Tokamak plasma rupture prediction method based on cWGAN and TCN

Abstract

The invention discloses a Tokamak plasma rupture prediction method based on cWGAN and TCN, which takes AUC as a measurement index to evaluate the effectiveness of a model, and adopts a conditional Wasserstein generative adversarial network (cWGAN)-based minority class sample generation method for solving the problem that the number of minority class samples is limited. Gradient penalty (GP) is introduced in a training stage to improve the quality of generated samples and training stability, and an inverse frequency mini-batch minority class equilibrium training strategy is adopted to enable a generator to fully learn distribution characteristics of broken samples, so that high-quality minority class samples are generated and a training set is expanded. A Transform structure is introduced on the basis of a time convolutional network (TCN), local time sequence features are efficiently extracted by using the TCN, and the Transform enhances the expression ability of a model for a long-time dependency relationship and global features, so that the prediction effect is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear fusion, and particularly relates to a tokamak plasma disruption prediction method based on cWGAN and TCN. BACKGROUND

[0002] In a large tokamak device (such as ITER), plasma disruption is an important event that affects the safety and stability of the device, and once it occurs, it may cause serious damage to the device. Therefore, it is of great engineering significance to predict the plasma disruption in advance. The existing disruption prediction methods mainly rely on deep learning models, which are trained by a large number of historical disruption samples to enable the model to automatically learn the complex time sequence features before the disruption. However, as the ITER and other large tokamak devices are operated, it is difficult to obtain disruption samples, resulting in a limited number of disruption samples available for training. Under such a small sample condition, the traditional deep learning model is difficult to fully learn the features of the minority class samples, resulting in a decline in the generalization ability of the model and a significant limitation of the prediction performance.

[0003] To alleviate the small sample problem, existing methods attempt to use data enhancement means such as oversampling and noise disturbance to expand the size of the minority class samples. However, these methods have obvious limitations: on the one hand, the generated samples usually lack the distribution characteristics consistent with the real events and cannot accurately reflect the complex time sequence patterns before the disruption; on the other hand, noise disturbance often destroys the structural features of the original data, making it difficult for the trained model to achieve good prediction results in actual application. Therefore, how to generate high-quality minority class samples through effective data enhancement means under the condition of limited number of minority class samples and improve the modeling ability of the model for complex time sequence features is still an important technical problem in the research of disruption prediction. SUMMARY

[0004] To address the technical problem of existing rupture prediction methods struggling to fully learn minority class features under small sample conditions, leading to degraded prediction performance, this invention provides a tokamak plasma rupture prediction method based on cWGAN and TCN. The effectiveness of the model is evaluated using AUC as a metric. To address the limited number of minority class samples, a minority class sample generation method based on Conditional Wasserstein Generative Adversarial Network (cWGAN) is employed. Gradient penalty (GP) is introduced during the training phase to improve the quality of generated samples and training stability. An inverse-frequency mini-batch minority class balancing training strategy is used to enable the generator to fully learn the distribution characteristics of rupture samples, thereby generating high-quality minority class samples and expanding the training set. To address the insufficient ability of prediction models to model complex temporal features, a Transformer structure is introduced on top of the Temporal Convolutional Network (TCN). TCN efficiently extracts local temporal features, while the Transformer enhances the model's ability to express long-term dependencies and global features, thus improving prediction performance. This invention's method can expand training data under limited minority class sample conditions, improve the model's learning ability for minority class features, and increase rupture prediction accuracy.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] A tokamak plasma breakup prediction method based on cWGAN and TCN includes the following steps:

[0007] Step 1: Preprocess the tokamak plasma time series data and divide it into training set, validation set and test set according to a certain ratio. The broken samples are the minority class samples and the non-broken samples are the majority class samples to ensure that each dataset is independent of the others.

[0008] Step 2: Construct a minority class sample generation model based on Conditional Wasserstein Generative Adversarial Network (cWGAN). Introduce gradient penalty (GP) during the training phase to improve the quality of generated samples and training stability. Adopt an inverse frequency mini-batch minority class balanced training strategy to weight the training samples according to class frequency, thereby enhancing the generator's ability to learn the distribution characteristics of minority class samples.

[0009] Step 3: Use the trained generator to generate minority class samples, and merge the generated minority class samples with the original minority class and majority class samples to expand the training set, thereby increasing the proportion of minority class samples in the training data.

[0010] Step 4: Construct a rupture prediction model. The model uses a temporal convolutional network (TCN) as the main structure to extract local temporal features of plasma time series signals. Based on this, a Transformer structure is introduced to enhance the model's ability to model long-term dependencies and global features.

[0011] Step 5: Input the expanded training set into the TCN-Transformer prediction model for training, and obtain a rupture prediction model that can take into account both minority and majority class features.

[0012] Step 6: Use the model weights obtained during training to predict on the validation set and compare the prediction results with the true labels. Optimize the model using the loss function of the validation set and incorporate an early stopping strategy to prevent overfitting. Finally, perform predictions on the test set using the AUC as the primary metric to evaluate the rupture prediction performance.

[0013] Furthermore, in step 1, data preprocessing includes resampling and normalization.

[0014] Furthermore, in step 2, the minority sample generation model introduces gradient penalty (GP) during the training phase and adopts an inverse frequency mini-batch minority balanced training strategy. By adjusting the ratio of minority and majority class samples in the batch through class frequency weighted sampling, the generator's ability to learn the distribution characteristics of minority samples is enhanced.

[0015] Furthermore, in step 4, the introduced Transformer module employs a multi-head attention mechanism and a feedforward neural network to globally model the temporal features extracted by the TCN, thereby improving the model's performance in complex temporal prediction tasks.

[0016] Furthermore, in step 6, in addition to AUC, evaluation metrics can also include recall, F1 score, etc., to comprehensively measure the model's performance in the rupture prediction task.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] The cWGAN minority class sample generation method proposed in this invention combines gradient penalty (GP) and inverse frequency mini-batch minority class balancing training strategy, which can generate high-quality rupture samples under the condition of limited minority class samples, thereby effectively expanding the training data. The TCN-Transformer prediction model built on this basis has both local feature extraction and global feature modeling capabilities, thereby improving the accuracy of rupture prediction. This provides operators with sufficient time to take corresponding control and intervention measures to avoid or mitigate the impact of rupture, thereby ensuring the safe operation of the tokamak device. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the tokamak plasma breakup prediction method using cWGAN and TCN of the present invention;

[0020] Figure 2 This is a flowchart of the training process of cWGAN in this invention;

[0021] Figure 3 This is a graph showing the quality assessment of the generated samples using cWGAN from this invention.

[0022] Figure 4 This is a schematic diagram of the TCN-Transformer rupture prediction model structure of the present invention;

[0023] Figure 5 This is a schematic diagram of the ROC curve of the present invention; Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0026] Figure 1 This is a flowchart illustrating the overall process of the tokamak plasma breakage prediction method based on cWGAN and TCN of this invention. The specific implementation steps are as follows:

[0027] Step 1: Preprocessing of Tokamak plasma time-series data. Resampling unifies continuous time-series signals with different sampling rates to the same sampling frequency to ensure consistency of the time axis. Then, cubic spline interpolation is used to regenerate the signal data on the unified time axis, resulting in a multi-channel time-series signal with consistent time resolution. After resampling, each channel signal is normalized, scaling the values ​​to the [0,1] range to eliminate dimensional differences between different physical quantities. Finally, the resampled and normalized signal is sliced ​​according to a fixed time window, with each sample having a shape of (T,M), where T represents the number of time steps within the window, and M is the number of diagnostic signal channels. After the above preprocessing, the obtained signal maintains consistency in sampling rate, time scale, and numerical range, and can be directly used for subsequent minority class sample generation and prediction model training.

[0028] Step 2: Construct a minority class sample generation model based on Conditional Wasserstein Generative Adversarial Network (cWGAN). Gradient penalty (GP) is introduced during the training phase to improve the quality and training stability of generated samples. An inverse-frequency mini-batch minority class balanced training strategy is employed, weighting the training samples by class frequency to enhance the generator's ability to learn the distribution characteristics of minority class fragmented samples. The sampling weight formula is:

[0029]

[0030] Where w c The sampling weights for this category in the mini-batch, where N is the total number of samples in the training set, n c Let be the number of samples in category c. The objective function of cWGAN is:

[0031]

[0032] in To generate samples, x represents the real sample, y represents the class condition, and D(·,y) represents the discriminator function. To compare the real sample x with the generated sample Samples obtained by random interpolation between them For the discriminator pair The gradient is denoted by λ, which is the gradient penalty coefficient used to control the constraint strength. The input sample has a shape of (T, M) and corresponds to the minority class label. This policy generator can fully capture the distribution characteristics of the minority class samples.

[0033] Step 3: Use the trained generator to generate minority class fragmented samples, merge them with the original minority and majority class samples to expand the training set, thereby increasing the proportion of minority class samples in the training data. The generated samples have the same shape as the original samples, which are (T,M) tensors, and do not require further normalization or pruning.

[0034] Step 4: Construct a rupture prediction model based on a Temporal Convolutional Network (TCN), and introduce a Transformer structure on top of it. The TCN module is used to extract local temporal features, with input sample shape (T, M); the Transformer module uses a multi-head attention mechanism and a feedforward neural network to globally model the temporal features extracted by the TCN, enhancing the model's ability to express long-term dependencies and global features. The formula is expressed as:

[0035] H output =Transformer(H TCN ),y=σ(WH output +b)

[0036] Where σ is the Sigmoid function and y is the output probability of breakage. The expanded training set is input into the TCN-Transformer model for training. The input shape is (N,T,M). During training, the model weights learn both majority and minority class features simultaneously, and the binary cross-entropy loss function is used for optimization.

[0037] Step 5: After training, adjust the model hyperparameters based on the loss on the validation set to optimize model performance. Finally, predict samples on an independent test set and calculate the AUC evaluation metric.

[0038]

[0039] In summary, this invention provides a tokamak plasma rupture prediction method based on cWGAN and TCN. By generating high-quality rupture samples under the condition of limited minority class samples, and combining the TCN-Transformer model to model local and global temporal features, high-precision prediction of plasma rupture is achieved. This method effectively expands the number of minority class samples in the training data, improves the model's learning ability for minority class features, and enhances the model's prediction accuracy.

[0040] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting Tokamak plasma disruption based on cWGAN and TCN, characterized in that, The method comprises the following steps: Step 1: preprocessing tokamak plasma time series data, and dividing into training set, validation set and test set according to a certain proportion, wherein the disruption sample is taken as minority class sample, and the non-disruption sample is taken as majority class sample, so as to ensure that each data set is independent; Step 2: constructing a minority class sample generation model based on a conditional Wasserstein generative adversarial network cWGAN, and introducing a gradient penalty GP in the training stage to improve the quality and stability of the generated samples; meanwhile, an inverse frequency mini-batch minority class balanced training strategy is adopted, and the training samples are sampled according to the category frequency, so as to enhance the learning ability of the generator to the distribution characteristics of the minority class samples; Step 3: generating minority class samples by using the trained generator, and merging the generated samples with the original minority class samples and majority class samples to expand the training set, so as to improve the proportion of minority class samples in the training data; Step 4: constructing a disruption prediction model, taking a time convolution network TCN as the main structure to extract local time sequence features of plasma time series signals, introducing a Transformer structure on this basis, and using a multi-head attention mechanism and a feedforward neural network to model the time sequence features extracted by the TCN to enhance the expression ability of the model to long-time dependence and global features; Step 5: inputting the expanded training set into the TCN-Transformer prediction model for training to obtain a disruption prediction model that can consider both minority class and majority class features; Step 6: using the model weight obtained in the training stage to predict the validation set, and comparing the prediction result with the true label. The model is optimized through the loss function of the validation set, and an early stopping strategy is combined to prevent overfitting. Finally, the trained model is used to predict the test set, and the AUC is taken as the main index to evaluate the disruption prediction performance.

2. The cWGAN and TCN-based tokamak plasma disruption prediction method of claim 1, wherein, In step 1, the data preprocessing includes resampling and normalization.

3. The cWGAN and TCN-based tokamak plasma disruption prediction method of claim 1, wherein, In step 2, the minority class sample generation model introduces a gradient penalty (GP) in the training stage, and adopts an inverse frequency mini-batch minority class balanced training strategy, which adjusts the proportion of minority class and majority class samples in the batch by weighted sampling according to the category frequency, so as to enhance the learning ability of the generator to the distribution characteristics of the minority class samples.

4. The cWGAN and TCN-based tokamak plasma disruption prediction method of claim 1, wherein, In step 4, the introduced Transformer module adopts a multi-head attention mechanism and a feedforward neural network to globally model the time sequence features extracted by the TCN, so as to improve the performance of the model in complex time series prediction tasks.

5. The cWGAN and TCN-based tokamak plasma disruption prediction method of claim 1, wherein, In step 6, in addition to the AUC, the evaluation index can also include recall rate, F1 value, etc., so as to comprehensively measure the performance of the model in the disruption prediction task.

Images