A method for reconstructing a core flow field of a large-scale pressurized water reactor

Abstract

The application discloses a kind of across scale pressurized water reactor core flow field reconstruction method, applied to core flow field reconstruction technical field, by the calculation result of low fidelity CFD model of large-scale drainage area is reconstructed, obtain the flow field data that in certain accuracy approximates the calculation result of high fidelity CFD model, the flow field data of large-scale drainage area is reconstructed using the above method, avoid directly establishing the high fidelity calculation model of large-scale drainage area and carrying out calculation.The calculation cost and the calculation burden of small-scale drainage area of a small amount of high, low fidelity CFD model are relatively small, and the calculation efficiency of low fidelity CFD model of large-scale drainage area is much higher than high fidelity CFD calculation model.At the same time, compared with the high fidelity CFD calculation process of large-scale drainage area, the time consumption proportion of POD decomposition process, TCA migration process and neural network training process is very small.Therefore, the above method can reduce the memory requirement needed in each calculation, reduce the time used for calculation.

Description

Technical Field

[0001] This invention relates to the field of core flow field reconstruction technology, and more specifically to a method for reconstructing the core flow field of a multi-scale pressurized water reactor. Background Technology

[0002] With the significant improvements in computer performance and computational model accuracy over the past few decades, CFD analysis technology has made great strides in accuracy and reliability. It provides researchers with high-fidelity coolant flow and heat transfer data and is gradually becoming an important component supporting the digital and intelligent development of nuclear power technology. However, the complex structure and cross-scale flow phenomena within the reactor core necessitate large meshes and complex models for CFD calculations. The time-consuming process of solving high-dimensional partial differential equations (PDEs) in the calculations presents technical challenges such as high computational cost, long computation time, and heavy computational burden when applying CFD technology in large flow domains within the reactor core.

[0003] Currently, there are few studies on high-fidelity CFD calculations based on large-scale flow domains within the reactor core. Among them, B. Liu et al. from the University of Sheffield, UK, published "Sub-channel CFD for nuclear fuel bundles" in Volume 355, Nuclear Engineering and Design, 2019, constructing a dual-grid model by embedding a fine grid within a coarse grid model. The coarse grid is used in large-scale regions of the reactor core, while a fine grid is used to simulate important local component regions, achieving calculations of large-scale flow domains within the reactor core. However, this method still uses a coarse grid in the large-scale flow domain area of ​​the reactor core, failing to obtain refined core calculation results. In contrast, Chinese scholars Chen Guangliang et al., published "Distributed-parallel CFD computation for all fuel assemblies in PWR core" in Volume 141, Annals of Nuclear Energy, 2020, based on the repetitive structure and periodic flow characteristics of the reactor core. They delineated multiple important sub-regions from the overall core region, used a distributed parallel computing scheme to perform refined CFD calculations on the important sub-regions, and employed extrapolation methods for non-important sub-regions, completing refined calculations of large-scale flow domains within the reactor core. However, this method calculates all sub-regions independently, lacking consideration for the overall core effect. Furthermore, the sub-regions represent important areas within the core, while non-critical areas are greatly simplified. Additionally, this scheme requires enormous computational resources and relies on large-scale computing clusters for implementation. It is evident that high-fidelity CFD calculations or data acquisition methods for large-scale flow domains within the reactor core are still in the exploratory stage.

[0004] Data-driven flow field reconstruction methods can utilize data to obtain the relationship between model inputs and outputs, avoiding the PDEs solution process in CFD calculations and achieving rapid computational response. This efficient computational approach provides a very attractive alternative for problems requiring rapid high-fidelity flow field reconstruction and key parameter prediction using CFD technology. However, high-fidelity CFD computation in large-scale flow fields within reactor cores is very costly, and it is difficult to collect enough data to model traditional data-driven flow field reconstruction methods. To address this issue, this method leverages the similarity in geometric structural features across flow fields of different sizes, proposing a "cross-scale" flow field reconstruction scheme that extracts flow field features from flow field data of small-scale flow fields for reconstruction of flow field data in larger-scale flow fields. Summary of the Invention

[0005] In view of this, the present invention provides a method for reconstructing the core flow field of a multi-scale pressurized water reactor to solve the problems existing in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Step 1: Based on the structural characteristics of the pressurized water reactor core rod bundle channels, different characteristic channels are identified. The model is divided into combinations of characteristic channels such as center channels, side channels, and corner channels.

[0008] After dividing the feature channels, basic grids for various feature channels, both high-fidelity and low-fidelity, are established. The high-fidelity grid is a suitable grid model selected through grid independence analysis, while the low-fidelity grid is a coarser, discretized grid based on the high-fidelity grid model, with its coarsening determined according to the user's acceptable computational cost and burden. For the source domain (small-scale watershed), high-fidelity and low-fidelity CFD models are established by copying and stitching the high-fidelity and low-fidelity basic grids. The overall grid model is established by copying and stitching the basic grids for the central channel, side channels, and corner channels.

[0009] Step 2: Establish the sample space for the boundary conditions of the source domain.

[0010] Θ=(Θ1,Θ2,...,Θ L )∈R s×L

[0011] Where, Θ i Let be the i-th boundary condition in the sample space of the source domain, L be the number of samples in the sample space, and s be the dimension of the boundary condition. The sample space of the source domain should be as close as possible to or contain the boundary conditions of the target domain to be computed.

[0012] We use high-fidelity CFD models and low-fidelity CFD models of the source domain to perform calculations based on the sample space of the source domain, and obtain high-fidelity and low-fidelity calculation data sample sets of the source domain, respectively.

[0013]

[0014] Where x represents the spatial discrete information of the overall model. To compute data for a vectorized model, n f The data dimension represents the computational model, and the subscript f represents the fidelity of the computational model, with hf for high fidelity and lf for low fidelity.

[0015] The high-fidelity and low-fidelity computational data of the source domain are segmented according to feature channels, and feature channel data of the same type are combined into corresponding high-fidelity and low-fidelity feature channel datasets.

[0016]

[0017] In this context, the subscript type represents the feature channel, ce is the center channel, si is the side channel, and co is the corner channel. For the i-th overall model data sample, the data set of the type-class feature channels after segmentation. The data vector for the feature channels. Let f be the number of feature channels in the source domain S. All feature channels of the same type and fidelity (f) are built from the same base grid and therefore have the same data dimension n. f,type .

[0018] Step 3: Use Proper Orthogonal Decomposition (POD) to decompose the source domain feature channel datasets of different types and fidelity obtained in Step 2 to obtain the POD basis vector set.

[0019]

[0020] in, is the k-th POD basis vector of the type feature channel with fidelity f, also known as the POD basis or mode. The superscript rf represents the number of POD basis vectors retained, rh is the number of high-fidelity POD basis vectors, and rl is the number of low-fidelity POD basis vectors.

[0021] Simultaneously, the coefficient vector set of the POD basis can be obtained through POD decomposition.

[0022]

[0023] Wherein, the POD basis coefficient matrix of the type feature channel of the f-fidelity of the i-th model sample data is:

[0024]

[0025] in, Let be the coefficient vector of the POD basis for the j-th feature channel data. This corresponds to the k-th coefficient in the coefficient vector.

[0026] Using the obtained POD basis vectors and vector coefficient set, the data sample of the j-th feature channel of the type class with f-fidelity for the i-th sample can be represented as follows:

[0027]

[0028] Step 4: By copying and stitching the low-fidelity base mesh from Step 1, a low-fidelity CFD model of the target domain (large-scale watershed) is established, and calculations are performed to obtain the overall calculation results of the low-fidelity model of the target domain.

[0029]

[0030] Where, x′ lf Let Θ′ represent the spatial discrete information of the low-fidelity model of the target domain, and let Θ′ represent the boundary conditions of the target domain. lf The data dimension of the low-fidelity model for the target domain.

[0031] The overall calculation result u′ lf (x′ lf ;Θ′) The overall computational data is segmented according to the feature channels to obtain low-fidelity feature channel data of each type of feature channel.

[0032]

[0033] Where type is the feature channel type, ce is the center channel, si is the side channel, and co is the corner channel; Let j represent the number of feature channels of type T in the target domain T, where j represents the data's index in that type of feature channel, and its value ranges from 1 to 1.

[0034] Using the low-fidelity POD basis vector set of the source domain in step 3 Obtain the coefficient vector set of the low-fidelity feature channels of the target domain. The calculation process is shown in the following formula.

[0035] α′ lf,type,j (Θ′)=Φ lf,type (x lf,type )T u′ lf,type,j (x′ lf,type ;Θ′)

[0036] in, For the type class of the low-fidelity CFD model in the target domain, there is the coefficient vector of the POD basis for the j-th feature channel.

[0037] Step 5: Use Transfer Component Analysis (TCA) to perform a nonlinear transformation on the low-fidelity POD basis vector coefficient set of the source domain obtained in Step 2 and the low-fidelity POD basis vector coefficient set of the target domain obtained in Step 4.

[0038] W TCA,type =F TCA,type (α′ lf,type (Θ′),α f,type (Θ))

[0039]

[0040]

[0041] Among them, W TCA For the TCA method at α′ lf,type (Θ′) and α f,type The transition matrix F obtained from (Θ) TCA (·) represents the TCA calculation process, K TCA (·) represents the vector coefficients based on the transition matrix W. TCA The migration process of kernel function methods. This is a set of low-fidelity coefficient vectors for the type class of the target domain after TCA migration. Let be the coefficient vector of the POD basis for the j-th feature channel.

[0042] Let be the set of low-fidelity coefficient vectors of the type class of the i-th sample in the source domain after TCA transfer. Let be the coefficient vector of the POD basis for the j-th feature channel.

[0043] Step 6: Using the transferred source domain coefficient vector set obtained in Step 5, use low-fidelity coefficient vectors... As input, high fidelity coefficient As output, a backpropagation neural network (BPNN) is trained to establish a mapping relationship between high-fidelity and low-fidelity coefficients under the same boundary conditions. The prediction process of BPNN can be represented as:

[0044]

[0045] Among them, F type,j (·) represents the BPNN neural network model, W type,j and b type,j Here are the weight matrix and bias vectors of the neural network. The subscript j represents the index of the high-fidelity basis vector coefficients, k = 1, ..., rh. To improve the computational accuracy of the neural network model, a non-shared weight approach is used to build the neural network model. Since all feature channels of type have rh POD basis vector coefficients that need to be predicted, type × rh BPNN models need to be built.

[0046] Step 7: Place α′ lf,type,j (Θ′) is used as input and substituted into the neural network model of the corresponding type established in step 6 to predict the high-fidelity coefficient vector of the target domain. The prediction process is shown below.

[0047]

[0048] Step 8: Obtain the high-fidelity coefficient vector α′ of the target domain from Step 7. hf,type,j (Θ′) and the high-fidelity POD basis vector set in step 3 The target domain flow field data can be reconstructed from the calculation results of an approximately high-fidelity CFD model. The reconstruction method is as follows:

[0049]

[0050] in, For the high-fidelity flow field data of the j-th feature channel of type class in the reconstructed target domain flow field.

[0051] Overall target domain flow field data u′ hf (x′ hf,type ;Θ′) can be obtained by stitching together the high-fidelity flow field data of the reconstructed feature channels.

[0052]

[0053] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method for reconstructing the core flow field of a multi-scale pressurized water reactor. By reconstructing the calculation results of a low-fidelity CFD model of a large-scale flow domain, flow field data that approximates the calculation results of a high-fidelity CFD model within a certain accuracy can be obtained, which can reduce the computational cost and computational burden required in large-scale flow domain calculations.

[0054] The method described above reconstructs flow field data for large-scale watersheds, avoiding the need to directly build high-fidelity computational models for such large-scale watersheds. The computational cost and burden of high- and low-fidelity CFD models for a small number of small-scale watersheds are relatively low, and the computational efficiency of low-fidelity CFD models for large-scale watersheds is significantly higher than that of high-fidelity CFD models. Furthermore, compared to the high-fidelity CFD computation process for large-scale watersheds, the time consumed by the POD decomposition process, TCA transfer process, and neural network training process is minimal. Therefore, this method reduces the memory requirements for each computation and lowers the computation time. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2a This is a schematic diagram of the 5×5 rod bundle flow domain of the source domain of the present invention;

[0058] Figure 2b This is a schematic diagram of the 17×17 rod bundle flow domain of the target domain of the present invention;

[0059] Figure 3a This is a schematic diagram of the characteristic channels of the source domain of the present invention;

[0060] Figure 3b This is a schematic diagram of the characteristic channels of the target domain of the present invention;

[0061] Figure 4a This is the basic mesh model for the high-fidelity feature channels of the present invention;

[0062] Figure 4b This is the basic mesh model for the low-fidelity feature channels of the present invention;

[0063] Figure 5a This is the overall high-fidelity mesh model of the source domain of the present invention;

[0064] Figure 5b This is the overall low-fidelity mesh model of the source domain of the present invention;

[0065] Figure 6 This is the overall low-fidelity mesh model of the target domain of the present invention;

[0066] Figure 7aThis is a contour plot of the velocity u of the reconstructed target domain of the present invention;

[0067] Figure 7b A contour plot of the velocity v of the reconstructed target domain in this invention;

[0068] Figure 7c A contour plot of the velocity w of the reconstructed target domain in this invention;

[0069] Figure 8a The error contour plot is the velocity u of the target domain flow field of this invention and the calculation of the high-fidelity CFD model.

[0070] Figure 8b The error contour plot is calculated by the high-fidelity CFD model for the velocity v of the target domain flow field in this invention.

[0071] Figure 8c This is an error contour plot of the velocity w in the target domain flow field of this invention and the calculation of the high-fidelity CFD model. Detailed Implementation

[0072] 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.

[0073] Step 1: For the cross-scale pressurized water reactor core flow field reconstruction problem in this embodiment, the 17×17 rod bundle flow field in the pressurized water reactor core is selected as the large-scale flow field to be predicted, and the 5×5 rod bundle flow field in the pressurized water reactor core is selected as the small-scale flow field to provide data, and the method is implemented. Figure 2a The target domain (large-scale watershed) is a 17×17 rod bundle watershed; Figure 2b It is a 5×5 rod bundle watershed in the source region (small-scale watershed).

[0074] See Figure 3a and Figure 3b The source and target domains are divided into a combination of feature channels such as center channel, side channel, and corner channel.

[0075] After dividing the feature channels, basic meshes for various feature channels at both high fidelity and low fidelity are established. The high-fidelity mesh is a suitable mesh model selected through mesh independence analysis, while the low-fidelity mesh is a coarser, discretized mesh based on the high-fidelity mesh model. In this embodiment, the basic meshes for each type of feature channel at different fidelities are as follows: Figure 4a and Figure 4b As shown.

[0076] For the source domain, high-fidelity CFD models and low-fidelity CFD models were established by copying and stitching the high-fidelity and low-fidelity base meshes, respectively. The overall CFD mesh model of the source domain is as follows: Figures 5a-5b As shown in the figure, the number of high-fidelity grids is 12,192,000; the number of low-fidelity grids is 1,523,000; and the ratio of high-fidelity to low-fidelity grids is 8.

[0077] Step 2: In this embodiment, the inlet velocity boundary condition of the target domain is 6.5 m / s. Based on the sample space of the boundary conditions of the target domain and the boundary conditions of the source domain, the inlet velocity sampling space of the source domain is 6.0 m / s to 7.0 m / s, with a sampling interval of 0.5 m / s, for a total of 3 sets.

[0078] Θ=(Θ1=6.0, Θ2=6.5, Θ3=7.0)∈R 1×L

[0079] Where, Θ i Let L be the i-th boundary condition in the sample space of the source domain, and L be the number of samples. In this example, L equals 3.

[0080] High-fidelity and low-fidelity CFD models of the source domain are used to perform calculations based on the sample space of the source domain, and high-fidelity and low-fidelity calculation data sample sets of the source domain are obtained respectively.

[0081]

[0082] Here, x represents the spatial discrete information of the overall model. To compute data for the vectorized model, n f This represents the data dimension of the computational model. The subscript f represents the fidelity of the computational model, with hf for high fidelity and lf for low fidelity. In the embodiment, n hf Equals 12,192,000, n lf It equals 1,523,000.

[0083] The high-fidelity and low-fidelity computational data of the source domain are segmented according to feature channels, and feature channel data of the same type are combined into corresponding high-fidelity and low-fidelity feature channel datasets.

[0084]

[0085] In this context, the subscript type represents the feature channel, ce is the center channel, si is the side channel, and co is the corner channel.

[0086] For the i-th overall model data sample, the data set of the type-class feature channels after segmentation. The data vector for the feature channels. Let f be the number of feature channels in the source domain S. All feature channels of the same type and fidelity (f) are built from the same base grid and therefore have the same data dimension n. f,type In the embodiment, n hf,ce Equals 550,000, n hf,si Equals 176,000, n hf,co Equals 144,000; n lf,ce Equals 68,750, n lf,si Equals 22,000, n lf,co It equals 18,000; Equals 16, Equals 16, It equals 4.

[0087] Step 3: Use Proper Orthogonal Decomposition (POD) to decompose the source domain feature channel datasets of different types and fidelity obtained in Step 2 to obtain the POD basis vector set.

[0088]

[0089] in, Let rf be the k-th POD basis vector of the type feature channel with fidelity f. It can also be called a POD basis or mode. The superscript rf represents the number of retained POD basis vectors, rh is the number of high-fidelity POD basis vectors, and rl is the number of low-fidelity POD basis vectors. In the example, rh equals 10 and rl equals 10.

[0090] Simultaneously, the coefficient vector set of the POD basis can be obtained through POD decomposition.

[0091]

[0092] Wherein, the POD basis coefficient matrix of the type feature channel of the f-fidelity of the i-th model sample data is:

[0093]

[0094] in, Let be the coefficient vector of the POD basis for the j-th feature channel data. This corresponds to the k-th coefficient in the coefficient vector.

[0095] Using the obtained POD basis vectors and vector coefficient set, the data sample of the j-th feature channel of the type class with f-fidelity for the i-th sample can be represented as follows:

[0096]

[0097] Step 4: By copying and stitching the low-fidelity base mesh from Step 1, a low-fidelity CFD mesh model of the target domain (large-scale watershed) is established. The low-fidelity mesh model of the target domain is as follows: Figure 6 As shown.

[0098] And perform calculations to obtain the overall calculation results of the low-fidelity model of the target domain.

[0099]

[0100] Where, x′ lf Let Θ′ represent the spatial discrete information of the low-fidelity model of the target domain, and let Θ′ represent the boundary conditions of the target domain. lf n′ represents the data dimension of the low-fidelity model for the target domain. In the example, n′ lf It equals 19,080,000.

[0101] The overall calculation result u′ lf (x′ lf ;Θ′) The overall computational data is segmented according to the feature channels to obtain low-fidelity feature channel data of each type of feature channel.

[0102]

[0103] Where type is the feature channel type, ce is the center channel, si is the side channel, and co is the corner channel; Let j represent the number of feature channels of type T in the target domain T, where j represents the data's index in that type of feature channel, and its value ranges from 1 to 1. In the embodiments, Equals 256, Equals 64, It equals 4.

[0104] Using the low-fidelity POD basis vector set of the source domain in step 3 Obtain the coefficient vector set of the low-fidelity feature channels of the target domain. The calculation process is shown in the following formula.

[0105] α′ lf,type,j (Θ′)=Φ lf,type (x lf,type ) T u′ lf,type,j (x′ lf,type ;Θ′)

[0106] in, For the type class of the target domain, the coefficient vector of the POD basis for the j-th feature channel.

[0107] Step 5: Use Transfer Component Analysis (TCA) to perform a nonlinear transformation on the low-fidelity POD basis vector coefficient set of the source domain obtained in Step 2 and the low-fidelity POD basis vector coefficient set of the target domain obtained in Step 4.

[0108] W TCA,type =F TCA,type (α′ lf,type (Θ′),α f,type (Θ))

[0109]

[0110]

[0111] Among them, W TCA For the TCA method at α′ lf,type (Θ′) and α f,type The transition matrix F obtained from (Θ) TCA (·) represents the TCA calculation process, K TCA (·) represents the vector coefficients based on the transition matrix W. TCA The migration process of kernel function methods. This is a set of low-fidelity coefficient vectors for the type class of the target domain after TCA migration. Let be the coefficient vector of the POD basis for the j-th feature channel.

[0112] Let be the set of low-fidelity coefficient vectors of the type class of the i-th sample in the source domain after TCA transfer. Let be the coefficient vector of the POD basis for the j-th feature channel.

[0113] Step 6: Using the transferred source domain coefficient vector set obtained in Step 5, use low-fidelity coefficient vectors... As input, high fidelity coefficient As output, a backpropagation neural network (BPNN) is trained to establish a mapping relationship between high-fidelity and low-fidelity coefficients under the same boundary conditions. The prediction process of BPNN can be represented as:

[0114]

[0115] Among them, F type,j (·) represents the BPNN neural network model, W type,j and b type,jHere are the weight matrix and bias vectors of the neural network. The subscript j represents the index of the high-fidelity basis vector coefficients, k = 1, ..., rh. To improve the computational accuracy of the neural network model, a non-shared weight approach is used to build the neural network model. Since all feature channels of type have rh POD basis vector coefficients that need to be predicted, type × rh BPNN models need to be built.

[0116] Step 7: Place α′ lf,type,j (Θ′) is used as input and substituted into the neural network model of the corresponding type established in step 6 to predict the high-fidelity coefficient vector of the target domain. The prediction process is shown below.

[0117]

[0118] Step 8: Obtain the high-fidelity coefficient vector α′ of the target domain from Step 7. hf,type,j (Θ′) and the high-fidelity POD basis vector set in step 3 The target domain flow field data can be reconstructed from the calculation results of the approximate high-fidelity calculation model. The reconstruction method is as follows:

[0119]

[0120] in, For the high-fidelity flow field data of the j-th feature channel of type class in the reconstructed target domain flow field.

[0121] Overall target domain flow field data u′ hf (x′ hf,type ;Θ′) can be obtained by stitching together the high-fidelity flow field data of the reconstructed feature channels.

[0122]

[0123] The reconstructed overall target domain flow field can be calculated using the above method. The velocity contours u, v, and w of the reconstructed overall target domain flow field are shown below. Figure 7a , Figure 7b , Figure 7c As shown. The error contour plots of the reconstructed overall target domain flow field velocities u, v, and w compared to those calculated by the high-fidelity CFD model are shown below. Figure 8a , Figure 8b , Figure 8c As shown.

[0124] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0125] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reconstructing the core flow field of a multi-scale pressurized water reactor, characterized in that, The specific steps are as follows: Establish high-fidelity CFD models and low-fidelity CFD models of the source domain; The sample space of the source domain boundary conditions is determined based on the boundary conditions of the target domain, and the high-fidelity and low-fidelity feature channel datasets of the source domain are obtained respectively. The source domain feature channel dataset is decomposed using the intrinsic orthogonal decomposition method to obtain the source domain POD basis vector set and the POD basis coefficient vector set; Construct a low-fidelity CFD model of the target domain and obtain a low-fidelity feature channel dataset of the target domain. Use the POD basis vector set of the source domain to obtain the coefficient vector set of the POD basis of the target domain. The migration component analysis method is used to perform nonlinear transformation on the coefficient vector set of the low-fidelity POD basis of the source domain and the coefficient vector set of the low-fidelity POD basis of the target domain. Using the coefficient vector set of the low-fidelity POD basis of the source domain after transfer as input and the coefficient vector set of the high-fidelity POD basis as output, a backpropagation neural network is trained to establish the mapping relationship between the high-fidelity coefficient vector and the low-fidelity coefficient vector under the same boundary condition. The coefficients of the low-fidelity POD basis vector of the target domain are used as input for prediction to obtain the coefficient vector of the high-fidelity POD basis of the target domain. The target domain flow field data are reconstructed by using the coefficient vectors of the high-fidelity POD basis of the target domain and the high-fidelity POD basis vector set of the source domain.

2. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 1, characterized in that, The specific steps for establishing high-fidelity CFD models and low-fidelity CFD models of the source domain are as follows: Based on the structural characteristics of the core rod bundle channels of pressurized water reactors, different characteristic channels are divided, including center channels, side channels and corner channels; After dividing the feature channels, basic grids for various feature channels of high fidelity and low fidelity are established respectively. For the source domain, high fidelity CFD models and low fidelity CFD models of the source domain are established by copying and splicing the high fidelity basic grids and low fidelity basic grids respectively.

3. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 2, characterized in that, The specific steps to obtain the source domain feature channel dataset are as follows: Establish the sample space of the boundary conditions of the source domain. in, For the sample space of the source domain, the first i One boundary condition L Let be the number of samples in the sample space. s The dimension of the boundary conditions; Using high-fidelity CFD models and low-fidelity CFD models in the source domain, calculations are performed based on the sample space of the source domain, and high-fidelity and low-fidelity calculation data sample sets in the source domain are obtained respectively: in, It represents the spatial discrete information of the overall model. Calculate data for the vectorized model. The subscript represents the data dimension of the computational model. f Represents the fidelity of the computational model. hf For high fidelity, lf For low fidelity; The high-fidelity and low-fidelity computational data of the source domain are segmented according to feature channels, and feature channel data of the same type are combined to form corresponding feature channel datasets for high-fidelity CFD models and low-fidelity CFD models. Among them, subscript type Representative feature channels, ce As the central passage, si For side passage, co Corner channel; For the first i After the overall model data samples are segmented type Data sets of class feature channels, The data vector for the feature channels. For source domain S The number of feature channels in the middle; j This is the feature channel number.

4. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 3, characterized in that, The specific steps to obtain the POD basis vector set and the coefficient vector set of the POD basis in the source domain are as follows: The intrinsic orthogonal decomposition method is used to decompose the computational data sample sets of different types of feature channels with different fidelities in the source domain to obtain the POD basis vector set: in, for f fidelity type The first feature channel of the class k POD basis vectors, superscript rf This represents the number of POD basis vectors retained. rh To achieve a high-fidelity POD base quantity, rl For low-fidelity POD base quantity; The numbers are the coefficients of the high-fidelity basis vectors; Obtain the coefficient vector set of the POD basis through POD decomposition: Among them, the i Model sample data f fidelity type The POD basis coefficient matrix of the feature channel is as follows in, For the first j The coefficient vector of the POD basis for each feature channel data. For the first in the corresponding coefficient vector k One coefficient; Using the obtained POD basis vectors and vector coefficients, the first i one sample f fidelity type The class of j The data samples for each feature channel are represented as follows: 。 5. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 4, characterized in that, The specific steps for obtaining the coefficient vector set of the POD basis in the target domain using the source domain POD basis vector set are as follows: By copying and stitching the low-fidelity base mesh, a low-fidelity CFD model of the target domain is established, and the overall calculation results of the low-fidelity CFD model of the target domain are obtained through calculation: in, For the spatial discrete information of the low-fidelity model of the target domain, For the boundary conditions of the target domain, For the data dimensions of the low-fidelity model of the target domain; The overall calculation results The overall computational data is segmented based on feature channels to obtain low-fidelity computational data for each type of feature channel: in, type For characteristic channel type, ce As the central passage, si For side passage, co Corner channel; For the target domain T In type The number of feature channels; Utilizing the low-fidelity POD basis vector set of the source domain Obtain the coefficient vector set of the basis of the low-fidelity feature channels of the target domain. The calculation process is shown in the following formula: in, For low-fidelity CFD models in the target domain type The class of j The coefficient vector of the POD basis for each feature channel; The data dimension of the low-fidelity model of the source domain.

6. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 5, characterized in that, The specific steps of the nonlinear transformation are as follows: The migration component analysis method is used to perform a nonlinear transformation on the obtained low-fidelity POD basis vector coefficients of the source domain and the obtained low-fidelity POD basis vector coefficients of the target domain. in, For the TCA method in and The migration matrix obtained from For the TCA calculation process, Representative vector coefficients are based on the transition matrix The migration process of kernel function methods; For the target domain after TCA migration type The coefficient vector set of the low-fidelity POD basis of the class. For the first j The coefficient vector of the POD basis for each feature channel; For the source domain after TCA migration, the first i one sample type The coefficient vector set of the low-fidelity POD basis of the class. For the first j The coefficient vector of the POD basis for each feature channel; For the coefficient vector set of the low-fidelity POD basis of the target domain; Let be the set of coefficient vectors of the low-fidelity POD basis of the source domain.

7. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 6, characterized in that, The mapping relationship between the high-fidelity coefficient vector and the low-fidelity coefficient vector under the same boundary conditions is expressed as: Using the obtained set of coefficient vectors from the low-fidelity POD basis of the transferred source domain, the low-fidelity coefficient vectors are... As input, the coefficient vector of the high-fidelity POD basis As output, the backpropagation neural network is trained to establish a mapping relationship between high-fidelity coefficient vectors and low-fidelity coefficient vectors under the same boundary conditions. The prediction process of the backpropagation neural network is represented as follows: in, This is a backpropagation neural network model. and Here are the weight matrix and bias vector of the backpropagation neural network; since all type The feature channels of the class all have rh The coefficients of each POD basis vector need to be predicted, and a system needs to be established. A backpropagation neural network model.

8. The method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 7, characterized in that, The specific steps for obtaining the coefficient vector set of the high-fidelity POD basis of the target domain are as follows: Obtain the coefficient vector set of the low-fidelity POD basis of the target domain... As input, substitute the corresponding established type In a backpropagation neural network model of a class, the coefficient vector for predicting high fidelity in the target domain. The prediction process is as follows: ； in, This is the weight matrix of the backpropagation neural network; This is the bias vector of the backpropagation neural network; This is a backpropagation neural network model.

9. A method for reconstructing the core flow field of a multi-scale pressurized water reactor according to claim 8, characterized in that, The specific steps for reconstructing the flow field data of the target domain are as follows: The high-fidelity coefficient vector obtained from the target domain High-fidelity POD basis vector set of source domain The target domain flow field data of the high-fidelity calculation model are reconstructed as follows: in, In the target domain flow field for reconstruction type The class of j High-fidelity flow field data for each characteristic channel; Overall target domain flow field data This can be obtained by stitching together the high-fidelity flow field data from the reconstructed feature channels. 。

Images