Method and system for predicting evolution of key macroscopic parameters of tokamak plasma
By employing a lightweight recurrent neural network and a progressive autoregressive training strategy, the problems of error accumulation and dynamic characteristics in tokamak plasma parameter prediction are solved, achieving high-precision, long-term parameter prediction that meets real-time control requirements and is applicable to tokamak devices and other long-term prediction fields.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-06-17
- Publication Date
- 2026-07-21
AI Technical Summary
Existing tokamak plasma parameter prediction models suffer from problems such as error accumulation during long-term autoregression, difficulty in balancing the dynamic characteristics of fast and slow variables, insufficient utilization of prior physical knowledge, and contradictions between computational resources and real-time performance, which cannot meet the requirements for long-pulse steady-state operation of tokamak.
A lightweight recurrent neural network and a progressive autoregressive training strategy are adopted. Through a physically decoupled dual-branch prediction head and an adaptive loss function, a progressive autoregressive training strategy is designed. Combined with a learnable amplitude scaling factor and an adaptive weighting strategy, the differential modeling of fast and slow variables and physical constraints are realized to meet the millisecond-level latency requirements of real-time control.
It achieves high-precision, long-term plasma parameter prediction, solves the error accumulation problem in autoregressive prediction, takes into account the dynamic characteristics of fast and slow variables, meets the time delay requirements of real-time control systems, and provides reliable support for plasma control and rupture early warning.
Smart Images
Figure CN122432677A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fusion plasma control, specifically to a method for predicting the evolution of key macroscopic parameters of tokamak plasma. Background Technology
[0002] A tokamak is a toroidal experimental device that uses a strong magnetic field to confine high-temperature plasma to achieve controlled nuclear fusion. During tokamak discharge, the confinement performance and stability of the plasma are characterized by several key dimensionless macroscopic parameters, mainly including poloidal specific pressure, normalized specific pressure, and internal inductance. The poloidal specific pressure reflects the ratio of plasma pressure to poloidal magnetic field pressure; the normalized specific pressure characterizes the degree to which the plasma approaches the pressure-driven magnetohydrodynamic instability threshold; and the internal inductance reflects the peaking degree of the plasma current distribution. Real-time acquisition of the evolution trends of these parameters over a period of time is crucial for realizing model predictive control, advanced control strategies based on reinforcement learning, and rupture early warning.
[0003] In existing technologies, the main approaches to obtaining the future evolution trajectory of the aforementioned parameters fall into two categories: simulation methods based on physical models and data-driven neural network methods. Simulation methods based on physical models (such as integrated simulation programs like RAPTOR and CRONOS) can provide physically consistent evolution predictions by self-consistently solving plasma transport and equilibrium equations. However, these methods are computationally intensive, with single simulations typically taking seconds or even longer, far from meeting the millisecond-level response requirements of real-time control systems. To address this, researchers have attempted to simplify physical models or linearize them near the operating point, but this limits model accuracy, making it difficult to accurately characterize the inherent nonlinear and strongly coupled dynamic characteristics of plasma, and resulting in insufficient generalization ability for different discharge modes. Data-driven neural network methods directly learn the mapping relationship between input and output from historical experimental data, offering advantages such as fast inference speed and the ability to implicitly capture nonlinear coupling relationships, gradually becoming a research hotspot in plasma modeling. Long Short-Term Memory (LSTM) networks, as recurrent neural networks capable of handling time-dependent processes, have been attempted for plasma parameter prediction. However, conventional LSTM prediction models suffer from the following prominent drawbacks: (1) Accumulation of autoregressive errors leads to long-term prediction divergence. In real-time control scenarios, the model needs to operate in an autoregressive manner, meaning that the model's current prediction will be used as the input for the next time step. Conventional models employ a teacher-forced strategy during the training phase, always using real historical data as input. This results in a significant difference between the training conditions (based on real history) and the inference conditions (based on the model's own predictions), known as exposure bias. When the model operates in an autoregressive manner for an extended period during the inference phase, the small errors in each prediction step gradually accumulate and amplify, eventually causing the predicted trajectory to deviate significantly from the true value, or even diverge, failing to meet the requirements of long-pulse steady-state operation of the tokamak for full-process parameter prediction.
[0004] (2) It is difficult to model the coupling of fast and slow dynamic variables. It responds rapidly (in milliseconds) to changes in auxiliary heating power, and Dominated by the current diffusion process, the evolution is slow (on the order of hundreds of milliseconds). Conventional models use a single output head to predict all variables simultaneously, making it difficult to simultaneously account for the high-frequency fluctuations of fast variables and the smooth evolution trend of slow variables. Forcing a unified modeling approach can lead to over-smoothing of fast variables or contamination of slow variables with high-frequency noise.
[0005] (3) Insufficient utilization of prior physical knowledge. Existing pure data-driven models usually only optimize the absolute error between the predicted value and the true value, without introducing constraints in plasma physics regarding variable continuity, evolution smoothness, etc., which may lead to the dynamic characteristics learned by the model contradicting the real physical laws, further aggravating the error accumulation in the autoregression process.
[0006] The main drawbacks of existing technologies can be summarized as follows: (1) Training-inference mismatch leads to long-term prediction instability. Existing data-driven models rely on real historical data input during training, but are forced to receive their own predictions during inference. Exposure bias causes the prediction error to amplify exponentially with time step, which cannot meet the requirements of long-pulse steady-state operation of tokamak for full-process parameter prediction.
[0007] (2) Single output head is difficult to take into account dynamics at multiple time scales. Existing models use the same network path to predict poloidal specific pressure, normalized specific pressure and induction, which cannot differentiate the dynamic characteristics of fast and slow variables, making it difficult to achieve both prediction accuracy and smoothness.
[0008] (3) Lack of effective use of physical priors. Existing models have a single training objective and do not impose constraints on the temporal derivatives (i.e. rate of change) of the predicted sequence, making it difficult to ensure that the predicted trajectory meets the physical smoothness requirements of plasma evolution.
[0009] (4) Conflict between computing resources and real-time performance. Some high-precision models (such as large-scale Transformers) have a large number of parameters, and the time taken for a single-step inference is far greater than the 1ms tokamak control cycle, making them unsuitable for deployment in real-time control systems. Summary of the Invention
[0010] To address the aforementioned technical problems, this invention provides a method and system for predicting the evolution of key macroscopic parameters in tokamak plasma. The specific objectives are as follows: (1) Achieve high-precision long-term autoregressive prediction. By using a progressive autoregressive training strategy, the model gradually adapts to the self-generated input during the training process, fundamentally alleviating exposure bias and ensuring that the trajectory does not diverge in thousands of autoregressive prediction steps.
[0011] (2) Decoupling the dynamic characteristics of fast and slow variables. By designing a physical decoupled dual-branch predictor, the pressure and induction are modeled independently, and a learnable residual scaling mechanism is used to balance the high-frequency response fidelity of fast variables and the smooth evolution trend of slow variables.
[0012] (3) Embed physical constraints to regulate prediction behavior. By constructing a composite loss function that includes a time derivative penalty term and adopting an adaptive weighting strategy based on homoscedastic uncertainty, the model is guided to learn dynamic evolution laws that conform to physical laws.
[0013] (4) Meets the millisecond-level latency requirements of real-time control. By adopting a lightweight LSTM encoder architecture (with approximately 800,000 parameters), the single-step inference time is compressed to less than 1ms, ensuring that the method can be directly embedded into existing plasma real-time control systems.
[0014] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the evolution of key macroscopic parameters of tokamak plasma, comprising: Historical discharge data of the tokamak device were acquired, and a multivariate time series including covariates and evolution target variables was constructed. A lightweight recurrent neural network prediction model is constructed. The prediction model includes: a shared encoder for extracting the latent state features of the input time series data; and a physically decoupled dual-head branch structure, wherein the first branch is used to predict the residual increment of the fast response variable, the second branch is used to predict the residual increment of the slow response variable, and each branch output is provided with a learnable amplitude scaling factor. A progressive autoregressive training strategy is adopted to train the prediction model. The training strategy includes: a course-based step size growth strategy that linearly increases the number of autoregressive expansion steps of the prediction model with each training round, and a planned sampling strategy that decays the teacher-forced probability according to an inverse S-shaped curve, so that the prediction model gradually adapts to the self-generated input distribution. During training, the absolute prediction error, the rate of change error of fast variables, and the rate of change error of slow variables are optimized simultaneously. Using the trained prediction model, the evolution target variable of the tokamak device at future time steps is continuously predicted in an autoregressive closed-loop manner.
[0015] In one embodiment, the historical discharge data includes: plasma current, low hybrid power, circumferential field coil current, ion cyclotron resonance heating power, electron cyclotron resonance heating power, neutral beam injection power, elongation ratio, electron density, radiation power, plasma-wall gap, ring voltage, poloidal specific pressure, normalized specific pressure, and inductance as model input channels; wherein, the poloidal specific pressure, normalized specific pressure, and inductance are simultaneously used as evolution target variables, and the rest are covariates.
[0016] In one embodiment, the shared encoder is used to extract latent state features of the input time-series data, specifically including: The shared encoder employs a two-layer stacked long short-term memory network to map the input multivariate time series into compact hidden state features.
[0017] In one embodiment, the first branch is used to predict the residual increment of the fast response variable, and the second branch is used to predict the residual increment of the slow response variable. Each branch output has a learnable magnitude scaling factor, specifically including: The hidden state features output from the shared encoder are fed into two independent branches. The first branch predicts the increments of the poloidal specific pressure and the normalized specific pressure; the second branch predicts the increment of the interotropic induction. Each branch is followed by a Tanh activation function to map the original increment prediction values of the two branches to a value within [...]. Within the interval [1,1], multiply by their respective learnable scaling factors. and .
[0018] In one embodiment, the prediction model further includes a residual output layer, specifically comprising: The incremental predicted values output by the first and second branches are added to the evolution target variable value at the previous time step to obtain the absolute predicted value at the current time step. In the training phase, the evolution target variable value at the previous time step is the true value; in the inference phase, the evolution target variable value at the previous time step is the historical predicted value output by the prediction model itself.
[0019] In one embodiment, the course-based step size growth strategy, which linearly increases the number of autoregressive expansion steps of the prediction model with each training round, specifically includes: Set two hyperparameters and During the training In each iteration, the actual number of autoregressive expansion steps of the prediction model Linear growth from small to large: ; in, This refers to the total number of training rounds.
[0020] In one embodiment, the planned sampling strategy of attenuating the teacher forced probability according to an inverse S-shaped curve specifically includes: At each step of the autoregressive expansion, the predictive model uses probability... Choosing to use the actual evolutionary target variable values as input for the next step, with probability... Use the self-predicted evolutionary target variable as input for the next step. Decrease with training rounds according to an inverse S-shaped curve: ; in It is the attenuation constant. For the first Index of each iteration round.
[0021] In one embodiment, the simultaneous optimization of absolute prediction error, fast variable rate of change error, and slow variable rate of change error during training specifically includes: Absolute prediction error ; This represents the target value of the evolutionary objective variable at time t+1. This represents the target value of the evolution target variable at time t+1; Error of rate of change of fast variable ; Indicates polar pressure or normalized specific pressure exist The predicted value of the time increment. Indicates polar pressure or normalized specific pressure exist The actual value of the time increment; Error in the rate of change of slow variables ; Indicating inner feelings exist The predicted value of the time increment. Indicating inner feelings exist The actual value of the time increment; Represents the SmoothL1 loss function; Total loss function for: ; in, It is a loss item Learnable noise parameters.
[0022] In one embodiment, the step of using a trained prediction model to continuously predict the evolution target variable of the tokamak device at future time steps in an autoregressive closed-loop manner specifically includes: The buffer is initialized with real discharge data within a historical window of length s. For each prediction step, the discharge data of the most recent s time steps in the buffer and the real covariate value of the current step are used as inputs to predict the next evolution target variable. The predicted evolution target variable is appended to the buffer and the window is slid. The above process is repeated to complete the multi-step autoregressive prediction.
[0023] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.
[0024] Compared with the prior art, the beneficial technical effects of the present invention are: This invention designs a method for predicting the evolution of key global parameters of tokamak plasma based on a lightweight recurrent neural network and progressive autoregressive training. This method employs a data-driven approach, utilizing historical plasma states and future control commands to predict the poloidal specific pressure (PSP) at multiple future time steps using an autoregressive method. ), normalized specific pressure ( ) and internal sensation ( This invention enables continuous prediction. It solves the problems of severe error accumulation and divergent prediction trajectories in long-term autoregressive prediction using traditional data-driven models. It also overcomes the difficulty in uniformly characterizing the dynamic characteristics of fast and slow variables in physical decoupling modeling, achieving high-precision, long-term autoregressive evolution prediction of key plasma parameters, thus providing reliable technical support for real-time plasma control and rupture early warning.
[0025] This invention, by designing a course-based autoregressive step size that dynamically increases with the training process and a planned sampling strategy with inverse S-shaped decay, is the first to systematically introduce the concept of course learning into plasma time series prediction tasks, effectively bridging the exposure bias between training and inference. Simultaneously, by predicting the residual increments of fast and slow variables separately through independent branches, and by introducing a learnable scaling factor and an adaptive multi-task uncertainty weighted loss function, prior knowledge of plasma physics is embedded into the network structure and optimization objective, significantly improving the stability and accuracy of long-term autoregressive prediction.
[0026] Traditional controllers or predictive models typically rely on mechanistic modeling and state-space equation construction of the controlled object. For highly nonlinear, strongly coupled complex systems like tokamak plasmas, obtaining an accurate physical model is extremely difficult. This invention starts entirely from experimental data, avoiding the need for explicit modeling of complex physical mechanisms. Compared to existing data-driven methods, this invention, for the first time, organically integrates multiple techniques such as planned sampling, curriculum-based autoregressive step size, residual learning, dual-head decoupled prediction, and adaptive uncertainty-weighted loss, forming a complete technical solution specifically designed for robust long-term autoregressive prediction.
[0027] The method proposed in this invention is applicable not only to the poloidal specific pressure in tokamak devices ( ), normalized specific pressure ( ) and internal sensation ( The core framework of this method can be directly transferred to the autoregressive evolution prediction of various slowly varying plasma parameters (such as energy storage, density, and current distribution) in other tokamak devices (such as HL-3, KSTAR, JET, ITER, etc.). Furthermore, the long-term autoregressive error accumulation problem solved by this method is widely present in many fields such as weather forecasting, financial time series analysis, and industrial process control. Therefore, the technical solution of this invention has cross-domain application potential and significant practical value. Attached Figure Description
[0028] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is the overall architecture of the plasma key macroscopic parameter evolution prediction model of the present invention; Figure 3 This invention relates to a progressive autoregressive training strategy; Figure 4 This is an example of a typical discharge autoregressive prediction result in an embodiment of the present invention; Figure 5 The test set for each gun in this embodiment of the invention Score distribution chart; Figure 6 This is a statistical chart showing the time consumption of single-step inference in an embodiment of the present invention. Detailed Implementation
[0029] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.
[0030] like Figure 1 As shown, a method for predicting the evolution of key macroscopic parameters of tokamak plasma according to the present invention includes the following steps: S1. Obtain historical discharge data of the tokamak device and construct a multivariate time series including covariates and evolution target variables; S2. Construct a lightweight recurrent neural network prediction model. The prediction model includes: a shared encoder for extracting the latent state features of the input time series data; a physically decoupled dual-head branch structure, wherein the first branch is used to predict the residual increment of the fast response variable, the second branch is used to predict the residual increment of the slow response variable, and each branch output is provided with a learnable amplitude scaling factor. S3 employs a progressive autoregressive training strategy to train the prediction model. The training strategy includes: a course-based step size growth strategy that linearly increases the number of autoregressive expansion steps of the model with each training round, and a planned sampling strategy that decays the teacher-forced probability according to an inverse S-shaped curve, so that the prediction model gradually adapts to the self-generated input distribution; during training, the absolute prediction error, the rate of change error of fast variables, and the rate of change error of slow variables are optimized simultaneously. S4 utilizes the trained prediction model to continuously predict the evolution target variable of the tokamak device at future time steps in an autoregressive closed-loop manner.
[0031] The present invention will be described in detail below in several parts.
[0032] 1. Discharge data collection.
[0033] The input and output variables of the prediction model were determined, and historical discharge data of the Experimental Advanced Superconducting Tokamak (EAST) nuclear fusion device were collected specifically. The influence of poloidal specific pressure was fully considered. ), normalized specific pressure ( ) and internal sensation ( The physical factors influencing the evolution of plasma were analyzed, and 30 physical signals related to plasma state and control commands were selected as model input channels. The names and classifications of each signal are detailed in Table 1.
[0034] Table 1. List of input and output variables for the prediction model
[0035] In this invention, all input variables are divided into two categories: covariates and evolution target variables. Covariates include control variables and other macroscopic parameters required for prediction. Control variables are auxiliary heating signals, including low-hybrid power, electron cyclotron resonance heating power, ion cyclotron resonance heating power, and neutral beam injection power, which directly drive changes in the target quantity. The remaining macroscopic parameters mainly include plasma current, loop voltage, electron density, circumferential field coil current, radiation power, and related configuration parameters. A total of 27 signals are involved. The evolution target variable is the variable to be predicted. , and There are a total of 3 signals. The input to the prediction model is a historical window of data of length s, containing all 30 signals mentioned above; and covariates for the next time step. The output of the prediction model is the evolution target variable for the next time step.
[0036] 2. Construction of evolution prediction dataset.
[0037] The raw experimental data is stored in the MDSplus database, and variable data is read according to tree name and node name. HDF5 files are constructed to store the data, unit by shot number. Each variable is a group, and its data is stored as a multidimensional array in the data set under that group. If an auxiliary heater is not used, it is represented by a one-dimensional array of all zeros representing the number of channels. If a signal in a channel is missing, the entire signal is set to 0.
[0038] 3. Data preprocessing.
[0039] The raw EAST experimental data suffered from inconsistent sampling frequencies and numerous outliers, necessitating detailed preprocessing for each shot. First, for the auxiliary heating data, the mean difference before and after time 0 was calculated. If the mean difference was small, the entire signal was considered sensor noise and not actually used, so the entire signal was set to 0. For other auxiliary heating signals, the stable power application phase was identified, and the remaining portions were set to 0. Subsequently, all signals underwent unified processing, including: removing outliers and outliers exceeding physically reasonable ranges; linear interpolation to supplement data; truncating the plasma current peak segment; and resampling to a unified frequency (1 kHz). The data with unified timestamps was then converted into CSV files named after the shot number. The preprocessed data was then divided into training, validation, and test sets in an 8:1:1 ratio according to the shot number.
[0040] 4. Prediction Model Construction The overall architecture of the prediction model is as follows Figure 2 As shown.
[0041] (1) Basic Prediction Model. This invention employs a two-layer stacked Long Short-Term Memory (LSTM) network to map the input multivariate time series into compact hidden state features. LSTM, through selective memory mechanisms of forget gates, input gates, and output gates, can effectively capture long-range dependencies in plasma parameter evolution while alleviating the gradient vanishing problem common in standard recurrent networks. The encoder has 30 input channels, corresponding to all covariates and target variables listed in Table 1; the hidden state dimension is set to 256, achieving a balance between expressive power and computational efficiency. The two-layer stacked structure enables the model to abstract temporal patterns layer by layer: the first layer extracts local dynamic features, and the second layer learns higher-level evolutionary rules based on these features. Dropout with a dropout rate of 0.2 is introduced between layers to reduce the risk of overfitting and improve generalization performance. Figure 2 middle This represents the covariates at time t+1, with 0 representing a placeholder for the evolutionary objective variable. express Time's up The hidden state feature vector at time t.
[0042] (2) Fast and slow variable branches. The hidden state of the last time step of the LSTM output is fed into two independent fully connected branches. The first branch ( (Head) used for prediction and The increment, with an output dimension of 2; the second branch ( (Head) used for prediction The increment is 1, and the output dimension is 1. Each branch is followed by a Tanh activation function to restrict the original mapping value to [...]. Within the interval [1,1], multiply by their respective learnable scaling factors. and This design enables the prediction model to automatically learn appropriate single-step change amplitude constraints for both fast and slow variables, thus serving as a regularization and stabilizing mechanism for the autoregressive process.
[0043] (3) Residual Output Layer. The prediction model does not directly predict the absolute value at the next time step, but rather predicts the increment (residual) relative to the previous time step. The incremental predicted value of the dual-head output is added to the true value (during training) or predicted value (during inference) of the target variable at the previous time step to obtain the final absolute predicted value. The residual learning method transforms the complex absolute value regression into incremental regression with a mean near zero, significantly reducing the learning difficulty, while naturally embedding the physical prior of temporal continuity.
[0044] (4) Adaptive Residual Constraint Loss Function. To guide the prediction model to learn changes that conform to physical laws, this invention designs a composite loss function, which includes three components: absolute error loss. , rate of change loss as well as rate of change loss Each component employs the SmoothL1 loss function, which exhibits squared loss (gradient smoothing) when the absolute value of the error is small, and linear loss (robust to outliers) when the error is large. This is particularly suitable for scenarios where plasma signals contain a small number of outliers. Specifically, for any predicted value... With target value The SmoothL1 loss is defined as: ; Based on this, the three losses are defined as follows: ① Loss Penalize absolute prediction error; ② Loss Used to constrain fast variables The time-series rate of change; ③ Loss Used to constrain slow variables The time-series rate of change.
[0045] To automatically balance the magnitude of the three losses and the convergence speed, this invention introduces an adaptive multi-task weighting mechanism based on homoscedastic uncertainty. The total loss function is defined as: .
[0046] in, These are the learnable noise parameters corresponding to the loss term, which are optimized together with the network weights during training: when the noise of a certain loss term is large ( When it is relatively large, its corresponding weighting coefficient Automatically decrease, thereby reducing the contribution of this term to the total loss; conversely, if a certain loss term is easy to fit ( (Smaller), its weight automatically increases. Meanwhile, the regularization term... prevent It degenerates to infinity. This mechanism completely eliminates the tedious process of manually adjusting the loss weights in traditional multi-task learning and can dynamically adapt to the natural changes in the magnitude of each loss term at different training stages. Experiments show that this adaptive mechanism enables... The weights automatically converge to higher values, effectively strengthening the support for slow variables. Constraints on smooth evolution, and The weights are relatively moderate, thus achieving both the response sensitivity of fast variables and the noise resistance and smoothness of slow variables.
[0047] 5. Progressive autoregressive training.
[0048] To overcome exposure bias, this invention designs a two-dimensional progressive training strategy, including curriculum-based autoregressive step size growth and planned sampling probability decay. The overall progressive autoregressive training strategy is as follows: Figure 3 As shown.
[0049] (1) Curriculum-based autoregressive step size. Two hyperparameters are set. and The training samples contain A long sequence of steps. In each iteration of training. In the model, the actual number of autoregressive expansion steps Linear growth from small to large: ; in This represents the total number of training epochs. In the early stages of training, the model only needs to process short autoregressive sequences, and the impact of error accumulation is small. As training progresses, the step size gradually increases, forcing the model to correct its own errors over longer time series.
[0050] (2) Planned sampling probability decay. At each step of the autoregressive expansion, the model uses probability decay. Choose to use the true target value as the historical input for the next step, with probability. Use the model's own predictions. Probability Decrease with training rounds according to an inverse S-shaped curve: ; in The decay constant is set as a hyperparameter. This scheduling prioritizes teacher-mandated inputs in the early stages of training to ensure convergence stability, and gradually transitions to an almost entirely self-generated input mode in the later stages, ensuring that the model is fully exposed to the same data distribution as in the inference stage during the training phase.
[0051] 6. Test set verification.
[0052] The model was evaluated during training using independent validation set data. Validation evaluation strictly followed the autoregressive closed-loop protocol: the history buffer was initialized only with the real data from the first s steps; for each subsequent step, the prediction model received only the last s steps' data from the history buffer (where the target variable had been completely replaced with the model's historical predictions) and the known covariates of the current step, predicted the target value for the current step, and appended it to the history buffer. The MAE (mean absolute error) between the predicted and true sequences was calculated. Indicators such as the coefficient of determination were used to select the model parameters with the best validation results as the final prediction model. Additionally, the time taken for the autoregressive prediction step was statistically analyzed to verify its feasibility as a feedback controller for real-time prediction training.
[0053] Example: This embodiment is based on over 400 discharge experimental data collected by the EAST device between 2021 and 2024. The data covers various auxiliary heating combinations and different plasma parameter ranges. All discharges were divided into training, validation, and test sets in an 8:1:1 ratio. Figure 4 The full-time autoregressive prediction results for a typical discharge (fire 139924) in the test set are presented. In the figure, the solid black line represents the actual diagnostic signal, the dashed red line represents the model's autoregressive prediction output, and the gray dashed line marks the prediction start time. In over 7000 autoregressive prediction steps, the predicted trajectory showed a high degree of agreement with the actual value, achieving an overall prediction accuracy of 93% for this discharge.
[0054] Figure 5 The complete test discharge was displayed. Score distribution. For , and Its average The values reached 0.844, 0.785, and 0.781 respectively, with MAE values of 0.062, 0.057, and 0.040. Most discharges... The clustering score is above 0.9, proving that the model has good generalization performance.
[0055] Figure 6 The statistics show the single-step inference time of the prediction model on the NVIDIA A800 GPU. The average inference time is 0.99ms, and the maximum value is less than 1.1ms, which fully meets the 1ms cycle requirement of the EAST real-time control system.
[0056] In summary, this embodiment fully verifies the effectiveness, stability, and real-time performance of the present invention on real tokamak experimental data.
[0057] This invention significantly improves the accuracy and stability of long-term autoregressive prediction. By simulating the closed-loop conditions of the inference phase during the training phase, the model can effectively suppress error accumulation in multi-step predictions, achieving [the desired improvement]. , and Continuous, non-divergent predictions over thousands of steps (corresponding to several seconds of physical time).
[0058] This invention accurately characterizes the multi-scale dynamic properties of fast and slow variables. The physically decoupled dual-head architecture enables the model to... Preserve high-frequency response details, while providing It provides a smooth evolution trajectory, avoiding the problem of a single output head failing to address all aspects.
[0059] This invention meets the millisecond-level latency requirements of real-time control systems. The model has only about 800,000 parameters, and the single-step inference time is less than 1ms, which is fully compatible with the stringent constraint of the 1ms control cycle of existing tokamak systems.
[0060] This invention reduces the reliance on complex physical modeling and improves the method's versatility. It is entirely data-driven, requiring no complex plasma state-space equations, and the method can be easily transferred to other tokamak devices and similar long-term prediction tasks.
[0061] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0062] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0063] In one embodiment, the present invention provides a computer system, which may be a server. The computer system includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.
[0064] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0065] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0066] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for predicting the evolution of key macroscopic parameters in tokamak plasma, characterized in that, include: Historical discharge data of the tokamak device were acquired, and a multivariate time series including covariates and evolution target variables was constructed. A lightweight recurrent neural network prediction model is constructed. The prediction model includes: a shared encoder for extracting the latent state features of the input time series data; and a physically decoupled dual-head branch structure, wherein the first branch is used to predict the residual increment of the fast response variable, the second branch is used to predict the residual increment of the slow response variable, and each branch output is provided with a learnable amplitude scaling factor. A progressive autoregressive training strategy is adopted to train the prediction model. The training strategy includes: a course-based step size growth strategy that linearly increases the number of autoregressive expansion steps of the prediction model with each training round, and a planned sampling strategy that decays the teacher-forced probability according to an inverse S-shaped curve, so that the prediction model gradually adapts to the self-generated input distribution. During training, the absolute prediction error, the rate of change error of fast variables, and the rate of change error of slow variables are optimized simultaneously. Using the trained prediction model, the evolution target variable of the tokamak device at future time steps is continuously predicted in an autoregressive closed-loop manner.
2. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 1, characterized in that, The historical discharge data includes: plasma current, low hybrid power, circumferential field coil current, ion cyclotron resonance heating power, electron cyclotron resonance heating power, neutral beam injection power, elongation ratio, electron density, radiation power, plasma-wall gap, ring voltage, poloidal specific pressure, normalized specific pressure, and inductance as model input channels; wherein, the poloidal specific pressure, normalized specific pressure, and inductance are simultaneously used as evolution target variables, and the rest are covariates.
3. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 1, characterized in that, The shared encoder is used to extract the latent state features of the input time-series data, specifically including: The shared encoder employs a two-layer stacked long short-term memory network to map the input multivariate time series into compact hidden state features.
4. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 2, characterized in that, The first branch is used to predict the residual increment of the fast-response variable, and the second branch is used to predict the residual increment of the slow-response variable. Each branch output has a learnable magnitude scaling factor, specifically including: The hidden state features output from the shared encoder are fed into two independent branches. The first branch predicts the increments of the poloidal specific pressure and the normalized specific pressure; the second branch predicts the increment of the interotropic induction. Each branch is followed by a Tanh activation function to map the original increment prediction values of the two branches to a value within [...]. Within the interval [1,1], multiply by their respective learnable scaling factors. and .
5. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 3, characterized in that, The prediction model also includes a residual output layer, specifically including: The incremental predicted values output by the first and second branches are added to the evolution target variable value at the previous time step to obtain the absolute predicted value at the current time step. In the training phase, the evolution target variable value at the previous time step is the true value; in the inference phase, the evolution target variable value at the previous time step is the historical predicted value output by the prediction model itself.
6. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 1, characterized in that, The course-based step size growth strategy, which linearly increases the number of autoregressive expansion steps of the prediction model with each training round, specifically includes: Set two hyperparameters and During the training In each iteration, the actual number of autoregressive expansion steps of the prediction model Linear growth from small to large: ； in, This refers to the total number of training rounds.
7. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 1, characterized in that, The planned sampling strategy of attenuating the teacher's forced probability according to the inverse S-shaped curve specifically includes: At each step of the autoregressive expansion, the predictive model uses probability... Choosing to use the actual evolutionary target variable values as input for the next step, with probability... Use the self-predicted evolutionary target variable as input for the next step. Decrease with training rounds according to an inverse S-shaped curve: ； in It is the attenuation constant. For the first Index of each iteration round.
8. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 2, characterized in that, The simultaneous optimization of absolute prediction error, fast variable rate of change error, and slow variable rate of change error during training specifically includes: Absolute prediction error ; This represents the target value of the evolutionary objective variable at time t+1. This represents the target value of the evolution target variable at time t+1; Error of rate of change of fast variable ; Indicates polar pressure or normalized specific pressure exist The predicted value of the time increment. Indicates polar pressure or normalized specific pressure exist The actual value of the time increment; Error in the rate of change of slow variables ; Indicating inner feelings exist The predicted value of the time increment. Indicating inner feelings exist The actual value of the time increment; Represents the SmoothL1 loss function; Total loss function for: ； in, It is a loss item Learnable noise parameters.
9. The method for predicting the evolution of key macroscopic parameters of tokamak plasma according to claim 1, characterized in that, The method of continuously predicting the evolution target variable of the tokamak device at future time steps using the trained prediction model in an autoregressive closed-loop manner specifically includes: The buffer is initialized with real discharge data within a historical window of length s. For each prediction step, the discharge data of the most recent s time steps in the buffer and the real covariate value of the current step are used as inputs to predict the next evolution target variable. The predicted evolution target variable is appended to the buffer and the window is slid. The above process is repeated to complete the multi-step autoregressive prediction.
10. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.