A data-driven method for modular small-break steam generator key parameter simulation

Abstract

This invention discloses a data-driven method for simulating key parameters of a modular small modular reactor (SMR) steam generator, relating to the field of nuclear reactor system simulation technology. The method includes: using the key CFD simulation parameter K of the steam generator as the model output, and using real-time monitored physical quantities P, time T, and spatial location information S as the model input; constructing a mapping relationship between the model output and model input as a pre-trained model; training the pre-trained model using a standardized training dataset; and determining the trained model as a data-driven model with nuclear working fluid characteristics, enabling the data-driven model to predict key parameters under multi-physics coupled operating conditions. This invention, by constructing a dedicated workflow system for steam generator simulation data processing and combining a model designed for key parameter prediction with optimized training strategies, can achieve efficient and accurate simulation and prediction of core parameters during steam generator operation.

Description

Technical Field

[0001] This invention relates to the field of nuclear reactor system simulation technology, and in particular to a data-driven method for simulating key parameters of a modular small reactor steam generator. Background Technology

[0002] Modular small modular reactors (SMRs), as a flexible, efficient, and adaptable advanced nuclear energy technology, are attracting widespread attention in the global energy market. With their modular design, rapid deployment, and scalable capabilities, SMRs have become a key option for solving energy supply and grid compatibility issues.

[0003] In the nuclear energy utilization cycle, the start-up and shutdown process of small modular reactor (SMR) systems is a core link in ensuring the safe and stable operation of nuclear facilities. Its safety not only directly affects the reliability of the entire nuclear system but is also a key component of the nuclear safety management system. Furthermore, accurate prediction of key SMR operating parameters (such as temperature, pressure, and flow rate) is crucial for optimizing system design, achieving optimal control strategies, and improving operational monitoring efficiency. However, the internal physical processes of SMR steam generators involve complex multiphase fluid dynamics and heat transfer phenomena, including the flow characteristics of single-phase / two-phase fluids, convective heat transfer and phase change heat transfer (such as boiling and condensation), and dynamic pressure changes. The construction of simulation models and numerical calculations for such multiphysics coupled processes face significant challenges: the coupling mechanism between multiphase flow and heat transfer must simultaneously consider fluid dynamics, thermodynamics, and phase change dynamics, leading to highly nonlinear mathematical models; traditional numerical methods require mesh partitioning to discretize the solution domain, resulting in exponentially increasing computational resource consumption for high-precision simulations.

[0004] Currently, computational fluid dynamics (CFD) is the mainstream method for simulating such complex physical processes, demonstrating high functionality and accuracy in fluid dynamics simulations. However, CFD methods face the following bottlenecks in practical applications: to capture subtle features of fluid flow (such as eddies and phase interface fluctuations), high-resolution meshes and time steps are required, leading to excessively long simulation times per run; high-precision CFD simulations rely on high-performance computing clusters, significantly increasing hardware costs and energy consumption; and in scenarios requiring rapid response, such as small reactor start-up and shutdown control, the latency characteristics of CFD make it difficult to meet the demands of real-time decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven method for simulating key parameters of a modular small reactor steam generator, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: A data-driven method for simulating key parameters of a modular small reactor steam generator includes: A pre-trained model is constructed. The pre-trained model uses the key CFD simulation parameter K of the steam generator as the model output, and the real-time monitored physical quantity P, time T and spatial location information S as the model input, and constructs the mapping relationship between the model output and the model input. A training dataset is obtained, the training dataset is standardized, and the standardized dataset is divided according to spatial and temporal dimensions to generate a fused training dataset; the fused training dataset includes primary and secondary side data of the steam generator inside the CFD simulation. Based on GPU acceleration technology and AMP acceleration technology, under the set training hyperparameter environment, the pre-trained model is trained using the fused training dataset, and the trained model is determined as a data-driven model with nuclear working fluid characteristics, enabling the data-driven model to predict key parameters under multi-physics coupling conditions; wherein, the predicted key parameters are used to provide real-time data support for reactor start-up detection, safety assessment and control optimization.

[0007] Optionally, the pre-trained model employs a modular deep neural network, comprising an embedding mapping layer, a hidden layer, and a linear projection layer connected in sequence; wherein, the embedding mapping layer is used to map the input data to a 128-dimensional space; the hidden layer is configured as two equal-dimensional hidden layers, each with 128 neurons, for learning complex features in the data; the linear projection layer is used to map the features learned by the hidden layer to the output space.

[0008] Optionally, the key CFD simulation parameter K includes: near-wall fluid temperature, center fluid temperature, and center fluid velocity; the physical quantity P includes: secondary inlet flow rate, secondary inlet temperature, secondary inlet pressure, primary inlet flow rate, primary inlet temperature, and main loop pressure.

[0009] Optionally, the Gaussian error linear unit function is used as the activation function in the pre-trained model.

[0010] Optionally, the standardization process employs feature standardization, which is used to eliminate the influence of differences in units among the feature data in the training dataset through mathematical transformation, so that all features are numerically of the same order of magnitude. The feature standardization process is described above. Represented as: In the formula, This represents the mean of the column vector of data. Represents the variance of the data column vector. This indicates subtraction by position.

[0011] Optionally, the spatial dimension partitioning process is as follows: based on the environmental and parameter change characteristics of various locations inside the steam generator during the start-up phase, the data in the training dataset is sampled and partitioned according to the z-axis order, as shown below: In the formula, For the training dataset, Indicates to according to The order is used for downsampling with an interval of 2 and the first term being 1. Representation and training dataset corresponding The value of , where n is the data dimension and R represents the real number field.

[0012] Optionally, the process of dividing by time dimension is as follows: data from different spatial locations at the same time are determined as the data after sampling and division.

[0013] Optionally, the pre-trained model uses mean squared error as the loss function during training.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a data-driven method for simulating key parameters of a modular small reactor (MSD) steam generator. The method includes: constructing a pre-trained model; the pre-trained model uses the CFD simulation key parameter K of the steam generator as the model output, and real-time monitored physical quantities P, time T, and spatial location information S as the model input, and constructs a mapping relationship between the model output and the model input; acquiring a training dataset, standardizing the training dataset, and dividing the standardized dataset according to spatial and temporal dimensions to generate a fused training dataset; the fused training dataset includes primary and secondary side data of the steam generator from CFD simulation; based on GPU acceleration technology and AMP acceleration technology, under a set training hyperparameter environment, the pre-trained model is trained using the fused training dataset, and the trained model is determined as a data-driven model with nuclear working fluid characteristics, enabling the data-driven model to predict key parameters under multi-physics coupled conditions; wherein, the predicted key parameters are used to provide real-time data support for reactor start-up detection, safety assessment, and control optimization.

[0015] This invention transforms traditional physics simulation problems into data-driven intelligent prediction problems by constructing a machine learning model specifically for predicting key parameters of small modular reactors (SMRs). This enables real-time, high-precision dynamic simulation of core parameters such as temperature, pressure, and flow rate during reactor startup. This method replaces the traditional CFD physics simulation model with a data-driven model, overcoming the inherent bottlenecks of CFD methods in terms of computational speed and resource consumption, and providing an efficient solution for rapid start-up and shutdown control of modular SMRs. Through data structure modeling and feature engineering, the complex simulation problem involving multiple coupled physics fields is transformed into a time-series prediction task that can be handled by machine learning, reducing the complexity of model construction. A neural network architecture tailored to the characteristics of nuclear working fluids is designed, combined with adaptive training algorithms and multi-source data fusion technology, to improve the model's generalization ability and prediction accuracy under extreme conditions. This provides real-time data support for reactor startup detection, safety assessment, and control optimization, significantly improving the operational efficiency and risk control capabilities of nuclear facilities. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steam generator rake line setup in this embodiment; Figure 2 This is a schematic diagram of spatially partitioning the dataset in this embodiment; Figure 3 This is a schematic diagram of the data-driven method in this embodiment; Figure 4 This is a schematic diagram of the neural network structure used in this embodiment; Figure 5 This is a graph showing the loss decrease during the model training process in this embodiment; Figure 6 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the first-order side parameter k1 on the left and the model prediction value on the right. Figure 7 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the first-order side parameter k2 on the left and the model prediction value on the right. Figure 8 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the primary side parameter k3 on the left and the model prediction value on the right. Figure 9 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the secondary side parameter k1 on the left and the model prediction value on the right. Figure 10 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the secondary side parameter k2 on the left and the model prediction value on the right. Figure 11 This is a comparison chart of the prediction effects in this embodiment, with the CFD simulation value of the secondary side parameter k3 on the left and the model prediction value on the right. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a data-driven method for simulating key parameters of a modular small reactor steam generator, aiming to solve or improve at least one of the above-mentioned technical problems.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] This invention provides a data-driven method for simulating key parameters of a modular small reactor steam generator, comprising: Step 1: Task Transition. Construct a pre-trained model; the pre-trained model uses the key CFD simulation parameter K of the steam generator as the model output, and the real-time monitored physical quantity P, time T, and spatial location information S as the model input, and constructs the mapping relationship between the model output and the model input.

[0022] This step specifically includes: The internal simulation model of a modular small modular reactor (SMR) steam generator is a tool that uses mathematical modeling and numerical calculations to simulate its internal physical processes, and is used to study its thermal-hydraulic behavior, structural performance, and multiphysics coupling effects. (Appendix) Figure 1 The internal structure of a steam generator is given, where the red and blue areas represent the secondary and primary flow domains, respectively. A CFD-based steam generator simulation system uses the secondary inlet flow rate (p1), secondary inlet temperature (p2), secondary inlet pressure (p3), primary flow rate (p4), primary inlet temperature (p5), and main loop pressure (p6) as inputs to calculate key parameters at each coordinate point at each moment for the fluid at the center (secondary side) and near the wall (primary side): near-wall fluid temperature (k1), center fluid temperature (k2), and center fluid velocity (k3).

[0023] The task transformation aims to convert the complex, detailed simulation process of a small, modular pressurized water reactor (PWR) direct-flow steam generator into a machine learning task. The complex key parameters K=(k1, k2, k3) calculated in the CFD simulation are used as the target for neural network prediction, and the real-time monitored physical quantities P=(p1, p2, p3, p4, p5, p6) are used as network inputs to construct a data-driven model f. Specifically, to deeply explore the spatial distribution characteristics of the key parameter K, spatial location information S=(x, y, z) and time T are used as core inputs to capture dynamic changes. This task utilizes a neural network model to construct the following mapping relationship (where θ represents all parameters in the neural network model): .

[0024] Step 2: Data Processing. Obtain the training dataset, standardize the training dataset, and divide the standardized dataset according to spatial and temporal dimensions to generate a fused training dataset; the fused training dataset includes primary and secondary side data from CFD simulations inside the steam generator.

[0025] This step specifically includes: (1) Data Preprocessing: In the data preprocessing stage, given that the key parameters K, time T, spatial location information S, and physical features P each carry different dimensions, which may lead to instability and bias accumulation during training, this invention adopts feature standardization. Mathematical transformations are used to eliminate the influence of dimensional differences between features, ensuring that all features are numerically on the same order of magnitude. This process first requires processing the CFD simulation data into... The form is where n is the amount of data obtained from the simulation calculation. The specific standardization process is actually to standardize each column of the above matrix. Taking the n-dimensional column vector p1 in P as an example: in Represents column vectors The mean, express variance This represents positional subtraction (i.e., vector subtraction). Subtract the constant from each element at each position. Therefore, the standardized result It is still an n-dimensional column vector. Therefore, the data standardization process can be represented in the following form: (2) Spatial Dataset Partitioning: To achieve accurate real-time prediction of the distribution of key parameters inside the steam generator, the model must be able to accurately derive the values ​​of key parameters corresponding to different locations inside the steam generator under given time characteristics and physical constraints. Therefore, unlike conventional machine learning tasks, this invention adopts a spatial partitioning strategy for the dataset. Based on the unique environment and parameter change characteristics of different locations inside the steam generator during the startup phase, the data in the dataset is sampled and partitioned according to the z-axis order (see Appendix). Figure 2 This location-oriented dataset partitioning method ensures that the model can fully learn the complex mapping relationship between the internal spatial distribution characteristics of the steam generator and key parameters during training. Specifically, this invention partitions the training set, validation set, and test set in a 2:1:1 ratio, where the training set is derived from downsampling with a z-interval of 2, and the validation and test sets are derived from downsampling with a z-interval of 4. The training set, validation set, and test set data can be represented as follows: in Indicates to according to The order is used for downsampling with an interval of 2 and the first term being 1. Representation and training set corresponding The value of , where n is the amount of data. It should be noted here that... The format is as follows: That is, a moment It corresponds to a set of spatial points S, while P and K are actually functions of time t and spatial location information S.

[0026] (3) Data batching by time: To improve the training efficiency of the neural network model and accelerate the evaluation process, this invention adopts a special data batching method. Data from different spaces at the same time are grouped into a small batch, and the model parameters are updated using a small-batch stochastic gradient descent algorithm. This strategy not only effectively utilizes the spatiotemporal characteristics of the data, but also significantly enhances the model's generalization ability and convergence speed during the learning process. Given time T and the corresponding physical feature P, a small batch of data is represented as: in, This represents a small batch of training subsets generated according to the strategy of this invention. MiniBatch This indicates the size of the small batch.

[0027] Step 3: Establish a modular deep fully connected neural network. Based on GPU acceleration and AMP acceleration technologies, under the set training hyperparameters, the pre-trained model is trained using the fused training dataset. The trained model is then defined as a data-driven model with nuclear working fluid characteristics, enabling it to predict key parameters under multi-physics coupling conditions. These predicted key parameters provide real-time data support for reactor startup detection, safety assessment, and control optimization.

[0028] This step specifically includes: Neural network structures often exhibit multi-layer architectures, with each layer flexibly configured with varying numbers of hidden units. Their complexity often stems from numerous and finely defined hyperparameters. Therefore, this invention employs the concept of modular deep neural networks, which are composed of stacked shallow neural network modules with identical or similar structures. Compared to traditional neural networks, modular deep neural networks significantly reduce the number of hyperparameters that need adjustment, thereby simplifying the parameter tuning process and making model optimization more efficient and direct. Secondly, this modular design enhances the network's scalability and flexibility, facilitating customization and adjustment according to specific task requirements. Generally, a shallow neural network module, i.e., a single-layer neural network, is as follows: in, express k Hidden variables, As a linear layer, It is a position-based activation function. This represents a shallow neural network module. Therefore, a modular deep neural network can be represented as a composition of shallow neural network modules: in, This represents a modular deep multilayer perceptron.

[0029] Step 4: Model training and evaluation.

[0030] (1) Loss Function: In machine learning and deep learning, Mean Squared Error (MSE) is a commonly used loss function, especially in regression problems. This invention uses MSE as the loss function to train the model, and its mathematical expression is: in, These are CFD simulation values. These are model predictions. It is an L2 norm.

[0031] (2) Training Algorithm: This invention uses Adam to update model parameters. The Adam algorithm is a widely used optimization algorithm for updating parameters in deep learning models. It combines the ideas of momentum and RMSprop algorithms, dynamically adjusting the learning rate of each parameter by calculating the first and second moment estimates of the gradient. In the Adam algorithm, the first and second moment estimates of the gradient are calculated. t The parameter updates in a round of iterations typically follow these steps: in, Indicates the first t The model parameters of DeepMLP during each iteration. This represents the initial learning rate. This represents the first moment estimate of the gradient. This represents the bias correction term for the first-order moment estimation. This represents the second-order moment estimate of the gradient. This represents the bias correction term for the second-order moment estimation. This represents the exponential decay rate estimated by the first moment. This represents the exponential decay rate estimated by the second moment. Indicates the first t The gradient of the loss function with respect to the model parameters during each iteration. This represents the numerical stability constant to prevent the denominator from being zero. The operator represents the gradient of a function. This represents the loss function.

[0032] (3) Evaluation metrics: This invention uses MAE, MSE, and R-squared to evaluate the final model. These metrics help the present invention understand the degree of difference between model predictions and actual values. MAE is the average of the absolute values ​​of the differences between predicted and actual values. It is less sensitive to outliers because outliers are not excessively amplified when their absolute values ​​are taken. The mathematical expression for MAE is: The R-squared score represents the goodness of fit between the model's predicted values ​​and the actual values. Its value ranges from negative infinity to 1. The closer the R-squared score is to 1, the better the model's fit; if the R-squared score is negative, the model's performance is even worse than a simple average prediction (i.e., always predicting the average of the actual values). The mathematical representation of the R-squared score is as follows: The overall flowchart of the above scheme is as follows: Figure 3 As shown.

[0033] Based on the above technical solution, the following embodiments are provided.

[0034] In this embodiment, the data used is refined simulation data of a small modular pressurized water reactor direct-flow steam generator under different control strategies. The total duration of the CFD simulation data is 4500 seconds.

[0035] Step 1: Task Switching This example utilizes real-time monitoring data and spatiotemporal information during the power enhancement process of a steam generator to predict CFD simulation data that is difficult to obtain directly. Figure 1 The internal structure of the steam generator is shown, with red and blue representing the secondary and primary flow domains, respectively. CFD simulation data is divided into two parts: the values ​​of key parameter K at the center of the pipe (secondary side) and near the wall (primary side). Data points are determined using three-dimensional coordinates (x, y, z, uniformly labeled S, unit: m), with different y values ​​defining the primary and secondary sides. To describe the distribution of key parameter K within the pipe, the z-axis is subdivided into 1000 points along the 1.8-meter pipe length, and key parameter values ​​are collected at these points during the power-up phase (4500 seconds). This embodiment predicts three key parameters K (k1, k2, and k3) using time information T, spatial location information S, and six observable physical information P (denoted as p1, p2, p3, p4, p5, and p6). Therefore, the input to this data-driven model is a 10-dimensional vector, and the output is a 3-dimensional vector.

[0036] Step 2: Data Processing (1) Data Preprocessing: As can be seen from Table 1, different variables exhibit diverse dimensional characteristics and are accompanied by a certain degree of complexity. To ensure the effectiveness and stability of model training, this invention performs standardization on all variables in the training set. Simultaneously, this invention records the mean and variance of each variable during the standardization process; this information is crucial in the subsequent model deployment stage. Specifically, when the model needs to be used for prediction, this invention uses these recorded means and variances to perform destandardization on the prediction results to ensure that the output prediction values ​​can be restored and conform to their original actual dimensions, thereby providing more realistic and accurate prediction results.

[0037] (2) Spatial division of the dataset: In stage (1), the present invention clarified the task requirements and data scope, and based on the primary and secondary side data inside the steam generator of CFD simulation, according to Figure 2The spatial location partitioning strategy shown was used to construct the dataset. Given that the CFD simulation covers two independent parts of data, each part covering 1000 spatial locations corresponding to K values ​​at the same time point, spanning 4500 time nodes, the total data volume in this example reaches 9 million. To ensure the effectiveness and generalization ability of the model training, this invention divides the data into a training set, a validation set, and a test set in a 2:1:1 ratio, where the training set contains 4.5 million data points, while the validation set and test set each contain 2.25 million data points.

[0038] (3) Data batching by time: In this example, all data at a single moment are used as a batch to train the neural network. This strategy merges the data from the two independent parts of the CFD simulation, ensuring that the model can fully learn the combined characteristics of the primary and secondary fluids inside the steam generator during training. Therefore, each batch contains 1000 data points during training, and 500 data points during evaluation and testing.

[0039] Step 3: Establish a modular deep fully connected neural network The neural network model constructed in this example uses the GELU (Gaussian Error Linear Unit) function as the activation function. The GELU function is represented as follows: The neural network model constructed in this example consists of three components (such as...). Figure 4 As shown): First, there is the embedding mapping layer, responsible for mapping the input data to a 128-dimensional space; then comes the equal-dimensional hidden layer, which consists of two hidden layers, each with 128 neurons, ensuring that the model can fully capture and learn the complex features in the data; finally, there is the linear projection layer, responsible for mapping the features learned by the hidden layers to the output space. In summary, this neural network model contains three non-linear layers (the embedding mapping layer and two equal-dimensional hidden layers), and the specific representation of this neural network is as follows: in, , Indicates the embedded mapping layer, This represents the first hidden layer. This indicates the second hidden layer. This indicates a linear projection layer.

[0040] Step 4: Model Training and Evaluation (1) Other training strategies: In order to improve training efficiency and speed, this embodiment uses two training acceleration strategies during the training of the neural network model. One is GPU (Graphics Processing Unit) acceleration technology. This embodiment utilizes the parallel processing characteristics of the GPU to achieve efficient data processing and rapid updating of model parameters, thereby significantly shortening the training time. In addition, this embodiment also adopts AMP (Automatic Mixed Precision) acceleration technology as another important training acceleration method. AMP technology dynamically adjusts the data precision during the model training process, that is, while ensuring the stability of model training, it appropriately reduces the precision requirements of certain calculation steps to reduce the consumption of computing resources and accelerate the calculation process. This strategy not only helps to alleviate GPU memory pressure, but also further improves the training speed, while maintaining or even improving the final performance of the model to a certain extent.

[0041] (2) Training hyperparameters: In this embodiment, the initial learning rate of the training algorithm is set to 0.005, and the model is trained for 10 rounds. The loss reduction trend during the training process is as follows: Figure 5 As shown, Figure 5 This indicates that the training set loss remained relatively stable during the last few training rounds, suggesting that the model had been sufficiently trained. Notably, the validation set loss did not increase with the number of training rounds, demonstrating the model's good generalization ability and effectively preventing overfitting.

[0042] (3) Model Evaluation: Tables 1 and 2 show the evaluation indicators for predicting the distribution of the core parameter K inside the steam generator at a fixed time interval of 500 seconds. Table 1 shows the evaluation of the prediction of the key parameters on the primary side. Table 2 shows the evaluation of the prediction of the key parameters on the secondary side.

[0043] The key data in these two tables demonstrate that the data-driven approach exhibits high accuracy in predicting the distribution of these key parameters K within the steam generator. Specifically, the mean absolute error (MAE) is strictly controlled within 0.38, and the root mean square error (MSE) is below 0.16, indicating minimal deviation between the predicted values ​​and CFD simulation values. Furthermore, the goodness of fit (R-squared) is not less than 0.95, further demonstrating a high degree of agreement between the predicted values ​​of the data-driven model and the CFD simulation values.

[0044] Figures 6 to 11The above evaluation results are further reinforced through intuitive comparison. These charts respectively show the baseline data obtained from CFD simulation (left) and the distribution of key parameters predicted by the data-driven method (right). The high degree of consistency between the two not only verifies the accuracy of the data-driven prediction model, but also highlights the powerful ability of this method to capture the complex physical processes inside steam generators.

[0045] Table 1 Evaluation of primary side prediction

[0046] Table 2 Evaluation of secondary side prediction

[0047] In summary, in practical applications, this invention demonstrates the following advantages over CFD simulation in predicting key parameters during the startup process of small reactors: Computational Efficiency and Speed: While CFD simulation technology can provide high-precision fluid dynamics simulations, its computational process is extremely complex and time-consuming. For scenarios requiring rapid response and immediate feedback, such as small reactor startup, the long computation time required by CFD simulation becomes a major bottleneck. In contrast, data-driven methods based on neural networks are renowned for their high computational efficiency. Once trained, neural network models can quickly process input data and output predictions, significantly reducing response time. This immediacy makes data-driven methods more suitable for industrial scenarios requiring rapid decision-making and control.

[0048] Scalability and Flexibility: Building and modifying CFD models typically involves complex physical modeling and programming, making them difficult to extend and adapt to new application scenarios. Data-driven approaches, on the other hand, are more flexible, adapting to different reactor types and operating conditions through simple data collection and model training. Furthermore, as new data accumulates, neural network models can continuously learn and optimize, further improving prediction and control performance.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0050] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A data driven method for modular small-break steam generator key parameter simulation, characterized in that, include: A pre-trained model is constructed. The pre-trained model uses the key CFD simulation parameter K of the steam generator as the model output, and the real-time monitored physical quantity P, time T and spatial location information S as the model input, and constructs the mapping relationship between the model output and the model input. Obtain the training dataset, standardize the training dataset, and divide the standardized dataset according to spatial and temporal dimensions to generate a fused training dataset. The fusion training dataset includes primary and secondary side data from CFD simulations of the steam generator's interior. Based on GPU acceleration technology and AMP acceleration technology, under the set training hyperparameter environment, the pre-trained model is trained using the fused training dataset, and the trained model is determined as a data-driven model with nuclear working fluid characteristics, enabling the data-driven model to predict key parameters under multi-physics coupling conditions; wherein, the predicted key parameters are used to provide real-time data support for reactor start-up detection, safety assessment and control optimization.

2. The data driven method for modular small bundle steam generator key parameter simulation as claimed in claim 1 wherein, The pre-trained model employs a modular deep neural network, comprising an embedding mapping layer, a hidden layer, and a linear projection layer connected in sequence. The embedding mapping layer maps the input data to a 128-dimensional space. The hidden layer consists of two equal-dimensional hidden layers, each with 128 neurons, used to learn complex features in the data. The linear projection layer maps the features learned by the hidden layers to the output space.

3. The data driven approach for modular small bundle steam generator key parameter simulation as claimed in claim 1 wherein, The key CFD simulation parameters K include: near-wall fluid temperature, center fluid temperature, and center fluid velocity; the physical quantities P include: secondary inlet flow rate, secondary inlet temperature, secondary inlet pressure, primary flow rate, primary inlet temperature, and main loop pressure.

4. The data driven method for modular small bundle steam generator key parameter simulation as claimed in claim 1 wherein, The pre-trained model uses the Gaussian error linear unit function as the activation function.

5. The data-driven method for simulating key parameters of a modular small reactor steam generator according to claim 1, characterized in that, The standardization process employs feature standardization, which uses mathematical transformations to eliminate the influence of differences in units among the feature data in the training dataset, ensuring that all features are numerically on the same order of magnitude. The feature standardization process is described above. Represented as: In the formula, This represents the mean of the column vector of data. Represents the variance of the data column vector. This indicates subtraction by position.

6. The data-driven method for simulating key parameters of a modular small reactor steam generator according to claim 1, characterized in that, The spatial dimension partitioning process is as follows: Based on the environmental and parameter change characteristics of various locations inside the steam generator during the startup phase, the data in the training dataset is sampled and partitioned according to the z-axis order, as shown below: In the formula, For the training dataset, Indicates to according to The order is used for downsampling with an interval of 2 and the first term being 1. Representation and training dataset corresponding The value of , where n is the data dimension and R represents the real number field.

7. The data-driven method for simulating key parameters of a modular small reactor steam generator according to claim 1, characterized in that, The process of dividing data according to the time dimension is as follows: data from different spatial locations at the same time are identified as the data after sampling and division.

8. The data-driven method for simulating key parameters of a modular small reactor steam generator according to claim 1, characterized in that, The pre-trained model uses mean squared error as the loss function during training.

Images