A method and system for graded protection against tokamak plasma disruption

Abstract

The application discloses a tokamak plasma disruption grading protection method and system, which realizes the joint prediction of the disruption occurrence probability and the harm degree of different harm types through a multi-task deep learning model, can not only early warn the disruption risk, but also can predict the harm degree of halo current, thermal load and escaped current; then, through attribution analysis on the multi-modal diagnostic signals, the key features in the multi-modal diagnostic signals are determined, and then the disruption cause is determined based on the key features, and the remaining time window from the current time node to the disruption occurrence is determined; in the grading protection decision stage, combined with the reversibility determined based on the disruption cause, the flow reduction stability determined based on the harm degree of each disruption harm type, and the flow reduction adequacy determined based on the remaining time window from the current time node to the disruption occurrence, different intensity inhibition means are constructed, and excessive response is avoided, so that the operation efficiency of the tokamak device is maximized.

Description

Technical Field

[0001] This invention belongs to the field of fracture protection and plasma control in the field of fusion, and more specifically, relates to a tokamak plasma fracture graded protection method and system. Background Technology

[0002] The key to the commercial application of magnetic confinement fusion reactors (such as tokamak) lies in the high-parameter, long-pulse steady-state operation of the device. However, in actual operation, plasma disruption is an unavoidable anomaly. It not only causes huge thermal loads and electromagnetic shocks to the first wall and supporting structure of the device, but is also the main cause of unplanned shutdowns, seriously threatening the safe operation of future fusion reactors.

[0003] Currently, machine learning-based rupture prediction methods have made significant progress, and these predictive models have been used to trigger rupture mitigation systems to protect the safety of the equipment. Most existing rupture prediction and mitigation schemes employ a binary classification approach to identify rupture precursors and trigger mitigation systems such as large-scale gas injection (MGI) or fragmentation shot injection (SPI). While this approach effectively protects the equipment, it often results in premature termination of discharge, leading to reduced operating efficiency and potentially significant economic losses for future tokamak fusion power plants.

[0004] To address these issues, the industry has explored various methods, but significant technical bottlenecks remain. First, most existing deep learning-based fracture prediction models are black-box models with poor interpretability, failing to reveal the specific physical causes of fracture (such as density limits, mode-locking, and vertical displacement). Operators cannot intervene specifically to prevent fracture. Second, existing models only predict "whether fracture will occur," failing to predict the specific types of hazards that fracture may cause (such as corona current, thermal load, and the generation of escape electrons). These specific hazard types determine the specific fracture mitigation design for large-scale equipment. Third, current fracture mitigation systems often employ a "one-size-fits-all" triggering mechanism, lacking tiered response logic tailored to different fracture causes and risk levels. Single-port or uneven injection can easily cause circumferential asymmetry, and delayed triggering may amplify local corona current peak forces. Failure to accurately select response methods based on fracture causes and hazard predictions may lead to premature discharge termination or insufficient mitigation, resulting in decreased operating efficiency and equipment damage. Summary of the Invention

[0005] To address the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a tokamak plasma fracture graded protection method and system, thereby solving the technical problems that existing tokamak fracture prediction models can only perform binary classification predictions and cannot identify fracture causes and hazard types, and that existing fracture protection strategies lack graded response mechanisms, resulting in low operating efficiency or poor protection effects.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a tokamak plasma rupture graded protection method is provided, comprising:

[0007] Multimodal diagnostic signals from the tokamak device at different time points are acquired in real time. The multimodal diagnostic signals from the current time point and historical time points are input into a pre-trained multi-task deep learning model to obtain the probability of rupture at the current time point and the severity of each hazard type predicted by the multi-task deep learning model. The multimodal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features. Each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load.

[0008] If the probability of rupture at the current time point exceeds a preset threshold, then attribution analysis is performed on the multimodal diagnostic signal to determine the key features in the multimodal diagnostic signal, and the rupture cause is determined based on the key features; based on the probability sequence of rupture occurrence probabilities at the current time point and historical time points, and the probability sequence of rupture occurrence probabilities at each time point in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point, the remaining time window from the current time point to the rupture occurrence is determined.

[0009] Based on the rupture trigger, the reversibility of the rupture trigger is determined; based on the severity of each hazard type, the flow reduction stability is determined; based on the remaining time window between the current time node and the rupture occurrence, the flow reduction sufficiency is determined; and based on the reversibility of the rupture trigger, the flow reduction stability, and the flow reduction sufficiency, a graded protection decision and response are made.

[0010] Based on the above-mentioned graded protection method for tokamak plasma rupture, the loss function used by the multi-task deep learning model during training is:

[0011]

[0012] in, The total loss of the model, For classifying losses, For the purpose of causing harm and loss, For physical loss; y represents the probability of rupture occurrence output by the multi-task deep learning model, and y represents the actual rupture label. , These represent corona current, escape electron current, and thermal load, respectively. To prevent damage and loss caused by corona current, To escape the damage caused by electron current, Losses due to heat load hazards; It refers to the actual degree of harm of the corresponding hazard type. It is the degree of harm of the corresponding hazard type output by the multi-task deep learning model; , These represent the density limit, the low q limit, and the tearing modulus limit, respectively. Represents the physical loss under the density limit. Represents the physical loss under the low q limit. Represents the physical loss under the tear modulus limit; For line average density, The value represents the Greenwald density limit; r and R are the small and large radii of the plasma, respectively. and These represent the circumferential magnetic field and the poloidal magnetic field, respectively. f is the amplitude of the tearing mode, and f is the characteristic frequency of the dominant mode; k, a and This is the default value.

[0013] According to the aforementioned tokamak plasma fracture graded protection method, the multi-task deep learning model includes an input layer, a convolutional layer, a recurrent network layer, an attention layer, and a multi-branch output layer. The input layer receives multimodal diagnostic signals from the current and historical time nodes. The convolutional layer extracts spatiotemporal features from the multimodal diagnostic signals to obtain feature maps for each time node. The recurrent network layer extracts temporal features from the feature maps to obtain contextual features for each time node. The attention layer performs attention transformation on the contextual features of each time node to obtain attention weights for each time node, and then weights and fuses the contextual features of each time node based on these attention weights to obtain a fusion vector. The multi-branch output layer predicts the probability of fracture at the current time node and the severity of each hazard type based on the fusion vector.

[0014] According to the above-mentioned tokamak plasma fracture graded protection method, the determination of the remaining time window from the current time node to the fracture occurrence based on the probability sequence of fracture occurrence based on the current time node and historical time nodes, and the probability sequence of fracture occurrence probability at each time node in the historical fracture events corresponding to the fracture trigger and the fracture occurrence time point, includes:

[0015] The probability sequence consisting of the probability of rupture occurrence at the current time node and historical time nodes is determined as the target sequence, and the probability sequence consisting of the probability of rupture occurrence at each time node in each historical rupture event corresponding to the rupture trigger is the sequence to be matched.

[0016] The target sequence is matched with each sequence to be matched, and the sequence to be matched that is most similar to the data change pattern of the target sequence is determined as the matching sequence. The matching time node that matches the current time node in the matching sequence and the historical rupture event corresponding to the matching sequence are also determined.

[0017] The time difference between the time point of the rupture occurrence of the historical rupture event corresponding to the matching sequence and the time point of the matching sequence that matches the current time point is determined as the remaining time window from the current time point to the rupture occurrence.

[0018] According to the above-mentioned graded protection method for tokamak plasma rupture, the step of determining the adequacy of the fall current based on the remaining time window before the rupture occurs, and making graded protection decisions based on the reversibility of the rupture induction, the stability of the fall current, and the adequacy of the fall current, includes:

[0019] If the reversibility of the rupture trigger indicates that it is reversible and the adequacy of the drop current indicates that the drop current time is sufficient, then the current protection strategy is determined to be a steady-state maintenance strategy; wherein, the steady-state maintenance strategy will select a physical quantity to be adjusted based on the rupture trigger in order to maintain plasma stability by adjusting the physical quantity to be adjusted.

[0020] If the reversibility of the rupture cause indicates that it is irreversible, the current reduction adequacy indicates that the current reduction time is sufficient, and the current reduction stability indicates that the current reduction process is stable, then the current protection strategy is determined to be a soft landing strategy; wherein, the soft landing strategy will perform current reduction operation on the tokamak device.

[0021] If the reversibility of the rupture cause indicates that it is irreversible, and the current reduction adequacy indicates that the current reduction time is insufficient or the current reduction stability indicates that the current reduction process is unstable, then the current protection strategy is determined to be a targeted mitigation strategy; wherein, the targeted mitigation strategy will inject impurities into the tokamak device or apply resonant magnetic disturbances.

[0022] According to the above-described graded protection method for tokamak plasma rupture, the method further includes:

[0023] During the response to the soft landing strategy, the stability of the tokamak device is also monitored.

[0024] If plasma instability is detected during stability monitoring, the targeted mitigation strategy will be switched to.

[0025] According to a second aspect of the present invention, a tokamak plasma rupture graded protection system is provided, comprising:

[0026] A multi-task joint prediction unit is used to collect multi-modal diagnostic signals of the tokamak device at different time points in real time. The multi-modal diagnostic signals of the current time point and historical time points are input into a pre-trained multi-task deep learning model to obtain the probability of failure at the current time point and the severity of each hazard type predicted by the multi-task deep learning model. The multi-modal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features. Each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load.

[0027] The interpretability analysis unit is used to perform attribution analysis on the multimodal diagnostic signal if the probability of rupture at the current time node exceeds a preset threshold, determine the key features in the multimodal diagnostic signal, and determine the rupture cause based on the key features; and determine the remaining time window from the current time node to the rupture occurrence based on the probability sequence of rupture occurrence probabilities at the current time node and historical time nodes, as well as the probability sequence of rupture occurrence probabilities at each time node in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point.

[0028] The hierarchical decision-making unit is used to determine the reversibility of the rupture trigger based on the rupture trigger, determine the flow reduction stability based on the severity of each hazard type, determine the flow reduction sufficiency based on the remaining time window between the current time node and the rupture occurrence, and make hierarchical protection decisions and responses based on the reversibility of the rupture trigger, the flow reduction stability, and the flow reduction sufficiency.

[0029] According to a third aspect of the present invention, an electronic device is provided, comprising: a computer-readable storage medium and a processor;

[0030] The computer-readable storage medium is used to store executable instructions;

[0031] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0032] According to a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the first aspect.

[0033] According to a fifth aspect of the invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method as described in the first aspect.

[0034] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0035] A multi-task deep learning model was used to jointly predict the probability of rupture and the severity of different hazard types, breaking the limitation of traditional binary classification models that only predict "whether rupture will occur". This model can not only warn of rupture risks, but also predict the specific severity of hazard such as corona current, thermal load, and escaped electrons, providing sufficient decision-making basis for subsequent graded protection. Subsequently, by performing attribution analysis on multimodal diagnostic signals, key features in the multimodal diagnostic signals were identified, and rupture inducing factors were determined based on these key features. The remaining time window from the current time point to the rupture occurrence was also determined, providing richer criteria for subsequent graded protection decisions. In the graded protection decision-making stage, by combining the reversibility determined based on the rupture inducing factors, the current reduction stability determined based on the severity of each hazard type, and the current reduction adequacy determined based on the remaining time window from the current time point to the rupture occurrence, targeted suppression measures of different intensities were constructed to avoid over-response, thereby maximizing the maintenance of the tokamak device's operating efficiency and minimizing physical damage to the tokamak device's first wall and support structure. Attached Figure Description

[0036] Figure 1 This is a schematic flowchart of a tokamak plasma rupture graded protection method provided in an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the structure of a multi-task deep learning model provided in an embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of the hierarchical protection decision-making process provided in an embodiment of the present invention. Detailed Implementation

[0039] 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention provides a tokamak plasma fracture protection method, such as... Figure 1 As shown, it includes:

[0041] Step 110: Real-time acquisition of multimodal diagnostic signals from the tokamak device at different time points; inputting the multimodal diagnostic signals from the current time point and historical time points into a pre-trained multi-task deep learning model to obtain the probability of rupture at the current time point and the severity of each hazard type predicted by the multi-task deep learning model; the multimodal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features; each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load;

[0042] Step 120: If the probability of rupture at the current time node exceeds a preset threshold, then perform attribution analysis on the multimodal diagnostic signal to determine the key features in the multimodal diagnostic signal, and determine the rupture cause based on the key features; based on the probability sequence of rupture occurrence probabilities at the current time node and historical time nodes, and the probability sequence of rupture occurrence probabilities at each time node in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point, determine the remaining time window from the current time node to the occurrence of rupture.

[0043] Step 130: Determine the reversibility of the rupture cause based on the rupture cause, determine the flow reduction stability based on the severity of each hazard type, determine the flow reduction sufficiency based on the remaining time window from the current time node to the rupture occurrence, and make graded protection decisions and responses based on the reversibility of the rupture cause, the flow reduction stability, and the flow reduction sufficiency.

[0044] Here, multimodal diagnostic signals reflecting the plasma core state at different time points are selected as input features, including but not limited to magnetohydrodynamic features (such as magnetic field signals and plasma current distribution), radiation features (such as radiation intensity and radiation distribution), plasma density features (such as density distribution in the plasma core and edge), and vertical displacement features. Based on these physical quantities reflecting the plasma core state, the multi-task deep learning model can predict the probability of rupture at the current time point and the severity of each hazard type. The probability of rupture at the current time point represents the likelihood of plasma rupture occurring after the current time point. Hazard types include corona current, escape electron current, and thermal load, and the severity is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load. Here, the corona current intensity is related to the electromagnetic shock during plasma rupture, the escape electron current intensity is a key factor leading to rupture, and the thermal load intensity affects the thermal structural safety of the tokamak device.

[0045] In some embodiments, such as Figure 2As shown, the multi-task deep learning model includes an input layer, convolutional layers, recurrent network layers, attention layers, and a multi-branch output layer. The input layer receives multimodal diagnostic signals from the current and historical time points. The convolutional layers extract spatiotemporal features from these signals to obtain feature maps for each time point. The recurrent network layers extract temporal features from the feature maps to obtain contextual features for each time point. The attention layer performs attention transformation on the contextual features of each time point to obtain attention weights, capturing semantic information from key time points. Based on these attention weights, the contextual features of each time point are weighted and fused to obtain a fusion vector. The multi-branch output layer predicts the probability of rupture at the current time point and the severity of each hazard type based on the fusion vector.

[0046] In other embodiments, to ensure that the aforementioned multi-task deep learning model possesses the ability to predict the probability of rupture and the severity of various hazard types, historical rupture experimental data of the tokamak device can be pre-collected. Sample multimodal diagnostic signals (containing the same physical quantities as the aforementioned multimodal diagnostic signals) collected at each time point during each experiment can be extracted as model input. The actual state of the plasma at each time point (ruptured or not ruptured) and the actual severity of each hazard type (i.e., actual corona current magnitude, actual escape electron current intensity, and actual thermal load intensity) can be obtained as labels. Based on these labels, the multi-task deep learning model is subjected to supervised training. This guides the model to learn the correlation between these physical quantities and the rupture event and the severity of various hazard types from the multiple physical quantities contained in the sample multimodal diagnostic signals at each time point. This allows the trained model to perform multi-task predictions based on the input multimodal diagnostic signals at each time point. To improve the prediction accuracy of the multi-task deep learning model, a physical constraint loss function is introduced during model training to penalize predictions that violate physical limits (such as density limits), thereby guiding the model to learn the correct physical laws.

[0047] Specifically, the loss function used during the training process of a multi-task deep learning model is:

[0048]

[0049] in, The total loss of the model, For classifying losses, For the purpose of causing harm and loss, For physical loss; y represents the probability of rupture occurrence output by the multi-task deep learning model, and y represents the actual rupture label. , These represent corona current, escape electron current, and thermal load, respectively. To prevent damage and loss caused by corona current, To escape the damage caused by electron current, Losses due to heat load hazards; It refers to the actual degree of harm of the corresponding hazard type. It is the degree of harm of the corresponding hazard type output by the multi-task deep learning model; , These represent the density limit, the low q limit, and the tearing modulus limit, respectively. Represents the physical loss under the density limit. Represents the physical loss under the low q limit. Represents the physical loss under the tear modulus limit; For line average density, The value represents the Greenwald density limit; r and R are the small and large radii of the plasma, respectively. and These represent the circumferential magnetic field and the poloidal magnetic field, respectively. f is the amplitude of the tearing mode, and f is the characteristic frequency of the dominant mode; k, a and This is the default value.

[0050] Furthermore, transfer learning or domain adaptation techniques can be employed during model training to improve the model's adaptability under different operating conditions. For example, transfer learning can be used to pre-train the model on the source dataset to learn general features, and then the model can be fine-tuned on the target dataset to adapt to the new data distribution. Secondly, domain adaptation techniques, such as feature alignment or adversarial training, can be used to reduce the data distribution differences between the source and target datasets, thereby enhancing the model's generalization ability under different operating conditions and improving prediction accuracy and robustness. The trained model is deployed in a high-performance computing unit to ensure that the inference speed meets the real-time control requirements.

[0051] After obtaining the probability of rupture at the current time point and the severity of each hazard type from the output of the multi-task deep learning model, it is determined whether the probability of rupture at the current time point exceeds a preset threshold. If the probability of rupture at the current time point exceeds the preset threshold, it indicates a high risk of plasma rupture occurring after the current time point. Attribution analysis is then performed on the multimodal diagnostic signal input to the multi-task deep learning model to determine the key features in the multimodal diagnostic signal that contribute to predicting the probability of rupture at the current time point, and the rupture inducing factors are determined based on these key features. The attribution analysis algorithm can employ methods such as SHAP analysis or class activation mapping; this embodiment of the invention does not specifically limit this method, as long as it can quantify the importance of each physical quantity in the multimodal diagnostic signal. After quantifying the importance of each physical quantity in the multimodal diagnostic signal using the attribution analysis algorithm, one or more physical quantities with higher importance can be selected as key features for output.

[0052] Subsequently, based on one or more selected key features, the rupture triggers causing the current high risk of rupture are determined. Here, the rupture trigger can be determined based on the feature type to which the key feature belongs. For example, if the key feature is a magnetohydrodynamic (MHD) related feature, it indicates that the plasma rupture may be caused by tearing modes, thus identifying tearing modes as the rupture trigger; if the key feature is a radiation related feature, it indicates that the plasma rupture may be caused by edge cooling, thus identifying edge cooling as the rupture trigger; if the key feature is a plasma density related feature, it indicates that the plasma rupture may be caused by density limit approximation, thus identifying density limit approximation as the rupture trigger; if the key feature is a vertical displacement related feature, it indicates that the plasma rupture may be caused by vertical displacement of the plasma, thus identifying vertical displacement as the rupture trigger. It should be noted that if multiple key features exist and belong to different feature types, multiple rupture triggers can be determined based on the above logic.

[0053] Furthermore, when the probability of rupture at the current time point exceeds a preset threshold, in addition to analyzing the rupture trigger, the remaining time window from the current time point to the rupture can be determined based on the probability sequence formed by the rupture probability at the current time point and historical time points, as well as the probability sequence formed by the rupture probability at each time point in the historical rupture events corresponding to the rupture trigger obtained from the above analysis and the rupture occurrence time point.

[0054] In some embodiments, a probability sequence consisting of the probability of rupture occurrence at the current time point and historical time points can be determined as the target sequence, and a probability sequence consisting of the probability of rupture occurrence at each time point in each historical rupture event corresponding to the same rupture trigger can be determined as the matching sequence. Each historical rupture event corresponds to one matching sequence. Subsequently, the target sequence is matched with each matching sequence, and the matching sequence with the most similar data change pattern to the target sequence is determined as the matching sequence. For example, a sliding window of the same length as the target sequence can be used to segment the matching sequence, and sequence similarity measurement methods such as dynamic time warping can be used to determine the similarity between the subsequences within each window and the target sequence. Then, the highest similarity is selected as the matching degree between the entire matching sequence and the target sequence. Based on this, the matching sequence with the highest matching degree can be selected as the matching sequence with the most similar data change pattern to the target sequence, and the last time point of the subsequence with the highest similarity to the target sequence in the matching sequence can be selected as the matching time point for matching the current time point. Additionally, the rupture occurrence time points of the historical rupture events corresponding to the matching sequence can also be obtained. Calculate the time difference between the rupture occurrence time of the historical rupture event corresponding to the matching sequence and the matching time node that matches the current time node in the matching sequence. Use this time difference as the remaining time window from the current time node to the rupture occurrence to provide a time series reference for subsequent decision-making.

[0055] Based on the aforementioned rupture causes, the severity of each hazard type, and the remaining time window from the current time point to the rupture occurrence, graded protection decisions and responses are made. Specifically, the reversibility of a rupture cause can be determined by whether it can be eliminated through closed-loop intervention. If the rupture cause can be eliminated through closed-loop intervention, it is reversible; otherwise, it is irreversible. Furthermore, based on the severity of each hazard type, it can be determined whether these hazard types pose a risk of exceeding their limits during current reduction. If such a risk exists, the current reduction stability is unstable; otherwise, the current reduction stability is stable. For example, if the rate at which the severity of any hazard type increases with time exceeds a preset value, or if the difference between the severity of any hazard type at the current time point and its limit value is less than a preset value, it can be determined that there is a risk of exceeding the limit value. Current reduction stability can also be determined based on experience; this embodiment of the invention does not specifically limit this. In addition, the adequacy of current reduction is determined based on the remaining time window from the current time point to the rupture occurrence. If the remaining time window is greater than a preset current reduction time, the current reduction adequacy is determined to be sufficient; otherwise, the current reduction adequacy is determined to be insufficient.

[0056] Based on the reversibility, current-dropping stability, and current-dropping sufficiency of the rupture cause, graded protection decisions and responses are made. Different levels of suppression measures are constructed to target the reversibility, current-dropping stability, and current-dropping sufficiency of the rupture cause, avoiding over-response, thereby maximizing the maintenance of the tokamak device's operating efficiency and reducing physical damage to the tokamak device's first wall and support structure.

[0057] In some embodiments, the following can be performed: Figure 3 The graded protection decision shown:

[0058] If the reversibility of the rupture trigger indicates its reversibility and the sufficient current reduction indicates sufficient current reduction time, then the current protection strategy is determined to be a steady-state maintenance strategy. The steady-state maintenance strategy selects adjustable physical quantities based on the aforementioned rupture triggers to eliminate the corresponding rupture triggers, thereby maintaining plasma stability. For example, if the rupture trigger is approaching the density limit, the plasma core or edge density or the density limit threshold can be selected as the adjustable physical quantity. The system steady state is restored by reducing the plasma core or edge density or raising the density limit threshold. If the rupture trigger is a tearing mode, the local plasma current density and electron temperature distribution can be selected as adjustable physical quantities. The missing bootstrap current is compensated by controlling the electron cyclotron resonance heating and current drive system, thereby suppressing magnetic island growth. If the rupture trigger is edge cooling, the heating power of the plasma core or edge can be determined, and the temperature distribution gradient of the plasma is maintained by adjusting the heating array. If the rupture trigger is vertical displacement, the plasma configuration can be selected as the adjustable physical quantity. The plasma configuration is corrected by calculating the optimal voltage adjustment matrix of the poloidal field control coil.

[0059] If the reversibility of the rupture initiation factor indicates its irreversibility, the adequacy of the current reduction indicates sufficient time for current reduction, and the stability of the current reduction indicates that the current reduction process is stable, then the current protection strategy is determined to be a soft landing strategy. The soft landing strategy performs current reduction operations on the tokamak device to maintain plasma stability by completing physical current reduction. In some embodiments, stability monitoring of the tokamak device is also performed during the response to the soft landing strategy. If plasma instability is detected during stability monitoring, a targeted mitigation strategy is switched to.

[0060] If the reversibility of the rupture cause indicates that it is irreversible, and either of the following conditions is met: (1) the adequacy of the current drawdown indicates that the current drawdown time is insufficient; or (2) the stability of the current drawdown indicates that the current drawdown process is unstable, then the current protection strategy is determined to be a targeted mitigation strategy. The targeted mitigation strategy will inject impurities into the tokamak device (e.g., large-scale gas injection or fragmented projectile injection) or apply resonant magnetic disturbances to prevent serious ruptures that could damage the device structure.

[0061] In summary, the method provided by this invention achieves joint prediction of rupture probability and severity of different hazard types through a multi-task deep learning model. This breaks the limitation of traditional binary classification models that only predict "whether rupture will occur." It not only provides early warning of rupture risks but also predicts the severity of specific hazard types such as corona current, thermal load, and escape electron current, providing sufficient decision-making basis for subsequent graded protection. Subsequently, by performing attribution analysis on multimodal diagnostic signals, key features in the multimodal diagnostic signals are identified, and rupture inducing factors are determined based on these key features. The remaining time window from the current time point to rupture occurrence is also determined, providing richer criteria for subsequent graded protection decisions. In the graded protection decision-making stage, by combining the reversibility determined based on rupture inducing factors, the current reduction stability determined based on hazard indicators of each hazard type, and the current reduction adequacy determined based on the remaining time window from the current time point to rupture occurrence, targeted suppression measures of different intensities are constructed to avoid over-response, thereby maximizing the maintenance of the tokamak device's operating efficiency and minimizing physical damage to the tokamak device's first wall and support structure.

[0062] The tokamak plasma rupture graded protection system provided by the present invention is described below. The tokamak plasma rupture graded protection system described below can be referred to in correspondence with the tokamak plasma rupture graded protection method described above.

[0063] This invention provides a tokamak plasma rupture graded protection system, comprising:

[0064] A multi-task joint prediction unit is used to collect multi-modal diagnostic signals of the tokamak device at different time points in real time. The multi-modal diagnostic signals of the current time point and historical time points are input into a pre-trained multi-task deep learning model to obtain the probability of failure at the current time point and the severity of each hazard type predicted by the multi-task deep learning model. The multi-modal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features. Each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load.

[0065] The interpretability analysis unit is used to perform attribution analysis on the multimodal diagnostic signal if the probability of rupture at the current time node exceeds a preset threshold, determine the key features in the multimodal diagnostic signal, and determine the rupture cause based on the key features; and determine the remaining time window from the current time node to the rupture occurrence based on the probability sequence of rupture occurrence probabilities at the current time node and historical time nodes, as well as the probability sequence of rupture occurrence probabilities at each time node in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point.

[0066] The hierarchical decision-making unit is used to determine the reversibility of the rupture trigger based on the rupture trigger, determine the flow reduction stability based on the severity of each hazard type, determine the flow reduction sufficiency based on the remaining time window between the current time node and the rupture occurrence, and make hierarchical protection decisions and responses based on the reversibility of the rupture trigger, the flow reduction stability, and the flow reduction sufficiency.

[0067] This invention provides an electronic device, including: a computer-readable storage medium and a processor;

[0068] The computer-readable storage medium is used to store executable instructions;

[0069] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any of the above embodiments.

[0070] This invention provides a computer-readable storage medium storing computer instructions that cause a processor to perform the method described in any of the above embodiments.

[0071] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the method described in any of the above embodiments.

[0072] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements 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 graded protection against plasma rupture in a tokamak, characterized in that, include: Multimodal diagnostic signals from the tokamak device at different time points are acquired in real time. The multimodal diagnostic signals from the current time point and historical time points are input into a pre-trained multi-task deep learning model to obtain the probability of rupture at the current time point and the severity of each hazard type predicted by the multi-task deep learning model. The multimodal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features. Each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load. If the probability of rupture at the current time point exceeds a preset threshold, then attribution analysis is performed on the multimodal diagnostic signal to determine the key features in the multimodal diagnostic signal, and the rupture cause is determined based on the key features; based on the probability sequence of rupture occurrence probabilities at the current time point and historical time points, and the probability sequence of rupture occurrence probabilities at each time point in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point, the remaining time window from the current time point to the rupture occurrence is determined. Based on the rupture trigger, the reversibility of the rupture trigger is determined; based on the severity of each hazard type, the flow reduction stability is determined; based on the remaining time window between the current time node and the rupture occurrence, the flow reduction sufficiency is determined; and based on the reversibility of the rupture trigger, the flow reduction stability, and the flow reduction sufficiency, a graded protection decision and response are made.

2. The tokamak plasma fracture protection method as described in claim 1, characterized in that, The loss function used in the training process of the multi-task deep learning model is: in, The total loss of the model, For classifying losses, For the purpose of causing harm and loss, For physical loss; y represents the probability of rupture occurrence output by the multi-task deep learning model, and y represents the actual rupture label. , These represent corona current, escape electron current, and thermal load, respectively. To prevent damage and loss caused by corona current, To escape the damage caused by electron current, Losses due to heat load hazards; It refers to the actual degree of harm of the corresponding hazard type. It is the degree of harm of the corresponding hazard type output by the multi-task deep learning model; , These represent the density limit, the low q limit, and the tearing modulus limit, respectively. Represents the physical loss under the density limit. Represents the physical loss under the low q limit. Represents the physical loss under the tear modulus limit; For line average density, The value represents the Greenwald density limit; r and R are the small and large radii of the plasma, respectively. and These represent the circumferential magnetic field and the poloidal magnetic field, respectively. f is the amplitude of the tearing mode, and f is the characteristic frequency of the dominant mode; k, a and This is the default value.

3. The tokamak plasma fracture protection method as described in claim 1 or 2, characterized in that, The multi-task deep learning model includes an input layer, a convolutional layer, a recurrent network layer, an attention layer, and a multi-branch output layer. The input layer receives multimodal diagnostic signals from the current and historical time points. The convolutional layer extracts spatiotemporal features from the multimodal diagnostic signals to obtain feature maps for each time point. The recurrent network layer extracts temporal features from the feature maps to obtain contextual features for each time point. The attention layer performs attention transformation on the contextual features of each time point to obtain attention weights, and then weights and fuses the contextual features based on these attention weights to obtain a fusion vector. The multi-branch output layer predicts the probability of rupture at the current time point and the severity of each hazard type based on the fusion vector.

4. The tokamak plasma fracture protection method as described in claim 1, characterized in that, The determination of the remaining time window from the current time node to the occurrence of rupture, based on the probability sequence of rupture occurrence probabilities at the current time node and historical time nodes, and the probability sequence of rupture occurrence probabilities at each time node in the historical rupture events corresponding to the rupture trigger and the rupture occurrence time point, includes: The probability sequence consisting of the probability of rupture occurrence at the current time node and historical time nodes is determined as the target sequence, and the probability sequence consisting of the probability of rupture occurrence at each time node in each historical rupture event corresponding to the rupture trigger is the sequence to be matched. The target sequence is matched with each sequence to be matched, and the sequence to be matched that is most similar to the data change pattern of the target sequence is determined as the matching sequence. The matching time node that matches the current time node in the matching sequence and the historical rupture event corresponding to the matching sequence are also determined. The time difference between the time point of the rupture occurrence of the historical rupture event corresponding to the matching sequence and the time point of the matching sequence that matches the current time point is determined as the remaining time window from the current time point to the rupture occurrence.

5. The tokamak plasma fracture protection method as described in claim 1, characterized in that, The graded protection decision based on the reversibility of the rupture cause, the stability of the downflow, and the adequacy of the downflow includes: If the reversibility of the rupture trigger indicates that it is reversible and the adequacy of the drop current indicates that the drop current time is sufficient, then the current protection strategy is determined to be a steady-state maintenance strategy; wherein, the steady-state maintenance strategy will select a physical quantity to be adjusted based on the rupture trigger in order to maintain plasma stability by adjusting the physical quantity to be adjusted. If the reversibility of the rupture cause indicates that it is irreversible, the current reduction adequacy indicates that the current reduction time is sufficient, and the current reduction stability indicates that the current reduction process is stable, then the current protection strategy is determined to be a soft landing strategy; wherein, the soft landing strategy will perform current reduction operation on the tokamak device. If the reversibility of the rupture cause indicates that it is irreversible, and the current reduction adequacy indicates that the current reduction time is insufficient or the current reduction stability indicates that the current reduction process is unstable, then the current protection strategy is determined to be a targeted mitigation strategy; wherein, the targeted mitigation strategy will inject impurities into the tokamak device or apply resonant magnetic disturbances.

6. The tokamak plasma fracture protection method as described in claim 5, characterized in that, The method further includes: During the response to the soft landing strategy, the stability of the tokamak device is also monitored. If plasma instability is detected during stability monitoring, the targeted mitigation strategy will be switched to.

7. A graded protection system for plasma rupture in a tokamak, characterized in that, include: A multi-task joint prediction unit is used to collect multi-modal diagnostic signals of the tokamak device at different time points in real time. The multi-modal diagnostic signals of the current time point and historical time points are input into a pre-trained multi-task deep learning model to obtain the probability of failure at the current time point and the severity of each hazard type predicted by the multi-task deep learning model. The multi-modal diagnostic signals include magnetohydrodynamic related features, radiation related features, plasma density related features, and vertical displacement related features. Each hazard type includes corona current, escape electron current, and thermal load, and the severity of the hazard is measured by the magnitude of the corona current, the intensity of the escape electron current, and the intensity of the thermal load. The interpretability analysis unit is used to perform attribution analysis on the multimodal diagnostic signal if the probability of rupture at the current time node exceeds a preset threshold, determine the key features in the multimodal diagnostic signal, and determine the rupture cause based on the key features; and determine the remaining time window from the current time node to the rupture occurrence based on the probability sequence of rupture occurrence probabilities at the current time node and historical time nodes, as well as the probability sequence of rupture occurrence probabilities at each time node in the historical rupture events corresponding to the rupture cause and the rupture occurrence time point. The hierarchical decision-making unit is used to determine the reversibility of the rupture trigger based on the rupture trigger, determine the flow reduction stability based on the severity of each hazard type, determine the flow reduction sufficiency based on the remaining time window between the current time node and the rupture occurrence, and make hierarchical protection decisions and responses based on the reversibility of the rupture trigger, the flow reduction stability, and the flow reduction sufficiency.

8. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the tokamak plasma fracture graded protection method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to execute the tokamak plasma fracture graded protection method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the tokamak plasma fracture graded protection method as described in any one of claims 1-6.

Images