Tokamak plasma configuration prediction system and method

Abstract

The invention relates to the technical field of magnetic confinement nuclear fusion, and particularly discloses a Tokamak plasma configuration prediction system and method.The Tokamak plasma configuration prediction method comprises the steps that time sequence data of multi-source parameters influencing the Tokamak plasma state are collected through multiple sensors and measuring equipment; a neural network model based on deep learning is adopted to carry out time sequence feature extraction and deep correlation analysis on the multi-source state parameters so as to capture a time sequence correlation change mode among the multi-source state parameters and reveal a dynamic behavior rule of the plasma under the time sequence synergistic effect of the multi-source state parameters; therefore, intelligent configuration prediction of the plasma is realized. In this way, the accuracy and real-time performance of plasma configuration prediction can be effectively improved, and a more reliable control basis is provided for operation of the Tokamak device.

Description

Technical Field

[0001] The present application relates to the field of magnetic confinement nuclear fusion technology, and more specifically, to a tokamak plasma configuration prediction system and method. Background Art

[0002] As the core equipment for magnetic confinement nuclear fusion research, the Tokamak device uses a strong magnetic field to confine high-temperature, high-density plasma to achieve controllable nuclear fusion reactions, providing possibilities for clean energy in the future. During the operation of the Tokamak, the configuration of the plasma (i.e., its shape, position, and stability) is one of the key factors that determine the efficiency of the fusion reaction and the safety of the device. Therefore, accurately predicting and controlling the plasma configuration is extremely important for optimizing fusion reaction conditions, improving energy output efficiency, and ensuring stable operation of the device.

[0003] Traditionally, the prediction and control of tokamak plasma configurations rely on complex physical models and a large amount of experimental data, such as the physical modeling method based on magnetohydrodynamics (MHD) equations, which analyzes plasma physical equations and combines experimental observation data to infer the state and behavior of plasma. However, plasma is a highly nonlinear, multivariable coupled complex system, and its dynamic behavior is affected by the coupling of multiple physical fields. There are strong coupling and time-varying correlations between the physical fields. Traditional mathematical models are difficult to accurately describe all dynamic details. With the continuous expansion of the scale of tokamaks and the continuous improvement of operating parameters, the accuracy and real-time requirements for plasma configuration prediction are becoming higher and higher, resulting in high computational complexity and response lag in real-time prediction of traditional methods.

[0004] Therefore, an optimized tokamak plasma configuration prediction system and method are desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a configuration prediction system and method for a tokamak plasma, which uses a variety of sensors and measuring equipment to collect time series data of multi-source parameters that affect the state of the tokamak plasma, and uses a neural network model based on deep learning to extract time series features and perform deep correlation analysis on the multi-source state parameters to capture the time series correlation change pattern between the multi-source state parameters, revealing the dynamic behavior law of the plasma under the time series synergy of the multi-source state parameters, thereby realizing intelligent configuration prediction of the plasma. In this way, the accuracy and real-time performance of plasma configuration prediction can be effectively improved, providing a more reliable control basis for the operation of the tokamak device.

[0006] Accordingly, according to one aspect of the present application, a method for predicting the configuration of a tokamak plasma is provided, comprising: Acquiring time series data of state parameters of the tokamak plasma, wherein the state parameters include a supply voltage of a vertical field coil, a supply voltage of a horizontal field coil, a loop current of a multi-stage field coil, and an ohmic field coil voltage; Performing time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature graph; Performing feature expression enhancement based on cross-level semantic synergy on the state parameter temporal feature deep association coding feature map to obtain an enhanced state parameter temporal feature deep association coding feature map; Based on the enhanced state parameter time series feature deep correlation coding feature map, a prediction result of the plasma configuration sequence is generated.

[0007] According to another aspect of the present application, a configuration prediction system for a tokamak plasma is provided, comprising: A plasma state parameter acquisition module, used to acquire time series data of state parameters of the tokamak plasma, wherein the state parameters include the supply voltage of the vertical field coil, the supply voltage of the horizontal field coil, the loop current of the multi-stage field coil, and the ohmic field coil voltage; A multi-source parameter time series correlation analysis module is used to perform time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature graph; A feature expression enhancement module, used for performing feature expression enhancement on the state parameter temporal feature deep correlation coding feature map based on cross-level semantic synergy to obtain a state parameter temporal feature deep correlation coding feature map; The plasma configuration prediction module is used to generate a prediction result of a plasma configuration sequence based on the enhanced state parameter time series feature depth correlation coding feature map.

[0008] Compared with the prior art, the configuration prediction system and method of the tokamak plasma provided by the present application utilizes a variety of sensors and measuring devices to collect time series data of multi-source parameters that affect the state of the tokamak plasma, and uses a neural network model based on deep learning to extract time series features and perform deep correlation analysis on the multi-source state parameters, so as to capture the time series correlation change pattern between the multi-source state parameters, reveal the dynamic behavior law of the plasma under the time series synergy of the multi-source state parameters, and thus realize the intelligent configuration prediction of the plasma. In this way, the accuracy and real-time performance of the plasma configuration prediction can be effectively improved, providing a more reliable control basis for the operation of the tokamak device. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 4 is a flow chart of a method for predicting the configuration of a tokamak plasma according to an embodiment of the present application.

[0011] Figure 2 Schematic diagram of data flow of a method for predicting the configuration of a tokamak plasma according to an embodiment of the present application.

[0012] Figure 3 Flow chart of step S2 in the method for predicting the configuration of a tokamak plasma according to an embodiment of the present application.

[0013] Figure 4 Flow chart of step S3 in the method for predicting the configuration of a tokamak plasma according to an embodiment of the present application.

[0014] Figure 5 4 is a block diagram of a tokamak plasma configuration prediction system according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0016] Figure 1 4 is a flow chart of a method for predicting the configuration of a tokamak plasma according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a method for predicting the configuration of a tokamak plasma according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to the embodiment of the present application, the configuration prediction method of the tokamak plasma includes the steps of: S1, obtaining the time series data of the state parameters of the tokamak plasma, wherein the state parameters include the power supply voltage of the vertical field coil, the power supply voltage of the horizontal field coil, the loop current of the multi-stage field coil and the ohmic field coil voltage; S2, performing time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature map; S3, performing feature expression enhancement based on cross-level semantic synergy on the state parameter time series feature deep correlation coding feature map to obtain an enhanced state parameter time series feature deep correlation coding feature map; S4, generating a prediction result of the plasma configuration sequence based on the enhanced state parameter time series feature deep correlation coding feature map.

[0017] In the above-mentioned configuration prediction method of tokamak plasma, the step S1 is to obtain time series data of state parameters of tokamak plasma, wherein the state parameters include the power supply voltage of the vertical field coil, the power supply voltage of the horizontal field coil, the loop current of the multi-stage field coil and the ohmic field coil voltage. It should be understood that the tokamak device is a complex system, and the behavior state of the plasma in the device is affected by a variety of factors, among which the parameters such as the power supply voltage of the vertical field coil, the power supply voltage of the horizontal field coil, the loop current of the multi-stage field coil and the ohmic field coil voltage play a key role in the confinement, heating and stability maintenance of the plasma.

[0018] Specifically, the magnetic field generated by the vertical field coil is mainly used to control the position and balance of the plasma in the vertical direction. When the supply voltage of the vertical field coil changes, the intensity and distribution of the magnetic field generated will also change accordingly, thereby exerting a vertical force on the plasma and affecting its vertical position and shape. For example, an increase in voltage may cause the plasma to move upward or become flatter in the vertical direction. The magnetic field generated by the horizontal field coil is used to constrain the movement and shape of the plasma in the horizontal direction. Different supply voltages will result in different intensities and distributions of the horizontal magnetic field, which in turn affect the expansion, contraction or distortion of the plasma in the horizontal direction. For example, a suitable supply voltage for the horizontal field coil can keep the plasma in a stable annular shape, and abnormal voltage may cause the shape of the plasma to deform or even break. Multi-level field coils are usually used to generate more complex and sophisticated magnetic field distributions to achieve more precise control of the plasma. Changes in the loop current will change the magnetic field pattern generated by the multi-level field coil, exert specific forces on the local area of ​​the plasma, affect the pressure distribution and current density distribution inside it, and thus affect the overall configuration and stability of the plasma. The ohmic field coil is mainly used to generate ohmic heating current to heat the plasma through the Joule heating effect. The size of the ohmic field coil voltage determines the level of heating power, which in turn affects the temperature, density and other physical parameters of the plasma. The temperature and density distribution of the plasma will in turn affect its behavior and configuration in the magnetic field. For example, high-temperature and high-density plasma may have a greater reaction to the magnetic field, causing changes in the magnetic field distribution, thereby changing the configuration of the plasma. Based on this, the present application accurately monitors and time-series analyzes the above-mentioned key state parameters to fully understand the environment in which the plasma is located and the forces it is subjected to, so as to achieve effective prediction of plasma behavior.

[0019] In practical applications, accurately obtaining the time series data of the above state parameters is of great significance for a comprehensive understanding of the environment in which the plasma is located, the forces it is subjected to, and its dynamic behavior laws. In actual operation, the selection and installation of sensors are key links in obtaining these data. For the measurement of the power supply voltage of the vertical field coil and the horizontal field coil, a voltage sensor with high sensitivity, high precision and wide dynamic range must be selected. For example, a voltage sensor based on the principle of capacitive voltage division has the characteristics of fast response speed and high measurement accuracy, and can accurately capture small changes in the power supply voltage. Installing this type of voltage sensor at one end of the power supply line of the vertical field coil and the horizontal field coil close to the coil can minimize the impact of line loss and interference on the measurement results, and ensure that the obtained voltage data truly reflects the actual working voltage of the coil.

[0020] For loop current measurement of multi-polar field coils, Rogowski coils are a common and effective choice. Based on the principle of electromagnetic induction, Rogowski coils have the advantages of non-contact measurement and wide response bandwidth, and can accurately measure alternating current. When installing the Rogowski coil, it is necessary to ensure that it is tightly wrapped in the loop of the multi-polar field coil, and the installation position should be selected as far as possible in the uniform magnetic field area to avoid measurement errors caused by uneven magnetic field. At the same time, pay attention to the installation direction of the Rogowski coil to ensure that the measured current direction is consistent with the actual current direction to ensure the accuracy of the measurement data.

[0021] The measurement of the voltage of the ohmic field coil also requires a high-precision voltage sensor. Considering that the ohmic field coil may generate high-frequency noise and electromagnetic interference during operation, a voltage sensor with good anti-interference performance should be selected, such as a fiber optic voltage sensor. The fiber optic voltage sensor uses the electro-optical effect of the optical fiber to convert the voltage signal into an optical signal for transmission and measurement, which can effectively avoid the influence of electromagnetic interference on the measurement results. When installing the fiber optic voltage sensor, attention should be paid to the connection and protection of the optical fiber to prevent the optical fiber from mechanical damage and bending, so as to ensure the stable transmission of the optical signal and the accuracy of the measurement.

[0022] Accordingly, the establishment of a data acquisition system is an important guarantee for accurately acquiring time series data. The system needs to have the performance of high-speed sampling, multi-channel synchronous acquisition and high resolution. Taking high-speed sampling as an example, due to the rapid change of the state of the tokamak plasma, the change frequency of some key parameters may reach kHz or even higher. Therefore, the sampling frequency of the data acquisition system should be at least twice the highest frequency of the signal to meet the requirements of the Nyquist sampling theorem, avoid aliasing, and ensure that the collected data can truly reflect the changes in the parameters. For example, for parameters with a maximum change frequency of 10kHz, the sampling frequency of the data acquisition system should be set at 20kHz or above.

[0023] At the same time, the multi-channel synchronous acquisition function is crucial for obtaining time series data of different state parameters. Because there are complex correlations and synergies between parameters such as the vertical field coil power supply voltage, the horizontal field coil power supply voltage, the multi-level field coil loop current and the ohmic field coil voltage, only by realizing multi-channel synchronous acquisition can we ensure that the acquired different parameter data have an accurate correspondence in time, and provide a reliable data basis for the subsequent analysis of the timing correlation and coordinated changes between these parameters. When actually building a data acquisition system, a multi-channel synchronous acquisition solution based on a field programmable gate array (FPGA) can be adopted. FPGA has the advantages of strong parallel processing capabilities and precise timing control, and can realize high-speed synchronous acquisition of multiple channel data. By programming the FPGA, set the parameters such as the sampling frequency, sampling time and trigger conditions of each channel to ensure that all channels start sampling at the same time and collect data according to the predetermined sampling frequency.

[0024] In addition, the resolution of the data acquisition system directly affects the accuracy of the measured data. A high-resolution data acquisition system can capture small changes in parameters and provide richer details for subsequent accurate analysis. Generally speaking, the resolution of the data acquisition system should be determined based on the actual measurement requirements and the accuracy of the sensor. For example, if the measurement accuracy of the sensor is 0.1V, in order to make full use of the accuracy of the sensor, the resolution of the data acquisition system should be at least 0.01V to ensure that the signal changes output by the sensor can be accurately measured and recorded.

[0025] Since the Tokamak device usually contains multiple subsystems and a large number of measuring devices, these devices may have time deviations during the data acquisition process. If time synchronization is not performed, the acquired time series data will not accurately reflect the actual changes in the plasma state parameters, thus affecting subsequent data analysis and model training. Therefore, time synchronization processing of multi-source parameters is the key to ensuring the accuracy and consistency of time series data. At present, commonly used time synchronization technologies include time synchronization based on the Global Positioning System (GPS) and time synchronization based on the Network Time Protocol (NTP).

[0026] Specifically, GPS-based time synchronization is a high-precision time synchronization method. The GPS system transmits precise time signals via satellites, and ground receiving equipment can receive these signals and obtain accurate time information. In the tokamak device, a GPS receiver can be installed at each node of the data acquisition system, and the clock of each node can be synchronized with the GPS time by receiving GPS satellite signals. This method can achieve high-precision time synchronization, and the synchronization accuracy can reach the nanosecond level, which can meet the strict requirements of time synchronization accuracy for the time series data acquisition of tokamak plasma state parameters. However, GPS-based time synchronization also has some limitations. For example, in indoor areas or areas with severe signal obstruction, GPS signals may be interfered with or cannot be received, thereby affecting the effect of time synchronization.

[0027] Time synchronization based on the Network Time Protocol (NTP) is a method of using the network to achieve time synchronization. The NTP protocol establishes a time server in the network, and each client synchronizes time with the time server to achieve the unification of the time of the entire network. In the network environment of the tokamak device, one or more high-precision time servers can be deployed, and each node of the data acquisition system synchronizes time with the time server as a client. The NTP protocol has the advantages of low cost, easy deployment and maintenance, but its synchronization accuracy is relatively low, generally reaching the millisecond level. In practical applications, GPS-based time synchronization and NTP-based time synchronization can be combined, using the high-precision synchronization advantage of GPS to ensure the accuracy of the time base, and using the network synchronization function of the NTP protocol to achieve time synchronization between devices, thereby improving the reliability and adaptability of the time synchronization system while ensuring synchronization accuracy.

[0028] In addition, since the amount of time series data of tokamak plasma state parameters is huge and needs to be preserved for a long time for subsequent analysis and research, it is crucial to choose a suitable data storage method and management system. At present, commonly used data storage methods include file system-based storage and database system-based storage. File system-based storage is a simple and intuitive data storage method. The collected time series data can be stored in a file in a certain format (such as CSV, HDF5, etc.), and the data can be organized and managed through file naming rules and directory structure. For example, files can be named and classified according to information such as acquisition time and parameter type, which is convenient for subsequent data query and reading. However, file system-based storage has certain limitations in data management and query, especially when processing large-scale data, the data retrieval and analysis efficiency is low. Storage based on database system can better meet the needs of large-scale data storage and management. Database system has the advantages of structured data storage, efficient query and transaction processing. In the storage of tokamak plasma state parameter data, relational databases (such as MySQL, Oracle, etc.) or non-relational databases (such as MongoDB, InfluxDB, etc.) can be selected. Relational databases are suitable for scenarios with fixed data structures and high data consistency requirements. They can achieve efficient storage and query of data by establishing data tables and indexes. For example, in a MySQL database, a data table containing fields such as timestamp, vertical field coil supply voltage, horizontal field coil supply voltage, multi-level field coil loop current, and ohmic field coil voltage can be created, and an index can be established for the timestamp field to improve the efficiency of data query. Non-relational databases have advantages in processing massive data and high-concurrency reading and writing, and are particularly suitable for storing time series data. For example, InfluxDB is a non-relational database designed specifically for time series data. It has efficient write performance, flexible data structure, and powerful query capabilities, and can quickly store and query large-scale plasma state parameter time series data.

[0029] In addition, during the data storage process, data backup and recovery strategies must also be considered. To prevent data loss or damage, data should be backed up regularly and stored in different geographical locations. At the same time, a complete data recovery mechanism should be established so that data can be restored in a timely manner when data problems occur, ensuring data integrity and availability.

[0030] In the above-mentioned configuration prediction method of tokamak plasma, the step S2 is to perform time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a time series feature deep correlation coding feature map of the state parameters. Figure 3FIG. 1 is a flow chart of step S2 in the configuration prediction method of a tokamak plasma according to an embodiment of the present application. Figure 3 As shown, the step S2 comprises: S21, data separation of the time series data of the state parameters to obtain the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-level field coil and the time series of the ohmic field coil voltage; S22, extracting the time series correlation feature of the state parameters from the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-level field coil and the time series of the ohmic field coil voltage to obtain the time series correlation feature of the vertical field line The implicit feature vector of the timing association of the vertical field coil power supply voltage, the implicit feature vector of the timing association of the horizontal field coil power supply voltage, the implicit feature vector of the timing association of the multi-level field coil loop current and the implicit feature vector of the timing association of the ohmic field coil voltage are encoded in timing association to obtain the state parameter timing feature deep association encoding feature map.

[0031] Specifically, in step S21, the time series data of the state parameters are demultiplexed to obtain the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-stage field coil, and the time series of the ohmic field coil voltage. It should be understood that different state parameters have different physical meanings and dynamic change laws. For example, the change of the vertical field coil power supply voltage may be related to the vertical displacement control of the plasma, showing a specific periodic or trend change; while the change of the horizontal field coil power supply voltage may mainly affect the horizontal confinement of the plasma, and its change law is different from that of the vertical field coil power supply voltage. Therefore, in order to improve the pertinence of data analysis and avoid confusion of timing information between different parameters, the present application first demultiplexes the time series data of the state parameters to obtain clear and independent time series of each state parameter, thereby helping to more accurately capture the timing change law within each state parameter.

[0032] Specifically, the step S22 extracts the internal time series correlation features of the state parameters from the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-level field coil, and the time series of the ohmic field coil voltage to obtain the time series correlation implicit feature vector of the vertical field coil power supply voltage, the time series correlation implicit feature vector of the horizontal field coil power supply voltage, the time series correlation implicit feature vector of the multi-level field coil loop current, and the time series correlation implicit feature vector of the ohmic field coil voltage. In a specific example of the present application, an LSTM (Long Short-Term Memory Network) model is used as a time series encoder to extract internal time series correlation features from the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-level field coil, and the time series of the ohmic field coil voltage after the columns are separated. It should be known to those skilled in the art that the LSTM model can effectively retain historical time series information and better learn and memorize the long-time span dependencies in the sequence data by introducing memory units and forget gate mechanisms. Based on this, in this application, the LSTM model is used to encode the time series of each state parameter, which can effectively extract the internal timing correlation characteristics of each state parameter, such as trend, periodicity, mutation point, etc., and generate the vertical field coil power supply voltage timing correlation implicit feature vector, the horizontal field coil power supply voltage timing correlation implicit feature vector, the multi-level field coil loop current timing correlation implicit feature vector and the ohmic field coil voltage timing correlation implicit feature vector, so as to reveal the inherent law of the state parameter changing over time.

[0033] Specifically, in step S23, the vertical field coil power supply voltage timing-related implicit feature vector, the horizontal field coil power supply voltage timing-related implicit feature vector, the multi-level field coil loop current timing-related implicit feature vector and the ohmic field coil voltage timing-related implicit feature vector are time-relatedly encoded to obtain the state parameter timing feature depth-related encoding feature map. In a specific example of the present application, after the vertical field coil power supply voltage timing-related implicit feature vector, the horizontal field coil power supply voltage timing-related implicit feature vector, the multi-level field coil loop current timing-related implicit feature vector and the ohmic field coil voltage timing-related implicit feature vector are aggregated into a state parameter timing feature aggregation matrix, the matrix is ​​input into a state parameter timing-related feature extractor based on a Mobile-Former model to obtain the state parameter timing feature depth-related encoding feature map. It should be understood that in the tokamak device, the parameters such as the vertical field coil supply voltage, the horizontal field coil supply voltage, the multi-level field coil loop current and the ohmic field coil voltage do not act independently, but interact and work together to determine the configuration of the plasma. Analyzing the timing correlation characteristics of each parameter separately can only understand the change law of the parameter itself, and cannot fully grasp the complex interactive relationship between the parameters. Therefore, the present application further aggregates the implicit feature vector of the vertical field coil supply voltage timing correlation, the implicit feature vector of the horizontal field coil supply voltage timing correlation, the implicit feature vector of the multi-level field coil loop current timing correlation and the implicit feature vector of the ohmic field coil voltage timing correlation to form a state parameter timing feature aggregation matrix, so as to facilitate a comprehensive analysis of multi-source parameters under a unified data framework, and to dig out the timing correlation change pattern between multi-source parameters, so as to more comprehensively understand the dynamic evolution mechanism of plasma behavior. Ordinary technicians in this field should know that the Mobile-Former model is a lightweight neural network model that combines the efficient convolution structure of MobileNet and the self-attention mechanism of Transformer. Among them, MobileNet has a strong local feature extraction capability and can effectively capture local correlation patterns and detail information in the data; while Transformer introduces an attention mechanism that can model global correlation information and capture long-distance dependencies between different locations in the data. By integrating the advantages of the MobileNet model and the Transformer, the Mobile-Former model can efficiently process the state parameter temporal feature aggregation matrix, while ensuring the accuracy of feature extraction of local temporal correlation changes between multi-source parameters, and effectively modeling the global correlation pattern, thereby learning the deep temporal correlation relationship between various parameters and generating a deep correlation encoding feature map of the state parameter temporal features.

[0034] In the above-mentioned configuration prediction method of tokamak plasma, the step S3 is to strengthen the feature expression of the state parameter time series feature deep correlation coding feature map based on cross-level semantic synergy to obtain a strengthened state parameter time series feature deep correlation coding feature map. Here, it is considered that the contribution and importance of different feature parts of the state parameter time series feature deep correlation coding feature map to the prediction of plasma behavior may not be the same. Therefore, in order to further improve the effectiveness and pertinence of feature representation, the present application proposes a feature expression strengthening method based on cross-level semantic synergy, which guides the refinement and enhancement of local features of each spatial position by mining the global key semantic information of the state parameter time series feature deep correlation coding feature map as a global semantic guide, so that the local features retain their own uniqueness while aligning with the global semantic consistency, thereby enhancing the characterization ability of the state parameter time series feature deep correlation coding feature map for plasma behavior prediction.

[0035] Figure 4 FIG. 1 is a flow chart of step S3 in the configuration prediction method of a tokamak plasma according to an embodiment of the present application. Figure 4 As shown, the step S3 includes: S31, inputting the state parameter temporal feature deep association coding feature map into the global key semantic information extraction network to obtain the state parameter temporal feature global key semantic information coding vector; S32, based on the state parameter temporal feature global key semantic information coding vector, performing fine-grained feature enhancement based on global information guidance on the channel feature vectors of each spatial position in the state parameter temporal feature deep association coding feature map to obtain the enhanced state parameter temporal feature deep association coding feature map.

[0036] Specifically, the step S31 is expressed by the formula:

[0037]

[0038] in, A deep correlation coding feature map representing the temporal characteristics of the state parameters, , and Respectively represent the height, width and number of channels of the state parameter temporal feature depth association coding feature map, Represents decoupled operation, , , and Respectively The first, second, and third and The state parameter time series deep correlation feature local decoupling feature vector, is the number of local decoupling feature vectors of state parameter temporal depth correlation features, represents the bias term, represents the weight matrix, represents matrix multiplication, represents the transpose of a vector, Represents the state parameter temporal depth correlation feature local feature semantic saliency score transformation vector, express The corresponding semantic significance scoring factor, represents the normalized exponential function, Represents the global key semantic information encoding vector of the temporal depth association feature of the state parameter.

[0039] That is, by constructing a global key semantic information extraction network to process the state parameter temporal feature deep association encoding feature map, based on cross-space position-by-position global dependency modeling, the global importance of each position in the state parameter temporal feature deep association encoding feature map is understood, and based on this, the key global semantic information is extracted to obtain the state parameter temporal deep association feature global key semantic information encoding vector, which provides global semantic guidance for subsequent local feature enhancement.

[0040] Specifically, the step S32 includes: first, extracting the channel feature vector of the spatial position (i, j) from the state parameter temporal feature deep association coding feature map as the state parameter temporal depth association feature to be processed channel feature vector, and based on the feature distribution characteristics of the state parameter temporal depth association feature to be processed channel feature vector, determining the local neighborhood semantic association window of the state parameter temporal depth association feature to be processed channel feature vector, wherein the state parameter temporal depth association feature to be processed channel feature vector is located at the center of the local neighborhood semantic association window, which is expressed by the formula:

[0041]

[0042] in, express The state parameter time series depth correlation feature of the (i, j) position is the channel feature vector to be processed, represents the calculation norm, represents the logarithmic function with base 2, Indicates rounding up. A local neighborhood semantic association window representing the channel feature vector to be processed of the state parameter temporal depth association feature, represents the half length of the local neighborhood semantic association window, and Represents the row and column indexes of the state parameter temporal depth association feature channel feature vector in the local neighborhood semantic association window.

[0043] That is, the channel feature vector of the spatial position (i, j) is extracted from the state parameter time series feature deep association coding feature map as the feature data to be processed, so as to perform refined enhancement on the features of each spatial position in the state parameter time series feature deep association coding feature map. In particular, here, the channel feature vector to be processed of the state parameter time series feature deep association coding feature map can be the channel feature vector of any spatial position in the state parameter time series feature deep association coding feature map. Then, by analyzing the feature distribution characteristics of the channel feature vector to be processed of the state parameter time series feature deep association coding feature, the highest semantic correlation point in the neighborhood of the current feature to be processed is dynamically perceived, and then the local neighborhood semantic association window is determined, so as to realize flexible adjustment of the receptive field range, thereby reducing unnecessary computational overhead and feature redundancy while capturing the local neighborhood context correlation.

[0044] Next, all state parameter temporal depth association feature channel feature vectors in the local neighborhood semantic association window are input into the local neighborhood association semantic information extraction network to obtain the state parameter temporal depth association feature local key semantic information encoding vector corresponding to the state parameter temporal depth association feature to be processed channel feature vector. In a specific example of the present application, the positional mean vector between all state parameter temporal depth association feature channel feature vectors in the local neighborhood semantic association window of the state parameter temporal depth association feature to be processed channel feature vector is calculated to obtain the state parameter temporal depth association feature local key semantic information encoding vector, which is expressed as:

[0045] in, Indicates the local neighborhood semantic association window The state parameters of the position, the temporal depth, the associated feature channel feature vector, Represents the state parameter temporal depth association feature local key semantic information encoding vector corresponding to the channel feature vector to be processed of the state parameter temporal depth association feature.

[0046] That is, after completing the division of the local window, the feature vectors of each channel within the window range are input into the local semantic information extraction network. By calculating the positional mean of each channel feature, the average semantic information within the local window range is extracted, and the feature commonality within the local window is captured. The state parameter temporal deep correlation feature local key semantic information encoding vector is generated, providing more detailed semantic guidance for subsequent local feature enhancement.

[0047] Furthermore, based on the semantic difference between the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the state parameter temporal depth association feature, the channel feature vector of the state parameter temporal depth association feature to be processed is subjected to feature enhancement processing to obtain an enhanced state parameter temporal depth association feature channel feature vector, wherein the enhanced state parameter temporal depth association feature channel feature vector is the channel feature vector of the spatial position (i, j) of the enhanced state parameter temporal depth association feature encoding feature map. In particular, in a preferred example of the present application, before performing a comparative analysis on the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the state parameter temporal depth association feature, first, based on the channel feature vector of the state parameter temporal depth association feature to be processed, the local key semantic information encoding vector of the state parameter temporal depth association feature is subjected to feature interaction optimization driven by semantic similarity to obtain an optimized state parameter temporal depth association feature local key semantic information encoding vector, which is expressed as:

[0048] in, is a vector and The semantic association matrix between is a matrix The diagonal vector composed of the diagonal elements of is a matrix The eigenvalues ​​of eigenvectors are Indicates point addition, express The corresponding optimization state parameter temporal depth correlation feature local key semantic information encoding vector.

[0049] Here, for the state parameter temporal depth association feature local key semantic information encoding vector obtained by local semantic information extraction, the entities with similar semantic neighborhood types in the local semantic interaction system, that is, the interaction between the state parameter temporal depth association feature local key semantic information encoding vector and the state parameter temporal depth association feature to be processed channel feature vector, are further characterized based on the concept of isomorphism. That is, the state parameter temporal depth association feature local key semantic information encoding vector with similar semantic neighborhood window association is driven by the semantic similarity of the state parameter temporal depth association feature to be processed channel feature vector, so that the state parameter temporal depth association feature local key semantic information encoding vector and the state parameter temporal depth association feature to be processed channel feature vector with similar attributes are more inclined to form feature entity association. Specifically, the diagonal vector and eigenvector of the association representation of the local key semantic information encoding vector of the state parameter temporal deep association feature and the feature vector of the channel to be processed of the state parameter temporal deep association feature are used as the mediated mediation effect representation and the targeted mediation effect representation respectively, so as to represent the indirect influence of the association semantic similarity on the target variable through the mediation variable through the mediation effect modeling, thereby performing local semantic logic mediation inference through the same type of semantic network structure, and further improving the subsequent semantic mapping alignment.

[0050] Next, based on the semantic contrast analysis between the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the optimized state parameter temporal depth association feature, determine the semantic mapping guidance factor of the channel feature vector to be processed of the state parameter temporal depth association feature; based on the semantic mapping guidance factor, feature modulate the channel feature vector to be processed of the state parameter temporal depth association feature to obtain the enhanced state parameter temporal depth association feature channel feature vector. In a specific example of the present application, feature modulate the channel feature vector to be processed of the state parameter temporal depth association feature based on the semantic mapping guidance factor, including: calculating the product between the semantic mapping guidance factor and the channel feature vector to be processed of the state parameter temporal depth association feature to obtain the enhanced state parameter temporal depth association feature channel feature vector.

[0051] The above calculation process is expressed as follows:

[0052] in, express and The inverse matrix of the covariance matrix between express The semantic mapping guidance factor, express The corresponding enhanced state parameter temporal depth correlation feature channel feature vector.

[0053] Here, the global key semantic information encoding vector of the state parameter temporal deep association feature provides a key semantic guide in the global scope, while the local key semantic information encoding vector of the optimized state parameter temporal deep association feature defines the local semantic focus of the current feature. By further comparing and analyzing the two, the semantic mapping relationship in the global context in the feature vector of the to-be-processed channel of the state parameter temporal deep association feature is determined, and a semantic mapping guidance factor is generated as the core modulation basis for subsequent feature enhancement, which helps to retain the personalized detail information of the local feature while strengthening the global consistency of feature expression. Furthermore, based on the semantic mapping guidance factor, by multiplicatively adjusting the feature vector of the to-be-processed channel of the state parameter temporal deep association feature to guide the direction and intensity of feature enhancement, the distinguishability of different feature parts in the original feature can be enhanced, and its feature expression ability can be improved, so that the state parameter temporal feature deep association coding feature map after the enhanced processing maintains the consistency of global semantics while strengthening the personalized information of local features.

[0054] In the configuration prediction method of the tokamak plasma described above, the step S4 generates a prediction result of the plasma configuration sequence based on the enhanced state parameter time series feature deep correlation coding feature map. In a specific example of the present application, the step S4 includes: inputting the enhanced state parameter time series feature deep correlation coding feature map into a prediction module based on the RNN model to obtain the prediction result of the plasma configuration sequence. It should be understood that the RNN model, as a recurrent neural network, has unique advantages in processing time series data. It can effectively model the time evolution characteristics of plasma behavior through the time dependency in the input data. In the present application, the RNN model learns the dynamic behavior law of the plasma under the time synergy of the multi-source parameters by mining the time series correlation change pattern between the multi-source parameter features in the enhanced state parameter time series feature deep correlation coding feature map, so as to predict the plasma configuration sequence on this basis, thereby providing strong data support for plasma physics research and industrial applications.

[0055] In summary, the configuration prediction method of the tokamak plasma according to the embodiment of the present application is explained, which uses a variety of sensors and measuring equipment to collect time series data of multi-source parameters that affect the state of the tokamak plasma, and uses a neural network model based on deep learning to extract time series features and perform deep correlation analysis on the multi-source state parameters to capture the time series correlation change pattern between the multi-source state parameters, revealing the dynamic behavior law of the plasma under the time series synergy of the multi-source state parameters, thereby realizing the intelligent configuration prediction of the plasma. In this way, the accuracy and real-time performance of the plasma configuration prediction can be effectively improved, providing a more reliable control basis for the operation of the tokamak device.

[0056] Furthermore, the present application also provides a tokamak plasma configuration prediction system.

[0057] Figure 5 FIG. 1 is a block diagram of a configuration prediction system for a tokamak plasma according to an embodiment of the present application. Figure 5 As shown, according to the embodiment of the present application, the configuration prediction system 100 of the tokamak plasma includes: a plasma state parameter acquisition module 110, which is used to obtain the time series data of the state parameters of the tokamak plasma, wherein the state parameters include the power supply voltage of the vertical field coil, the power supply voltage of the horizontal field coil, the loop current of the multi-stage field coil and the ohmic field coil voltage; a multi-source parameter timing association analysis module 120, which is used to perform timing deep association coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter timing feature deep association coding feature map; a feature expression enhancement module 130, which is used to perform feature expression enhancement based on cross-level semantic synergy on the state parameter timing feature deep association coding feature map to obtain a state parameter timing feature deep association coding feature map; a plasma configuration prediction module 140, which is used to generate a prediction result of the plasma configuration sequence based on the enhanced state parameter timing feature deep association coding feature map.

[0058] Here, those skilled in the art will appreciate that the specific operations of each module in the above-mentioned Tokamak plasma configuration prediction system have been described in detail above. Figures 1 to 4 It has been described in detail in the description of the tokamak plasma configuration prediction method, and therefore, its repeated description will be omitted.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the configuration of a tokamak plasma, characterized in that: include: Acquiring time series data of state parameters of the tokamak plasma, wherein the state parameters include a supply voltage of a vertical field coil, a supply voltage of a horizontal field coil, a loop current of a multi-stage field coil, and an ohmic field coil voltage; Performing time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature graph; Performing feature expression enhancement based on cross-level semantic synergy on the state parameter temporal feature deep association coding feature map to obtain an enhanced state parameter temporal feature deep association coding feature map; Based on the enhanced state parameter time series feature deep correlation coding feature map, a prediction result of the plasma configuration sequence is generated.

2. The method for predicting the configuration of a tokamak plasma according to claim 1, characterized in that: Performing time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature map, including: Separating the time series data of the state parameters to obtain a time series of the power supply voltage of the vertical field coil, a time series of the power supply voltage of the horizontal field coil, a time series of the loop current of the multi-level field coil, and a time series of the ohmic field coil voltage; Extracting the internal time series correlation features of the state parameters from the time series of the power supply voltage of the vertical field coil, the time series of the power supply voltage of the horizontal field coil, the time series of the loop current of the multi-level field coil and the time series of the ohmic field coil voltage to obtain the time series correlation implicit feature vector of the vertical field coil power supply voltage, the time series correlation implicit feature vector of the horizontal field coil power supply voltage, the time series correlation implicit feature vector of the multi-level field coil loop current and the time series correlation implicit feature vector of the ohmic field coil voltage; The vertical field coil power supply voltage timing-associated implicit feature vector, the horizontal field coil power supply voltage timing-associated implicit feature vector, the multi-level field coil loop current timing-associated implicit feature vector and the ohmic field coil voltage timing-associated implicit feature vector are timing-associated encoded to obtain the state parameter timing feature deep association encoding feature map.

3. The method for predicting the configuration of a tokamak plasma according to claim 2, characterized in that: Performing time-series correlation coding on the vertical field coil power supply voltage time-series correlation implicit feature vector, the horizontal field coil power supply voltage time-series correlation implicit feature vector, the multi-level field coil loop current time-series correlation implicit feature vector and the ohmic field coil voltage time-series correlation implicit feature vector to obtain the state parameter time-series feature deep correlation coding feature map, including: After the vertical field coil power supply voltage timing-associated implicit feature vector, the horizontal field coil power supply voltage timing-associated implicit feature vector, the multi-level field coil loop current timing-associated implicit feature vector and the ohmic field coil voltage timing-associated implicit feature vector are aggregated into a state parameter timing feature aggregation matrix, it is input into a state parameter timing-associated feature extractor based on the Mobile-Former model to obtain a deep association coding feature map of the state parameter timing features.

4. The method for predicting the configuration of a tokamak plasma according to claim 3, characterized in that: The state parameter temporal feature deep correlation coding feature map is enhanced based on cross-level semantic synergy to obtain an enhanced state parameter temporal feature deep correlation coding feature map, including: Inputting the state parameter temporal feature deep correlation coding feature map into the global key semantic information extraction network to obtain the state parameter temporal feature global key semantic information coding vector; Based on the global key semantic information encoding vector of the state parameter temporal deep association feature, the channel feature vectors of each spatial position in the state parameter temporal feature deep association encoding feature map are respectively enhanced with fine-grained features guided by global information to obtain the enhanced state parameter temporal feature deep association encoding feature map.

5. The method for predicting the configuration of a tokamak plasma according to claim 4, characterized in that: Based on the global key semantic information encoding vector of the state parameter temporal deep association feature, the channel feature vectors at each spatial position in the state parameter temporal deep association encoding feature map are respectively enhanced based on global information guidance to obtain the enhanced state parameter temporal deep association encoding feature map, including: Extracting a channel feature vector at a spatial position (i, j) from the state parameter temporal feature deep correlation coding feature map as a channel feature vector to be processed for the state parameter temporal deep correlation feature; Based on the characteristic distribution characteristics of the channel feature vector to be processed of the state parameter time series depth association feature, determine the local neighborhood semantic association window of the channel feature vector to be processed of the state parameter time series depth association feature, wherein the channel feature vector to be processed of the state parameter time series depth association feature is located at the center of the local neighborhood semantic association window; Inputting all state parameter temporal depth association feature channel feature vectors in the local neighborhood semantic association window into a local neighborhood association semantic information extraction network to obtain a state parameter temporal depth association feature local key semantic information encoding vector corresponding to the state parameter temporal depth association feature to-be-processed channel feature vector; Based on the semantic difference between the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the state parameter temporal depth association feature, feature enhancement processing is performed on the channel feature vector to be processed of the state parameter temporal depth association feature to obtain an enhanced state parameter temporal depth association feature channel feature vector, wherein the enhanced state parameter temporal depth association feature channel feature vector is the channel feature vector of the spatial position (i, j) of the enhanced state parameter temporal depth association encoding feature map.

6. The method for predicting the configuration of a tokamak plasma according to claim 5, characterized in that: Inputting all state parameter temporal depth association feature channel feature vectors in the local neighborhood semantic association window into the local neighborhood association semantic information extraction network to obtain the state parameter temporal depth association feature local key semantic information encoding vector corresponding to the state parameter temporal depth association feature to-be-processed channel feature vector, including: The positional mean vector between all state parameter temporal depth association feature channel feature vectors in the local neighborhood semantic association window of the channel feature vector to be processed of the state parameter temporal depth association feature is calculated to obtain the local key semantic information encoding vector of the state parameter temporal depth association feature.

7. The method for predicting the configuration of a tokamak plasma according to claim 6, characterized in that: Based on the semantic difference between the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the state parameter temporal depth association feature, feature enhancement processing is performed on the to-be-processed channel feature vector of the state parameter temporal depth association feature to obtain an enhanced state parameter temporal depth association feature channel feature vector, including: Based on the to-be-processed channel feature vector of the state parameter temporal depth association feature, the local key semantic information encoding vector of the state parameter temporal depth association feature is optimized based on semantic similarity-driven feature interaction to obtain an optimized local key semantic information encoding vector of the state parameter temporal depth association feature; Determine the semantic mapping guidance factor of the channel feature vector to be processed of the state parameter temporal depth association feature based on the semantic comparison analysis between the global key semantic information encoding vector of the state parameter temporal depth association feature and the local key semantic information encoding vector of the optimized state parameter temporal depth association feature; Based on the semantic mapping guidance factor, feature modulation is performed on the channel feature vector to be processed of the state parameter temporal depth correlation feature to obtain the enhanced state parameter temporal depth correlation feature channel feature vector.

8. The method for predicting the configuration of a tokamak plasma according to claim 7, characterized in that: Based on the semantic mapping guidance factor, feature modulation is performed on the state parameter temporal depth correlation feature to-be-processed channel feature vector to obtain the enhanced state parameter temporal depth correlation feature channel feature vector, including: The product of the semantic mapping guidance factor and the state parameter temporal depth association feature to be processed channel feature vector is calculated to obtain the enhanced state parameter temporal depth association feature channel feature vector.

9. The method for predicting the configuration of a tokamak plasma according to claim 8, characterized in that: Based on the enhanced state parameter time series characteristic deep correlation coding characteristic map, a prediction result of a plasma configuration sequence is generated, including: The enhanced state parameter time series feature deep correlation coding feature map is input into a prediction module based on an RNN model to obtain a prediction result of the plasma configuration sequence.

10. A configuration prediction system for a tokamak plasma, characterized in that: include: A plasma state parameter acquisition module, used to acquire time series data of state parameters of the tokamak plasma, wherein the state parameters include the supply voltage of the vertical field coil, the supply voltage of the horizontal field coil, the loop current of the multi-stage field coil, and the ohmic field coil voltage; A multi-source parameter time series correlation analysis module is used to perform time series deep correlation coding between multi-source parameters on the time series data of the state parameters to obtain a state parameter time series feature deep correlation coding feature graph; A feature expression enhancement module, used for performing feature expression enhancement on the state parameter temporal feature deep correlation coding feature map based on cross-level semantic synergy to obtain a state parameter temporal feature deep correlation coding feature map; The plasma configuration prediction module is used to generate a prediction result of a plasma configuration sequence based on the enhanced state parameter time series feature depth correlation coding feature map.

Images