Water turbine impeller blade shape optimization method combining neural network and conformal geometry

Abstract

The invention discloses a water turbine impeller blade shape optimization method combining a neural network and conformal geometry, which comprises the following steps: selecting various blade shape parameters to construct a sample set, then mapping samples to a two-dimensional plane by applying a conformal mapping technology, and calculating conformal factors; performing fluid-solid coupling analysis on the sample, and collecting displacement field data; an improved U-NET neural network of a global attention mechanism is adopted for processing, and the network learns interaction between conformal factors and displacement field data to achieve the purpose of optimizing the impeller blade shape; and refining conformal factors by using a genetic algorithm. According to the method, the relation between the water turbine impeller conformal factor and the optimization function value is analyzed through the neural network, so that the conformal factor capable of enabling the target function to reach the optimal state is recognized, and the optimal impeller blade shape is determined according to the conformal factor. And the performance of the water turbine impeller is also remarkably improved, and the water turbine impeller has remarkable significance in improving the operation efficiency and economic value of water energy conversion equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of fluid machinery structure optimization, and in particular to a method for optimizing the blade shape of a water turbine impeller by combining a neural network with conformal geometry. Background Art

[0002] As an important water energy conversion equipment, hydraulic turbines are widely used in hydropower stations around the world. Their efficiency and reliability directly affect the energy conversion efficiency and economic benefits of hydropower stations. As one of the key components of hydraulic turbines, the blade shape of hydraulic turbine blades directly determines the performance of the impeller in the flow field. Therefore, the design of hydraulic turbine impellers is crucial to improving the performance of hydraulic turbines. At present, the optimization method for impeller blade shape is often optimized based on the overall shape parameters. This method can only produce a specific type of blade shape, so there are great limitations in the optimization flexibility. Therefore, it is very necessary to propose a highly flexible impeller blade shape optimization method.

[0003] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0004] The purpose of the present invention is to provide a method for optimizing the impeller blade shape of a water turbine by combining neural networks with conformal geometry. The method of combining neural networks with conformal mapping is adopted to accurately adjust the impeller blade shape, so as to improve the working performance and operating efficiency of the water turbine.

[0005] To achieve the above object, the present invention provides a method for optimizing the blade shape of a water turbine impeller by combining a neural network with conformal geometry, comprising the following steps:

[0006] (1) Generate a turbine impeller blade shape sample set according to different blade shape parameters;

[0007] (2) transforming each leaf shape sample into a two-dimensional plane through a conformal mapping method, thereby obtaining the corresponding two-dimensional grid and conformal factor;

[0008] (3) Perform fluid-structure coupling analysis on each sample to obtain the displacement field of the sample under specific flow field conditions;

[0009] (4) Replace the smooth transition layer of the U-NET neural network with the global attention layer to obtain the global attention U-NET neural network;

[0010] (5) using the conformal factor of step (2) as input and the displacement field of step (3) as output, training the global attention U-NET neural network of step (4);

[0011] (6) selecting samples from the sample set obtained in step (1) as initial samples, extracting conformal factors corresponding to the initial samples as initial optimization data, optimizing the initial data by a genetic algorithm, and obtaining a final optimized conformal factor;

[0012] (7) The conformal factor optimized in step (6) is interpolated back to the triangular mesh through the deep neural mesh to determine the length of each edge of the mesh, thereby optimizing the blade shape of the turbine impeller.

[0013] Preferably, in the above technical solution, the leaf shape parameters in step (1) include cylindrical, twisted blade, and swept shape, and a plurality of different leaf shape parameters are set for each leaf shape.

[0014] Preferably, in the above technical solution, the joint mapping method in step (2) includes searching for an inscribed quadrilateral with the largest area in the leaf-shaped grid as a mapping fixed point, and selecting a bounded rectangular plane as a mapping target surface.

[0015] Preferably, in the above technical solution, step (2) of obtaining the conformal factor comprises:

[0016] 1) Find the inscribed quadrilateral with the largest area in the leaf-shaped grid, and set these four points as fixed points in the mapping process;

[0017] 2) The target plane of the mapping is selected as a bounded rectangular plane, the sum of the target angles of the four points in step 1) is set to 90 degrees, and the target angles of the remaining points are set to 0 degrees;

[0018] 3) mapping the leaf-shaped mesh generated by the leaf-shaped sample in step (1) by using the RICCI flow algorithm for discrete triangle meshes;

[0019] 4) The mapped grid is converted into a rectangular grid through a barycentric interpolation algorithm, and then the conformal factor is extracted on the rectangular grid.

[0020] Preferably, in the above technical solution, the flow field condition in step (3) is assumed to be a uniform incompressible flow field, the speed at which the samples are placed is 2 m / S, and the material parameters include a Poisson's ratio of 0.3 and an elastic modulus of 2.1×10^7.

[0021] Preferably, in the above technical solution, the conformal factor and the displacement field in step (5) are both expressed in matrix form, and the size of the matrix remains consistent.

[0022] Preferably, in the above technical solution, step (5) of training the global attention U-NET neural network comprises:

[0023] The total sample data set is divided into training set and test set in a ratio of 6-9:1-4; the minimum batch size for training is 30-100 samples / batch; the number of iterations is set to 1000-5000 times;

[0024] The parameter optimizer used is Adam, and an adaptive adjustment strategy is adopted to adjust the network parameters;

[0025] The droop out method is used in the training process, and the cross entropy loss function is used to evaluate the prediction accuracy of the network; based on the above method, the global attention U-NET neural network is trained, and finally a trained global attention U-NET neural network is obtained.

[0026] Preferably, in the above technical solution, the method for optimizing step (6) includes:

[0027] 1) Select several leaf shapes with the best objective function value from the sample set, and average the conformal factors of these leaf shapes to obtain the optimized initial conformal factor;

[0028] 2) Inputting the initial conformal factor obtained in step 1) into the global attention U-NET neural network trained in step (5) to obtain the predicted displacement field, and calculating the value of the current objective function based on it;

[0029] 3) The gradient of the objective function with respect to the conformal factor is obtained by the chain rule, and then the input conformal factor is updated by the genetic algorithm;

[0030] 4) Input the latest conformal factor into the trained global attention U-NET neural network to obtain the latest objective function value, and then update the conformal factor through the process of step 3), and continue the above process until the objective function converges, so as to obtain the final optimized conformal factor.

[0031] Preferably, in the above technical solution, step (7) of optimizing the turbine impeller blade shape comprises:

[0032] 1) In the two-dimensional plane of step (2), the plane triangular mesh is divided by connecting the diagonal lines on the mesh vertices, and then the final optimized conformal factor is interpolated to the vertex of each triangular mesh by an interpolation algorithm;

[0033] 2) Taking the conformal factor of each triangular mesh node in step 1) as input, the DNN is used to predict the length of each edge of the triangular mesh under the conformal factor, thereby reproducing the predicted optimal turbine blade shape;

[0034] 3) Obtaining the conformal factor of the predicted optimal turbine blade shape through the process of step S2, comparing the conformal factor with the final optimized conformal factor generated in step (6), and finally updating the network parameters of the DNN in step 2) by means of a cross entropy loss function;

[0035] 4): Loop through step 3) until the absolute value of the difference between the conformal factor corresponding to the predicted leaf shape output by DNN and the conformal factor finally optimized in step (6) is no greater than 0.001, then the leaf shape predicted by DNN can be considered to be the final optimized leaf shape.

[0036] Preferably, in the above technical solution, the global attention U-NET neural network includes the following modules:

[0037] A tuning module for replacing the smooth transition layer of the standard U-NET neural network with a global attention layer to form a global attention U-NET neural network. This improvement is intended to enhance the network’s ability to handle leaf shape optimization problems.

[0038] The input module is designed to encode the conformal factors and displacement fields into a format compatible with the global attention U-NET neural network;

[0039] The training module is responsible for training the global attention U-NET neural network with the conformal factor as input and the displacement field as output until the network can accurately predict the optimized leaf shape, thereby obtaining the trained global attention U-NET neural network;

[0040] The output module receives the optimized conformal factors from the training module, iteratively optimizes them through the genetic algorithm method until the preset optimization conditions are met, and finally outputs the parameters for impeller blade shape optimization.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) The present invention combines a neural network with a conformal geometry method for optimizing the blade shape of a turbine impeller, and replaces the smooth transition layer of the U-NET neural network with a global attention U-NET neural network of a global attention layer. The global attention layer is a network architecture that performs convolution operations on matrices in different dimensions and restores the results by superimposing them. Compared with the smooth transition layer, the global attention layer performs convolution processing on all dimensions, thereby having a stronger analytical ability for the local information of the matrix, and is particularly suitable for processing topology optimization problems.

[0043] (2) The method of the present invention adjusts the shape of the turbine impeller by adjusting the size of the conformal factor. The conformal factor is a surface metric that can accurately characterize the local curvature and ranging length of the surface. Therefore, by adjusting the conformal factor, the local surface can be adjusted. Therefore, the optimization method proposed in this application has higher flexibility.

[0044] (3) The method of the present invention uses a neural network to calculate the relationship between the conformal factor of the turbine blade shape and the displacement field, and presents the relationship in the form of an empirical formula. Based on the empirical formula, the gradient of the objective function relative to the conformal factor can be calculated, and then the conformal factor can be updated through a genetic algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of a method for optimizing a water turbine impeller blade shape by combining a neural network with conformal geometry according to the present invention;

[0046] Figure 2 is a schematic diagram of establishing a sample training database in the method according to the present invention;

[0047] Figure 3 is a flow chart of leaf conformal factor extraction in the method according to the present invention;

[0048] Figure 4 A schematic diagram showing a neural network calculation process in the method according to the present invention;

[0049] Figure 5 Schematic diagram of the structure of the global attention U-Net neural network in the method according to the present invention;

[0050] Figure 6 Detailed structural diagram of the global attention layer in the method according to the present invention is shown;

[0051] Figure 7 A process schematic diagram of a genetic algorithm in the method according to the present invention is shown;

[0052] Figure 8 FIG. 4 is a diagram showing the optimal grid restoration process in the method according to the present invention. DETAILED DESCRIPTION

[0053] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.

[0054] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0055] like Figures 1 to 8 As shown, according to a specific embodiment of the present invention, a method for optimizing the blade shape of a water turbine impeller combining a neural network and conformal geometry comprises the following steps:

[0056] S1: In the industrial design software, a sample set of turbine impeller blade shapes is generated by selecting different blade shape parameters.

[0057] In this embodiment, the main purpose of step S1 is to provide enough samples for the subsequent training of the neural network. Step S1 is specifically performed as follows: in the leaf shape design function of the AxSTREAM software, by selecting 20 common leaf shapes such as cylindrical, twisted blades, and swept shapes, each leaf shape is set with 100 different leaf shape parameters, and finally a total of 2000 different leaf shapes are generated. These leaf shapes are represented as triangular meshes in discrete space and used as data sample sets.

[0058] S2: Please refer to Figure 2 , each leaf shape sample is transformed into a two-dimensional plane through a conformal mapping method to obtain the corresponding two-dimensional grid and conformal factor.

[0059] In this embodiment, the main purpose of step S2 is to provide the input part of the training set for the subsequent training of the neural network. Step S2 can be specifically divided into the following steps:

[0060] S21: Find the inscribed quadrilateral with the largest area in the leaf-shaped grid, and set these four points as fixed points in the mapping process;

[0061] S22: The mapped target plane is selected as a bounded rectangular plane, the sum of the target angles of the four points in step S21 is set to 90 degrees, and the target angles of the remaining points are set to 0 degrees;

[0062] S23: mapping the leaf-shaped mesh generated by the leaf-shaped sample in step S1 by using the RICCI flow algorithm for discrete triangle meshes;

[0063] S24: The mapped grid is converted into a rectangular grid by a barycentric interpolation algorithm, and then a conformal factor is extracted on the rectangular grid.

[0064] S3: Perform fluid-solid coupling analysis on each sample to obtain the external force field of the sample under specific flow field conditions.

[0065] In this embodiment, the material parameters of all blades are: Poisson's ratio is 0.3; elastic modulus is 2.1*10 7 All samples were placed at a speed of 2m / S and a density of 1000KG / m 3 Numerical experiments are performed in a uniform incompressible flow field. The displacement field and elastic stiffness matrix corresponding to each sample are obtained through the experiments. The displacement field is converted into a 150x150 matrix.

[0066] S4: Replace the smooth transition layer of the U-NET neural network with a global attention layer to obtain a global attention U-NET neural network.

[0067] The present invention improves the U-NET neural network. Figure 3As shown in FIG. 1 , the present invention replaces the smooth transition layer at the bottom of the U-NET neural network with a global attention layer. The global attention layer is a network architecture that performs convolution operations on the matrix in different dimensions and then restores the results of these operations by superimposing them. Compared with the smooth transition layer, the global attention layer performs convolution processing in all dimensions, so it has a better ability to parse the local information of the matrix, and is particularly suitable for leaf shape optimization problems.

[0068] S5: Taking the conformal factor of step S2 as input and the external force field of step S3 as output, train the global attention U-NET neural network of step S4.

[0069] The conformal factor and displacement field are both represented in the form of a matrix of size 150x150. The network training conditions are: the training set and the test set are differentiated and divided from the total sample data set in a ratio of 7:3. The minimum batch of training is 50 samples / batch; the iteration round is set to 3000 times. The parameter optimizer used is Adam; and an adaptive adjustment strategy is adopted to adjust the network parameters. The droop out method is used during the training process. The cross entropy loss function is used to evaluate the prediction accuracy of the network. Based on the above method, the global attention U-NET neural network is trained, and finally a trained global attention U-NET neural network is obtained.

[0070] S6: Select samples from the sample set obtained in step S1 as initial samples, extract conformal factors corresponding to the initial samples as initial optimization data, and optimize the initial data using a genetic algorithm.

[0071] In this embodiment, the strain energy of the blade shape under given flow field conditions is used as the objective function, and the minimum strain energy is the optimization target. Figure 4 As shown, step S6 can be divided into the following small steps:

[0072] S61: Select three leaf shapes with the best objective function value from the sample set, and average the conformal factors of the three leaf shapes to obtain the optimized initial conformal factor.

[0073] S62: Input the initial conformal factor obtained in step S61 into the global attention U-NET neural network trained in step S5 to obtain the predicted displacement field, and calculate the value of the current objective function based on it.

[0074] S63: The gradient of the objective function with respect to the conformal factor is obtained by the chain rule, and then the input conformal factor is updated by a genetic algorithm.

[0075] S64: Input the latest conformal factor into the trained global attention U-NET neural network to obtain the latest objective function value, and then update the conformal factor through the process described in step S63. The above process is repeated until the objective function converges, so as to obtain the final optimized conformal factor.

[0076] S7: interpolate the conformal factor optimized in step S6 back to the triangular mesh, and then determine the length of each edge of the triangular mesh through DNN, so as to optimize the blade shape of the turbine impeller.

[0077] In this embodiment, the length of each side of the triangular mesh that can satisfy the conformal factor obtained in step S64 is found through continuous iteration of DNN, so as to finally determine the blade shape of the optimized turbine blade. Figure 5 As shown, step S7 can be divided into the following small steps:

[0078] S71: In the target plane described in step S24, the plane triangular mesh is divided by connecting the diagonal lines on the mesh vertices, and then the final optimized conformal factor is interpolated to the vertex of each triangular mesh by an interpolation algorithm.

[0079] S72: Taking the conformal factor of each triangular mesh node in step S71 as input, the DNN is used to predict the length of each edge of the triangular mesh under the conformal factor, thereby reproducing the predicted optimal turbine blade shape.

[0080] S73: Obtain the conformal factor of the predicted optimal turbine blade shape through the process described in step S2, compare the conformal factor with the final optimized conformal factor generated in step S64, and finally update the network parameters of the DNN in step S72 by means of a cross entropy loss function.

[0081] S74: Loop through step S73 until the absolute value of the difference between the conformal factor corresponding to the predicted leaf shape output by the DNN and the conformal factor generated in step S64 is no greater than 0.001, and the leaf shape predicted by the DNN can be considered to be the final optimized leaf shape.

[0082] The global attention U-NET neural network framework of the present invention is used for optimizing the blade shape of a turbine impeller, and the neural network framework includes the following modules: an adjustment module, which is used to replace the smooth transition layer of the U-NET neural network with the global attention layer to obtain the global attention U-NET neural network; an input module, which encodes the conformal factor obtained by conformal mapping of the sample in step S1 and the displacement field obtained by fluid-solid coupling analysis in step S2 to match the sizes of the input channel and the output channel preset by the global attention U-NET neural network; a training module, which trains the global attention U-NET neural network with the conformal factor as input and the displacement field as output to obtain a trained global attention U-NET neural network; an output module: a plurality of samples with better target values ​​are selected from the sample set obtained in step S1 to form an initial sample, the conformal factors corresponding to the initial samples are averaged and then input into the trained global attention U-NET neural network, and then iterates through a genetic algorithm until the optimization condition is finally met and the optimized conformal factor is output.

[0083] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A method for optimizing the blade shape of a hydraulic turbine impeller by combining neural network and conformal geometry, characterized in that: The following steps are involved: (1) Generate a turbine impeller blade shape sample set according to different blade shape parameters; (2) transforming each leaf shape sample into a two-dimensional plane through a conformal mapping method, thereby obtaining the corresponding two-dimensional grid and conformal factor; (3) Perform fluid-structure coupling analysis on each sample to obtain the displacement field of the sample under specific flow field conditions; (4) Replace the smooth transition layer of the U-NET neural network with the global attention layer to obtain the global attention U-NET neural network; (5) using the conformal factor of step (2) as input and the displacement field of step (3) as output, training the global attention U-NET neural network of step (4); (6) selecting samples from the sample set obtained in step (1) as initial samples, extracting conformal factors corresponding to the initial samples as initial optimization data, optimizing the initial data by a genetic algorithm, and obtaining a final optimized conformal factor; (7) The conformal factor optimized in step (6) is interpolated back to the triangular mesh through the deep neural mesh to determine the length of each edge of the mesh, thereby optimizing the blade shape of the turbine impeller.

2. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: Step (1) The leaf shape parameters include cylindrical, twisted blade, and swept shape, and a plurality of different leaf shape parameters are set for each leaf shape.

3. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: The common mapping method in step (2) includes searching for the largest inscribed quadrilateral in the leaf-shaped grid as a mapping fixed point, and selecting a bounded rectangular plane as a mapping target surface.

4. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 3, characterized in that: Step (2) of obtaining the conformal factor includes: 1) Find the inscribed quadrilateral with the largest area in the leaf-shaped grid, and set these four points as fixed points in the mapping process; 2) The target plane of the mapping is selected as a bounded rectangular plane; 3) mapping the leaf-shaped mesh generated by the leaf-shaped sample in step (1) by using the RICCI flow algorithm for discrete triangle meshes; 4) The mapped grid is converted into a rectangular grid through a barycentric interpolation algorithm, and then the conformal factor is extracted on the rectangular grid.

5. The method for optimizing the blade shape of a hydraulic turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: The flow field conditions in step (3) are assumed to be a uniform incompressible flow field, the sample is placed at a speed of 2 m / s, and the material parameters include a Poisson's ratio of 0.3 and an elastic modulus of 2.1×10^7.

6. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: In step (5), the conformal factor and the displacement field are both expressed in matrix form, and the size of the matrix remains consistent.

7. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: Step (5) training the global attention U-NET neural network includes: The total sample data set is divided into training set and test set in a ratio of 6-9:1-4; the minimum batch size for training is 30-100 samples / batch; the number of iterations is set to 1000-5000 times; The parameter optimizer used is Adam, and an adaptive adjustment strategy is adopted to adjust the network parameters; The droop out method is used in the training process, and the cross entropy loss function is used to evaluate the prediction accuracy of the network; based on the above method, the global attention U-NET neural network is trained, and finally a trained global attention U-NET neural network is obtained.

8. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: The method for optimizing step (6) includes: 1) Select several leaf shapes with the best objective function value from the sample set, and average the conformal factors of these leaf shapes to obtain the optimized initial conformal factor; 2) Inputting the initial conformal factor obtained in step 1) into the global attention U-NET neural network trained in step (5) to obtain the predicted displacement field, and calculating the value of the current objective function based on it; 3) The gradient of the objective function with respect to the conformal factor is obtained by the chain rule, and then the input conformal factor is updated by the genetic algorithm; 4) Input the latest conformal factor into the trained global attention U-NET neural network to obtain the latest objective function value, and then update the conformal factor through the process of step 3), and continue the above process until the objective function converges, so as to obtain the final optimized conformal factor.

9. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: Step (7) of optimizing the turbine impeller blade shape comprises: 1) In the two-dimensional plane of step (2), the plane triangular mesh is divided by connecting the diagonal lines on the mesh vertices, and then the final optimized conformal factor is interpolated to the vertex of each triangular mesh by an interpolation algorithm; 2) Taking the conformal factor of each triangular mesh node in step 1) as input, the DNN is used to predict the length of each edge of the triangular mesh under the conformal factor, thereby reproducing the predicted optimal turbine blade shape; 3) Obtaining the conformal factor of the predicted optimal turbine blade shape through the process of step S2, comparing the conformal factor with the final optimized conformal factor generated in step (6), and finally updating the network parameters of the DNN in step 2) by means of a cross entropy loss function; 4): Loop through step 3) until the absolute value of the difference between the conformal factor corresponding to the predicted leaf shape output by DNN and the conformal factor finally optimized in step (6) is no greater than 0.001, then the leaf shape predicted by DNN can be considered to be the final optimized leaf shape.

10. The method for optimizing the blade shape of a water turbine impeller by combining neural network and conformal geometry according to claim 1, characterized in that: The global attention U-NET neural network includes the following modules: A tuning module for replacing the smooth transition layer of the standard U-NET neural network with a global attention layer to form a global attention U-NET neural network. This improvement is intended to enhance the network’s ability to handle leaf shape optimization problems. The input module is designed to encode the conformal factors and displacement fields into a format compatible with the global attention U-NET neural network; The training module is responsible for training the global attention U-NET neural network with the conformal factor as input and the displacement field as output until the network can accurately predict the optimized leaf shape, thereby obtaining the trained global attention U-NET neural network; The output module receives the optimized conformal factors from the training module, iteratively optimizes them through the genetic algorithm method until the preset optimization conditions are met, and finally outputs the parameters for impeller blade shape optimization.

Images