strings, polar pressure ratio, plasma internal inductance, energy storage, plasma boundary, radiation power, ring voltage, plasma boundary, auxiliary heating ECRH power, auxiliary heating ICRF power, and auxiliary heating LHW power.

3. A data-driven ring voltage control method in fusion according to claim 1, characterized in that: The mainstream framework is either TensorFlow or Pytorch.

4. The data-driven ring voltage control method in fusion according to claim 1, characterized in that: In the constructed neural network model, the LSTM neural network is configured as a three-layer structure, and each layer contains 35, 10, and 5 neurons respectively.

5. The data-driven ring voltage control method in fusion according to claim 1, characterized in that: In the test of the model, the error is required to be less than 5%.

Images