Tokamak plasma magnetic response model construction method and system and medium

Abstract

The invention discloses a Tokamak plasma magnetic response model construction method, a Tokamak plasma magnetic response model construction system and a medium. Relates to the technical field of nuclear fusion plasma simulation. According to the scheme, improvement is carried out on the basis of the prior art, modeling is carried out on the magnetic response of the Tokamak plasma on the basis of basic data of the Tokamak plasma, and plasma behaviors are directly learned from experimental data; the simulator taking the plasma magnetic configuration evolution autoregression model as a carrier can process nonlinear terms and high-order terms, and is low in calculation cost, high in speed and capable of being applied in real time; according to the method, the behavior of the plasma under magnetic control can be well simulated, the plasma behavior is directly learned from experimental data without simplifying and assuming plasma properties, and the method can be used for off-line simulation and can also be used for control strategy optimization or fault detection in real time.

Description

Technical Field

[0001] This invention relates to the field of nuclear fusion plasma simulation technology, specifically to a method, system, and medium for constructing a tokamak plasma magnetic response model. Background Technology

[0002] Plasma magnetic response models are mathematical models that describe how plasma responds to external disturbances (such as changes in electromagnetic fields) to its state (e.g., current distribution, position, shape). In tokamak plasma simulations, common plasma magnetic response models include the RZIP matrix model, which assumes the plasma is an axisymmetric rigid body, and the CREATE-L model. These models aim to describe the interaction between the plasma current in the poloidal coil, the plasma current center, and the magnetic flux at the control point. They have advantages such as strong robustness and easy acquisition of dynamic responses, which are beneficial for analyzing the dynamic behavior of plasma in control systems. However, these models have the following problems:

[0003] First-principles plasma response models require assumptions about plasma properties, followed by mathematical modeling based on these assumptions. However, plasma behavior and the experimental environment are complex and variable, and factors ignored during mathematical modeling can cause significant deviations between simulation and actual experimental results. First-principles plasma response models are complex and computationally expensive, limiting their application in real-time computing and making them unsuitable for real-time control and fault monitoring. While linear approximations can improve computational speed and enable real-time processing, this reduces model accuracy, resulting in inaccurate simulations, especially in long-cycle experiments where such errors accumulate over time. Summary of the Invention

[0004] The technical problem this invention aims to solve is that traditional plasma response models require assumptions about plasma properties, resulting in significant discrepancies between simulation and actual experimental results. This invention provides a method, system, and medium for constructing a tokamak plasma magnetic response model. Based on existing technologies, it improves upon existing methods by modeling the magnetic response of tokamak plasma using fundamental data. It directly learns plasma behavior from experimental data, and the simulator, based on an autoregressive model of plasma magnetic configuration evolution, can handle nonlinear and higher-order terms. Furthermore, it boasts low computational cost, high speed, and real-time application. It can effectively simulate plasma behavior under magnetic control without simplifying or assuming plasma properties, directly learning plasma behavior from experimental data. This allows for both offline simulation and real-time application in control strategy optimization or fault detection.

[0005] This invention is achieved through the following technical solution:

[0006] This solution provides a method for constructing a tokamak plasma magnetic response model, including:

[0007] Basic data of the tokamak plasma are collected and preprocessed; the basic data includes coil current control parameters and plasma magnetic configuration parameters.

[0008] A plasma magnetic configuration evolution autoregressive model is constructed, and the model is trained based on a dataset built from preprocessed basic data. The model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment.

[0009] The target initial plasma magnetic configuration state parameters and coil current control parameters of the tokamak plasma are obtained. The plasma magnetic configuration evolution autoregressive model is input to predict the plasma magnetic configuration state parameters of the tokamak plasma under the target initial plasma magnetic configuration state parameters and coil current control parameters.

[0010] A further optimization scheme is that the preprocessing includes: first removing errors from the basic data and then performing normalization processing.

[0011] A further optimized scheme is that the coil current control parameters include: the poloidal field coil current and the central solenoid coil current;

[0012] The plasma magnetic configuration parameters include at least: plasma current, plasma position, plasma small radius, plasma elongation ratio, and plasma three angles.

[0013] A further optimization scheme is that the plasma magnetic configuration evolution autoregressive model is a WaveNet autoregressive model; in the WaveNet autoregressive model, a learnable weight matrix W is embedded in the skip connection path of each dilated convolutional layer. skip :h skip = W skip ⊙ h layer ; where h skip h represents the output feature vector after the skip connection. skip h represents the learnable skip connection weight matrix. layer ⊙ represents the original output features of each dilated convolutional layer; ⊙ represents the matrix multiplication operator.

[0014] A further optimization scheme is that the plasma magnetic configuration evolution autoregressive model also includes a random walk noise layer and a Gaussian noise layer at the input end; wherein, the random walk noise layer is added to the plasma magnetic configuration to simulate the error accumulation of the plasma magnetic configuration evolution autoregressive model in the autoregressive usage mode; the Gaussian noise layer is added to the coil current to simulate the system noise present in the experiment.

[0015] A further optimized solution is that the formula for generating the random walk noise is:

[0016] X t =X t-1 +δ t ;

[0017] Among them, X t X represents the cumulative error noise value of the time series at time point t; t-1 δ represents the cumulative error noise value of the time series at time point t-1; t This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise.

[0018] A further optimization scheme is that the loss function L of the plasma magnetic configuration evolution autoregressive model is... tot for:

[0019] L tot = L1 + ;

[0020] Where L1 represents the first-step prediction error; L i λ represents the error at the i-th step after the initial prediction error; i This represents the weight of the error at step i.

[0021] A further optimization scheme is that the autoregressive path of the plasma magnetic configuration evolution autoregressive model is weighted and fused with a teacher-forced path; the teacher-forced path uses real plasma magnetic response data as input; the autoregressive path is:

[0022] L = λ TF ·LTF + λ AR ·LAR;

[0023] Where L represents the loss function value used in the backpropagation algorithm; LTF represents the autoregressive path of the WaveNet autoregressive model; λ TF The weights of the autoregressive path in the WaveNet autoregressive model are represented, gradually decaying from the initial values ​​to the minimum weight limit according to the cosine annealing curve; LAR represents the teacher-forced path; λ AR This represents the weight of the teacher-mandated path.

[0024] This solution also provides a tokamak plasma magnetic response model construction system for implementing the above-mentioned tokamak plasma magnetic response model construction method; the system includes:

[0025] The acquisition module is used to acquire basic data of the tokamak plasma and preprocess the basic data; the basic data includes coil current control parameters and plasma magnetic configuration parameters.

[0026] The model building module is used to construct a plasma magnetic configuration evolution autoregressive model. It constructs a dataset based on preprocessed basic data to train the plasma magnetic configuration evolution autoregressive model. The plasma magnetic configuration evolution autoregressive model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The plasma magnetic configuration evolution autoregressive model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment.

[0027] The prediction module is used to obtain the target coil current control parameters of the tokamak plasma, input the plasma magnetic configuration evolution autoregressive model, and predict the plasma magnetic configuration state parameters of the tokamak plasma under the coil current control parameters and the target initial plasma magnetic configuration state parameters.

[0028] This solution also provides a computer-readable medium storing a computer program thereon, which, when executed by a processor, can implement a tokamak plasma magnetic response model construction method as described above.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] 1. This invention provides a method, system, and medium for constructing a magnetic response model of tokamak plasma. Based on existing technologies, it improves upon existing methods by modeling the magnetic response of tokamak plasma using fundamental data. It directly learns plasma behavior from experimental data. The simulator, based on an autoregressive model of plasma magnetic configuration evolution, can handle nonlinear and higher-order terms, and features low computational cost, high speed, and real-time application. It can effectively simulate plasma behavior under magnetic control without simplifying or making assumptions about plasma properties, directly learning plasma behavior from experimental data. This allows for both offline simulation and real-time application for control strategy optimization or fault detection.

[0031] 2. The present invention provides a method, system and medium for constructing a tokamak plasma magnetic response model. By using random walk noise enhancement and a multi-step autoregressive loss function, it effectively solves the error accumulation problem of traditional autoregressive models in long-term prediction of plasma current configuration control, ensuring the long-term predictive stability of the model for the evolution of tokamak plasma current and configuration changes, and providing a reliable guarantee for steady-state operation control of plasma.

[0032] 3. The present invention provides a method, system, and medium for constructing a tokamak plasma magnetic response model. Through the design of a composite loss function, the model can simultaneously and accurately predict two different control scenarios: normal current-driven operation and rapid configuration change of the plasma. This achieves unified modeling of the complex current configuration response process of the tokamak plasma magnetic control system, improving the model's versatility and practicality. The dynamic teacher-forced strategy effectively solves the training instability problem in complex plasma current configuration response modeling, ensuring rapid convergence and final performance of the model training, and laying the foundation for the engineering application of the tokamak plasma magnetic control model. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0034] Figure 1 A schematic diagram illustrating the process of constructing a magnetic response model for tokamak plasma;

[0035] Figure 2 This is a schematic diagram of the structure of an autoregressive model for plasma magnetic configuration evolution.

[0036] Figure 3 This is a schematic diagram of the autoregressive process of the model;

[0037] Figure 4 This is a schematic diagram of the single-step prediction results of the trained model on the 3956 gun.

[0038] Figure 5 This is a schematic diagram of the autoregressive prediction results of the trained model on the 3956 gun. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0040] Traditional plasma response models require assumptions about plasma properties, resulting in significant discrepancies between simulations and actual experimental results. Therefore, this solution provides the following embodiments to address the aforementioned technical problems:

[0041] Example 1

[0042] This embodiment provides a method for constructing a tokamak plasma magnetic response model, such as... Figure 1 As shown, it includes:

[0043] Step 1: Collect basic data of the tokamak plasma and preprocess the basic data; the basic data includes coil current control parameters and plasma magnetic configuration parameters.

[0044] The coil current control parameters include: poloidal field coil current and central solenoid coil current;

[0045] The plasma magnetic configuration parameters include at least: plasma current, plasma position, plasma minima, plasma elongation ratio, and plasma three angles (plasma inductance, plasma density, and plasma ring pressure).

[0046] Constructing an autoregressive model of plasma magnetic configuration evolution requires preparing a corresponding dataset. For magnetic control, the inputs to the autoregressive model are all poloidal (PF) coil currents and the central solenoid (CS) coil currents, while the outputs vary depending on the application scenario. When modeling the entire discharge process, the outputs include the plasma currents and all their configuration parameters. However, when applied only to a specific stage of the discharge process, where there are no additional inputs such as gas injection, the outputs can be expanded to include signals such as density and inductance. In this case, plasma inductance, plasma density, and plasma ring pressure can all be used as plasma magnetic configuration state parameters.

[0047] Step 2: Construct a plasma magnetic configuration evolution autoregressive model. Based on the preprocessed basic data, construct a dataset to train the plasma magnetic configuration evolution autoregressive model. The plasma magnetic configuration evolution autoregressive model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The plasma magnetic configuration evolution autoregressive model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment.

[0048] The autoregressive model of plasma magnetic configuration evolution needs to learn a mapping function from X (coil current control parameters) to Y (plasma magnetic configuration state parameters) from the dataset. Considering that there is not only spatial coupling but also temporal coupling between different parameters of tokamak plasma, the formula learned by the autoregressive model of plasma magnetic configuration evolution is Y. T =F(X (T-t, T) Y (T-t-1, T-1) In the formula, T represents the time that the plasma magnetic configuration evolution autoregressive model needs to predict, in milliseconds; t represents the input time length of the plasma magnetic configuration evolution autoregressive model, in milliseconds; and 1 in the formula represents 1 millisecond. The plasma magnetic configuration evolution autoregressive model requires input X data at time [Tt, T] and Y data at time [Tt-1, T-1] to obtain Y data at time T. The construction of the plasma magnetic configuration evolution autoregressive model can refer to mainstream autoregressive models such as Wavenet, but it is necessary to customize the parameter configuration of the model by combining characteristic time information of plasma magnetic configuration evolution.

[0049] In this embodiment, the plasma magnetic configuration evolution autoregressive model is the WaveNet autoregressive model; in the WaveNet autoregressive model, a learnable weight matrix W is embedded in the skip connection path of each dilated convolutional layer. skip :h skip = W skip ⊙h layer ; where h skip h represents the output feature vector after the skip connection. skip h represents the learnable skip connection weight matrix. layer ⊙ represents the original output features of each dilated convolutional layer; ⊙ represents the matrix multiplication operator.

[0050] This paper presents a frequency-aware skip connection weight matrix for the plasma magnetic response. Tokamak plasma magnetic response signals contain rich spectral components, ranging from low-frequency magnetic configuration drift (~Hz) to high-frequency magnetohydrodynamic oscillations (~kHz). Traditional WaveNet architectures with hard connections cannot distinguish the varying importance of these different frequency components to plasma magnetic control. This proposed method, based on a learnable skip connection weight matrix, enables the model to adaptively learn the contribution weights of different frequency components to plasma magnetic control. After training, the weight matrix can be analyzed as the frequency response characteristics of the plasma magnetic response, intuitively demonstrating the physical importance of different frequency bands to magnetic configuration evolution. This frequency-aware mechanism is particularly suitable for tokamak plasma magnetic control applications, automatically identifying and enhancing frequency components critical to magnetic confinement performance while suppressing the influence of irrelevant noise.

[0051] The autoregressive model of plasma magnetic configuration evolution also includes a random walk noise layer and a Gaussian noise layer;

[0052] The random walk noise layer is used to add cumulative error noise to the plasma magnetic configuration state parameters of the input plasma magnetic configuration evolution autoregressive model; the formula for generating the random walk noise is:

[0053] X t =X t-1 +δ t ;

[0054] Among them, X t X represents the cumulative error noise value of the time series at time point t; t-1 δ represents the cumulative error noise value of the time series at time point t-1; t This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise.

[0055] A Gaussian noise layer is used to add high-frequency noise perturbation to the coil current control parameters of the input plasma magnetic configuration evolution autoregressive model. The high-frequency noise perturbation is simulated based on the actual coil current measurement process.

[0056] A random walk noise layer is added to the plasma magnetic configuration state parameters input to the plasma magnetic configuration evolution autoregressive model, and a Gaussian noise layer is added to the input coil current control parameters. This is used to further train the plasma magnetic configuration evolution autoregressive model to suppress the prediction error that accumulates over simulation time. The motivation for adding accumulated error noise is to reduce the prediction error that accumulates over time when the output Y at time t is measured. t During calculation, the output Y at time t−1 t-1 It will be used as the input to the model, and Y t-1 The prediction errors already carried over will accumulate in further predictions and amplify as the simulation time increases. By adding random walk noise, the autoregressive model of plasma magnetic configuration evolution can be exposed to this gradually accumulating noise during training and adapt, attempting to correct the accumulated errors that may arise during autoregression, thereby improving the model's ability to perform long-term simulations. In the formula for generating random walk noise, σ... 2 The size was determined to be 0.1 times the variance of the plasma state data.

[0057] Adding a Gaussian noise layer to the coil current input is to simulate the high-frequency noise disturbances that often occur in actual coil current measurements, thus allowing the model to adapt during the training phase and improving its performance in the application phase.

[0058] The autoregressive model for plasma magnetic configuration evolution is based on a multi-step autoregressive loss function. Considering the autoregressive prediction characteristics of tokamak plasma magnetic control, a multi-step autoregressive loss function is designed to address the error propagation problem in plasma magnetic response prediction. In plasma magnetic control applications, the accumulation of prediction errors can lead to a decrease in magnetic configuration control accuracy and may even trigger serious consequences such as plasma rupture. Traditional single-step loss functions only evaluate instantaneous prediction errors and cannot effectively constrain the error propagation of plasma magnetic configuration parameters during the autoregression process. Therefore, the loss function L in this scheme for the plasma magnetic configuration evolution autoregressive model... tot for:

[0059] L tot = L1 + ;

[0060] Where L1 represents the first-step prediction error; L i λ represents the error at the i-th step after the initial prediction error; i This represents the weight of the error at step i. The loss function design considers the time-scale characteristics of tokamak plasma magnetic control: the initial prediction error (L1) corresponds to the immediate response capability of the magnetic control system, with a weight of 1.0; the cumulative errors of the subsequent five steps (L2-L6) reflect the medium- to long-term prediction accuracy of the plasma magnetic configuration evolution, each with a decay weight of 0.1. This design enables the model to learn the long-term evolution law of the plasma magnetic configuration while ensuring the real-time performance of magnetic control, thus improving the predictive stability of magnetic confinement performance.

[0061] The autoregressive model of plasma magnetic configuration evolution incorporates a teacher-forced path in a weighted fusion of autoregressive paths; the teacher-forced path uses real plasma magnetic response data as input; the autoregressive path is as follows:

[0062] L = λ TF ·LTF + λ AR ·LAR;

[0063] Where L represents the loss value used in the backpropagation algorithm; LTF represents the autoregressive path of the WaveNet autoregressive model; λ TF The weights of the autoregressive path in the WaveNet autoregressive model are represented, gradually decaying from the initial values ​​to the minimum weight limit according to the cosine annealing curve; LAR represents the teacher-forced path; λ AR This represents the weight of the teacher-mandated path.

[0064] To address the gradient explosion problem in tokamak plasma magnetic control training, this scheme designs a dynamic teacher-forced strategy. In the early stages of model training, due to the lack of stable multi-step magnetic response prediction capabilities, the autoregressive loss function is prone to gradient explosion, affecting model convergence. This scheme simultaneously employs a teacher-forced path (LTF) based on real plasma magnetic response data input and an autoregressive path (LAR) based on the model's self-predicted magnetic response input, and performs weighted fusion through dynamic weights. This "guided-autonomous" training strategy simulates the learning process of the tokamak plasma magnetic control system: in the early stages of training, it relies on historical plasma magnetic response data for guided learning, gradually transitioning to modeling the magnetic control response based on autonomous prediction, and ultimately achieving accurate prediction of the tokamak plasma magnetic confinement behavior.

[0065] The autoregressive model architecture for plasma magnetic configuration evolution is as follows: Figure 2 As shown, X and Y represent the input and output physical quantities, respectively; T indicates the time the model needs to predict, in milliseconds; t indicates the time length of the model input, in milliseconds; and the subscript 1 indicates 1 millisecond. In the one-dimensional convolutional layer, k is the kernel size, and f is the number of feature layers. The autoregressive process of the model is as follows: Figure 3 As shown, where n represents the length of the sequence, in this process, the model first uses the initial input X1 to X... t and Y0 to Y t-1 To generate the predicted output Y t Subsequently, Y t Incorporated into the input sequence, along with the new input data X2 to X t+1 and Y1 to Y t Together, they are used to predict the output Y at the next time point. t+1 This process is repeated until a complete prediction of the entire Y sequence is generated.

[0066] Step 3: Obtain the target initial plasma magnetic configuration state parameters and coil current control parameters of the tokamak plasma, input them into the plasma magnetic configuration evolution autoregressive model, and predict the plasma magnetic configuration state parameters of the tokamak plasma under the target initial plasma magnetic configuration state parameters and coil current control parameters.

[0067] Example 2

[0068] This embodiment provides a tokamak plasma magnetic response model construction system for implementing the tokamak plasma magnetic response model construction method described in Embodiment 1; the system includes:

[0069] The acquisition module is used to acquire basic data of the tokamak plasma and preprocess the basic data; the basic data includes coil current control parameters and plasma magnetic configuration parameters.

[0070] The model building module is used to construct a plasma magnetic configuration evolution autoregressive model. It constructs a dataset based on preprocessed basic data to train the plasma magnetic configuration evolution autoregressive model. The plasma magnetic configuration evolution autoregressive model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The plasma magnetic configuration evolution autoregressive model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment.

[0071] The prediction module is used to obtain the target initial plasma magnetic configuration state parameters and coil current control parameters of the tokamak plasma. It inputs the plasma magnetic configuration evolution autoregressive model and predicts the plasma magnetic configuration state parameters of the tokamak plasma under the target initial plasma magnetic configuration state parameters and coil current control parameters.

[0072] Example 3

[0073] This embodiment provides a computer-readable medium storing a computer program thereon. The computer program, when executed by a processor, can implement a tokamak plasma magnetic response model construction method as described in Embodiment 1; specifically, it performs the following steps:

[0074] Step 1: Collect basic data of the tokamak plasma and preprocess the basic data; the basic data includes coil current control parameters and plasma magnetic configuration parameters.

[0075] Step 2: Construct a plasma magnetic configuration evolution autoregressive model. Based on the preprocessed basic data, construct a dataset to train the plasma magnetic configuration evolution autoregressive model. The plasma magnetic configuration evolution autoregressive model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The plasma magnetic configuration evolution autoregressive model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment.

[0076] Step 3: Obtain the target initial plasma magnetic configuration state parameters and coil current control parameters of the tokamak plasma, input them into the plasma magnetic configuration evolution autoregressive model, and predict the plasma magnetic configuration state parameters of the tokamak plasma under the target initial plasma magnetic configuration state parameters and coil current control parameters.

[0077] This embodiment takes the magnetic response model construction process of the HL-3 circulator device No. 3 as an example. First, the data of shots #2000 to #6698 on HL-3 are collected as a historical experimental database. 100 shots are randomly selected as the test set, and the rest are used as the training set.

[0078] Then, normalization is performed on all data samples X in the training set, that is, the maximum value X of all X data is calculated. max With minimum value X min According to formula X norm =(XX min ) / (X max -X min The data is scaled to between 0 and 1 to eliminate the influence of different feature units and improve the efficiency of model training. The test set is normalized using the same normalization parameters as the training set to ensure the accuracy and reliability of model evaluation.

[0079] Next, the improved Wavenet network model is used to fit Y on the training set. T = F(X (T-t, T) Y (T-t-1, T-1) The function is then used to check the model's fitting results on the test set. This patent uses the area under the curve (AUC) as a metric to evaluate accuracy; the formula is |Area|. p -Area a Area a Area p The area refers to the area enclosed by the predicted curve and the horizontal axis. a This refers to the area enclosed by the true curve and the horizontal axis. For each feature of each sample in the test set, the accuracy is calculated separately, and then the average is taken to obtain the final accuracy evaluation result. In single-step prediction mode, the model's accuracy on the test set is 99.5%, and in autoregressive mode, the accuracy is 96.5%. The figure below shows the prediction results of the trained model on the 3986 gun. Figure 4 This is the result of a single-step prediction, specifically the single-step prediction result of the trained model on the 3956 gun. In the graph, the yellow line represents the model's predicted value, and the blue line represents the actual value. Therefore, the similarity between the actual signal value and the model's predicted value both reach over 99.5%. Figure 5 These are the results of autoregressive predictions, specifically the autoregressive predictions of the trained model on shot 3956. The similarity between the true values ​​and the model predictions for all signals reached over 97.1%. It can be seen that the model fits the experimental data well in both modes.

[0080] This scheme achieves a breakthrough by reducing the prediction time of plasma current and configuration response in a single time step to less than 1 ms, which is three orders of magnitude faster than the second-level calculation time of traditional numerical simulation. It meets the real-time requirements of the tokamak plasma magnetic control system and provides a technical foundation for real-time plasma current control and configuration control. It breaks away from the limitations of idealized assumptions and geometric approximations in traditional theoretical models, and directly models based on experimental data of tokamak plasma discharge. It achieves high-fidelity prediction of plasma current and configuration response under coil current control, and the consistency between the prediction accuracy and experimental results is significantly better than that of traditional numerical simulation.

[0081] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a tokamak plasma magnetic response model, characterized in that, include: Basic data of the tokamak plasma are collected and preprocessed; the basic data includes coil current control parameters and plasma magnetic configuration parameters. A plasma magnetic configuration evolution autoregressive model is constructed, and the model is trained based on a dataset built from preprocessed basic data. The model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment. The target initial plasma magnetic configuration state parameters and coil current control parameters of the tokamak plasma are obtained. The plasma magnetic configuration evolution autoregressive model is input to predict the plasma magnetic configuration state parameters of the tokamak plasma under the target initial plasma magnetic configuration state parameters and coil current control parameters.

2. The method for constructing a tokamak plasma magnetic response model according to claim 1, characterized in that, The preprocessing includes: first removing errors from the basic data and then normalizing it.

3. The method for constructing a tokamak plasma magnetic response model according to claim 1, characterized in that, The coil current control parameters include: poloidal field coil current and central solenoid coil current; The plasma magnetic configuration parameters include at least: plasma current, plasma position, plasma small radius, plasma elongation ratio, and plasma three angles.

4. The method for constructing a tokamak plasma magnetic response model according to claim 3, characterized in that, The plasma magnetic configuration evolution autoregressive model is a WaveNet autoregressive model; in the WaveNet autoregressive model, a learnable weight matrix W is embedded in the skip connection path of each dilated convolutional layer. skip :h skip = W skip ⊙ h layer ; where h skip h represents the output feature vector after the skip connection. skip h represents the learnable skip connection weight matrix. layer ⊙ represents the original output features of each dilated convolutional layer; ⊙ represents the matrix multiplication operator.

5. The method for constructing a tokamak plasma magnetic response model according to claim 4, characterized in that, The plasma magnetic configuration evolution autoregressive model also includes a random walk noise layer and a Gaussian noise layer at the input end; wherein, the random walk noise layer is added to the plasma magnetic configuration to simulate the error accumulation of the plasma magnetic configuration evolution autoregressive model in the autoregressive usage mode; the Gaussian noise layer is added to the coil current to simulate the system noise present in the experiment.

6. The method for constructing a tokamak plasma magnetic response model according to claim 5, characterized in that, The formula for generating the random walk noise is: X t =X t-1 +δ t ; Among them, X t X represents the cumulative error noise value of the time series at time point t; t-1 δ represents the cumulative error noise value of the time series at time point t-1; t This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise.

7. The method for constructing a tokamak plasma magnetic response model according to claim 4, characterized in that, The loss function L of the autoregressive model of plasma magnetic configuration evolution tot for: L tot = L1 + ; Where L1 represents the first-step prediction error; L i λ represents the error at the i-th step after the initial prediction error; i This represents the weight of the error at step i.

8. The method for constructing a tokamak plasma magnetic response model according to claim 4, characterized in that, The autoregressive model of plasma magnetic configuration evolution incorporates a teacher-forced path in a weighted fusion of the autoregressive paths; the teacher-forced path uses real plasma magnetic response data as input; the autoregressive path is as follows: L = λ TF ·LTF + λ AR ·LAR; Where L represents the loss function value used in the backpropagation algorithm; LTF represents the autoregressive path of the WaveNet autoregressive model; λ TF The weights of the autoregressive path in the WaveNet autoregressive model are represented, gradually decaying from the initial values ​​to the minimum weight limit according to the cosine annealing curve; LAR represents the teacher-forced path; λ AR This represents the weight of the teacher-mandated path.

9. A tokamak plasma magnetic response model construction system, characterized in that, A method for constructing a tokamak plasma magnetic response model according to any one of claims 1-8; the system comprises: The acquisition module is used to acquire basic data of the tokamak plasma and preprocess the basic data; the basic data includes coil current control parameters and plasma magnetic configuration parameters. The model building module is used to construct a plasma magnetic configuration evolution autoregressive model. It constructs a dataset based on preprocessed basic data to train the plasma magnetic configuration evolution autoregressive model. The plasma magnetic configuration evolution autoregressive model learns the mapping function from coil current to plasma magnetic configuration from the dataset. The plasma magnetic configuration evolution autoregressive model is also used to predict the plasma magnetic configuration state parameters at the next moment based on the coil current control parameters at the current moment and the plasma magnetic configuration state parameters at the previous moment. The prediction module is used to obtain the target coil current control parameters of the tokamak plasma, input the plasma magnetic configuration evolution autoregressive model, and predict the plasma magnetic configuration state parameters of the tokamak plasma under the coil current control parameters and the target initial plasma magnetic configuration state parameters.

10. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement a method for constructing a tokamak plasma magnetic response model as described in any one of claims 1-8.

Images