Compressor flow field model order reduction method and device

Abstract

The invention belongs to the technical field of nuclear power, and particularly relates to a compressor flow field model order reduction method and device. According to the method disclosed by the invention, data is obtained from a high-resolution numerical simulation result or a measurement result of a reactor compressor flow field, and main modals of original data are extracted by adopting intrinsic orthogonal decomposition (POD). And finally, taking the time-varying boundary condition as input, taking the variable data after order reduction as output, carrying out deep neural network training, and generating an order reduction model. Taking the time-varying boundary condition as input, outputting a predicted compressor flow field reduced-order variable by the reduced-order model, and finally reducing the reduced-order variable to obtain a compressor full-flow-field variable. According to the method and the system, rapid prediction of the three-dimensional flow field of the compressor in reactor system simulation can be realized, the simulation precision of the flow problem is greatly improved on the premise of ensuring the simulation efficiency of the reactor system, and timely and reliable data support is provided for research and development, operation and maintenance and the like of the reactor system.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear power, and particularly relates to a method and device for reducing the order of a compressor flow field model. Background Art

[0002] The accurate simulation of the three-dimensional flow field in a reactor is crucial for the analysis and evaluation of the overall performance of the reactor. Usually, the CFD (Computational Fluid Dynamics) method is used to carry out high-precision numerical simulations on the reactor core and various components of the system. The CFD method can very detailedly display the thermohydraulic characteristics of the fluid in the complex structure of the reactor. The compressor is a core component of the reactor system. Its flow field structure is complex and the physical properties change violently. It is one of the most difficult components in the reactor system for flow field simulation. In related technologies, the CFD operation scale for the compressor is very large, consuming huge amounts of computing resources and time costs.

[0003] Model order reduction is a common method in optimization design. Its essence is to perform a low-dimensional approximate description of a multi-dimensional physical process that changes with time. By capturing the system, the most optimized processing can be carried out, so as to achieve the effects of reducing the computational dimension, reducing the amount of calculation, and saving the calculation time. POD (Proper Orthogonal Decomposition) can screen out the most representative modes from many data snapshots at different times and different parameters to achieve system dimensionality reduction. The machine learning method is that the neural network model improves its own performance using empirical data, and realizes the classification, recognition, or prediction of complex problems through the trained neural network. The so-called deep neural network is a neural network model with many hidden layers. However, the huge variable data in the numerical calculation of the compressor flow field makes it difficult to carry out the neural network training process for compressor flow field prediction. One training often consumes a large amount of time and even fails to obtain a convergent model. How to improve the generation efficiency and application scope of the reduced-order model and ensure the prediction accuracy has become an urgent problem to be solved. Summary of the Invention

[0004] To overcome the problems existing in the related technologies, a method and device for reducing the order of a compressor flow field model are provided.

[0005] According to one aspect of the embodiments of the present disclosure, a method for reducing the order of a compressor flow field model is provided. The method includes:

[0006] Step 101, obtaining local flow field data of a reactor compressor, where the local flow field data of the reactor compressor includes the coordinates, velocity, and pressure of the local flow field of the compressor;

[0007] Step 102: After cleaning the obtained local compressor flow field data, organize and form a local compressor flow field matrix in the form of an instantaneous image matrix and save it;

[0008] Step 103: Use the POD decomposition to process the saved local compressor flow field matrix for data reduction to obtain reduced-order modes. Among the reduced-order modes, select the first k singular vectors as the main mode data according to the order of the singular values of the singular vectors from large to small, and the sum of the singular values of the first k singular vectors is greater than 99% of the total singular values;

[0009] Step 104: Use the preset time-varying boundary conditions as the input training set and the main mode data as the output test set. Use the training set to train the neural network model to obtain the trained neural network model. Use the test set to evaluate the trained neural network model, and continuously repeat the model training and model evaluation until the trained neural network model converges to obtain a reduced-order model;

[0010] Step 105: Input the unknown time-varying boundary conditions as excitation data into the trained reduced-order model and output the predicted first k main modes;

[0011] Step 106: Use the POD-based flow field reconstruction algorithm to process the first k main modes output in Step 105, restore to obtain the full compressor flow field variables in matrix form, and reconstruct the predicted compressor flow field.

[0012] In a possible implementation, in Step 103, use the POD decomposition to process the saved local compressor flow field matrix for data reduction to obtain reduced-order modes: corresponding to the flow field problem, decompose the velocity field u(x, t) of the flow field variables into the superposition of the average velocity u 0 and the pulsating velocity u′(x, t), as shown in Equation 1:

[0013] u(x, t) = u 0 + u′(x, t) Equation 1

[0014] In addition to changing in the time scale, the velocity field also contains changes in different spatial scales. As shown in Equation 2, decouple the pulsating velocity u′(x, t) in the time domain and the spatial domain:

[0015]

[0016] Substitute Equation 1 into Equation 2 to obtain Equation 3:

[0017]

[0018] where, a 0 (t) = 1, The pulsation of the velocity is expressed as a set of spatial functions independent of time According to the corresponding coefficient a n (t) varies with time and is the result of linear superposition in the end; set the spatial function has orthonormality on the calculated fluid domain Ω, as shown in Equation 4:

[0019]

[0020] Define the inner product operator R of any flow field variable in the fluid domain Ω, as shown in Equation 5:

[0021]

[0022] Determine the R statistic of all points in the flow field at time step m, and obtain the fluctuating kinetic energy E of the flow field at this moment T , Equation 6 can be obtained according to Equation 5:

[0023]

[0024] Substitute Equation 2 into 6 and simplify according to Equation 4 to obtain Equation 7:

[0025]

[0026] where, λ n is used to characterize the magnitude of the fluctuating kinetic energy contained in each eigen-space basis (mode); extract the orthonormal flow field basis from the existing flow field according to the magnitude of the fluctuating kinetic energy component; apply the inner product operator to all time points and all variable points, and obtain Equation 8 from the orthonormality and Equation 7:

[0027]

[0028] If the number of discrete elements in the flow field is m and the number of time steps is n, traverse all discrete elements and time nodes, and obtain Equation 9 from Equation 8:

[0029]

[0030] where, U ∈ R m×n , U ij = u′(x i , t j )(i = 1, …, m, j = 1, ……, n), W ∈ R m×m , W ij = ΔV i δ ij , ΔV i is the discrete element volume; is called the covariance matrix, and the corresponding λ and can be obtained through SVD nArrange in descending order to obtain a new sequence: λ n (λ 1 > λ 2 > λ 3 >…), and the bases corresponding to these terms are the POD bases.

[0031] In a possible implementation, the neural network model in step 104 is the pytorch neural network framework; the activation function is ReLU, the loss function is MSELoss, the AdamW optimization algorithm is used, and the initial learning rate is set to 0.005.

[0032] In a possible implementation, in step 104, a deep feedforward neural network is adopted, where the neural network framework is the pytorch framework, a 6-layer deep feedforward neural network is used, and the number of input and output neurons corresponds to all the reduced-order modes and boundary conditions of the variables of the learning object; the loss function is MSELoss, the activation function is ReLU, the number of training times is 5000; the network optimization algorithm is AdamW, and the initial learning rate is set to 0.005; the training environment uses GPU acceleration for training.

[0033] In a possible implementation, in step 103, k is 15.

[0034] According to another aspect of the embodiments of the present disclosure, a compressor flow field model reduction device is provided, and the device includes:

[0035] An acquisition module, configured to acquire local flow field data of a reactor compressor, where the local flow field data of the reactor compressor includes the coordinates, velocities, and pressures of the local flow field of the compressor;

[0036] A data processing module, configured to perform data cleaning on the acquired local flow field data of the compressor, and then organize and form a local flow field matrix of the compressor in the form of an instantaneous image matrix and save it;

[0037] A reduction processing module, configured to perform data reduction on the saved local flow field matrix of the compressor by using POD decomposition to obtain reduced-order modes, and select the first k-order singular vectors as the main mode data according to the order of the singular values of the singular vectors from large to small in the reduced-order modes, and the sum of the singular values of the first k-order singular vectors is greater than 99% of the total singular values;

[0038] A training and evaluation module, configured to use the preset time-varying boundary conditions as the input training set, use the main mode data as the output test set, train the neural network model with the training set to obtain a trained neural network model, evaluate the trained neural network model with the test set, and continuously repeat the model training and model evaluation until the trained neural network model converges to obtain a reduced-order model;

[0039] A prediction module, configured to input unknown time-varying boundary conditions as excitation data into a trained reduced-order model, and output the predicted first k dominant modes.

[0040] A reconstruction module, configured to process the first k dominant modes output in step 105 by using a POD-based flow field reconstruction algorithm, restore the full compressor flow field variables in matrix form, and reconstruct the predicted compressor flow field.

[0041] In a possible implementation, in the reduced-order processing module, the local flow field matrix of the compressor saved by POD decomposition is used for data reduction to obtain reduced-order modes. Corresponding to the flow field problem, the velocity field u(x, t) of the flow field variables is decomposed into the sum of the average velocity u 0 and the pulsating velocity u′(x, t), as shown in Equation 1:

[0042] u(x, t) = u 0 + u′(x, t) Equation 1

[0043] In addition to changing on the time scale, the velocity field also contains changes on different spatial scales. As shown in Equation 2, the pulsating velocity u′(x, t) is decoupled in the time domain and the spatial domain:

[0044]

[0045] Substituting Equation 1 into Equation 2 gives Equation 3:

[0046]

[0047] where a 0 (t) = 1, The pulsation of the velocity is expressed as a set of spatial functions independent of time changing in time according to the corresponding coefficient a n (t) and finally the result of linear superposition; it is assumed that the spatial function has orthonormality on the calculated fluid domain Ω, as shown in Equation 4:

[0048]

[0049] Define the inner product operator R of any flow field variable in the fluid domain Ω, as shown in Equation 5:

[0050]

[0051] Determine the R statistic of all points in the flow field at time step m, and obtain the pulsating kinetic energy E T of the flow field at this moment. According to Equation 5, Equation 6 can be obtained:

[0052]

[0053] Substituting Equation 2 into Equation 6 and simplifying according to Equation 4, we can obtain Equation 7:

[0054]

[0055] where λ n is used to characterize the magnitude of the pulsating kinetic energy contained in the basis (mode) of each feature space; the orthonormal flow field basis is extracted from the existing flow field according to the magnitude of the pulsating kinetic energy component; applying the inner product operator to all time points and all variable points, Equation 8 is obtained from the orthonormality and Equation 7:

[0056]

[0057] If the number of discrete elements in the flow field is m and the number of time steps is n, by traversing all discrete elements and time nodes, Equation 9 is obtained from Equation 8:

[0058]

[0059] where U ∈ R m×n , U ij = u′(x i , t j )(i = 1, …, m, j = 1, ……, n), W ∈ R m×m , W ij = ΔV i δ ij , ΔV i is the volume of the discrete element; is called the covariance matrix, and the corresponding λ and can be obtained through SVD. Sorting λ n in descending order to obtain a new sequence: λ n (λ 1 > λ 2 > λ 3 > …), and the basis corresponding to these terms is the POD basis.

[0060] In a possible implementation, the neural network model in the training and evaluation module is the pytorch neural network framework; the activation function is ReLU, the loss function is MSELoss, and the AdamW optimization algorithm is used, with the initial learning rate set to 0.005.

[0061] In a possible implementation, in the training and evaluation module, a deep feedforward neural network is adopted, where the neural network framework is the PyTorch framework, and a 6-layer deep feedforward neural network is used. The number of input and output neurons corresponds to all reduced-order modes and boundary conditions of the variables of the learning object; the loss function is MSELoss, the activation function is ReLU, and the number of training times is 5000; the network optimization algorithm is AdamW, and the initial learning rate is set to 0.005; the training environment uses GPU acceleration for training.

[0062] In a possible implementation, in the reduced-order processing module, k is 15.

[0063] According to another aspect of the embodiments of the present disclosure, a compressor flow field model reduction device is provided, and the device includes:

[0064] A processor;

[0065] A memory for storing instructions executable by the processor;

[0066] Wherein, the processor is configured to execute the above method.

[0067] According to another aspect of the embodiments of the present disclosure, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above method is implemented.

[0068] The beneficial effects of the present disclosure are as follows: The compressor flow field model reduction method provided by the present disclosure realizes using the main modes after proper orthogonal decomposition as the training data of the neural network, thereby effectively improving the generation efficiency and applicable range of the reduced-order model, ensuring the prediction accuracy, and realizing the rapid solution of the compressor flow field. Thus, under the premise of ensuring the system simulation efficiency, refined results of the local compressor three-dimensional flow field can be obtained, greatly improving the calculation accuracy of the flow problems in the reactor system simulation, reducing the time cost of the reactor system refined simulation, providing timely and reliable data support for the reactor system research and development, operation and maintenance, etc., and supporting the efficient research and development of the reactor system. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flowchart of a compressor flow field model reduction method shown according to an exemplary embodiment.

[0070] Figure 2 is a flowchart of prediction using a compressor flow field model shown according to an exemplary embodiment.

[0071] Figures 3a to 3d is a schematic diagram of the accuracy of directly reconstructing the flow field by POD with different truncation numbers.

[0072] Figure 4aComparison of the prediction results obtained by the method of the present disclosure and the velocity map of CFD at the z = 0 cross-section.

[0073] Figure 4b Comparison of the prediction results obtained by the method of the present disclosure and the pressure contour map of CFD at the z = 0 cross-section.

[0074] Figure 5a Statistical chart of the relative error of velocity prediction for quantitative analysis.

[0075] Figure 5b Statistical chart of the relative error of pressure value prediction for quantitative analysis.

[0076] Figure 6 Block diagram of a compressor flow field model reduction device shown according to an exemplary embodiment. Detailed implementation manners

[0077] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] Figure 1 Flowchart of a compressor flow field model reduction method shown according to an exemplary embodiment, Figure 2 Flowchart of prediction using a compressor flow field model shown according to an exemplary embodiment, as Figure 1 and Figure 2 shown, this method can be executed by a terminal device, where the terminal device can be a server, a desktop computer, etc., and the present disclosure embodiment does not limit the type of the terminal device. As Figure 1 shown, this method includes:

[0079] Step 101, obtain the local flow field data of the reactor compressor. For example, the local flow field data of the reactor compressor can be obtained through a high-resolution numerical simulation model of the reactor flow field (such as CFD); for another example, the flow field data of the compressor at different times can be obtained through experimental measurement. The local flow field data of the reactor compressor includes the coordinates, velocity, pressure, etc. of the local flow field of the compressor.

[0080] Step 102, after cleaning the obtained local flow field data of the compressor, organize and form a local flow field matrix of the compressor in the form of an instantaneous image matrix and save it.

[0081] Step 103, perform data reduction on the saved local flow field matrix of the compressor by using POD decomposition to obtain reduced-order modes. In the reduced-order modes, the first k singular vectors are selected as the main mode data according to the order of the singular values of the singular vectors from large to small, and the sum of the singular values of the first k singular vectors is greater than 99% of the total singular values.

[0082] Step 104: Use the preset time-varying boundary conditions as the input training set and the main modal data as the output test set. Train the neural network model using the training set to obtain the trained neural network model. Evaluate the trained neural network model using the test set, and continuously repeat the model training and model evaluation until the trained neural network model converges to obtain the reduced-order model.

[0083] In a possible implementation, the neural network model in Step 104 is the pytorch neural network framework; the activation function is ReLU, the loss function is MSELoss, the AdamW optimization algorithm is used, and the initial learning rate is set to 0.005.

[0084] Step 105: Input the unknown time-varying boundary conditions as excitation data into the trained reduced-order model, and output the predicted first k main modes.

[0085] Step 106: Process the first k main modes output in Step 105 using the POD flow field reconstruction algorithm to restore the full compressor flow field variables in matrix form and reconstruct the predicted compressor flow field.

[0086] As an example of this embodiment, the snapshot matrix of the present disclosure can be expressed as a data matrix for training constructed by the snapshot (snapshot) technique. The snapshot matrix is the spatial distribution value of the physical field at different times. For incompressible fluid problems, a snapshot matrix A of the velocity U in the first vector direction can be constructed U and a snapshot matrix A of the velocity V in the second vector direction V and a snapshot matrix A of the pressure P. P .

[0087]

[0088] Where L is the number of snapshot times or the number of snapshots. After finite element discretization, the degrees of freedom are M, then A ∈ R MxL , each column of A is called a snapshot, the number of columns of A is the number of snapshots, and M >> L.

[0089] In Step 103, the compressor local flow field matrix saved by POD decomposition is used for data reduction to obtain the reduced-order mode: corresponding to the flow field problem, in order to better focus on the relevant structures of the flow problem, the velocity field u(x,t) of the flow field variables can be decomposed into the sum of the average velocity u 0 and the pulsating velocity u′(x,t), as shown in Equation 1.

[0090] u(x,t) = u 0 + u′(x,t) Equation 1

[0091] In addition to changing on the time scale, the velocity field also contains changes on different spatial scales. Therefore, as shown in Equation 2, the fluctuating velocity u′(x,t) is decoupled in the time domain and the spatial domain.

[0092]

[0093] Substituting Equation 1 into Equation 2 gives Equation 3.

[0094]

[0095] where a 0 (t) = 1, The fluctuating quantity of velocity is expressed as a set of spatial functions independent of time changing in time according to the corresponding coefficient a n (t) and finally the result of linear superposition; for the convenience of mathematical form, it is set that the spatial function has orthonormality on the fluid domain Ω under calculation, as shown in Equation 4.

[0096]

[0097] Define the inner product operator R of any flow field variable in the fluid domain Ω, as shown in Equation 5:

[0098]

[0099] In the flow field, the magnitude of this statistic R is used to characterize the fluctuating kinetic energy of the compressor flow field. Therefore, by calculating the R statistic for all points in the flow field at time step m, the fluctuating kinetic energy E of the flow field at this moment can be obtained T , and according to Equation 5, Equation 6 can be obtained.

[0100]

[0101] Substituting Equation 2 into 6 and simplifying according to Equation 4 gives Equation 7.

[0102]

[0103] where λ n is used to characterize the magnitude of the fluctuating kinetic energy contained in each eigen - space basis (mode). Extract the orthonormal flow field basis from the existing flow field according to the magnitude of the fluctuating kinetic energy components. Apply the inner product operator to all time points and all variable points, and from the orthonormality and Equation 7, Equation 8 is obtained.

[0104]

[0105] If the number of discrete elements in the flow field is m and the number of time steps is n, by traversing all discrete elements and time nodes, Equation 9 is obtained from Equation 8.

[0106]

[0107] Among them, U∈R m×n , U ij =u′(x i ,t j )(i=1,…,m, j=1,…,n), W∈R m×m , W ij =ΔV i δ ij , ΔV i is the volume of discrete unit; It is called the covariance matrix, and the corresponding λ and λ n Arrange in descending order to get a new sequence: n (λ 1 >λ 2 >λ 3 >…), the basis corresponding to these terms That is the POD substrate.

[0108] In one possible implementation, the sum of the first 10% or even 1% of the singular values ​​accounts for more than 99% of the sum of all singular values. In other words, the largest k singular values ​​and the corresponding left and right singular vectors can be used to approximate the matrix, a process called POD truncation. The above process projects the high-dimensional complex fluid variable system onto a projection space consisting of a few POD reduced-order modes, and contains the main dynamic information in the original data.

[0109] See also Figure 1 In step 104, a deep feedforward neural network is used, the preset time-varying boundary conditions are used as the input training set, the k-order main modal data are used as the output test set, the training set is used to train the neural network model, and the test set is used to evaluate the trained neural network model. This step is repeated until the neural network model converges to obtain the trained neural network model, and the establishment stage of the reduced-order model is completed. Among them, the neural network framework: pytorch framework, using a 6-layer deep feedforward neural network, the number of input and output neurons corresponds to all the reduced-order modes and boundary conditions of the variables of the learning object; the loss function is MSELoss, the activation function is ReLU, and the number of training times is 5000. The network optimization algorithm is AdamW, and the initial learning rate is set to 0.005; the training environment uses GPU accelerated training.

[0110] The method for using the reduced-order model of the present invention is as shown in the attached Figure 2As shown, the time-varying boundary condition is used as an excitation and input into the neural network model, and the reduced-order principal mode is predicted by the neural network. Through the reduction process of POD decomposition and combined with the modal information, it is restored to the full-order compressor flow field data matrix.

[0111] Taking the supercritical carbon dioxide compressor as the calculation and verification object, since the supercritical carbon dioxide compressor operates near the critical point, there are complex flow and physical property change processes inside it. The variables that the compressor model needs to predict include density, pressure, and velocities in three directions. In this embodiment, the data volume reaches 4.63 million. The original data is obtained from 22 different combinations of inlet temperature and flow rate, and 20 transient image data are used as the training set.

[0112] Figures 3a to 3d It is a schematic diagram of the accuracy of directly reconstructing the flow field after POD adopts different truncation numbers, which is used to determine the appropriate truncation number k during the iterative process. It can be seen that there is a relatively reasonable accuracy when truncated at the 15th order. At this time, the reconstruction error of the principal mode does not exceed 5%, and the sum of the k-order principal singular values reaches 0.999 of the total singular values.

[0113] In an application example, the method of the present disclosure is used to predict the flow field under unknown conditions within the compressor training set range. Figure 4a It is a comparison of the velocity map at the z = 0 section between the prediction result obtained by using the method of the present disclosure and CFD. Figure 4b It is a comparison of the pressure contour map at the z = 0 section between the prediction result obtained by using the method of the present disclosure and CFD. From Figure 4a and 4b It can be seen that except for the inlet pipe, the results of the velocity difference and the pressure field are very consistent with CFD. Figure 5a It is a statistical chart of the relative error of velocity prediction for quantitative analysis. Figure 5b It is a statistical chart of the relative error of pressure value prediction for quantitative analysis. See Figure 5a and Figure 5b The average relative error of velocity prediction is 9.7%, and the average relative error of pressure value is 3.05%.

[0114] In terms of efficiency, the time-consuming of a single CFD under a single working condition is about 3 hours, while the time-consuming of using the method of the present disclosure under the same configuration is less than 1 second, and it can give any flow field within the time range of transient image data in real time.

[0115] Figure 6 It is a block diagram of a device for reducing the order of a compressor flow field model shown according to an exemplary embodiment. For example, the device 1900 can be provided as a server. Refer to Figure 6, Device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by a memory 1932 for storing instructions executable by the processing component 1922, such as application programs. The application programs stored in the memory 1932 may include one or more modules each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above-described method.

[0116] Device 1900 may further include a power component 1926 configured to perform power management of the device 1900, a wired or wireless network interface 1950 configured to connect the device 1900 to a network, and an input / output (I / O) interface 1958. Device 1900 may operate based on an operating system stored in the memory 1932, such as Windows ServerTM, MacOS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.

[0117] In an exemplary embodiment, a non-transitory computer-readable storage medium is also provided, such as the memory 1932 including computer program instructions, and the computer program instructions may be executed by the processing component 1922 of the device 1900 to complete the above-described method.

[0118] The present disclosure may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0119] A computer-readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing. The computer-readable storage medium used herein is not construed as being a transitory signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0120] The computer-readable program instructions described herein can be downloaded to various computing / processing devices from a computer-readable storage medium or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0121] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider). In some embodiments, by using the state information of the computer-readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer-readable program instructions to implement various aspects of the present disclosure.

[0122] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0123] These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in one or more boxes of the flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that causes a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer-readable medium storing the instructions comprises a manufacture including instructions which implement various aspects of the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0124] The computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other devices to produce a computer-implemented process such that the instructions executed on the computer, other programmable data processing apparatus, or other devices implement the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0125] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of code, or a portion of an instruction, and the module, segment of code, or portion of an instruction may include one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the boxes may occur out of the order noted in the figures. For example, two consecutive blocks may in fact be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by special-purpose hardware-based systems that perform the specified functions or acts, or by combinations of special-purpose hardware and computer instructions.

[0126] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or improvements made to the technology in the market, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A compressor flow field model reduction method, characterized in that: The method comprises: Step 101, obtaining reactor compressor local flow field data, the reactor compressor local flow field data including the coordinates, velocity and pressure of the compressor local flow field; Step 102, after cleaning the acquired compressor local flow field data, organize it into a compressor local flow field matrix in the form of a snapshot matrix and save it; Step 103, using the compressor local flow field matrix saved by POD decomposition to perform data reduction to obtain reduced-order modes, wherein the first k-order singular vectors are selected as main modal data in the reduced-order modes according to the order of the singular values ​​of the singular vectors from large to small, and the sum of the singular values ​​of the first k-order singular vectors is greater than 99% of the total singular values; Step 104, using the preset time-varying boundary conditions as the input training set and the main modal data as the output test set, using the training set to perform model training on the neural network model to obtain a trained neural network model, using the test set to perform model evaluation on the trained neural network model, and continuously repeating the model training and model evaluation until the trained neural network model converges to obtain a reduced-order model; Step 105, inputting unknown time-varying boundary conditions as excitation data into the trained reduced-order model and outputting the predicted first k-order main modes; Step 106, using the POD flow field reconstruction algorithm to process the first k-order main modes output from step 105, restore the full compressor flow field variables in matrix form, and reconstruct the predicted compressor flow field.

2. The method according to claim 1, characterized in that In step 103, the compressor local flow field matrix saved by POD decomposition is used to reduce the data order and obtain the reduced-order mode: corresponding to the flow field problem, the velocity field u(x, t) of the flow field variable is decomposed into the average velocity u0 and the pulsating velocity u ′ The superposition of (x, t) is shown in Equation 1: u(x,t) = u0 + u ′ (x,t) Equation 1 The velocity field not only changes in time scale, but also includes changes in different spatial scales, as shown in Equation 2. ′ (x,t) is decoupled in the time domain and space domain: Substituting equation 1 into equation 2 yields equation 3: Where a0(t) = 1, The velocity fluctuations are expressed as a set of time-independent spatial functions According to the corresponding coefficient a n (t) changes in time and is finally the result of linear superposition; set the spatial function It has orthogonal normalization in the calculated fluid domain Ω, as shown in Equation 4: Define the inner product operator R of any flow field variable in the fluid domain Ω, as shown in Equation 5: Determine the R statistics of all points in the flow field with a time step of m, and obtain the pulsating kinetic energy E of the flow field at that moment T , according to formula 5, we can get formula 6: Substituting equation 2 into equation 6 and simplifying equation 4, we can get equation 7: Among them, λ n It is used to characterize the magnitude of the pulsating kinetic energy contained in each feature space basis (mode); extract the orthogonal normalized flow field basis from the existing flow field according to the magnitude of the pulsating kinetic energy component; and convert the inner product operator Acting on all time points and all variable points, the orthogonal normalization and equation 7 give equation 8: If the number of discrete units in the flow field is m and the number of time steps is n, traverse all discrete units and time nodes, and obtain Equation 9 from Equation 8: Among them, U∈R m×n , U ij =u ′ (x i ,t j )(i=1,…,m, j=1,…,n), W∈R m×m , W ij =ΔV i δ ij , ΔV i is the discrete unit volume; It is called the covariance matrix, and the corresponding λ and λ n Arrange in descending order to get a new sequence: n (λ1>λ2>λ3>…), the basis corresponding to these terms That is the POD substrate.

3. The method according to claim 1, characterized in that: The neural network model in step 104 is a pytorch neural network framework; the activation function is ReLU, the loss function is MSELoss, the AdamW optimization algorithm, and the initial learning rate is set to 0.

005.

4. The method according to claim 1, characterized in that In step 104, a deep feedforward neural network is used, wherein the neural network framework is the pytorch framework, a 6-layer deep feedforward neural network is used, the number of input and output neurons corresponds to all reduced-order modes and boundary conditions of the variables of the learning object; the loss function is MSELoss, the activation function is ReLU, the number of training times is 5000 times; the network optimization algorithm is AdamW, and the initial learning rate is set to 0.005; the training environment uses GPU accelerated training.

5. The method according to claim 1, characterized in that In step 103 , k is 15.

6. A compressor flow field model reduction device, characterized in that: The device comprises: An acquisition module is used to acquire the local flow field data of the reactor compressor, wherein the local flow field data of the reactor compressor includes the coordinates, velocity and pressure of the local flow field of the compressor; A data processing module is used to clean the acquired compressor local flow field data, organize it into a compressor local flow field matrix in the form of a snapshot matrix, and save it; The order reduction processing module is used to reduce the order of the compressor local flow field matrix saved by POD decomposition to obtain the reduced-order mode. In the reduced-order mode, the first k-order singular vectors are selected as the main modal data according to the order of the singular values ​​of the singular vectors from large to small, and the sum of the singular values ​​of the first k-order singular vectors is greater than 99% of the total singular values; A training and evaluation module is used to use the preset time-varying boundary conditions as an input training set and the main modal data as an output test set, use the training set to perform model training on the neural network model to obtain a trained neural network model, use the test set to perform model evaluation on the trained neural network model, and continuously repeat the model training and model evaluation until the trained neural network model converges to obtain a reduced-order model; A prediction module is used to input unknown time-varying boundary conditions as excitation data into the trained reduced-order model and output the predicted first k-order main modes; The reconstruction module is used to process the first k-order main modes outputted from step 105 using the POD flow field reconstruction algorithm, restore the full compressor flow field variables in matrix form, and reconstruct the predicted compressor flow field.

7. The device according to claim 6, characterized in that In the order reduction processing module, the local flow field matrix of the compressor saved by POD decomposition is used to reduce the data order and obtain the reduced-order mode: corresponding to the flow field problem, the velocity field u(x, t) of the flow field variable is decomposed into the average velocity u0 and the pulsating velocity u ′ The superposition of (x, t) is shown in Equation 1: u(x,t) = u0 + u ′ (x,t) Equation 1 The velocity field not only changes in time scale, but also includes changes in different spatial scales, as shown in Equation 2. ′ (x,t) is decoupled in the time domain and space domain: Substituting equation 1 into equation 2 yields equation 3: Where a0(t) = 1, The velocity fluctuations are expressed as a set of time-independent spatial functions According to the corresponding coefficient a n (t) changes in time and is finally the result of linear superposition; set the spatial function It has orthogonal normalization in the calculated fluid domain Ω, as shown in Equation 4: Define the inner product operator R of any flow field variable in the fluid domain Ω, as shown in Equation 5: Determine the R statistics of all points in the flow field with a time step of m, and obtain the pulsating kinetic energy E of the flow field at that moment T , according to formula 5, we can get formula 6: Substituting equation 2 into equation 6 and simplifying equation 4, we can get equation 7: Among them, λ n It is used to characterize the magnitude of the pulsating kinetic energy contained in each feature space basis (mode); extract the orthogonal normalized flow field basis from the existing flow field according to the magnitude of the pulsating kinetic energy component; and convert the inner product operator Acting on all time points and all variable points, the orthogonal normalization and equation 7 give equation 8: If the number of discrete units in the flow field is m and the number of time steps is n, traverse all discrete units and time nodes, and obtain Equation 9 from Equation 8: Among them, U∈R m×n , U ij =u ′ (x i ,t j )(i=1,…,m, j=1,…,n), W∈R m×m , W ij =ΔV i δ ij , ΔV i is the discrete unit volume; It is called the covariance matrix, and the corresponding λ and λ n Arrange in descending order to get a new sequence: n (λ1>λ2>λ3>…), the basis corresponding to these terms That is the POD substrate.

8. The device according to claim 6, characterized in that The neural network model in the training and evaluation module is the pytorch neural network framework; the activation function is ReLU, the loss function is MSELoss, the AdamW optimization algorithm, and the initial learning rate is set to 0.

005.

9. The device according to claim 6, characterized in that In the training and evaluation module, a deep feedforward neural network is used. The neural network framework is the pytorch framework, and a 6-layer deep feedforward neural network is adopted. The number of input and output neurons corresponds to all reduced-order modes and boundary conditions of the variables of the learning object; the loss function is MSELoss, the activation function is ReLU, and the number of training times is 5000; the network optimization algorithm is AdamW, and the initial learning rate is set to 0.005; the training environment uses GPU accelerated training.

10. The device according to claim 6, characterized in that In the order reduction processing module, k is 15.

11. A compressor flow field model reduction device, characterized in that: The device comprises: processor; a memory for storing processor-executable instructions; The processor is configured to execute the method according to any one of claims 1 to 5.

12. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images