Intelligent prediction method for complex large-scale unsteady flow field

Abstract

The invention discloses an intelligent prediction method for a complex large-scale unsteady flow field, and relates to the technical field of fluid mechanics, and the method comprises the steps: carrying out the numerical simulation of unsteady flow fields under various geometric shapes and different flow conditions, so as to construct a data set; performing coordinate transformation on the data set, and calculating and subtracting an average flow field to obtain a flow field snapshot containing pulsating quantity; the flow field snapshots are decoupled through eigenorthogonal decomposition, so that all space modals and time coefficients are obtained; inputting the time coefficient into an attention-enhanced neural network model to predict a time coefficient at a future moment, and performing training and parameter optimization on the model based on a well-designed physical information loss function; and combining and reconstructing the predicted time coefficient, the corresponding spatial mode and the average flow field to obtain a predicted flow field at a future moment. According to the invention, the long-term evolution of the unsteady flow field can be efficiently and accurately captured.

Description

Technical Field

[0001] The present application relates to the technical field of fluid mechanics, and in particular to an intelligent prediction method for complex, large-scale unsteady flow fields. Background Art

[0002] Unsteady flow is a common fluid mechanics phenomenon, whose physical properties evolve dynamically with space and time. It plays an important role in science and engineering fields such as aerospace, meteorology and oceanography, automobile industry and energy. Revealing the key transient behavior of unsteady flow has irreplaceable value for the optimization design and technological innovation in these fields. However, unsteady flow usually exhibits complex and strongly nonlinear characteristics, such as flow separation and vortex shedding, which brings great challenges to accurate simulation. For the prediction of unsteady flow field: traditional numerical methods have long calculation time, high overhead and difficult convergence; data-driven modeling methods based on deep learning have become a potential means of efficient simulation. The core is to use the powerful learning ability of neural networks to approximate the nonlinear mapping relationship between input and output, and replace expensive numerical iterations with fast predictions. However, in the existing intelligent prediction methods, the prediction results of the previous time step will be used as the input of the subsequent prediction, and the error between the predicted value and the true value will accumulate over time, resulting in a sharp drop in the prediction accuracy of the model. Error accumulation not only hinders the model's ability to capture the dynamic changes of the flow field, but also severely limits its generalization in complex physical systems.

[0003] It can be seen from the above that how to achieve spatiotemporal prediction of complex and large-scale unsteady flow fields and efficiently and accurately simulate the long-term evolution of unsteady flow fields are key issues that need to be urgently solved in this field. Summary of the invention

[0004] In view of this, the purpose of this application is to provide an intelligent prediction method for complex large-scale unsteady flow fields, which can achieve efficient and accurate spatiotemporal prediction. The specific scheme is as follows:

[0005] The present application discloses an intelligent prediction method for complex large-scale unsteady flow fields, comprising:

[0006] Numerical simulations of unsteady flow fields under various geometries and flow conditions are performed to construct data sets;

[0007] Performing coordinate transformation on the data set, calculating and subtracting the average flow field, so as to obtain a flow field snapshot including pulsation;

[0008] Decoupling the flow field snapshots using intrinsic orthogonal decomposition to obtain spatial modes and time coefficients;

[0009] Inputting the time coefficient into an attention-enhanced neural network model to predict the time coefficient at a future moment; wherein the neural network model is trained and parameter optimized based on a physical information loss function;

[0010] The predicted time coefficient at the future moment, the spatial mode corresponding to the time coefficient at the future moment, and the average flow field are combined and reconstructed to obtain the predicted flow field at the future moment.

[0011] Optionally, the unsteady flow field under various geometric shapes and different flow conditions is numerically simulated to construct a data set, including:

[0012] The data set is constructed by numerically simulating unsteady flow fields under various geometric shapes and flow conditions using physical solvers, including ANSYS Fluent, OpenFOAM, and SU2; the geometric shapes include airfoils, cylinders, and blades; and the flow conditions include angle of attack, Mach number, and Reynolds number.

[0013] Optionally, the coordinate transformation of the data set, calculating and subtracting the average flow field to obtain a flow field snapshot containing a pulsating amount, includes:

[0014] The data set is mapped from the non-uniform physical domain to the uniform computational domain using the structured grid feature, and then the average flow field is calculated and subtracted to obtain a flow field snapshot including the pulsating volume.

[0015] Optionally, the decoupling of the flow field snapshot by using intrinsic orthogonal decomposition to obtain various spatial modes and time coefficients includes:

[0016] The flow field snapshot minus the mean flow field is decoupled using the intrinsic orthogonal decomposition method to separate the spatial mode and the corresponding time coefficient.

[0017] Optionally, the intelligent prediction method for complex large-scale unsteady flow fields further includes:

[0018] An attention-enhanced neural network model is constructed based on the Mamba-2 module, the bidirectional gated recurrent unit module, and the channel attention mechanism module.

[0019] Optionally, inputting the time coefficient into an attention-enhanced neural network model to predict the time coefficient at a future moment includes:

[0020] Inputting the time coefficients into an attention-enhanced neural network model so as to model the long-term dependencies of the time coefficients using a Mamba-2 module in the neural network model to extract key features;

[0021] Using the bidirectional gated recurrent unit module in the neural network model to capture forward and reverse timing information of the key features, and modifying the number of channels through the channel mapping layer in the neural network model to obtain updated timing information;

[0022] The channel attention mechanism module in the neural network model is used to extract the frequency domain features in the updated time series information, and the predicted sequence length is adjusted through the length mapping layer in the neural network model to obtain the time coefficient of the predicted future moment.

[0023] Optionally, the time coefficient is input into the attention-enhanced neural network model so as to use the Mamba-2 module in the neural network model to model the long-term dependency of the time coefficient to extract key features; the bidirectional gated recurrent unit module in the neural network model is used to capture the forward and reverse timing information of the key features, and the number of channels is modified through the channel mapping layer in the neural network model to obtain updated timing information; the channel attention mechanism module in the neural network model is used to extract the frequency domain features in the updated timing information, and the length mapping layer in the neural network model is used to adjust the predicted sequence length to obtain the predicted time coefficient of the future moment, including:

[0024] Inputting the time coefficient into the attention-enhanced neural network model, and performing a reshaping operation and a layer normalization operation on the time coefficient to obtain normalized data;

[0025] Inputting the normalized data into the Mamba-2 module so that the Mamba-2 module models the long-term dependency of the time coefficients using a selective state space model to extract key features;

[0026] Using the bidirectional gated recurrent unit module in the neural network model to capture the forward and reverse timing information of the key features, and modifying the number of channels through the channel mapping layer in the neural network model, and then performing a flipping operation to obtain updated timing information;

[0027] The channel attention mechanism module in the neural network model is utilized, discrete cosine transform is adopted to extract the frequency domain features in the updated time series information, and the predicted sequence length is adjusted through the length mapping layer in the neural network model to obtain the time coefficient of the predicted future moment.

[0028] Optionally, the reconstructing by combining the predicted time coefficient at the future moment, the spatial mode corresponding to the time coefficient at the future moment, and the average flow field to obtain the predicted flow field at the future moment includes:

[0029] The product between the time coefficient at the predicted future moment and the spatial mode corresponding to the time coefficient at the future moment is calculated, and then the product is added to the average flow field to reconstruct the predicted flow field at the future moment.

[0030] Optionally, the process of constructing the physical information loss function includes:

[0031] A physical information loss function is constructed based on the predicted flow field and the actual flow field; the physical information loss function includes a Smooth L1 loss function, a gradient loss function and a wall loss function.

[0032] Optionally, the physical information loss function is:

[0033]

[0034] in, is the physical information loss function, Loss SL1 is the Smooth L1 loss function, λ 1 is the weight of the gradient loss, Loss Grad is the gradient loss function, λ 2 is the weight of the wall loss, Loss wall is the wall loss function;

[0035] The Smooth L1 loss function is:

[0036]

[0037] Where C is the number of input or output channels, H is the height of the flow field snapshot, W is the width of the flow field snapshot, and X k，i，j is the actual flow field, To predict the flow field;

[0038] The gradient loss function is:

[0039]

[0040] in, and are the partial derivatives of the flow field variables in the horizontal and vertical directions respectively;

[0041] The wall loss function is:

[0042]

[0043] It can be seen that the present application provides an intelligent prediction method for complex large-scale unsteady flow fields, including numerical simulation of unsteady flow fields under various geometric shapes and different flow conditions to construct a data set; coordinate transformation of the data set, calculation and subtraction of the average flow field to obtain a flow field snapshot containing pulsating quantities; decoupling the flow field snapshot using intrinsic orthogonal decomposition to obtain various spatial modes and time coefficients; inputting the time coefficient into an attention-enhanced neural network model to predict the time coefficient at future moments; wherein the neural network model is trained and parameter optimized based on a physical information loss function; and the time coefficient at the predicted future moment, the spatial mode corresponding to the time coefficient at the future moment, and the average flow field are combined and reconstructed to obtain the predicted flow field at the future moment. The present application performs numerical simulations on unsteady flow fields under various geometric shapes and different flow conditions to construct a data set; then the data set is transformed in coordinates to completely retain all flow information in the near-wall area in the calculation domain, thereby achieving accurate acquisition of flow field snapshots at all time steps; the flow field snapshots are decoupled using intrinsic orthogonal decomposition to represent complex flows as a small number of spatial modes and time coefficients that vary with time, thereby reducing data dimensions and improving modeling efficiency; the time coefficients are input into an attention-enhanced neural network model to predict the time coefficients at future moments, which significantly improves the accuracy of long-term predictions, thereby effectively alleviating the problem of error accumulation, and the present application trains and optimizes the parameters of the attention-enhanced neural network model based on a physical information loss function, thereby being able to more accurately capture the dynamic evolution of the flow field. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0045] Figure 1 A flow chart of an intelligent prediction method for complex large-scale unsteady flow fields disclosed in this application;

[0046] Figure 2 A physical domain diagram of a hydrofoil and cylinder case disclosed in this application;

[0047] Figure 3 A coordinate transformation process diagram disclosed in this application;

[0048] Figure 4 A horizontal velocity field diagram in a physical domain and a computational domain disclosed in the present application;

[0049] Figure 5 This is a diagram of the architecture of an attention-enhanced neural network model disclosed in this application;

[0050] Figure 6 This is an overall framework diagram of ST-FlowNet disclosed in this application. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0052] Unsteady flow is a common fluid mechanics phenomenon, whose physical properties evolve dynamically with space and time. It plays an important role in science and engineering fields such as aerospace, meteorology and oceanography, automobile industry and energy. Revealing the key transient behavior of unsteady flow has irreplaceable value for the optimization design and technological innovation in these fields. However, unsteady flow usually exhibits complex and strongly nonlinear characteristics, such as flow separation and vortex shedding, which brings great challenges to accurate simulation. For the prediction of unsteady flow field: traditional numerical methods have long calculation time, high overhead and difficult convergence; data-driven modeling methods based on deep learning have become a potential means of efficient simulation. The core is to use the powerful learning ability of neural networks to approximate the nonlinear mapping relationship between input and output, and replace expensive numerical iterations with fast predictions. However, in the existing intelligent prediction methods, the prediction results of the previous time step will be used as the input of the subsequent prediction, and the error between the predicted value and the true value will accumulate over time, resulting in a sharp drop in the prediction accuracy of the model. Error accumulation not only hinders the model's ability to capture the dynamic changes of the flow field, but also severely limits its generalization in complex physical systems. It can be seen from the above that how to achieve spatiotemporal prediction of complex and large-scale unsteady flow fields and efficiently and accurately simulate the long-term evolution of unsteady flow fields are key issues that need to be urgently solved in this field.

[0053] See also Figure 1 As shown, the embodiment of the present application discloses an intelligent prediction method for complex large-scale unsteady flow fields, which may specifically include:

[0054] Step S11: numerical simulation of unsteady flow fields under various geometric shapes and different flow conditions is performed to construct a data set.

[0055] In this embodiment, a physical solver is used to perform numerical simulation on an unsteady flow field under various geometric shapes and different flow conditions to construct a data set; the physical solver includes ANSYS Fluent, OpenFOAM and SU2; the geometric shapes include airfoils, cylinders and blades; and the flow conditions include angle of attack, Mach number and Reynolds number.

[0056] In addition, the physical solver, geometric shape and flow conditions proposed in this application include but are not limited to the above contents.

[0057] This application takes a cylinder and an airfoil as an example. The physical domain settings of the hydrofoil and cylinder cases are as follows: Figure 2 As shown. The cylinder diameter and the airfoil chord length are both 1 meter. Boundary condition settings: the inlet velocity is fixed at 1 m / s, and the wall is a no-slip condition. The distance from the flow field inlet and outlet to the center is 15 meters, and the flow field diameter is 30 meters. Set the positive direction of the x-axis to the flow direction when the angle of attack is equal to 0°.

[0058] Step S12: coordinate transformation is performed on the data set, and the average flow field is calculated and subtracted to obtain a flow field snapshot including the pulsation amount.

[0059] In this embodiment, the data set is mapped from the non-uniform physical domain to the uniform computational domain using the structured grid characteristic, and then the average flow field is calculated and subtracted to obtain a flow field snapshot including the pulsating quantity.

[0060] In many studies, data is usually obtained by projecting the flow field onto a uniform Cartesian grid through a pixelation preprocessing method. However, this method introduces roughness and may cause the loss of flow details in the near-wall region, thereby increasing the prediction error. To address this problem, this application adopts a coordinate transformation strategy that is faster than point cloud and graph neural networks. This method uses the properties of structured grids to map the flow field from a non-uniform physical domain (x, y) to a uniform computational domain (ξ, η). Figure 3 It is a coordinate transformation process, which ensures a one-to-one correspondence between points in the physical domain and the computational domain. Figure 4 is the horizontal velocity field in the physical domain and the computational domain. Obviously, in the computational domain, all the flow information in the near-wall region is completely preserved, and accurate acquisition of the flow field snapshots at all time steps is achieved.

[0061] Step S13: Decoupling the flow field snapshots by using intrinsic orthogonal decomposition to obtain various spatial modes and time coefficients.

[0062] In this embodiment, POD (Proper Orthogonal Decomposition) is used to decouple the flow field snapshot minus the average flow field, so as to separate the spatial mode and the corresponding time coefficient.

[0063] This method represents complex flows as a small number of spatial modes and time-varying time coefficients, significantly reducing the data dimension and improving the modeling efficiency.

[0064] The principle of POD is as follows: Assume that physical quantities (such as velocity, pressure, and density) at m computational grid points are collected at n moments as samples, and are represented in chronological order as To capture the dynamic behavior, it is necessary to extract and separate the static features (mean flow field), which are calculated as:

[0065]

[0066] where represents the 0th order mode. Subtracting from the initial snapshot sequence X, a matrix composed of flow field pulsations is obtained:

[0067]

[0068] The original POD is calculated using the covariance matrix. Since n << m, this usually involves large memory requirements and computational costs. Snapshot POD is an effective alternative method that significantly reduces the size of the matrices involved and lowers the computational cost. Specifically, the covariance matrix is calculated as follows:

[0069]

[0070] where The eigenvector matrix is obtained through eigenvalue decomposition, and the corresponding eigenvalue matrix is The two satisfy:

[0071] CW = λW;

[0072] Next, the mode matrix is calculated as follows:

[0073]

[0074] Therefore, the POD time coefficient matrix can be expressed as:

[0075]

[0076] where a i = [a 1，i , a 2，i ,..., a n，i T (i = 1, 2,..., n) is the corresponding time coefficient vector. Select the first r order modes to reconstruct the original flow field. At this time, a i = [a1，i , a 2，i , ..., a r，i ] T The reconstructed flow field is calculated as:

[0077]

[0078] In order to determine the appropriate number of spatial modes, this application defines the reconstructed MAPE (Mean Absolute Percentage Error) as:

[0079]

[0080] Among them, x i，j and x′ i，i They represent the true value and the reconstructed value of the flow field respectively. This application sets the modal number according to the requirement that MAPE is less than 5%.

[0081] Step S14: Input the time coefficient into the attention-enhanced neural network model to predict the time coefficient at a future moment; wherein the neural network model is trained and parameter optimized based on a physical information loss function.

[0082] In this embodiment, an attention-enhanced neural network model is constructed based on the Mamba-2 module, the bidirectional gated recurrent unit module, and the channel attention mechanism module. The time coefficient is input into the attention-enhanced neural network model so that the long-term dependency of the time coefficient is modeled using the Mamba-2 module in the neural network model to extract key features; the bidirectional gated recurrent unit module in the neural network model is used to capture the forward and reverse timing information of the key features, and the number of channels is modified through the channel mapping layer in the neural network model to obtain updated timing information; the channel attention mechanism module in the neural network model is used to extract the frequency domain features in the updated timing information, and the length of the predicted sequence is adjusted through the length mapping layer in the neural network model to obtain the predicted time coefficient of the future moment.

[0083] Among them, the specific process of predicting the time coefficient of the future moment is: inputting the time coefficient into the attention-enhanced neural network model, and performing a reshaping operation and a layer normalization operation on the time coefficient to obtain normalized data; inputting the normalized data into the Mamba-2 module, so that the Mamba-2 module uses a selective state-space model to model the long-term dependency of the time coefficient to extract key features; using the bidirectional gated recurrent unit module in the neural network model to capture the forward and reverse timing information of the key features, and modifying the number of channels through the channel mapping layer in the neural network model, and then performing a flipping operation to obtain updated timing information; using the channel attention mechanism module in the neural network model, using discrete cosine transform to extract frequency domain features in the updated timing information, and adjusting the prediction sequence length through the length mapping layer in the neural network model to obtain the predicted time coefficient of the future moment.

[0084] In view of the fact that existing intelligent methods are difficult to effectively deal with the problem of error accumulation, this application designs an attention-enhanced neural network model, whose architecture is as follows: Figure 5 As shown. The size of the input data is [B, L in , C, M], where B represents the batch size, L in Indicates the length of the input sequence, C indicates the number of channels, and M indicates the number of modes. Taking the prediction of the horizontal and vertical velocity components and pressure as an example, there are three prediction quantities, so C takes the value of 3. After the reshape operation, the input is converted to [B, L in , C×M], using layer normalization to accelerate model convergence and improve training stability. The normalized data is input into the Mamba-2 module, which can extract key features from time series and use SSM (State Space Model, Selective State Space Model) to model long-term dependencies. Its structure is as follows Figure 5 As shown in the Mamba-2 module in the

[0085] Next, the extracted features are passed to the Bidirectional Gated Recurrent Unit (BiGRU) module, which is used to capture both forward and reverse timing information. The output size of this stage is [B, L in , 2×D], where D represents the number of features in the hidden state. The output is then mapped back to [B, L in , C×M], and perform a flip operation to transform the data into [B, C×M, L in]. Subsequently, the CAM (Channel Attention Mechanism) module captures frequency domain information through DCT (Discrete Cosine Transform). This module is used to improve the feature selection ability of the model, effectively suppress noise and enhance the expression of key channel features. Finally, the length mapping layer is used to adjust the length of the predicted sequence. The final output size is [B, L out , C, M], where L out Indicates the length of the predicted sequence. Both the channel mapping layer and the length mapping layer are fully connected layers.

[0086] GRU neural network is widely used in time series. Its information storage and internal state update can be expressed as:

[0087] z t =σ(W z ·[h t-1 , x t ]+b z );

[0088] r t =σ(W r ·[h t-1 , x t ]+b r );

[0089]

[0090] Among them, x t represents the input data at time t; σ represents the sigmoid activation function, which converts the state value to the range of [0, 1]; h t-1 and h t are the hidden states at time t-1 and time t respectively; is the candidate hidden state; W z , W r and are all weight matrices; b z 、b r and are all bias terms. Tanh is the hyperbolic tangent function, and ⊙ represents the Hadamard product. t Represents the reset gate, which is responsible for determining the degree of forgetting previous information and handling short-term dependencies; trepresents the update gate, which is responsible for controlling the retention and update of information in the hidden state and handling long-term dependencies. In the process of unsteady flow field changes, there is a strong correlation between flow field data at different times. The BiGRU network can capture more comprehensive time information for prediction. It extends the standard GRU by processing data forward and backward, enabling the model to capture time series information from past and future states at the same time. The structure of BiGRU is as follows Figure 5 As shown in the BiGRU module in, its expression is as follows:

[0091]

[0092] Among them, GRU(·) is a unidirectional GRU network, W t is the weight matrix, b t is the bias term, and are the forward hidden state and reverse hidden state of GRU at time t respectively.

[0093] In order to enhance the modeling capability of temporal coefficients, this application adopts a DCT-based channel attention mechanism, such as Figure 5 For the input sequence Where m represents the number of channels and n represents the length of the input sequence. First, DCT is applied to each channel (each row) to extract frequency features:

[0094] F i =DCT(X i ), i=, 1, ..., m-1;

[0095] Among them, X i represents the i-th row of X, F i Represents the frequency feature of the i-th channel. The frequency features are stacked and the attention weight is obtained through a fully connected layer and a nonlinear activation function:

[0096] F=stack(F 0 , F 1 , ..., F n-1 );

[0097]

[0098] Among them, W 1 and W 2 is the weight of the fully connected layer, σ and All represent nonlinear activation functions, such as Sigmoid, Swish, and Mish, and the number of fully connected layers can be flexibly adjusted. Finally, these attention weights are used to rescale the original features:

[0099]

[0100] Where ⊙ represents the Hadamard product. In this way, CAM combines the time domain and frequency domain information, thereby improving the model's ability to model the time coefficient.

[0101] Mathematically speaking, the training of a neural network can be viewed as an optimization problem, the goal of which is to minimize the loss function. In previous studies on flow field prediction, MAE (Mean Absolute Error) and MSE (Mean Square Error) are most commonly used, but their respective defects may affect the prediction accuracy. MAE converges slowly and is not sensitive enough to small error changes; while MSE is too sensitive to outliers, which may lead to inadequate robustness of the prediction results. Therefore, this application develops a physical information loss function, which includes Smooth L1 loss, gradient loss (GradientLoss) and wall loss (Wall Loss). The Smooth L1 loss function is particularly suitable for regression tasks, and its goal is to strike a balance between MAE and MSE. It effectively combines the advantages of both, and can enhance the robustness of the model to outliers while maintaining small error fitting, which helps to improve the long-term prediction accuracy and robustness of the model. In order to improve the smoothness and continuity of the prediction results, this application designs a gradient loss. This application uses the second-order central difference method to calculate the gradient. For unsteady flow prediction tasks, gradient loss ensures that the model generates physically reasonable flow fields and prevents numerical mutations, which is particularly important for long-term prediction tasks. In addition, this application also constructs a wall loss to improve the accuracy of predictions near wall boundaries. This loss focuses on near-wall data points to ensure that the model accurately captures the flow characteristics in these key areas.

[0102] The physical information loss function developed in this application is:

[0103]

[0104] in, is the physical information loss function, Loss SL1 is the Smooth L1 loss function, λ 1 is the weight of the gradient loss, Loss Grad is the gradient loss function, λ 2 is the weight of the wall loss, Loss Wall is the wall loss function;

[0105] The Smooth L1 loss function is:

[0106]

[0107] Where C is the number of input or output channels, H is the height of the flow field snapshot, W is the width of the flow field snapshot, and X k，i，j is the actual flow field, To predict the flow field;

[0108] The gradient loss function is:

[0109]

[0110] in, and are the partial derivatives of the flow field variables in the horizontal and vertical directions respectively;

[0111] The wall loss function is:

[0112]

[0113] Step S15: Reconstructing the predicted time coefficient at the future time, the spatial mode corresponding to the time coefficient at the future time, and the average flow field to obtain the predicted flow field at the future time.

[0114] In this embodiment, the product between the predicted time coefficient and the corresponding spatial mode is calculated, and then the product is added to the average flow field to reconstruct the predicted flow field at the future moment.

[0115] This application designs an intelligent prediction method for complex large-scale unsteady flow fields. The specific framework is ST-FlowNet. Figure 6 As shown in Figure 1, the framework aims to ensure high accuracy and efficiency in long-term prediction of unsteady flow fields while alleviating the problem of error accumulation. The framework is divided into three stages: In the data generation and processing stage, the physical solver is first used to simulate the real flow field, and then the coordinate transformation is used to transform them from the physical domain to the computational domain, and the average flow field is calculated and subtracted to obtain the flow field snapshot containing the pulsating quantity. These snapshots are decomposed into spatial modes and time coefficients using POD to reduce the data dimension and improve modeling efficiency. In the prediction stage, the attention-enhanced neural network model takes the time coefficient of the previous moment as input to predict the time coefficient of the future moment. In the flow field reconstruction stage, the predicted time coefficient is combined with the corresponding spatial mode to realize the reconstruction of the flow field. Finally, the physical information loss function is used to guide the training and parameter optimization of the neural network model.

[0116] The innovation of this application lies in: designing an intelligent prediction method for complex large-scale unsteady flow fields, the specific framework is ST-FlowNet, coupling the reduced-order model with the neural network, accurately retaining the near-wall flow field information through coordinate transformation, and realizing high-precision modeling of unsteady flow fields at a low computational cost; designing an attention-enhanced neural network model to improve the accuracy of long-term prediction, thereby effectively alleviating the problem of error accumulation; constructing a physical information loss function, combining Sm oo th L1, gradient loss and wall loss, improve the smoothness and physical consistency of the prediction, and better capture the dynamic evolution of the flow field.

[0117] In this embodiment, numerical simulations are performed on unsteady flow fields under various geometric shapes and different flow conditions to construct a data set; the data set is subjected to coordinate transformation, and the average flow field is calculated and subtracted to obtain a flow field snapshot containing a pulsating quantity; the flow field snapshot is decoupled using intrinsic orthogonal decomposition to obtain various spatial modes and time coefficients; the time coefficient is input into an attention-enhanced neural network model to predict the time coefficient at a future moment; wherein the neural network model is trained and parameter optimized based on a physical information loss function; the predicted time coefficient at a future moment, the spatial mode corresponding to the time coefficient at a future moment, and the average flow field are combined and reconstructed to obtain a predicted flow field at a future moment. The present application performs numerical simulations on unsteady flow fields under various geometric shapes and different flow conditions to construct a data set; then the data set is transformed in coordinates to completely retain all flow information in the near-wall area in the calculation domain, thereby achieving accurate acquisition of flow field snapshots at all time steps; the flow field snapshots are decoupled using intrinsic orthogonal decomposition to represent complex flows as a small number of spatial modes and time coefficients that vary with time, thereby reducing data dimensions and improving modeling efficiency; the time coefficients are input into an attention-enhanced neural network model to predict the time coefficients at future moments, which significantly improves the accuracy of long-term predictions, thereby effectively alleviating the problem of error accumulation, and the present application trains and optimizes the parameters of the attention-enhanced neural network model based on a physical information loss function, thereby being able to more accurately capture the dynamic evolution of the flow field.

[0118] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0119] The above is a detailed introduction to an intelligent prediction method for complex large-scale unsteady flow fields provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. An intelligent prediction method for complex large-scale unsteady flow fields, characterized in that: include: Numerical simulations of unsteady flow fields under various geometries and flow conditions are performed to construct data sets; Performing coordinate transformation on the data set, calculating and subtracting the average flow field, so as to obtain a flow field snapshot including pulsation; Decoupling the flow field snapshots using intrinsic orthogonal decomposition to obtain spatial modes and time coefficients; Inputting the time coefficient into an attention-enhanced neural network model to predict the time coefficient at a future moment; wherein the neural network model is trained and parameter optimized based on a physical information loss function; The predicted time coefficient at the future moment, the spatial mode corresponding to the time coefficient at the future moment, and the average flow field are combined and reconstructed to obtain the predicted flow field at the future moment.

2. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: The unsteady flow fields under various geometric shapes and different flow conditions are numerically simulated to construct a data set, including: The data set is constructed by numerically simulating unsteady flow fields under various geometric shapes and flow conditions using physical solvers, including ANSYS Fluent, OpenFOAM, and SU2; the geometric shapes include airfoils, cylinders, and blades; and the flow conditions include angle of attack, Mach number, and Reynolds number.

3. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: The coordinate transformation of the data set is performed, and the average flow field is calculated and subtracted to obtain a flow field snapshot containing a pulsating amount, including: The data set is mapped from the non-uniform physical domain to the uniform computational domain using the structured grid feature, and then the average flow field is calculated and subtracted to obtain a flow field snapshot including the pulsating volume.

4. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: The flow field snapshot is decoupled by using intrinsic orthogonal decomposition to obtain various spatial modes and time coefficients, including: The flow field snapshot minus the mean flow field is decoupled using the eigenorthogonal decomposition method to separate the spatial mode and the corresponding time coefficient.

5. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: Also includes: An attention-enhanced neural network model is constructed based on the Mamba-2 module, the bidirectional gated recurrent unit module, and the channel attention mechanism module.

6. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: The step of inputting the time coefficient into the attention-enhanced neural network model to predict the time coefficient at a future moment includes: Inputting the time coefficients into an attention-enhanced neural network model so as to model the long-term dependencies of the time coefficients using a Mamba-2 module in the neural network model to extract key features; Using the bidirectional gated recurrent unit module in the neural network model to capture forward and reverse timing information of the key features, and modifying the number of channels through the channel mapping layer in the neural network model to obtain updated timing information; The channel attention mechanism module in the neural network model is used to extract the frequency domain features in the updated time series information, and the predicted sequence length is adjusted through the length mapping layer in the neural network model to obtain the time coefficient of the predicted future moment.

7. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 6 is characterized in that: The time coefficient is input into the attention-enhanced neural network model so as to use the Mamba-2 module in the neural network model to model the long-term dependency of the time coefficient to extract key features; the bidirectional gated recurrent unit module in the neural network model is used to capture the forward and reverse timing information of the key features, and the number of channels is modified through the channel mapping layer in the neural network model to obtain updated timing information; the channel attention mechanism module in the neural network model is used to extract the frequency domain features in the updated timing information, and the length mapping layer in the neural network model is used to adjust the predicted sequence length to obtain the predicted time coefficient of the future moment, including: Inputting the time coefficient into the attention-enhanced neural network model, and performing a reshaping operation and a layer normalization operation on the time coefficient to obtain normalized data; Inputting the normalized data into the Mamba-2 module so that the Mamba-2 module models the long-term dependency of the time coefficients using a selective state space model to extract key features; Using the bidirectional gated recurrent unit module in the neural network model to capture the forward and reverse timing information of the key features, and modifying the number of channels through the channel mapping layer in the neural network model, and then performing a flipping operation to obtain updated timing information; The channel attention mechanism module in the neural network model is utilized, discrete cosine transform is adopted to extract the frequency domain features in the updated time series information, and the predicted sequence length is adjusted through the length mapping layer in the neural network model to obtain the time coefficient of the predicted future moment.

8. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 1 is characterized in that: The reconstructing by combining the predicted time coefficient at the future moment, the spatial mode corresponding to the time coefficient at the future moment, and the average flow field to obtain the predicted flow field at the future moment includes: The product between the time coefficient at the predicted future moment and the spatial mode corresponding to the time coefficient at the future moment is calculated, and then the product is added to the average flow field to reconstruct the predicted flow field at the future moment.

9. The intelligent prediction method for complex large-scale unsteady flow fields according to any one of claims 1 to 8, characterized in that: The process of constructing the physical information loss function includes: A physical information loss function is constructed based on the predicted flow field and the actual flow field; the physical information loss function includes a Smooth L1 loss function, a gradient loss function and a wall loss function.

10. The intelligent prediction method for complex large-scale unsteady flow fields according to claim 9, characterized in that: The physical information loss function is: in, is the physical information loss function, Loss SL1 is the Smooth L1 loss function, λ1 is the weight of the gradient loss, Loss Grad is the gradient loss function, λ2 is the weight of the wall loss, Loss Wall is the wall loss function; The Smooth L1 loss function is: Where C is the number of input or output channels, H is the height of the flow field snapshot, W is the width of the flow field snapshot, and X k，i，j , is the actual flow field, To predict the flow field; The gradient loss function is: in, and are the partial derivatives of the flow field variables in the horizontal and vertical directions respectively; The wall loss function is:

Images