Method and device for generating simulation model of wind turbine component

Abstract

The application discloses a wind turbine component simulation model generation method and device, and relates to the technical field of simulation. The method is as follows: based on a pre-established basic three-dimensional model, a design variable is determined; according to the design variable, a target sample space is obtained; based on each sample point in the target sample space, a corresponding model sample is generated; based on the finite element analysis result of the model sample, a finite element graph carrying a stress value label is determined; based on the finite element graph, a corresponding training sample is generated; and the training sample is used to train a graph neural network model to obtain a target simulation model. The method uses the design variable to generate the training sample, ensures that the training data required in the model training process is reliable and sufficient, generates the training sample in combination with the finite element graph, makes the target simulation model different from a traditional neural network model, realizes direct prediction of a stress field simulation result from geometric input, and makes the target simulation model meet the simulation requirements of wind turbine components.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of simulation, in particular to a wind turbine component simulation model generation method and device. BACKGROUND

[0002] In the development process of modern wind turbines, computer-aided engineering technology has become an important tool for design and optimization. In the structural strength checking of wind turbine components, finite element technology is widely used, forming a complete simulation process. Due to the complexity of the modeling process and the long solution time, the iterative cycle of the design process is prolonged, which becomes a bottleneck for rapid iteration of the model. In recent years, AI (Artificial Intelligence) technology has made significant progress in various fields, especially in data processing and pattern recognition. In the engineering field, artificial intelligence is mostly used in the development of various proxy models to replace conventional finite element simulation and improve iteration speed. Such proxy models are usually designed based on full connection layers to realize the prediction of regression problems from parameterized input to parameterized output.

[0003] However, the proxy model only uses engineering problems that can be parameterized with input and output, and cannot directly predict physical quantities related to spatial position (such as stress contour, strain contour, etc.) from geometric input, so it cannot replace the wind turbine finite element simulation. And for wind turbines, due to the flexible structure deformation mode of the components and the complex load, the training set for training the AI model is huge, which brings additional challenges to the calculation resources and model training. SUMMARY

[0004] The present application provides a wind turbine component simulation model generation method and device, aiming to realize the simulation of wind turbine components.

[0005] In order to achieve the above purpose, the present application provides the following technical solutions:

[0006] A wind turbine component simulation model generation method, comprising:

[0007] Based on the pre-established basic three-dimensional model, determine the design variables;

[0008] According to the design variables, obtain a target sample space; the target sample space includes sample points for representing structural changes and load changes of the basic three-dimensional model;

[0009] Based on each sample point in the target sample space, generate a corresponding model sample; the model sample is a component model after structural change and load change;

[0010] Based on the finite element analysis result of the model sample, a finite element graph carrying a stress value label is determined; the stress value label is used to represent the stress value of each vertex in the finite element graph;

[0011] Based on the finite element graph, a corresponding training sample is generated;

[0012] The training sample is used to train a graph neural network model to obtain a target simulation model; the target simulation model takes a geometric model and global attributes as input and outputs a stress field simulation result of the geometric model, the stress field simulation result including a stress field distribution of the geometric model, and the stress field distribution including a cloud chart composed of stress values of each vertex.

[0013] Optionally, the design variables include geometric design variables and load design variables, and the design variables are determined based on a pre-established basic three-dimensional model, including:

[0014] A two-dimensional grid is used to establish a basic three-dimensional model of a sample wind turbine component, and a structure change region of the sample wind turbine component is determined;

[0015] A control body deformation technique is used to change the structure characteristics of the basic three-dimensional model in the structure change region to obtain a structure-changed part model;

[0016] Based on the deformation mode of the structure-changed part model, a corresponding geometric design variable is established;

[0017] Based on the load of the structure-changed part model, a corresponding load design variable is established.

[0018] Optionally, according to the design variables, a target sample space is obtained, including:

[0019] The geometric design variables and the load design variables corresponding to a plurality of structure-changed part models are obtained;

[0020] A space filling method is used to generate a first sample point corresponding to each geometric design variable;

[0021] For each first sample point, based on the stress of the first sample point as a target response, an interaction analysis of each geometric design variable is performed in combination with an experimental design method to obtain an interaction level of each geometric design variable;

[0022] According to the interaction level of each geometric design variable, each geometric design variable is recombined to obtain a third sample point after structure change;

[0023] The space filling method is used to generate a second sample point corresponding to each of the load design variables;

[0024] For each of the second sample points, a main effect analysis is performed on each of the load design variables based on the stress of the second sample point as a target response in combination with an experimental design method to obtain an effect level of each of the load design variables;

[0025] According to the effect level of each of the load design variables, each of the load design variables is recombined to obtain a fourth sample point after load change;

[0026] Based on the third sample point and the fourth sample point, a target sample space is generated.

[0027] Optionally, based on the finite element analysis result of the model sample, a finite element graph carrying a stress value label is determined, including:

[0028] Performing finite element analysis on the model sample to obtain a corresponding finite element analysis result; the finite element analysis result includes a finite element model and a corresponding global attribute; the finite element model includes a plurality of nodes, and a connection relationship and a stress value between each of the nodes;

[0029] Based on the nodes as vertices, the connection relationship as edges, and the stress value of the nodes as the stress value label of the vertices, an initial finite element graph is constructed;

[0030] An edge corresponding to a vertex pair meeting a preset condition in the initial finite element graph is added; the preset condition is that both vertices in the vertex pair are in a designated grid encryption area, and a spatial distance between the two vertices meets a preset distance threshold;

[0031] A main vertex corresponding to the global attribute is generated, and the main vertex is connected with all vertices of the initial finite element graph to obtain a finite element graph carrying a stress value label.

[0032] Optionally, based on the finite element graph, a corresponding training sample is generated, including:

[0033] Determining a vertex attribute tensor corresponding to each vertex in the finite element graph, and a global attribute tensor corresponding to a main vertex;

[0034] The vertex attribute tensors and the global attribute tensor are integrated to obtain a graph attribute tensor;

[0035] Based on each edge in the finite element graph and a corresponding vertex pair, an adjacency matrix is determined;

[0036] Based on the graph attribute tensor and the adjacency matrix, a corresponding training sample is generated.

[0037] Optionally, the graph neural network model adopts an encoder-decoder architecture, the encoder-decoder architecture comprising a plurality of network layers, the plurality of network layers comprising two multi-layer perceptrons and three graph convolution layers.

[0038] Optionally, in the process of training the graph neural network model, an Adam optimizer is adopted, the mean square error of the predicted node stress value and the node stress value is calculated, the gradient direction at which the mean square error decreases fastest is solved, and the weights of the graph convolution layers and the fully connected layers in the multi-layer perceptrons are updated according to the gradient direction.

[0039] A wind turbine component simulation model generation device, comprising:

[0040] A variable determination unit is configured to determine design variables based on a pre-established basic three-dimensional model.

[0041] A space determination unit is configured to obtain a target sample space according to the design variables; the target sample space comprises sample points for representing structural changes and load changes of the basic three-dimensional model.

[0042] A model generation unit is configured to generate corresponding model samples based on each sample point in the target sample space; the model samples are part models after structural changes and load changes.

[0043] An original drawing determination unit is configured to determine a finite element graph carrying a stress value label based on a finite element analysis result of the model samples; the stress value label is used to represent stress values of each vertex in the finite element graph.

[0044] A sample generation unit is configured to generate corresponding training samples based on the finite element graph.

[0045] A model training unit is configured to train a graph neural network model using the training samples to obtain a target simulation model; the target simulation model takes a geometric model and global attributes as input and outputs a stress field simulation result of the geometric model, the stress field simulation result comprises a stress field distribution of the geometric model, and the stress field distribution comprises a cloud chart composed of stress values of each vertex.

[0046] A storage medium comprising a stored program, wherein the program is run by a processor to perform the wind turbine component simulation model generation method.

[0047] An electronic device, comprising: a processor, a memory and a bus; the processor and the memory are connected through the bus;

[0048] The memory is configured to store a program, and the processor is configured to run the program, wherein the program, when executed by the processor, implements the wind turbine component simulation model generation method.

[0049] The technical scheme provided in the application determines a design variable based on a pre-established basic three-dimensional model. According to the design variable, a target sample space is obtained. Based on each sample point in the target sample space, a corresponding model sample is generated. Based on the finite element analysis result of the model sample, a finite element graph carrying a stress value label is determined. Based on the finite element graph, a corresponding training sample is generated. The training sample is used to train a graph neural network model to obtain a target simulation model. The application generates a training sample by using a design variable, ensures that the training data required in the model training process is reliable and sufficient, and generates a training sample in combination with a finite element graph, so that the target simulation model can be distinguished from a traditional neural network model, the stress field simulation result is directly predicted from the geometric input, and the target simulation model meets the simulation requirements of wind turbine components. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Figure 1 A flowchart of a wind turbine component simulation model generation method provided by an embodiment of the application is shown in the figure;

[0052] Figure 2 A flowchart of another wind turbine component simulation model generation method provided by an embodiment of the application is shown in the figure;

[0053] Figure 3 A flowchart of another wind turbine component simulation model generation method provided by an embodiment of the application is shown in the figure;

[0054] Figure 4 A flowchart of another wind turbine component simulation model generation method provided by an embodiment of the application is shown in the figure;

[0055] Figure 5 A flowchart of another wind turbine component simulation model generation method provided by an embodiment of the application is shown in the figure;

[0056] Figure 6 An architectural diagram of a wind turbine component simulation model generation device provided by an embodiment of the application is shown in the figure;

[0057] Figure 7 A schematic diagram of an encoder-decoder architecture provided in an embodiment of this application;

[0058] Figure 8 A schematic diagram illustrating a model training method provided in an embodiment of this application;

[0059] Figure 9 A schematic diagram of a control body provided in an embodiment of this application;

[0060] Figure 10 This is a schematic diagram of another control body provided in an embodiment of this application;

[0061] Figure 11 A schematic diagram of an initial finite element diagram provided for an embodiment of this application;

[0062] Figure 12 A schematic diagram of a finite element diagram provided for an embodiment of this application;

[0063] Figure 13 This is a schematic diagram illustrating the acquisition of graph attribute tensors provided in an embodiment of this application. Detailed Implementation

[0064] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0066] like Figure 1 The diagram shown is a flowchart illustrating a method for generating simulation models of wind turbine components according to an embodiment of this application, including the following steps.

[0067] S101: Determine design variables based on the pre-established basic 3D model.

[0068] The definition of the design variable is an independent parameter that needs to be determined in advance in the design process, and is used to construct the model sample.

[0069] Optionally, the design variable includes a geometric design variable and a load design variable. Based on a pre-established basic three-dimensional model, the implementation process of determining the design variable can be referred to as the steps shown in Figure 2 and the corresponding explanation.

[0070] S102: Obtain a target sample space according to the design variable.

[0071] The target sample space includes sample points for representing structural changes and load changes of the basic three-dimensional model.

[0072] Optionally, the implementation process of obtaining the target sample space according to the design variable can be referred to as the steps shown in Figure 3 and the corresponding explanation.

[0073] S103: Generate a corresponding model sample based on each sample point in the target sample space.

[0074] The model sample is a part model after structural changes and load changes.

[0075] S104: Determine a finite element graph carrying a stress value label based on the finite element analysis result of the model sample.

[0076] The stress value label is used to represent the stress value of each vertex in the finite element graph.

[0077] Optionally, the implementation process of determining the finite element graph carrying the stress value label based on the finite element analysis result of the model sample can be referred to as the steps shown in Figure 4 and the corresponding explanation.

[0078] S105: Generate a corresponding training sample based on the finite element graph.

[0079] The finite element graph carries a stress value label, and therefore the generated training sample also carries a stress value label.

[0080] Optionally, the implementation process of generating the corresponding training sample based on the finite element graph can be referred to as the steps shown in Figure 5 and the corresponding explanation.

[0081] S106: Train the graph neural network model using the training sample to obtain a target simulation model.

[0082] The target simulation model takes the geometric model and global attributes as input and outputs the stress field simulation results of the geometric model. The stress field simulation results include the stress field distribution of the geometric model, which includes a cloud map composed of the stress values ​​of each vertex.

[0083] Optionally, the graph neural network model adopts an encoder-decoder architecture, which includes multiple network layers, including two multilayer perceptrons and three graph convolutional layers.

[0084] Optionally, during the training of the graph neural network model, the Adam optimizer is used to calculate the mean square error of the predicted node stress value and the node stress value, and to solve for the gradient direction in which the mean square error decreases the fastest. The weights of the graph convolutional layer and the fully connected layer in the multilayer perceptron are updated according to the gradient direction.

[0085] In some examples, the working principle of a multilayer perceptron (MLP) mainly involves two processes: forward propagation of the signal and backward propagation of the error.

[0086] See in some examples Figure 7 The encoder-decoder architecture shown consists of a multilayer perceptron, three graph convolutional layers, and another multilayer perceptron. The number of neurons (E) in each layer satisfies E > 3*n / 700 for the number of grid nodes (n). An additional dropout layer with a dropout rate of 0.1 is added after each graph convolutional layer. Data normalization is performed before the first multilayer perceptron, using different normalization strategies for different dimensions of the vertex attribute tensor. Except for the last fully connected layer, all other layers use the ReLU activation function.

[0087] In some examples, when starting the graph neural network model, the training and validation sets are randomly divided at a ratio of 0.85 / 0.15. The order of the training sets is randomly shuffled each training iteration, and an early stopping strategy is implemented. During training, the Adam optimizer is used. The gradient direction with the fastest descent is calculated by determining the mean squared error between the predicted node stress value and the node stress value. The weights of the graph convolutional layers and the fully connected layers in the multilayer perceptron are updated accordingly. The initial learning rate lr0 used to solve for the gradient satisfies lr0 = 0.0005 * bs, and the training batch bs is the number of samples required for each gradient calculation by the graph neural network model. The learning rate uses a fixed step decay strategy, satisfying the relationship: lrk+1 = lrk * decay_rate. When k%step == 0, step is the decay step size, which satisfies epoch_num / 8 <= step <= epoch_num / 15. k is the current training epoch, decay_rate is 0.96, and epoch_num is the total number of training epochs. After 5000 training rounds, the model with the best performance on the validation set is selected, saved, and used for subsequent predictions.

[0088] It is important to note that the convolution operation in a graph convolutional neural network is different from that in a traditional convolutional neural network: since the neighbor information of a node is uncertain, the update of node attributes depends on the adjacency matrix generated from the finite element mesh, rather than the directional sweep of the convolution kernel. How to construct this adjacency matrix and update the node attributes so that it learns the intrinsic relationship of unstructured data is the point of difference between the graph neural network model in this embodiment and other neural network models.

[0089] It should be noted that in the application scenario of the target simulation model, users only need to input the geometric model and global attributes (such as boundary conditions, material properties, etc.) into the target simulation model. The target simulation model, combined with automated mesh generation technology, transforms the geometric model and global attributes into graph attribute tensors and adjacency matrices, which are then input into the target simulation model to perform stress field prediction and obtain stress field simulation results. After the prediction is completed, the graph label tensor of the stress field simulation results can also be visualized to draw a 3D stress cloud map, which facilitates users' assessment of structural strength.

[0090] Combination Figures 2-5 The method shown, the method for generating the target simulation model in the embodiments of this application, can be summarized as follows: Figure 8As shown, and compared with a conventional finite element model, the target simulation model has the following advantages: (1) fast prediction speed of the stress field, the result of 4 hours of conventional finite element calculation can be obtained in only tens of seconds, facilitating designers to quickly iterate the model in the early development stage; (2) no need to perform complex modeling, only the geometry needs to be input to obtain the stress field result, reducing the technical threshold of the user; (3) good precision and generalization ability, capable of predicting different topological structures of the geometry and load, with an average error of <10%.

[0091] In some examples, Figures 3-5 The method shown can be regarded as a training sample generation method, and also has the following advantages: (1) solving the dependence on historical data when training the model, even when designing a completely new product, the training set can be trained based on the basic three-dimensional model to quickly guide the design; (2) through DOE analysis of the design variables, the sample space is maximized while reflecting the sample changes, ensuring the prediction accuracy while improving the training efficiency.

[0092] The above-mentioned processes S101-S106 use the design variables to generate training samples, ensure that the training data required in the model training process is reliable and sufficient, and generate training samples in combination with the finite element diagram, so that the target simulation model can be distinguished from the traditional neural network model, realize direct prediction of the physical quantity (i.e. stress field simulation result) related to the spatial position from the geometry input, and make the target simulation model meet the simulation requirements of the wind turbine components.

[0093] As Figure 2 As shown, the flowchart of another method for generating a wind turbine component simulation model provided by the embodiment of the present application, comprising the following steps.

[0094] S201: Use a two-dimensional grid to establish a basic three-dimensional model of the sample wind turbine component, and determine the structure change area of the sample wind turbine component.

[0095] Among them, the two-dimensional grid can adopt the form of triangular facets, specifically, the original three-dimensional model of the sample wind turbine component is converted into the form of triangular facets (such as STL format) using a structural design software to obtain the basic three-dimensional model, facilitating subsequent free deformation of the basic three-dimensional model. Generally speaking, if the original three-dimensional model is directly deformed, deformation failure may occur due to the interference of different parameters.

[0096] It should be noted that the structure change area can be understood as: determining which structural features of the sample wind turbine component need to be changed, such as the size of the hole, the wall thickness, the size of the chamfer, etc. Since the deformation mode of such structural features is complex and cannot be directly parameterized, it needs to be processed in combination with the control body deformation technology.

[0097] S202: using a control body deformation technique to change the structural features of the basic three-dimensional model in the structural change region to obtain a part model after structural change.

[0098] The control body deformation technique can use free-form surface deformation, which is a geometric modeling technique for deforming a three-dimensional model through a control point grid system. The volume of the part model feature change created by human beings can be regarded as a corresponding control body. The control body is any fixed and invariable volume in space with a certain scale (finite value or infinitesimal).

[0099] S203: establishing a corresponding geometric design variable based on the deformation mode of the part model after structural change.

[0100] The geometric design variable corresponding to the deformation mode of the part model after structural change can be recorded by the three-dimensional modeling software as a design variable in the design environment of the three-dimensional modeling software, as shown in Figure 9 , the shape of the control body deformation hole edge rounding (which can be understood as a deformation body) is used to control the size of the rounding by dragging the anchor point on the deformation body. The deformation can be recorded by the three-dimensional modeling software as a design variable, and the deformation amplitude can be adjusted, as shown in Figure 10 . Similarly, design variables can be created for other structural features by this method, and a plurality of design variables with different values can be generated by an external script, and the design variables with different values are referred to as geometric design variables.

[0101] S204: establishing a corresponding load design variable based on the load of the part model after structural change.

[0102] The load is an external force and other factors that cause internal force and deformation of the structure or component.

[0103] In some examples, the load is applied to the finite element model corresponding to the part model after structural change, and the load of the part model after structural change is obtained, so that the corresponding load design variable is established based on the load of the part model after structural change.

[0104] The above-mentioned S201-S204 process can use the basic three-dimensional model of the sample wind turbine part to obtain corresponding geometric design variables and load design variables, and provide effective data support for subsequent model sample establishment.

[0105] As shown in Figure 3 , another flowchart of a wind turbine part simulation model generation method provided by the embodiment of the application is shown, which includes the following steps.

[0106] S301: Obtain geometric design variables and load design variables corresponding to a plurality of part models after structural changes.

[0107] S302: Generate first sample points corresponding to each geometric design variable using a space-filling method.

[0108] The essence of the space-filling method is a hierarchical sampling method, such as Latin hypercube sampling. This method can ensure that sample points are uniformly distributed in multi-dimensional space, and different sampling densities can be set in different dimensions to make sample points representative.

[0109] In some examples, the first sample points refer to samples with different geometric design variable values generated using the space-filling sampling method. After finite element calculation, corresponding stress value labels are obtained and converted into a graph form as a training set to facilitate model training.

[0110] It should be noted that the space-filling method is used to generate first sample points corresponding to geometric design variables, which can obtain a sample space composed of geometric components after structural changes of different degrees. The sample space is composed of a plurality of first sample points.

[0111] S303: For each first sample point, based on the stress of the first sample point as the target response, and combining the experimental design method, the interaction of each geometric design variable is analyzed to obtain the interaction level of each geometric design variable.

[0112] The DOE (DESIGN OF EXPERIMENT, experimental design) method is a method for efficiently determining the relationship between influencing factors and response variables by systematically planning experimental schemes.

[0113] In some examples, the DOE method is used to study whether different levels of variable A have an impact on variable B, with maximum stress as the response. If the maximum stress of the part is independent of the level of B when A is at level 1 or 2, it is considered that A and B have no interaction.

[0114] In some examples, the interaction level can be used to evaluate the strength of the interaction to achieve classification of the geometric design variables. The higher the interaction level, the stronger the interaction, and the lower the interaction level, the weaker the interaction.

[0115] S304: According to the interaction level of each geometric design variable, each geometric design variable is recombined to obtain third sample points after structural changes.

[0116] Wherein, according to the interaction level of each geometric design variable, each geometric design variable can be classified, for the design variable with weak interaction, the sample set is constructed by adopting the synchronous increase and decrease mode, and for the design variable with strong interaction, the sample set is constructed by adopting the orthogonal design mode, and the sample set comprises the third sample points.

[0117] It should be noted that the interaction refers to the phenomenon that the difference between the reaction quantities of each level of a factor changes with different levels of other factors.

[0118] In some examples, the principle of synchronous increase and decrease is that if the interaction of each design variable is weak, the influence of the combination mode of different levels on the response can be ignored, and only one combination mode is considered, that is, synchronous increase and decrease.

[0119] In some examples, the principle of orthogonal design is that representative points are selected from the experimental design of all possible combination modes of the levels of the geometric design variables for experiment, and the points cover the combination modes of different levels of each set of design variables and can represent the interaction between the geometric design variables.

[0120] In possible embodiments, it is assumed that there are five geometric design variables, A, B, C, D, E, and F, the variable ABC has weak interaction, and the variable DE has strong interaction, under the premise of considering two interaction levels, the final sample points are shown in Table 1.

[0121] Table 1

[0122]

[0123] In the content shown in Table 1 above, level 1 and level 2 both represent the interaction level, and level 2 is higher than level 1, 1, 2, 3, and 4 are the third sample points.

[0124] S305: Using a space-filling method, the second sample points corresponding to each load design variable are generated.

[0125] Wherein, the second sample points refer to samples with different load design variable values generated by using the space-filling sampling method, and the samples are subjected to finite element calculation to obtain corresponding stress value labels and are converted into a graph form as a training set to facilitate model training.

[0126] S306: For each second sample point, based on the stress of the second sample point as the target response, the main effect analysis of each load design variable is performed by combining the experimental design method, to obtain the effect level of each load design variable.

[0127] Wherein, the load effect refers to the internal force or displacement of the structure under the action of a certain load.

[0128] S307: recombine each load design variable according to the effect level of each load design variable to obtain a fourth sample point after the load is changed.

[0129] wherein, based on the effect level of each load design variable, the sample point distribution on the corresponding dimension satisfies 1.1*(N+1)*(N+2)*(C_rate*2), N is the total number of load design variables, and C_rate is the contribution of the load design variable to the response, which can be obtained by solving the eigenvalue of the normalized covariance matrix.

[0130] In some examples, N load design variables and a plurality of geometric design variables are combined to obtain the fourth sample point using a space-filling method, or N+1 design variables are combined to obtain the fourth sample point.

[0131] S308: generate a target sample space based on the third sample point and the fourth sample point.

[0132] wherein, the third sample point is not additionally encrypted.

[0133] The above-mentioned S301-S308 shows the flow, which uses a space-filling method and an experimental design method to combine geometric design variables and load design variables to obtain a target sample space.

[0134] As shown in Figure 4 , another flowchart of a wind turbine component simulation model generation method provided by the embodiments of the present application is shown, which includes the following steps.

[0135] S401: perform finite element analysis on the model sample to obtain the corresponding finite element analysis result.

[0136] wherein, the finite element analysis result includes a finite element model and corresponding global attributes, and the finite element model includes a plurality of nodes and the connection relationship and stress value between each node.

[0137] In some examples, the finite element model relies on the spatial coordinates and connection relationship of the nodes to predict the stress value of the nodes, and the spatial coordinates of the nodes are marked as corresponding attributes.

[0138] S402: construct an initial finite element graph based on nodes as vertices, connection relationship as edges, and stress value of nodes as stress value label of vertices.

[0139] wherein, the initial finite element graph is a kind of non-euclidean data, has irregular structure, and the neighbor number of each basic unit is uncertain, in the embodiment, due to the complexity of the spatial structure such as stress nephogram and boundary condition, it is also transformed into the form of graph for processing.

[0140] S403: An edge corresponding to a vertex pair meeting a preset condition in the initial finite element graph is added.

[0141] The preset condition is that both vertices in the vertex pair are in a designated mesh encryption region, and a spatial distance between the two vertices meets a preset distance threshold.

[0142] In some examples, referring to FIG. 5, a designated encryption region is a region composed of 3, 5, 6, 7, and 8, and the vertex pairs meeting the preset condition include 2 and 8, and 5 and 8. Therefore, an edge of 2 and 8 and an edge of 5 and 8 are added. Figure 11 It should be noted that adding an additional edge to the vertex pair meeting the preset condition can achieve the transmission of node information and improve the prediction accuracy of the graph neural network model on the encryption region.

[0143] S404: A main vertex corresponding to a global attribute is generated, and the main vertex is connected with all vertices of the initial finite element graph to obtain a finite element graph carrying a stress value label.

[0144] The global attribute includes but is not limited to boundary conditions, material properties, and mesh density. The main vertex is used to separately represent the global attribute, and the main vertex is connected with all vertices to transmit the global attribute to all vertices.

[0145] In some examples, referring to FIG. 6, M is a main vertex, and 1-6 are vertices.

[0146] Figure 12 The above-mentioned S401-S404 process can generate a finite element graph carrying a stress value label based on the finite element analysis result of the model sample.

[0147] As shown in FIG. 7, another flowchart of a wind turbine component simulation model generation method provided by an embodiment of the present application is shown, which includes the following steps.

[0148] As shown in FIG. 7, another flowchart of a wind turbine component simulation model generation method provided by an embodiment of the present application is shown, which includes the following steps. Figure 5 S501: Determine a vertex attribute tensor corresponding to each vertex in the finite element graph and a global attribute tensor corresponding to a main vertex.

[0149] The vertex attribute tensor and the global attribute tensor can be determined by a method of filling a gradient influence minimum value to the same dimension.

[0150] S502: Synthesize each vertex attribute tensor and the global attribute tensor to obtain a graph attribute tensor.

[0151] As shown in FIG. 8, each vertex attribute tensor and the global attribute tensor are synthesized to obtain a graph attribute tensor.

[0152] Figure 13 As shown in FIG. 8, each vertex attribute tensor and the global attribute tensor are synthesized to obtain a graph attribute tensor. ​​

[0153] S503: Determine an adjacency matrix based on each edge and the corresponding vertex pair in the finite element graph.

[0154] The adjacency matrix is stored in the form of an edge vector, so that the adjacency matrix can be represented in a sparse form, thereby maximizing the reduction of memory occupation and improving the training efficiency of the graph neural network model.

[0155] S504: Generate a corresponding training sample based on the graph attribute tensor and the adjacency matrix.

[0156] The training sample also carries a stress value label.

[0157] The above S501-S504 shows the flow, which can obtain the corresponding training sample based on the finite element graph.

[0158] As shown in the following, it is an architecture schematic diagram of a wind turbine component simulation model generation device provided by an embodiment of the present application, which comprises the following units. Figure 6 The variable determination unit 100 is configured to determine the design variable based on the pre-established basic three-dimensional model.

[0159] Optionally, the design variable includes a geometric design variable and a load design variable, and the variable determination unit 100 is specifically configured to: establish a basic three-dimensional model of a sample wind turbine component by using a two-dimensional grid, and determine a structure change area of the sample wind turbine component; use a control body deformation technology to change the structure characteristics of the basic three-dimensional model in the structure change area to obtain a part model after structure change; establish a corresponding geometric design variable based on the deformation mode of the part model after structure change; and establish a corresponding load design variable based on the load of the part model after structure change.

[0160] The space determination unit 200 is configured to obtain a target sample space according to the design variable; the target sample space includes a sample point for representing the structure change and load change of the basic three-dimensional model.

[0161]

[0162] ​Optionally, the space determining unit 200 is specifically configured to: acquire the geometric design variables and the load design variables corresponding to the plurality of part models after the structure is changed; generate, by using a space filling method, a first sample point corresponding to each geometric design variable; for each first sample point, based on stress of the first sample point as a target response, in combination with an experimental design method, perform interaction analysis on each geometric design variable to obtain an interaction level of each geometric design variable; recombine each geometric design variable according to the interaction level of each geometric design variable to obtain a third sample point after the structure is changed; generate, by using the space filling method, a second sample point corresponding to each load design variable; for each second sample point, based on stress of the second sample point as a target response, in combination with the experimental design method, perform main effect analysis on each load design variable to obtain an effect level of each load design variable; recombine each load design variable according to the effect level of each load design variable to obtain a fourth sample point after the load is changed; and generate a target sample space based on the third sample point and the fourth sample point.

[0163] The model generating unit 300 is configured to generate a corresponding model sample based on each sample point in the target sample space; and the model sample is a part model after the structure and the load are changed.

[0164] The original graph determining unit 400 is configured to determine a finite element graph carrying a stress value label based on a finite element analysis result of the model sample; and the stress value label is used to represent a stress value of each vertex in the finite element graph.

[0165] Optionally, the original graph determining unit 400 is specifically configured to: perform finite element analysis on the model sample to obtain a corresponding finite element analysis result; the finite element analysis result includes a finite element model and a corresponding global attribute; the finite element model includes a plurality of nodes, a connection relationship between each node, and a stress value; based on the nodes as vertices, the connection relationship as edges, and the stress value of the nodes as stress value labels of the vertices, an initial finite element graph is constructed; an edge corresponding to a vertex pair meeting a preset condition in the initial finite element graph is added; the preset condition is that both vertices in the vertex pair are in a designated grid densification area, and a spatial distance between the two vertices meets a preset distance threshold; a main vertex corresponding to the global attribute is generated, and the main vertex is connected with all vertices of the initial finite element graph to obtain the finite element graph carrying the stress value label.

[0166] The sample generating unit 500 is configured to generate a corresponding training sample based on the finite element graph.

[0167] Optionally, the sample generation unit 500 is specifically configured to: determine a vertex attribute tensor corresponding to each vertex in the finite element graph, and a global attribute tensor corresponding to the main vertex; synthesize the vertex attribute tensor and the global attribute tensor to obtain a graph attribute tensor; determine an adjacency matrix based on each edge in the finite element graph and the corresponding vertex pair; and generate a corresponding training sample based on the graph attribute tensor and the adjacency matrix.

[0168] The model training unit 600 is configured to train the graph neural network model using the training sample to obtain a target simulation model; the target simulation model takes the geometric model and the global attribute as input and outputs a stress field simulation result of the geometric model, and the stress field simulation result includes a stress field distribution of the geometric model, and the stress field distribution includes a cloud picture formed by stress values of each vertex.

[0169] Optionally, the graph neural network model adopts an encoder-decoder architecture, and the encoder-decoder architecture includes a plurality of network layers, and the plurality of network layers include two multilayer perceptrons and three graph convolution layers.

[0170] Optionally, the model training unit 600 adopts an Adam optimizer in the process of training the graph neural network model, calculates the mean square error of the predicted node stress value and the node stress value, solves the gradient direction at which the mean square error decreases fastest, and updates the weights of the graph convolution layer and the fully connected layer in the multilayer perceptron according to the gradient direction.

[0171] The above-mentioned various units generate training samples using design variables to ensure that the training data required in the model training process is reliable and sufficient, and generate training samples in combination with finite element graphs, so that the target simulation model can be distinguished from traditional neural network models, directly predict physical quantities related to spatial positions (i.e., stress field simulation results) from geometric inputs, and make the target simulation model meet the simulation requirements of wind turbine components.

[0172] The application also provides a computer-readable storage medium, which includes a stored program, wherein the program executes the wind turbine component simulation model generation method provided by the application.

[0173] The application also provides an electronic device, which includes a processor, a memory and a bus. The processor is connected to the memory through the bus, the memory is used to store a program, and the processor is used to run the program, wherein the program runs to execute the wind turbine component simulation model generation method provided by the application.

[0174] While several inventive embodiments have been described above, it should be appreciated that many modifications can be made of these embodiments in light of the above disclosure. Therefore, the disclosed embodiments are not intended to limit the scope of the application to the particular embodiments disclosed herein but can be practiced with modifications and changes by those having ordinary skill in the art without departing from the scope of the following claims.

[0175] The above description is merely illustrative of the application and the application should not be limited to the specific embodiments that have been described, which are to be regarded as illustrative rather than restrictive. Variations and modifications and other implementations can be possible. The order or sequence of any process or method can be different from that as described. Various "steps" as recited in the process or method claims can be carried out simultaneously or in another order than as described. Suitable results can be achieved with the described technology even when deviations are made from the specific arrangements described. It is intended that the specification and examples be considered as exemplary only, with a true scope of the application being indicated by the following claims.

Claims

1. A method for generating simulation models of wind turbine components, characterized in that, include: Based on a pre-established basic 3D model, design variables are determined; Based on the design variables, the target sample space is obtained; The target sample space includes sample points used to characterize structural and load changes in the basic three-dimensional model; Based on each sample point in the target sample space, a corresponding model sample is generated; the model sample is a part model after structural and load changes. Based on the finite element analysis results of the model sample, a finite element diagram carrying stress value labels is determined; The stress value label is used to characterize the stress value of each vertex in the finite element diagram; Based on the finite element diagram, corresponding training samples are generated; The graph neural network model is trained using the training samples to obtain the target simulation model; The target simulation model takes a geometric model and global attributes as input and outputs the stress field simulation results of the geometric model. The stress field simulation results include the stress field distribution of the geometric model, which includes a cloud map composed of the stress values ​​of each vertex. The global attributes include a combination of boundary conditions, material properties, and mesh density.

2. The method according to claim 1, characterized in that, The design variables include geometric design variables and load design variables. Based on a pre-established basic three-dimensional model, the design variables are determined, including: A basic three-dimensional model of the sample wind turbine components was established using a two-dimensional mesh, and the structural change areas of the sample wind turbine components were determined. Within the structural change area, controlled volume deformation technology is used to modify the structural features of the basic three-dimensional model to obtain a part model with structural changes. Based on the deformation mode of the part model after the structural change, establish the corresponding geometric design variables; Based on the loads of the component model after the structural changes, corresponding load design variables are established.

3. The method according to claim 2, characterized in that, Based on the design variables, the target sample space is obtained, including: Obtain the geometric design variables and load design variables corresponding to the multiple structurally modified part models; Using the space filling method, the first sample point corresponding to each of the geometric design variables is generated; For each of the first sample points, based on the stress of the first sample point as the target response, and combined with the experimental design method, an interaction analysis is performed on each of the geometric design variables to obtain the interaction level of each of the geometric design variables; Based on the interaction level of each geometric design variable, the geometric design variables are recombined to obtain the third sample point after the structural change. Using the space filling method, a second sample point corresponding to each load design variable is generated; For each of the second sample points, based on the stress of the second sample point as the target response, and combined with the experimental design method, a main effect analysis is performed on each of the load design variables to obtain the effect level of each of the load design variables; Based on the effect level of each load design variable, the load design variables are recombined to obtain the fourth sample point after the load change. Based on the third and fourth sample points, a target sample space is generated.

4. The method according to claim 1, characterized in that, Based on the finite element analysis results of the model samples, finite element diagrams carrying stress value labels are determined, including: Finite element analysis is performed on the model sample to obtain the corresponding finite element analysis results; the finite element analysis results include the finite element model and the corresponding global properties; the finite element model includes multiple nodes, as well as the connection relationships and stress values ​​between the nodes; An initial finite element graph is constructed based on the nodes as vertices, the connections as edges, and the stress values ​​of the nodes as the stress value labels of the vertices. Add corresponding edges to the vertex pairs in the initial finite element diagram that meet the preset conditions; the preset conditions are: both vertices in the vertex pair are located in the specified mesh encryption area, and the spatial distance between the two vertices meets the preset distance threshold. Generate the main vertex corresponding to the global attribute, and connect the main vertex to all vertices of the initial finite element diagram to obtain a finite element diagram carrying stress value labels.

5. The method according to claim 4, characterized in that, Based on the finite element diagram, corresponding training samples are generated, including: Determine the vertex attribute tensor corresponding to each vertex in the finite element diagram, and the global attribute tensor corresponding to the main vertex; By combining the vertex attribute tensors and the global attribute tensor, a graph attribute tensor is obtained; Based on each edge and its corresponding vertex pair in the finite element diagram, determine the adjacency matrix; Based on the graph attribute tensor and the adjacency matrix, corresponding training samples are generated.

6. The method according to claim 1, characterized in that, The graph neural network model adopts an encoder-decoder architecture, which includes multiple network layers, including two multilayer perceptrons and three graph convolutional layers.

7. The method according to claim 6, characterized in that, During the training of the graph neural network model, the Adam optimizer is used to calculate the mean square error of the predicted node stress value and the node stress value, and to solve for the gradient direction in which the mean square error decreases the fastest. The weights of the graph convolutional layer and the fully connected layer in the multilayer perceptron are updated according to the gradient direction.

8. A device for generating simulation models of wind turbine components, characterized in that, include: The variable determination unit is used to determine design variables based on a pre-established basic 3D model; A spatial determination unit is used to obtain the target sample space based on the design variables; The target sample space includes sample points used to characterize structural and load changes in the basic three-dimensional model; The model generation unit is used to generate corresponding model samples based on each sample point in the target sample space; the model samples are part models after structural changes and load changes. The original drawing determination unit is used to determine the finite element drawing carrying stress value labels based on the finite element analysis results of the model sample. The stress value label is used to characterize the stress value of each vertex in the finite element diagram; The sample generation unit is used to generate corresponding training samples based on the finite element diagram. The model training unit is used to train the graph neural network model using the training samples to obtain the target simulation model; The target simulation model takes a geometric model and global attributes as input and outputs the stress field simulation results of the geometric model. The stress field simulation results include the stress field distribution of the geometric model, which includes a cloud map composed of the stress values ​​of each vertex. The global attributes include a combination of boundary conditions, material properties, and mesh density.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein the program is executed by a processor to perform the wind turbine component simulation model generation method according to any one of claims 1-7.

10. An electronic device, characterized in that, include: Processor, memory, and bus; The processor and the memory are connected via the bus; The memory is used to store the program, and the processor is used to run the program, wherein the program is executed by the processor to perform the wind turbine component simulation model generation method according to any one of claims 1-7.

Images