Artificial intelligence based fast prediction method for turbine aerodynamic excitations

Abstract

The application discloses a turbine aerodynamic excitation rapid prediction method based on artificial intelligence, and the prediction method comprises the following steps: dividing a turbine flow field into a main flow area and a complex flow area; for the complex flow area, a mathematical model is established to predict the flow field parameter distribution of a wake area, and the boundary parameters of the interface between the complex flow area and the main flow area are determined; a physical information neural network is constructed to predict the flow field parameters of the main flow area; the flow field parameters of the main flow area and the flow field parameters of the wake area in the complex flow area are merged to obtain the flow parameters of the whole flow field, and the aerodynamic excitation is calculated based on the predicted pressure values of the extracted downstream blade surface. The present application can rapidly predict the turbine aerodynamic excitation by coupling the two methods through the transmission of the wake boundary values, and the calculation speed is greatly improved compared with the existing aerodynamic excitation calculation and prediction methods under the premise of ensuring the calculation accuracy.

Description

Technical Field

[0001] This invention relates to the field of turbine aerodynamic excitation prediction, and more particularly to a rapid prediction method for turbine aerodynamic excitation based on artificial intelligence. Background Technology

[0002] With the continuous improvement of aero-engine performance requirements, the strength margin of turbine components is decreasing, making the aerodynamic excitation problem caused by the unsteady sweep of the upstream guide vane wake increasingly prominent. In the current design process, obtaining the aerodynamic excitation characteristics of the turbine relies heavily on unsteady numerical simulation technology, especially for multi-stage turbines with non-uniform designs, which typically require full-ring unsteady calculations. However, the computational cost of such high-precision full-ring multi-stage turbine unsteady numerical simulations is extremely high, requiring dozens of days on high-performance computers to obtain results. This huge time overhead is significantly contradictory to the iterative engineering requirements of the aero-engine design phase, severely restricting design efficiency and optimization space.

[0003] In recent years, Physics-Informe Neural Networks (PINN), as a machine learning model that integrates deep learning and physics constraints, has demonstrated excellent generalization ability in data-scarce and noisy engineering problems by embedding partial differential equations into the loss function to guide network training. After training, PINN inference is extremely time-efficient, theoretically meeting the need for rapid prediction of turbine aerodynamic excitations. However, PINN commonly suffers from the "AI illusion" problem in practical applications, where the network output tends to converge to non-physical interpretations lacking physical meaning and failing to reflect engineering realities. Directly applying it to the global prediction of turbine flow fields makes reliability difficult to guarantee. Therefore, effectively mitigating the non-physical interpretation defects while fully leveraging PINN's computational speed advantage has become a key technical bottleneck for achieving fast and reliable prediction of turbine aerodynamic excitations. Summary of the Invention

[0004] This invention provides an artificial intelligence-based method for rapid prediction of turbine aerodynamic excitation, which addresses the technical problem of how to significantly improve the calculation speed of aerodynamic excitation while ensuring calculation accuracy, so as to achieve rapid prediction of turbine aerodynamic excitation.

[0005] A rapid prediction method for turbine aerodynamic excitation based on artificial intelligence includes:

[0006] S1. Divide the turbine flow field into the mainstream region and the complex flow region. The complex flow region includes the wake region, while the mainstream region is the inviscid flow region far away from the boundary layer and the wake region.

[0007] S2. For complex flow regions, establish a mathematical model based on fluid dynamics boundary layer theory and high-precision numerical simulation results to predict the flow field parameter distribution in the wake region and determine the boundary parameters at the interface between the complex flow region and the mainstream region.

[0008] S3. Construct a physical information neural network, using the Euler equation as a physical constraint and the boundary parameters of the interface between the complex flow region and the mainstream region as boundary condition inputs, to train the physical information neural network in order to predict the flow field parameters of the mainstream region.

[0009] S4. Merge the flow field parameters of the main flow region with the flow field parameters of the wake region in the complex flow region to obtain the flow parameters of the entire flow field, and calculate the aerodynamic excitation based on the extracted pressure prediction values ​​of the downstream blade surface.

[0010] In this invention, a partitioned modeling strategy is adopted to divide the turbine flow field into a main flow region and a complex flow region (wake region). Appropriate solution methods are selected for the differences in flow characteristics in different regions: a mathematical model for the complex flow region is established based on fluid dynamics boundary layer theory and high-precision numerical simulation results, while the Euler equations for the main flow region are solved using a Physical Information Neural Network (PINN). The two methods are coupled through the propagation of wake boundary values. This strategy fully utilizes the smooth flow and small gradient characteristics of the main flow region, significantly reducing the complexity and difficulty of solving the PINN governing equations. It effectively avoids the AI ​​illusion problem that easily occurs when PINN is directly applied to the entire flow field, i.e., convergence to a non-physical solution that lacks physical meaning and does not conform to engineering reality. Simultaneously, the complex flow region is described using a mathematical model with a clear physical basis, ensuring the reliability and accuracy of the wake region prediction. The two methods work together to achieve a dual guarantee of computational accuracy and physical consistency.

[0011] In terms of computational efficiency, this invention achieves an order-of-magnitude improvement over existing unsteady numerical simulation methods. Existing high-precision full-ring multi-stage turbine unsteady numerical simulations typically require hours to days on high-performance computers, while the PINN inference of this invention, after training, is extremely time-efficient, with overall prediction completed within minutes or even seconds, fully meeting the iterative engineering requirements of aero-engine design. Furthermore, this invention strictly adheres to the fundamental principles of fluid mechanics and the PINN training paradigm, employing a mature and feasible technical approach that combines significant computational speed advantages with engineering application value, providing a practical and reliable technical path for rapid and reliable prediction of turbine aerodynamic excitations. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a rapid prediction method for turbine aerodynamic excitation based on artificial intelligence in one embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based method for rapid prediction of turbine aerodynamic excitation is provided, comprising the following steps:

[0016] S1. Divide the turbine flow field into the mainstream region and the complex flow region. The complex flow region includes the wake region, while the mainstream region is the inviscid flow region far away from the boundary layer and the wake region.

[0017] S2. For complex flow regions, establish a mathematical model based on fluid dynamics boundary layer theory and high-precision numerical simulation results to predict the flow field parameter distribution in the wake region and determine the boundary parameters at the interface between the complex flow region and the mainstream region.

[0018] In one embodiment, step S2 further includes the following sub-steps:

[0019] S201. Establish a velocity distribution model for the wake region of the turbine flow field, and define the mathematical expression for the wake velocity distribution of the turbine guide vane:

[0020]

[0021] in, This represents the velocity of the wake to be measured. Indicates the mainstream speed. The dimensionless maximum velocity deficit of the wreath is only related to... Related unknowns; Indicates the width of the trail, which is only related to... Related unknowns; and These represent the horizontal and vertical coordinates, respectively, and are the input values ​​for the model.

[0022] S202. Assume the dimensionless maximum velocity deficit of the wake. The distribution expression is:

[0023]

[0024] in, Indicates the width of the trail. Indicates the feature length; This represents the drag coefficient of the turbine guide vane; This represents the constant to be determined (which can be obtained through experiments or CFD).

[0025] S203, Set the tail width The distribution expression is:

[0026]

[0027] in, The x-axis is the horizontal axis. Let be a constant to be determined, and define the wake boundary position as 99.5% of the mainstream velocity.

[0028] In one embodiment, step S2 further includes the following sub-steps:

[0029] S204. Obtain representative flow field data of a typical turbine. The flow field data is obtained through experiments or high-precision CFD calculations, including pressure field, temperature field, and velocity field.

[0030] S205, Calculate the wake width The distribution of values ​​is used to extract the abscissas from the velocity field. The ordinate corresponding to the position of the tail boundary ,have to ; with the width of the trail Distribution expression The form is fitted using the least squares method to obtain the constant. , ,c;

[0031] S206. Calculate the dimensionless maximum velocity deficit of the wake. The distribution (to calculate the dimensionless maximum velocity loss) To determine the distribution, it is necessary to calculate the constants in the formula. and ),

[0032] The formula The equivalent transformation is:

[0033]

[0034] The right side of the equation and The value can be read from the velocity field obtained from experiments on typical turbines or high-precision CFD; the characteristic length L is taken as the guide vane chord length. That is, the constant to be determined... With drag coefficient The product of known quantities It indicates that the obtained By performing a least-squares fit in constant form, the result can be obtained. The value;

[0035] S207, in determining and After determining the constant in the equation, substitute it into the following formula:

[0036]

[0037] in, By approximating the absolute velocity of the turbine guide vane exit velocity triangle, we obtain the wake velocity distribution function.

[0038] S3. Construct a physical information neural network, using the Euler equation as the physical constraint and the boundary parameters of the interface between the complex flow region and the mainstream region as the boundary condition input, to train the physical information neural network to predict the flow field parameters of the mainstream region.

[0039] In one embodiment, step S3 further includes the following sub-steps:

[0040] S301. Assume the flow in the main flow region is a compressible, inviscid, unsteady ideal fluid motion, and its governing equations are expressed in the conservation form of the Euler equations as follows:

[0041]

[0042]

[0043] in, Indicates density, Represents the velocity component. Represents pressure, related to the ideal gas law. Find out, Represents total energy. It indicates the specific heat ratio.

[0044] S302, Define the neural network For parameters A deep neural network, with spatiotemporal coordinates as input. With time The output is the predicted conserved variables. ;

[0045] S303. Achieve joint constraints of physical consistency and data-driven approaches by minimizing the composite loss function to train the neural network:

[0046]

[0047] Train the neural network, wherein , , , >0 represents the weighting coefficient;

[0048] S304. Force the network output to satisfy the Euler equation by using partial differential equation loss:

[0049]

[0050] in, Configuration points for random sampling within the mainstream area;

[0051] Dirichlet-type boundary conditions are applied at the interface between the mainstream region and the complex flow region using boundary condition loss:

[0052]

[0053] in, For the first The flow field parameters calculated by the complex flow region model at each boundary point are used as the physical boundary condition input for PINN.

[0054] The initial conditional loss is:

[0055]

[0056] The data loss is:

[0057]

[0058] Data-driven constraints are applied by comparing the results with those of high-precision numerical simulations.

[0059] S305, When the total loss Less than the preset threshold Training ends when the time is right.

[0060] Understandably, PINN's input includes the initial flow parameters (including boundary and initial conditions) and geometric conditions of the dataset, and its output is the predicted values ​​of the flow parameters. During training, physical consistency of the network output is ensured through four joint constraints: Partial Differential Equation Loss (PDELoss), Boundary Condition Loss (BCLoss), Initial Condition Loss (ICLoss), and Data Loss (DataLoss). The total loss is the sum of these four losses, and training terminates when the total loss converges to below a preset threshold ε. Afterward, the boundary and geometric conditions of the turbine to be predicted are input into the trained PINN to obtain the predicted flow parameters of the main flow region. By merging the flow field predictions of the main flow region and the complex flow region, the complete turbine full-field flow parameters can be reconstructed; further extraction of the predicted pressure values ​​on the downstream blade surface completes the calculation of aerodynamic excitation.

[0061] S4. Merge the flow field parameters of the main flow region with the flow field parameters of the wake region in the complex flow region to obtain the flow parameters of the entire flow field, and calculate the aerodynamic excitation based on the extracted pressure prediction values ​​of the downstream blade surface.

[0062] In one embodiment, step S4 further includes the following sub-steps:

[0063] S401, the absolute velocity of the guide vane exit velocity triangle In the approximate replacement of the wake velocity distribution model Take the guide leaf chord length as the characteristic length ;

[0064] S402, Take a series of coordinate points ( Substituting the wake velocity distribution function into the equation, we can obtain the dimensionless velocity distribution of the wake. , thereby obtaining a prediction of the wake velocity field;

[0065] S403. Extract the velocity value at the wake boundary and use it as the boundary condition for the physical information neural network to predict the flow parameters in the main flow area.

[0066] S404. Merge the flow field prediction results of the main flow region and the wake region to obtain the predicted velocity field value of the entire flow field; calculate other flow parameters based on the predicted velocity field value, and finally perform aerodynamic excitation prediction.

[0067] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0068] In one embodiment, an AI-based rapid prediction device for turbine aerodynamic excitation is provided, which corresponds one-to-one with the AI-based rapid prediction method for turbine aerodynamic excitation described in the above embodiments. The AI-based rapid prediction device for turbine aerodynamic excitation includes:

[0069] The flow field partitioning module is used to divide the turbine flow field into a mainstream region and a complex flow region. The complex flow region includes the wake region, while the mainstream region is the inviscid flow region far away from the boundary layer and the wake region.

[0070] The complex flow region modeling module is used to establish mathematical models based on fluid dynamics boundary layer theory and high-precision numerical simulation results for complex flow regions, in order to predict the flow field parameter distribution in the wake region and determine the boundary parameters of the interface between the complex flow region and the mainstream region.

[0071] The PINN training module for the main flow region is used to construct a physical information neural network. It uses the Euler equation as a physical constraint and the boundary parameters of the interface between the complex flow region and the main flow region as boundary condition inputs to train the physical information neural network in order to predict the flow field parameters of the main flow region.

[0072] The results merging and aerodynamic excitation calculation module is used to merge the flow field parameters of the mainstream region and the flow field parameters of the wake region in the complex flow region to obtain the flow parameters of the entire flow field, and calculate the aerodynamic excitation based on the extracted pressure prediction values ​​of the downstream blade surface.

[0073] Specific limitations regarding the AI-based rapid prediction device for turbine aerodynamic excitation can be found in the limitations of the AI-based rapid prediction method for turbine aerodynamic excitation described above, and will not be repeated here. Each module in the aforementioned AI-based rapid prediction device for turbine aerodynamic excitation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0075] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A rapid prediction method for turbine aerodynamic excitation based on artificial intelligence, characterized in that, include: S1. Divide the turbine flow field into the mainstream region and the complex flow region. The complex flow region includes the wake region, while the mainstream region is the inviscid flow region far away from the boundary layer and the wake region. S2. For complex flow regions, establish a mathematical model based on fluid dynamics boundary layer theory and high-precision numerical simulation results to predict the flow field parameter distribution in the wake region and determine the boundary parameters at the interface between the complex flow region and the mainstream region. S3. Construct a physical information neural network, using the Euler equation as a physical constraint and the boundary parameters of the interface between the complex flow region and the mainstream region as boundary condition inputs, to train the physical information neural network in order to predict the flow field parameters of the mainstream region. S4. Merge the flow field parameters of the main flow region with the flow field parameters of the wake region in the complex flow region to obtain the flow parameters of the entire flow field, and calculate the aerodynamic excitation based on the extracted pressure prediction values ​​of the downstream blade surface.

2. The rapid prediction method for turbine aerodynamic excitation based on artificial intelligence according to claim 1, characterized in that, Step S2 further includes the following sub-steps: S201. Establish a velocity distribution model for the wake region of the turbine flow field, and define the mathematical expression for the wake velocity distribution of the turbine guide vane: ， in, This represents the velocity of the wake to be measured. Indicates the mainstream speed. The dimensionless maximum velocity deficit of the wreath is only related to... Related unknowns; Indicates the width of the trail, which is only related to... Related unknowns; and These represent the x-axis and y-axis, respectively, and are the input values ​​for the model; S202. Assume the dimensionless maximum velocity deficit of the wake. The distribution expression is: ， in, Indicates the width of the trail. Indicates the characteristic length; This represents the drag coefficient of the turbine guide vane; Denotes the constant to be determined; S203, Set the tail width The distribution expression is: ， in, The x-axis is the horizontal axis. Let be a constant to be determined, and define the wake boundary position as 99.5% of the mainstream velocity.

3. The rapid prediction method for turbine aerodynamic excitation based on artificial intelligence according to claim 2, characterized in that, Step S2 further includes the following sub-steps: S204. Obtain representative flow field data of a typical turbine. The flow field data is obtained through experiments or high-precision CFD calculations, including pressure field, temperature field and velocity field. S205, Calculate the wake width The distribution of values ​​is used to extract the abscissas from the velocity field. The ordinate corresponding to the position of the tail boundary ,have to ; with the width of the trail Distribution expression The form is fitted using the least squares method to obtain the constant. , ,c; S206. Calculate the dimensionless maximum velocity deficit of the wake. The distribution requires calculating the constants in the formula. and , will The equivalent transformation is: ， The right side of the equation and The value can be obtained from the velocity field obtained by experiments on typical turbines or by high-precision CFD; the characteristic length L is taken as the guide vane chord length; that is, the constant to be determined. With drag coefficient The product of known quantities It indicates that the obtained By performing a least-squares fit in constant form, the result can be obtained. The value; S207, in determining and After determining the constant in the equation, substitute it into the following formula: ， in, By approximating the absolute velocity of the turbine guide vane exit velocity triangle, we obtain the wake velocity distribution function.

4. The rapid prediction method for turbine aerodynamic excitation based on artificial intelligence according to claim 3, characterized in that, Step S3 further includes the following sub-steps: S301. Assume the flow in the main flow region is a compressible, inviscid, unsteady ideal fluid motion, and its governing equations are expressed in the conservation form of the Euler equations as follows: ， ， in, Indicates density, Represents the velocity component. Represents pressure, related to the ideal gas law. Find out, Represents total energy. Indicates specific heat ratio; S302, Define the neural network For parameters A deep neural network, with spatiotemporal coordinates as input. With time The output is the predicted conserved variables. ; S303. Achieve joint constraints of physical consistency and data-driven approaches by minimizing the composite loss function to train the neural network: ， Train the neural network, wherein , , , >0 represents the weighting coefficient; S304. Force the network output to satisfy the Euler equation by using partial differential equation loss: ， in, Configuration points for random sampling within the mainstream area; Dirichlet-type boundary conditions are applied at the interface between the mainstream region and the complex flow region using boundary condition loss: ， in, For the first The flow field parameters calculated by the complex flow region model at each boundary point are used as the physical boundary condition input for PINN. The initial conditional loss is: ， The data loss is: ， Data-driven constraints are applied by comparing the results with high-precision numerical simulations. S305, When the total loss Less than the preset threshold Training ends when the time is right.

5. The rapid prediction method for turbine aerodynamic excitation based on artificial intelligence according to claim 4, characterized in that, Step S4 further includes the following sub-steps: S401, the absolute velocity of the guide vane exit velocity triangle In the approximate replacement of the wake velocity distribution model Take the guide leaf chord length as the characteristic length ; S402, Take a series of coordinate points ( Substituting the wake velocity distribution function into the equation, we can obtain the dimensionless velocity distribution of the wake. , thereby obtaining a prediction of the wake velocity field; S403. Extract the velocity value at the wake boundary and use it as the boundary condition for the physical information neural network to predict the flow parameters in the main flow area. S404. Merge the flow field prediction results of the main flow region and the wake region to obtain the predicted velocity field value of the entire flow field; calculate other flow parameters based on the predicted velocity field value, and finally perform aerodynamic excitation prediction.

Images