Reactor core transient CFD flow field reconstruction technology based on data driving

Abstract

The invention discloses a reactor core transient CFD flow field reconstruction technology based on data driving, which is applied to the technical field of reactor core flow field reconstruction, and reconstructs a calculation result of a low-fidelity transient CFD model by combining a multi-fidelity model and an LSTM neural network to obtain transient flow field data approaching a calculation result of a high-fidelity transient CFD model to a certain extent; when the method is used for predicting the high-resolution and high-precision transient data of the reactor core, the situation that a high-fidelity CFD model is directly used for transient calculation can be avoided, and the calculation cost and the calculation burden of a low-fidelity CFD model and an LSTM network are small. Therefore, by the adoption of the method, the calculation burden and cost of the transient CFD calculation process can be reduced, researchers are helped to understand the thermal hydraulic behaviors in the reactor more deeply, the safety accident process is predicted more accurately, and safer reactor core design optimization and more effective emergency response plan making are facilitated.

Description

Technical Field

[0001] The invention relates to the technical field of core flow field reconstruction, which is a core transient CFD flow field reconstruction technology based on data drive. Background Art

[0002] The contradiction between the rapid growth of energy demand and energy shortage has become one of the main problems in the world today. The development and utilization of new clean energy can not only alleviate the increasingly serious energy crisis, but also curb the problem of global warming to a certain extent. Nuclear energy, as the only green and clean energy that can generate electricity stably and participate in peak load regulation, has been widely used for half a century. At present, various types of nuclear reactors have been widely used in the fields of electricity, heating and power engineering. Based on the extreme importance of nuclear reactor safety, thermal safety analysis in the reactor core design stage and thermal state prediction in the operation stage are crucial. It is necessary to focus on ensuring that the reactor coolant can timely remove the core heat under any circumstances to prevent the occurrence of core meltdown. At the same time, the thermal hydraulic state inside the core also affects the economy of the entire nuclear power system.

[0003] In recent decades, with the substantial improvement in computer performance and computational model accuracy, the accuracy and reliability of CFD analysis technology have made great progress, providing researchers with high-fidelity coolant flow and heat transfer data for refined core thermal hydraulic analysis, and gradually becoming an important part of supporting the digital and intelligent development of nuclear power technology. Through the calculation of the CFD program for refined core thermal hydraulics, it is possible to simulate and predict the heat release, heat transfer and flow of the core on a fine spatial scale. This simulation is conducive to reducing the safety margin, supporting the increase in power plant power and the extension of the fuel cycle, and ensuring the safety and economy of nuclear power. At the same time, refined CFD transient calculations of the core can help researchers have a deeper understanding of the thermal hydraulic behavior in the reactor, more accurately predict the process of safety accidents, and facilitate safer core design optimization and more effective emergency response plan formulation. However, for thermal hydraulic equipment such as reactor core, which has complex structure and serious cross-scale problems, it is still difficult to implement full-scale or even single fuel assembly numerical calculations using conventional numerical calculation methods under existing computer hardware conditions. Even if relevant calculations can be barely performed on super-large parallel computing clusters, there are still problems such as low efficiency, high cost and poor popularity of pre-processing, numerical solution and post-processing. Especially under transient conditions, the cost, time and storage space required for CFD calculations are much greater than those of steady-state processes. Therefore, in order to promote the engineering application of thermal hydraulic CFD calculation analysis in the core component flow basin, it is necessary to develop a calculation method that reduces the resources and time required for CFD calculations in the core component flow basin. Summary of the invention

[0004] In view of this, the present invention provides a data-driven core transient CFD flow field reconstruction technology to solve the problems existing in the background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Step S1: For the target core rod bundle channel geometry, according to the required accuracy and flow field resolution, a suitable grid model is selected through grid independence analysis to establish a CFD calculation model, which is called a high-fidelity CFD transient calculation model and is represented by HF-CFD. The low-fidelity CFD transient calculation model is established by a rough grid that is roughened on the basis of the high-fidelity model grid, and its roughness can be adjusted according to the calculation time requirements, affordable costs, etc. of the method application object.

[0007] Step S2: Establish a sample space of transient boundary conditions and obtain a transient flow field data set.

[0008]

[0009] Among them, Θ i is the i-th transient boundary condition in the sample space, L is the number of samples in the sample space, s is the number of boundary conditions, and τ is the time dimension of the boundary conditions; the number of samples and the time dimension in the sample space should be determined according to the affordable calculation burden, cost limit, and accuracy requirements, and the selection range should preferably include the boundary conditions of the target type to be calculated.

[0010] Using the high-fidelity and low-fidelity transient CFD models obtained in S1, calculations are performed according to the boundary conditions in the sample space, and high-fidelity and low-fidelity transient calculation sample data sets U f :

[0011]

[0012] Among them, U f,1 is the result calculated according to the i-th transient boundary condition in the sample space, and the subscript f represents the fidelity of the calculation model used, hf is high-fidelity, and lf is low-fidelity. n f represents the data dimension of the model;

[0013] Step S3: Obtain the modal vector group and transient modal coefficient vector set of the high-fidelity and low-fidelity transient flow field data sets through the proper orthogonal decomposition method, and establish a transient modal coefficient vector sample set for training the LSTM neural network model. The process of obtaining the modal vector group Φ f and α f is as follows,

[0014] (Φ f ,αf ) = P(U f (Θ))

[0015] Among them, P(·) is the proper orthogonal decomposition process.

[0016] The modal vector group Φ f is the geometry of the transient data modal vectors, representing the flow field characteristics of the flow field data.

[0017]

[0018] Among them, is the j-th modal vector, rf represents the number of modal vectors retained by different fidelity modal vector groups, rh is the number of high-fidelity modal vectors, and rl is the number of low-fidelity modal vectors;

[0019] The modal coefficient vector set α f is the set of coefficients corresponding to the modal vectors.

[0020]

[0021] Among them, the modal coefficient vector group of the i-th transient data sample of the model fidelity f is

[0022]

[0023] Among them, is the coefficient vector of the time series of the k-th modal vector.

[0024]

[0025] Among them, is the modal vector coefficient of the t-th timestamp.

[0026] Deep autoencoders for low-fidelity and high-fidelity flow field data sets are established for low-fidelity and high-fidelity flow field data respectively. The autoencoder is divided into two parts: an encoder (Encoder) and a decoder (Decoder).

[0027] h f = f f (U f,L )

[0028] U f,L ≈ U′ f,L = g f (h f )

[0029] Among them, f f (·) and g f (·) are the encoder and the decoder respectively, and h fThe low-dimensional features extracted by the encoder in the deep autoencoder from the flow field data; U' f,L The approximation of the encoder input U reconstructed by the decoder in the deep autoencoder based on the low-dimensional feature h f , the result of f,L .

[0030] Step S4: Use the high- and low-fidelity transient modal coefficient vector sets to establish and train an LSTM neural network model to achieve the mapping between low- and high-fidelity transient modal coefficient vectors. Through the obtained high- and low-fidelity transient modal coefficient vector sets, use the low-fidelity modal vector coefficients boundary conditions from the (t - m)-th to the t-th time stamps in the i-th sample as the input, and the high-fidelity modal vector coefficients as the output, establish and train an LSTM neural network model to establish the mapping between low- and high-fidelity transient modal coefficient vectors under the same boundary conditions and at the same moment. The prediction process of the LSTM is expressed as

[0031]

[0032] where f t , i t , C t , o t , h t are respectively the "forget gate output", "input gate output", "cell state", "output weight", and "LSTM network output" in the "gate" structure of the LSTM network at time t, and are expressed as

[0033] f t = σ(W f · [h t-1 , x t + b f )

[0034] i t = σ(W i · [h t-1 , x t + b i )

[0035] C t = f t * C t-1 + i t * tanh(W C · [h t-1 , x t + b C )

[0036] o t = σ(W o · [ht-1 , x t + b o )

[0037] h t = o t * tanh(C t )

[0038] Among them, W and b are the weight matrix and the bias vector, and the subscripts f, i, C, and o represent the forget gate, the input gate, the cell state, and the output gate respectively. σ(·) and tanh(·) are the Sigmoid and Tanh activation functions respectively, and the mathematical symbol * represents the pointwise multiplication operation (element-wise addition or multiplication at the same position).

[0039] Step S5: Use the low-fidelity CFD transient calculation model obtained in S1 to perform calculations for the target operating conditions to obtain the low-fidelity transient flow field data of the target operating conditions

[0040]

[0041] Among them, τ′ is the number of transient timestamps of the target operating conditions.

[0042] Use the obtained low-fidelity modal vector group Φ lf Extract the modal coefficient vector group of the low-fidelity transient flow field data under the target operating conditions

[0043]

[0044]

[0045] Among them, is the modal coefficient vector at the t-th timestamp of the target operating conditions.

[0046] Step S6: Establish the input data set of the LSTM network model. Split the low-fidelity modal vector group into data sets with multiple timestamps in the interval of m

[0047]

[0048]

[0049] Among them, is the content in the input vector of the LSTM network at the current moment, carrying the information of the low-fidelity model.

[0050] Split the boundary condition time series of the target operating conditions into boundary condition sets with multiple timestamps in the interval of m

[0051]

[0052]

[0053] Among them, is the content in the input vector of the LSTM network at the current moment, carrying the information of the boundary conditions.

[0054] At the current moment t, and are subjected to a vector concatenation operation ([·,·], Concatenate operation) to obtain the input vector group x t ,

[0055]

[0056] Step S7: Predict the high-fidelity modal coefficients and reconstruct the high-fidelity transient flow field data. Input x t into the LSTM model obtained in S4 to predict the high-fidelity modal vector group of the target working condition at the current moment t The calculation process is as follows:

[0057]

[0058] Utilize the high-fidelity modal vector group Φ hf obtained in S3 and the predicted high-fidelity modal vector group of the target working condition to reconstruct the flow field data of the target working condition approximating the calculation result of the high-fidelity CFD transient calculation model The reconstruction calculation process is as follows:

[0059]

[0060] Among them, is the calculation result of the high-fidelity CFD transient calculation model at the current moment t.

[0061] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a data-driven core transient CFD flow field reconstruction technology, which reconstructs the calculation result of the low-fidelity CFD transient model to obtain flow field data that approximates the calculation result of the high-fidelity CFD model within a certain accuracy, and can reduce the calculation cost and calculation burden required in transient calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0063] Figure 1 It is the flowchart of the method of the present invention;

[0064] Figure 2 It is the schematic diagram of the 1×2 rod bundle model and measuring points of the present invention;

[0065] Figure 3 It is the comparison diagram of the high-fidelity grid and low-fidelity grid models of the present invention;

[0066] Figure 4 It is the axial velocity prediction results of the step-by-step and parallel schemes of the present invention at the measuring point P3;

[0067] Figure 5 It is the temperature prediction results of the step-by-step and parallel schemes of the present invention at the measuring point P3. Specific embodiments

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0069] Step S1: For the problem of reconstructing the core transient flow field data of this embodiment, a 1×2 sub-channel model is intercepted from the geometry of the fuel rod bundle in the pressurized water reactor core, and the embodiment of the method is carried out. Figure 2 It is the intercepted 1×2 sub-channel model.

[0070] First, according to the required accuracy and flow field resolution, a suitable grid model is selected through grid independence analysis to establish a high-fidelity CFD calculation model. The grid density of this model is relatively high, and it can accurately capture the flow and heat transfer characteristics in the flow field, but the calculation cost of this model is high and the calculation burden is large.

[0071] Secondly, based on the established high-fidelity grid model, the grid model is coarsened to obtain a coarse grid model, and a low-fidelity CFD transient calculation model is established. Since only the low-fidelity CFD transient calculation model is used in the prediction process, the coarsening degree of the coarse model can be determined according to the calculation time requirement and calculation accuracy requirement of the target object. In this example, the grid models of the CFD transient calculation models with different fidelities are compared as Figure 3 shown. The number of grids in the high-fidelity CFD transient calculation model is about 1.03 million, and the number of grids in the low-fidelity calculation model is about 0.11 million. The high- and low-fidelity models adopt the same periodic boundary conditions, SST turbulence model and calculation convergence criterion (<10 -4 ).

[0072] Step S2: The present invention selects the transient starting state and step size to describe different step transient processes. A total of 5 groups of step transient processes are selected for training and testing. The CFD data of 3 groups of transient processes are selected as the training samples of the data reconstruction model, and the boundary conditions are the initial inlet flow velocity of 3 m / s, 5 m / s and 7 m / s, and the flow rate step sizes of -50%, -20% and -80% respectively. The other 2 groups are used as test samples, and the boundary conditions are the initial inlet flow velocity of 3 m / s and 7 m / s, and the flow rate step sizes are both -20%.

[0073] The coolant inlet temperature at steady state for all working conditions is 565.15 K, and the average heat flux density is 8.473×10 8 W / m 2 , and the operating pressure is set to 15.5 MPa. In the transient calculation, the channel inlet maintains the initial inlet flow velocity and is calculated until 20 s. At the 20th second, the inlet flow velocity is directly stepped to the preset value, and the other boundary conditions remain unchanged. After the inlet velocity step occurs, the CFD calculation uses a time step of 0.001 s, and a total of 1.0 s of transient process is simulated. The transient turbulent motion is described by the SST turbulence model of URANS, and the convergence residuals of each parameter within a single time step are set to 10 -4 , and the maximum number of iterations within a single time step is set to 40 times.

[0074] Using the high-fidelity and low-fidelity CFD transient calculation models obtained in step S1, calculations are performed according to the 3 transient boundary conditions in the sample space, and high-fidelity and low-fidelity transient calculation data sets U f are obtained respectively:

[0075]

[0076] Among them, is the result calculated according to the i-th transient boundary condition in the sample space, is the flow field data vector at the t-th moment. The subscript f represents the fidelity of the calculation model used, hf for high fidelity, lf for low fidelity, and n f represents the data dimension of the model;

[0077] Step S3: Obtain the modal vector group and transient modal coefficient vector set of the high-fidelity and low-fidelity transient flow field data sets through the proper orthogonal decomposition method, and establish a transient modal coefficient vector sample set for training the LSTM neural network model. Perform proper orthogonal decomposition to obtain the modal vector group Φ f and α f The process is as follows:

[0078] First, obtain the correlation matrix R, expressed as

[0079]

[0080] The eigenvalues λ f and eigenvectors of the correlation matrix R can be calculated according to the following formula:

[0081] R f Ψ f = λΨ f

[0082] Based on the above eigenvalues and eigenvectors, the modal vector α f and the corresponding modal vector coefficients Φ f ,

[0083]

[0084]

[0085] Among them, the modal vector group is the geometry of the transient data modal vector, representing the flow field characteristics of the flow field data, is the j-th modal vector, and rf represents the number of modal vectors retained in the modal vector groups of different fidelities, rh is the number of high-fidelity modal vectors, and rl is the number of low-fidelity modal vectors;

[0086] The modal coefficient vector set is a set of coefficients corresponding to the modal vectors, is the modal coefficient vector group of the i-th transient data sample of the model fidelity f, is the coefficient vector of the time series of the k-th modal vector,

[0087]

[0088] Among them, is the modal vector coefficient at the t-th timestamp.

[0089] Step S4: Use the high- and low-fidelity modal coefficient vector sets α obtained in S3 f , and use the low-fidelity modal vector coefficients from the (t - m)-th to the t-th time stamps in the i-th sample Boundary conditions as a set of inputs, and use the high-fidelity modal vector coefficients as the corresponding outputs to establish a training sample library (x, y),

[0090] x = (x 1 , …, x L )

[0091] y = (y 1 , …, y L )

[0092]

[0093]

[0094]

[0095]

[0096] where [·, ·] is the vector concatenation operation (Concatenate operation), and in this embodiment, m is 10.

[0097] Build an LSTM neural network model, which realizes the mapping between low- and high-fidelity transient modal coefficient vectors under the same boundary conditions and at the same moment,

[0098]

[0099] where f t , i t , C t , o t , h t are respectively the "forget gate output", "input gate output", "cell state", "output weight", and "LSTM network output" in the "gate" structure of the LSTM network at time t, and are expressed as

[0100] f t = σ(W f · [h t-1 , x t + b f )

[0101] i t = σ(W i · [h t-1 , x t + bi )

[0102] C t = f t * C t-1 + i t * tanh(W C · [h t-1 , x t + b C )

[0103] o t = σ(W o · [h t-1 , x t + b o )

[0104] h t = o t * tanh(C t )

[0105] Among them, W and b are the weight matrix and bias vector. The subscripts f, i, C, and o represent the forget gate, input gate, cell state, and output gate respectively. σ(·) and tanh(·) are the Sigmoid and Tanh activation functions respectively. The mathematical symbol * represents the pointwise multiplication operation (element-wise addition or multiplication at the same position).

[0106] The LSTM model utilizes the established training sample library (x, y), and updates and corrects the weights and biases in the LSTM model through the Loss of the MSE loss function MSE

[0107]

[0108]

[0109]

[0110] Among them, η is the learning rate, which is set to 0.0005 in this embodiment. At the same time, in this embodiment, 5-fold cross-validation is adopted for training convergence determination. The convergence determination in this example is epoch > 2000 and Loss MSE <10 -4 , and the Adam optimizer is adopted.

[0111] Step S5: Use the low-fidelity CFD transient calculation model obtained in S1 to calculate for the transient boundary conditions of the verification working condition to obtain the low-fidelity transient flow field data of the target working condition

[0112] ​

[0113] Among them, τ′ is the number of transient timestamps of the target operating condition transient boundary condition .

[0114] Use the low-fidelity modal vector group Φ obtained in step S3 lf to perform an inner product on the target operating condition transient flow field data to obtain the modal coefficient vector group of the target operating condition transient flow field data

[0115]

[0116] Among them, is the transpose of Φ lf . is the modal coefficient vector at the t-th timestamp of the target operating condition

[0117] Step S6: Split the modal coefficient vector group of the target operating condition transient flow field data in step S5 into data sets with multiple timestamps in an interval of m

[0118]

[0119]

[0120] Among them, is the content in the input vector of the LSTM network at the current moment, carrying the information of the low-fidelity model

[0121] Split the boundary condition time series of the target operating condition into boundary condition sets with multiple timestamps in an interval of m

[0122]

[0123]

[0124] Among them, is the content in the input vector of the LSTM network at the current moment, carrying the information of the boundary condition

[0125] At the current t moment, and are subjected to a vector concatenation operation to obtain the input vector group of the LSTM model

[0126]

[0127] Step S7: The input vector group Input it into the LSTM model trained in step S4 to achieve the high-fidelity modal vector group of the target working condition at the current moment t The calculation process is as follows:

[0128]

[0129] Utilize the high-fidelity modal vector group Φ obtained in step S3 hf and the high-fidelity modal vector group of the predicted target working condition to reconstruct the flow field data of the target working condition that approximates the calculation result of the high-fidelity CFD transient calculation model The reconstruction calculation process is as follows:

[0130]

[0131] Among them, is the calculation result of the high-fidelity CFD transient calculation model at the current moment t

[0132] In particular, for the transient flow field data U f (Θ) and two data processing schemes of step-by-step and parallel can be carried out: The step-by-step processing scheme does not directly perform eigen-decomposition on the transient flow field data in steps S2 and S5, but performs proper orthogonal decomposition on the change amount of the time steps of the transient flow field data to obtain the modal vector group Φ f and α f The process is as follows,

[0133]

[0134]

[0135]

[0136] Among them, P(·) is the proper orthogonal decomposition process

[0137] The parallel scheme does not need to adjust the above steps S1 - S7 Figure 4 、 Figure 5 show the predicted trend charts of the axial velocity and temperature changes at the P3 measuring point in the step-by-step scheme and the parallel scheme in this embodiment. It can be seen from the verification results that good calculation results can be obtained in the transient working condition prediction using the above two schemes

[0138] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section. The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven CFD data reconstruction technology for core transient flow fields, characterized in that, the specific steps are as follows: For the target object, establish a high-fidelity CFD transient calculation model with a fine grid of the reactor core and a low-fidelity CFD transient calculation model with a coarse grid; Establish a boundary condition sample space according to the working conditions to be analyzed, and calculate using the high- and low-fidelity CFD models respectively based on the sample space to obtain high-fidelity and low-fidelity flow field data sets; Use the proper orthogonal decomposition method to extract features from the two transient data sets respectively, and obtain the mode vector groups and coefficient vector sets of the mode vectors (hereinafter referred to as mode coefficient vector sets) of the high-fidelity transient data set and the low-fidelity transient data set; Construct a Long Short Term Memory (LSTM) recurrent neural network model. The LSTM network model takes the low-precision transient mode coefficient vector as the input and the high-fidelity transient mode coefficient vector as the output, and trains and establishes a dynamic mapping relationship from low-fidelity to high-fidelity transient data in the low-dimensional space. Use the low-fidelity transient CFD model to calculate the low-fidelity transient flow field data for the target working condition. Use the mode vector group of the obtained low-precision transient data set to extract the transient mode coefficient vector set corresponding to the low-fidelity transient flow field data, and input this mode coefficient vector group into the established LSTM model to predict the mode coefficient vector of the high-fidelity transient flow field data; Reconstruct the high-fidelity transient flow field data under the target working condition through the predicted high-fidelity transient mode coefficient vector and the obtained high-fidelity mode vector group.

2. A data-driven CFD data reconstruction model for reactor core transient flow fields according to claim 1, characterized in that, Establish a high-fidelity transient CFD model and a low-fidelity transient CFD model for the core, and solve to obtain a training data set under the same boundary conditions (sample space). The specific steps are as follows: According to the geometry of the rod bundle channels in the pressurized water reactor core, establish a high-fidelity transient CFD model with a fine grid, and on the basis of this grid model, reduce the grid density to establish a low-fidelity transient CFD model with a coarse grid; Establish a sample space for transient boundary conditions, where θ i is the i-th transient boundary condition in the sample space, L is the number of samples in the sample space, s is the number of boundary conditions, and τ is the time dimension of the boundary conditions; Using the obtained high-fidelity and low-fidelity transient CFD models, calculations are performed according to the sample space, and high-fidelity and low-fidelity transient calculation sample data sets U are obtained respectively f : Among them, U f,1 is the result calculated according to the i-th transient boundary condition in the sample space. The subscript f represents the fidelity of the calculation model used, hf for high fidelity, and lf for low fidelity. n f represents the data dimension of the model.

3. A data-driven CFD data reconstruction model for reactor core transient flow fields according to claim 1, characterized in that, Obtain the mode vector groups and transient mode coefficient vector sets of the high-fidelity and low-fidelity transient flow field data sets through the proper orthogonal decomposition method. The specific steps are as follows: The proper orthogonal decomposition method is used to perform POD decomposition on the high- and low-fidelity transient flow field data sets respectively to obtain the modal vector group Φ f , Among them, is the j-th modal vector, rf represents the number of modal vectors retained in different fidelity modal vector groups, rh is the number of high-fidelity modal vectors, and rl is the number of low-fidelity modal vectors; Obtain the modal coefficient vector set α corresponding to the fidelity data by using the proper orthogonal decomposition method f , Among them, the mode coefficient vector group of the i-th transient data sample of the model fidelity f is, Among them, is the coefficient vector of the time series of the k-th modal vector, Among them, is the modal vector coefficient at the t-th timestamp.

4. A data-driven CFD data reconstruction model for reactor core transient flow fields according to claim 1, characterized in that, Use the high- and low-fidelity transient mode coefficient vector sets to establish and train an LSTM neural network model to realize the mapping between low- and high-fidelity transient mode coefficient vectors. The specific steps are as follows: With the obtained high- and low-fidelity transient modal coefficient vector sets, the low-fidelity modal vector coefficients from the (t - m)-th to the t-th time stamps in the i-th sample Boundary condition and the high-fidelity modal vector coefficient at the t-th time stamp As inputs, and the high-fidelity modal vector coefficient As the output, an LSTM neural network model is established and trained to establish the mapping between low- and high-fidelity transient modal coefficient vectors under the same boundary conditions and at the same moment. The prediction process of the LSTM is expressed as Among them, f t , i t , C t , o t , h t are respectively the "forget gate output", "input gate output", "cell state", "output weight" and "LSTM network output" in the "gate" structure of the LSTM network at time t, which are expressed as, f t = σ(W f · [h t-1 , x t + b f ) i t = σ(W i · [h t-1 , x t + b i ) C t = f t * C t-1 + i t * tanh(W C · [h t-1 , x t + b C ) o t = σ(W o · [h t-1 , x t + b o ) h t = o t *tanh(C t ) Among them, W and b are the weight matrix and the bias vector, and the subscripts f, i, C, and o represent the forget gate, the input gate, the cell state, and the output gate respectively. σ(·) and tanh(·) are the Sigmoid and Tanh activation functions respectively, and the mathematical symbol * represents the pointwise (element-by-element addition or multiplication) multiplication operation.

5. A data-driven CFD data reconstruction model for the transient flow field of a reactor core according to claim 1, characterized in that the specific steps for predicting the transient high-fidelity flow field approximate data under the target working conditions are as follows: Using the established low-fidelity CFD transient calculation model, perform calculations for the target working conditions to obtain the low-fidelity transient flow field data of the target working conditions Using the obtained low-fidelity modal vector group Φ lf Extract the modal coefficient vector group of the low-fidelity transient flow field data under the target working conditions Bring this low-fidelity modal vector group Into the obtained LSTM model to obtain the predicted high-fidelity modal vector group of the target working conditions The calculation process is shown as follows: Using the obtained high-fidelity modal vector group Φ hf and the high-fidelity modal vector group of the predicted target operating condition reconstruct the flow field data of the target operating condition that approximates the calculation results of the high-fidelity CFD transient calculation model The reconstruction calculation process is as follows: Among them, is the calculation result of the high-fidelity CFD transient calculation model.

Images