Fluid-structure interaction multi-physics field prediction method for offshore wind plant

Abstract

The invention provides a fluid-structure interaction multi-physics field prediction method for an offshore wind plant. According to the method, any Lagrangian-Euler mapping is introduced, a computational domain containing a structure and fluid is expressed as a deformation computational grid dynamically evolved along with structural vibration or free interface change, time sequence prediction is conducted on a grid displacement field through a first generator, and then the computational grid is updated; efficient prediction of fluid field or multi-physical field state variables is achieved through a second generator; moreover, a discrimination model is introduced in a prediction model training stage, and combined constraint is carried out on statistical consistency and physical rationality of a prediction result in combination with a physical constraint loss function. According to the decoupling prediction method, the moving boundary calculation precision is improved by adding physical constraints, so that the prediction precision is improved, powerful technical support can be provided for key links of offshore wind power engineering, and the decoupling prediction method can be applied to optimization of structural design, fine evaluation of the operation state and prospective prediction and early warning of extreme environment working conditions.

Description

Technical Field

[0001] This invention relates to multiphysics analysis and prediction, particularly to multiphysics reconstruction and prediction in the context of fluid-structure interaction in offshore wind farms, and belongs to the field of wind power generation technology. Background Technology

[0002] As the global energy structure shifts towards low-carbon and clean energy, offshore wind farms have become an important direction for the development of renewable energy. In offshore wind power projects, a large number of key structures (such as monopile foundations, jacket foundations, floating wind power platforms and their ancillary components) are in a complex marine environment for a long time and need to withstand multiple loads from wind, waves, currents and other sources. Their dynamic response is essentially a typical multiphysics coupling problem.

[0003] In practical engineering, offshore wind power structures often have the following characteristics: (1) Significant structural flexibility: Slender support components and auxiliary components are prone to vibration or large deformation under the action of incoming flow; (2) Boundary conditions change over time: Structural vibration and free liquid surface fluctuations will cause the computational domain boundary to move continuously; (3) Strong load uncertainty: Under extreme wind and wave conditions, the instantaneous pressure on the structural surface and the evolution of the flow field are highly nonlinear; (4) High requirements for the timeliness of engineering assessment: Wind farm design, operation and maintenance and risk assessment urgently need rapid prediction methods. At present, fluid-structure interaction analysis in offshore wind power engineering mainly relies on high-fidelity numerical simulation methods (e.g., CFD-FEM coupling), but the computational cost of such methods is extremely high, which makes it difficult to meet the engineering needs of wind farm group analysis, real-time early warning and digital twins.

[0004] In recent years, deep learning methods have demonstrated great potential in flow field reconstruction and prediction, providing a new approach for fluid-structure interaction (FSI) multiphysics prediction in offshore wind farms. However, existing methods relying on convolutional neural networks (CNNs) or convolutional long short-term memory networks (ConvLSTMs) are mostly based on fixed Eulerian grids, typically representing solid or boundary locations using masks. These methods introduce significant interpolation errors in critical areas such as structural surfaces and free liquid surfaces, making it difficult to accurately predict important engineering quantities that directly affect the safety of offshore wind farm structures, such as surface pressure and shear force. Therefore, there is an urgent need for a multiphysics prediction method that is suitable for complex engineering scenarios, can naturally incorporate moving boundary information into the deep learning framework, and balances physical consistency and prediction accuracy. Summary of the Invention

[0005] This invention addresses the problems of high computational cost and difficulty in meeting the rapid evaluation needs of engineering projects by existing moving boundary multiphysics numerical analysis methods, as well as the difficulty of accurately describing the effects of computational domain deformation and boundary motion by existing data-driven prediction methods. It proposes a fluid-structure interaction multiphysics prediction method for offshore wind farms.

[0006] It should be noted that the relational terms such as "first" and "second" used in this invention are only used to distinguish different entities or operations and do not imply any actual relationship or order between them. Furthermore, unless otherwise specified, the terms "comprising," "including," or similar expressions refer to non-exclusive inclusion, meaning that the process, method, product, or apparatus may include other elements not explicitly listed, or elements inherent to the process, method, product, or apparatus, in addition to the listed elements. Unless further limited, if an element is described as "comprising one…", it does not exclude the presence of other identical elements in the corresponding process, method, product, or apparatus.

[0007] The technical solution of this invention is: a fluid-structure interaction multiphysics prediction method for offshore wind farms, which can be used to handle multiphysics problems with moving boundaries. Its core is to decouple the deformation process of the computational domain from the multiphysics evolution process through modeling, and to explicitly introduce physical constraint mechanisms into a deep learning framework, thereby achieving efficient, stable, and generalizable prediction of multiphysics problems in complex engineering scenarios. Specifically, it includes the following steps:

[0008] Step (1) Data acquisition and preprocessing: Taking the offshore wind farm as the target object, the multi-physics related data of the target structure of the offshore wind farm at multiple historical moments are acquired through structural surface sensor measurement or high-fidelity numerical simulation. The data includes structural surface displacement, velocity, pressure or a combination thereof, and the data is reconstructed into a time series to form an input dataset suitable for deep learning model training. The output dataset consists of multi-physics data, and the output target value is temporally correlated with the input dataset.

[0009] Step (2) Prediction model architecture construction: Introduce arbitrary Lagrange-Euler mapping to uniformly describe the computational domain containing structure and fluid. Define the time-evolving mesh displacement field in the computational domain to characterize the computational domain deformation caused by structural vibration, free interface changes or boundary motion, in order to realize the reconstruction prediction task based on sparse data. The dual generators (including generator one and generator two) are designed to handle the mesh deformation caused by moving boundaries and flow field prediction respectively, forming a network framework that inputs the sequence of past sparse monitoring point information to obtain the future flow field prediction.

[0010] Step (3) Parameter adjustment and training of the prediction model: A discriminant model is introduced to distinguish the predicted multiphysics results from the real data at the statistical distribution level. By setting a physical constraint loss function, the smoothness of the grid, the continuity of field variables and the conservation characteristics are constrained to improve the physical consistency and stability of the prediction results. Next, an appropriate optimization algorithm is used to train the neural network. During the training process, the weights of the physical constraint terms are dynamically adjusted to strengthen the network's learning of physical characteristics.

[0011] Step (4) Results Output and Application: Using the trained generator one, input historical multi-time data to predict the grid displacement field at future times, thereby realizing a rapid estimation of the deformation process of the computational domain; based on the predicted grid displacement field, update the computational grid, and input the updated deformation computational grid into generator two to predict the spatial distribution and temporal evolution characteristics of fluid field or multiphysics state variables at future times, output the multiphysics prediction results at one or more future times, and use them for engineering structure response analysis, operational status assessment, or rapid prediction and early warning under extreme working conditions.

[0012] Further, in step (1): the input dataset includes multiphysics-related data of the target structure at multiple historical moments. The data includes at least one or more of the following: structural surface displacement, velocity, and pressure. The monitoring points are generally arranged on the surface of the solid structure. For deformable structures, the moving boundary position is naturally recorded. For the output target value, the flow field variables are projected onto the deformable mesh defined by the moving boundary displacement. The mesh deformation displacement is controlled by the weighted Laplace equation. Dirichlet boundary conditions are applied to all mesh deformation boundaries: the computational domain boundary displacement is zero, while the displacement at the moving boundary is equal to its actual motion displacement.

[0013] Further, in step (2): the dual generators adopt the same encoder-decoder architecture and integrate a recurrent neural network. The encoder uses a 5×5 convolutional kernel with a stride of 1, combined with pooling or strided convolution for progressive downsampling, enabling the network to capture multi-scale spatial features from local details to global structure; at the same time, the reduction in the spatial dimension of the feature map also directly reduces the computational complexity of subsequent layers, thereby enabling the extraction of multi-scale spatial features and reducing computational costs. In each decoder layer, its parameters correspond one-to-one with the parameters of the encoder layer. After each convolution operation, a ReLU activation function is used for nonlinear transformation. In addition, for each spatial dimension, two convolution operations are used to gradually reduce the spatial dimension while gradually increasing the number of channels, thereby deepening the neural network and enhancing its ability to capture features. Among them, except for the initial feature matrix which uses a 1×1 convolutional kernel to expand the dimension, the other dimension expansion convolutional layers use a 3×3 convolutional kernel. The convolutional long short-term memory neural network unit sequentially receives the multi-time step features extracted by the encoder and generates n predictions for future time steps in an autoregressive manner, which are provided to the decoder as input to the decoder. The decoder path then reconstructs the flow field through feature fusion via a series of upsampling operations and skip connections, ensuring the recovery of fine-scale details and accurate field prediction. The encoder receives past sensor data and incorporates the corresponding computational domain information into the prediction model.

[0014] Furthermore, in the above-mentioned fluid-structure interaction multiphysics prediction method for offshore wind farms, in step (2): the historical sparse monitoring point data is processed through three stacked fully connected layers to expand its dimension and reshape it into a two-dimensional matrix consistent with the dimension of the physical field; the amplified sparse monitoring point data is combined with the domain mask of the physical field to form a feature matrix, wherein the domain mask distinguishes different physical regions by assigning different numbers to different physical fields.

[0015] Furthermore, in the above-mentioned fluid-structure interaction multiphysics prediction method for offshore wind farms, step (3) includes the following loss function of the generator: the mean square error loss term is used to measure the difference between the variable values ​​of the reconstructed physical field and the real labeled physical field at each grid point; the adversarial loss term evaluates the probability that the generated result comes from real data through a discriminator, so as to improve the overall distribution fidelity of the reconstructed physical field by the generator; the continuity equation residual loss term is used to quantify the deviation of the incompressible conditions in the reconstructed flow field, and can also be discretized by the finite difference method, and the influence caused by grid deformation is offset by the inverse operation of the Jacobian matrix; the grid smoothness constraint error term is used to limit the degree of displacement change between adjacent grid nodes by applying Laplace-type or weighted Laplace-type constraints to the grid displacement field.

[0016] Furthermore, in the above-mentioned fluid-structure interaction multiphysics prediction method for offshore wind farms, in step (3): on the one hand, the training environment of the constructed prediction model is based on PyTorch, and the generator loss function is minimized by using the Adam optimizer. After introducing the optimal transmission distance (Wasserstein distance), the discriminator can effectively quantify the difference between the real data distribution and the data distribution generated by the generator, thereby overcoming the gradient vanishing problem caused by KL divergence; on the other hand, in order to ensure that the 1-Lipschitz constraint condition is met, a gradient penalty (GP) term is introduced into the loss function of the discriminator. This gradient penalty term is used to apply a penalty when the gradient norm of the interpolated sample deviates from the target value of 1, thereby ensuring the stability of the training process.

[0017] Furthermore, in the above-mentioned fluid-structure interaction multiphysics prediction method for offshore wind farms, in step (4), generator 2 can use a small sample transfer training method to achieve efficient and flexible prediction and evaluation of fluid-structure interaction problems of moving boundaries under various flow conditions or structural parameters, and then apply it to the accurate calculation of the hydrodynamic load and response on the wall of the structure.

[0018] Compared with the prior art, the technical solution of this invention has the following beneficial effects:

[0019] (1) This case introduces an arbitrary Lagrange-Euler mapping into the deep learning framework to uniformly describe and model the deformation process of the computational domain caused by structural vibration or free liquid surface changes based on sparse data. This method realizes the direct processing of moving boundary problems, avoids the boundary ambiguity and accuracy loss caused by the reliance on masks or interpolation in the traditional fixed Euler mesh method, and significantly improves the accuracy of predicting physical quantities of the structural surface and its adjacent areas.

[0020] (2) This case decouples the deformation prediction of the computational domain from the multiphysics state prediction, establishing a grid displacement field prediction model and a multiphysics prediction model respectively. This method enables the model to learn the "geometric deformation evolution law" and the "physical field evolution law" in stages, thereby reducing the training difficulty brought about by a single model simultaneously learning complex geometric changes and nonlinear physical processes. This not only improves the convergence speed and prediction stability of the model, but also enhances its generalization ability to different working conditions.

[0021] (3) In this case, a physical constraint loss function is introduced during the model training process. These functions will explicitly constrain the grid smoothness, physical field continuity and conservation properties, so that the prediction results not only closely resemble the real data, but also conform to the rationality of physical principles. This method reduces the non-physical oscillation phenomena commonly found in pure data-driven methods and improves the credibility and interpretability of the prediction results in engineering applications.

[0022] Other features and advantages of the present invention will be further described in the following description, some of which will be understood after reading the specification and can also be gained through actual practice of the invention. The objectives and other advantages of the present invention can be achieved by means of the designs clearly pointed out in the specification and its drawings. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0024] Figure 2 This is a schematic diagram of a multiphysics prediction framework based on sparse data provided in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the flow field label output of the dynamic mesh mapping data provided in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of a dual-generator multi-scale neural network structure provided in an embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram illustrating the prediction effect of the present invention under small-sample transfer training;

[0028] Figure 6This is a schematic diagram comparing the flow field prediction results reconstructed by the model in this embodiment of the invention with those of the traditional method model. Detailed Implementation

[0029] This invention addresses the problems of high computational cost and difficulty in meeting the needs of rapid engineering evaluation in complex engineering scenarios such as offshore wind farms, as well as the difficulty of existing data-driven prediction methods in accurately describing the effects of computational domain deformation and boundary motion. It proposes a multiphysics spatiotemporal prediction method for fluid-structure interaction problems in offshore wind farms based on arbitrary Lagrange-Euler mapping.

[0030] like Figure 1 As shown, this invention provides a multiphysics spatiotemporal prediction method for fluid-structure interaction problems in offshore wind farms. By introducing an arbitrary Lagrange-Euler mapping, the flow development of offshore wind farms can be reconstructed and predicted using a small number of sparse monitoring points. According to this invention, the computational domain containing both structure and fluid is represented as a deformable computational grid that dynamically evolves with structural vibration or changes in the free interface. Specifically, this includes: obtaining sparse sensor inputs and accurately labeled flow fields based on Lagrange-Euler mappings using high-fidelity CFD calculations to form a training dataset; constructing a dual-generator adversarial network with embedded physical constraints for prediction; training the neural network model using appropriate optimization algorithms and dynamic adjustment of physical losses; in the application stage, generator one can be directly applied to sample transfer under different parameter spaces, while generator two can be applied to small-sample transfer training.

[0031] This method utilizes a deep learning prediction generator pair to perform temporal predictions of the displacement field of a grid. Based on this, the computational grid is updated, and a second deep learning prediction generator is constructed on the deformed grid, achieving efficient prediction of fluid field or multi-physics state variables. Simultaneously, a discriminative model is introduced during the model training phase, combined with a physical constraint loss function, to jointly constrain the statistical consistency and physical rationality of the prediction results. This further improves the accuracy and physical consistency of the predictions while maintaining computational efficiency. Therefore, this improved scheme is expected to be applied in the engineering of offshore wind farms, with specific applications including but not limited to improving structural design, evaluating operational status, and rapidly predicting and issuing warnings under extreme environmental conditions.

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the described embodiments are only a part of the implementation of the present invention, and not all of them. The components of the embodiments shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents specific implementations of the 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.

[0033] This invention selects the fluid-structure interaction problem of a flexible plate deforming and vibrating under fluid action as the verification object. The boundary of the flexible plate moves over time, causing continuous deformation of the fluid computational domain, which is a typical moving boundary multiphysics problem. To achieve the goal of "reconstructing and predicting the future multiphysics evolution based on only a small amount of sparse monitoring point information," this embodiment uses an arbitrary Lagrange-Eulerian (ALE) mapping to describe the computational domain deformation and designs a dual-generator network to handle mesh deformation and flow field prediction respectively. A discriminant model and a physical constraint loss function are also introduced to improve the physical consistency and stability of the prediction results. The architecture is as follows: Figure 2 As shown. The specific steps of the embodiment are as follows:

[0034] Step 1: Data Acquisition and Preprocessing (Building Input / Output Paired Datasets)

[0035] In an embodiment of the present invention, step (1) includes:

[0036] (1a) High-fidelity data acquisition: The response information of the flexible plate and the corresponding fluid field data at multiple historical moments were obtained through high-fidelity numerical simulation of fluid-structure interaction. The length of the flexible plate was selected as 3D, where D is the diameter of the front cylinder, and the instantaneous flow field data under different compliance coefficients with Reynolds number Re=150 were used to build a dataset.

[0037] (1b) Sparse monitoring point setup: A limited number (e.g., 100) of monitoring points (e.g., sensor points) are arranged on the upper and lower surfaces of the flexible plate. The monitoring points move with the flexible plate and the input includes the position information, velocity, and pressure information of the moving boundary. The output of the monitoring points constitutes a sparse time series input signal.

[0038] (1c) Time series reconstruction and sample organization: The sparse monitoring point data of the 8 historical moments are organized into input sequence samples in chronological order; the corresponding supervision output is multiphysics data that is temporally correlated with the input, and the output time is multiple future moments (e.g., 4 or 8), where the time interval of each moment is a dimensionless time unit of 0.1.

[0039] (1d) Mesh Consistency Processing of Output Target: To ensure consistency between the learning target and the moving boundary, the output target value (flow field variable) is projected onto a deformed mesh (with an accuracy of 200×100) defined by the displacement of the moving boundary. This ensures geometric consistency between the training labels and the prediction results, reducing interpolation errors near the moving boundary under a fixed Eulerian mesh. The output flow field labels are as follows: Figure 3 As shown.

[0040] Step 2: Construction of the dual-generator network framework (mesh deformation prediction + flow field prediction)

[0041] To achieve the reconstruction prediction task based on sparse data, this embodiment designs a dual-generator network framework to handle mesh deformation caused by moving boundaries and flow field prediction respectively, thus forming an end-to-end framework of "inputting a sequence of past sparse monitoring point information → outputting future multiphysics prediction".

[0042] (2a) Generator 1: Mesh Displacement Field Prediction

[0043] Generator 1 is used as input to a sequence of historical sparse monitoring points to predict the mesh displacement field at future moments, which describes the deformation of the computational domain. Generator 1 employs an encoder-decoder architecture and integrates recurrent neural network units to characterize the spatiotemporal evolution of mesh deformation.

[0044] (2b) Generator 2: Flow field prediction on deformable mesh

[0045] Generator 2 is used to predict the spatial distribution and temporal evolution characteristics of fluid field or multiphysics state variables at future time points on the deformed computational grid predicted by Generator 1. The structure of Generator 2 is consistent with or similar to that of Generator 1 in order to share the same feature extraction and temporal modeling capabilities.

[0046] Preferably, in step (2), the dual generators employ the same encoder-decoder architecture and integrate a recurrent neural network such as Figure 4 As shown, the specific steps include the following:

[0047] (2b-1) Sparse input dimensionality expansion and reshaping: Historical sparse monitoring point data is first processed through three stacked fully connected layers to expand its dimension, and then reshaped into a two-dimensional matrix consistent with the physical field dimension so as to match the input form of the convolutional network.

[0048] (2b-2) Domain mask fusion: The magnified two-dimensional matrix of sparse monitoring points is combined with the domain mask of the physical field to form a feature matrix; the domain mask distinguishes different regions (e.g., solid region and fluid region) by assigning different numbers to different physical regions.

[0049] (2b-3) Encoder feature extraction: The encoder uses a 5×5 convolution kernel with a stride of 1 to extract multi-scale spatial features layer by layer and downsamples them step by step; preferably, the encoder uses 1×1 and 3×3 convolution kernels for feature extraction, and gradually increases the number of channels as the spatial dimension decreases to enhance expressive power; a ReLU activation function is set after each convolution layer.

[0050] (2b-4) Temporal modeling and autoregressive prediction: Convolutional Long Short-Term Memory (ConvLSTM) network units sequentially receive multi-time-step features extracted by the encoder, generate predicted features for the next n time steps in an autoregressive manner, and then pass them to the decoder.

[0051] (2b-5) Decoder reconstruction and skip connections: The decoder path restores the spatial resolution through a series of upsampling operations and fuses features with the corresponding layer of the encoder through skip connections to ensure fine-scale detail recovery and prediction accuracy; preferably, the decoder parameters correspond to the encoder, and an additional convolution operation is added during the feature dimension aggregation process to cope with the increase in feature dimension caused by the upsampling dimension and the skip connection feature fusion.

[0052] Step 3: Train the model using appropriate optimization algorithms and dynamic adjustment techniques for the physical loss term.

[0053] In this embodiment of the invention, in step (3), the Adam optimizer is used to minimize the generator loss function and the discriminator loss is monitored simultaneously to stabilize the adversarial training process.

[0054] Specifically, the weights and biases of each layer of the generator and discriminator are first initialized. Then, in each training iteration, a number of paired samples are randomly selected from the training set to form a batch, which is then input into the generator to obtain the reconstruction result. This reconstructed result is then fed into the discriminator along with the corresponding labeled flow field for adversarial training.

[0055] In the generator update phase, the gradient is calculated based on the total loss function, and the Adam optimizer is used to update the parameters with a learning rate of 10. -4 ~10 -3 The range is 8 to 32 in batch size. To prevent the discriminator from being too strong or too weak, which could lead to an imbalance in adversarial training, the discriminator loss is monitored synchronously in each training round. The stability of adversarial training is maintained by appropriately adjusting the discriminator update frequency (e.g., one update of the discriminator corresponds to five updates of the generator).

[0056] Step 4: Prediction and Application Process

[0057] (4a) Mesh displacement prediction and mesh update: In the prediction stage, the sparse monitoring point data of historical multiple time moments are input into generator one, and the mesh displacement field of future time moments is output; the computational mesh is updated accordingly to obtain the deformation computational mesh, so as to realize the rapid estimation of the deformation process of the computational domain.

[0058] (4b) Flow field prediction on deformable mesh: The updated deformable computational mesh and corresponding input features are fed into Generator 2 to predict the spatial distribution and temporal evolution characteristics of the fluid field or multiphysics state variables at one or more future time points. Moreover, Generator 2 adopts a "small sample migration" method to quickly adapt the prediction model under different working conditions or different structural parameters, achieving highly flexible prediction and evaluation for fluid-structure interaction problems with moving boundaries. The results are as follows: Figure 5 As shown.

[0059] (4c) Engineering output: The output multiphysics prediction results can be used for post-processing to calculate the hydrodynamic loads on the wall of the structure (e.g., pressure load distribution and its time evolution), for structural response analysis, operational status assessment or rapid prediction and early warning under extreme conditions.

[0060] like Figure 6 As shown, in the comparison between the prediction results of the embodiment model and the traditional method model, the prediction accuracy of this embodiment is significantly improved compared with the traditional method due to its smaller prediction error. In summary, this case verifies the effectiveness and superiority of the proposed multiphysics spatiotemporal prediction method based on arbitrary Lagrange-Euler mapping in the fluid-structure interaction problem of flexible plates, providing strong support for the rapid evaluation and optimization design of fluid-structure interaction problems in offshore wind farms.

[0061] In the technical solution of this invention, the key technologies are: describing the computational domain deformation based on the Arbitrary Lagrange-Euler (ALE) method, designing a dual generator network to handle mesh deformation and flow field prediction respectively, and introducing a series of improvements to the discriminant model and physical constraint loss function. Figure 1 and Figure 2 The key features shown are the implementation process of this case and the multiphysics prediction framework based on sparse data. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, including but not limited to ROM, RAM, disk, and optical disk.

[0062] Thus, by adopting the technical solution of this invention, the problem of moving boundary is directly addressed, avoiding the boundary ambiguity and accuracy loss caused by relying on masks or interpolation in the traditional fixed Eulerian mesh method. This significantly improves the accuracy of predicting physical quantities of the structural surface and its adjacent areas, not only enhancing the convergence speed and prediction stability of the prediction model and strengthening its generalization ability to different working conditions, but also reducing the non-physical oscillation phenomena commonly found in pure data-driven methods, thereby improving the credibility and interpretability of the prediction results in engineering applications.

[0063] As can be seen from the above description, compared with the prior art, the technical solution of this invention introduces an arbitrary Lagrange-Euler mapping into the deep learning framework, and uses sparse data to uniformly describe and model the deformation process of the computational domain caused by structural vibration or free liquid surface changes; it decouples the deformation prediction of the computational domain from the multiphysics state prediction, and establishes a mesh displacement field prediction model and a multiphysics field prediction model respectively. This method enables the model to learn the "geometric deformation evolution law" and the "physical field evolution law" in stages, thereby reducing the training difficulty caused by a single model simultaneously learning complex geometric changes and nonlinear physical processes; moreover, physical constraint loss functions are introduced during the model training process. These functions will explicitly constrain the mesh smoothness, physical field continuity, and conservation properties, making the prediction results not only close to the real data, but also consistent with the rationality of physical principles.

[0064] The technical solution, working process, and implementation effects of the present invention have been described in detail above. It should be noted that the described examples are only typical examples of the present invention. In addition, the present invention may have many other specific implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A fluid-structure interaction multiphysics prediction method for offshore wind farms, based on the arbitrary Lagrange-Euler mapping method, designs a dual-generator network to handle mesh deformation and flow field prediction respectively, characterized in that... Includes the following steps: Step (1) Data acquisition and preprocessing: Taking offshore wind farms as the target object, multi-physics field related data of the target structure of offshore wind farms at multiple historical moments are obtained through sensor measurements on the surface of solid structures or high-fidelity numerical simulation. The data is reconstructed in time series to form an input dataset. The output dataset consists of multi-physics field data, and the output target value is time-related to the input dataset. Step (2) Prediction model architecture construction: Introduce arbitrary Lagrange-Euler mapping to uniformly describe the computational domain containing structure and fluid, define the time-evolving grid displacement field in the computational domain, design a dual generator for reconstructing the prediction task, and predict the future flow field by constructing a network framework and inputting historical sparse monitoring point data. Step (3) Parameter adjustment and training of the prediction model: Introduce a discriminant model to evaluate the consistency between the multiphysics prediction results and the real data in terms of statistical distribution, and set physical constraint loss function to constrain grid smoothness, field variable continuity and conservation characteristics. Then use optimization algorithm to train the neural network and strengthen the network's learning of physical characteristics by dynamically adjusting the weight of physical constraint terms. Step (4) Results Output and Application: Using the trained generator one, input historical multi-time data to predict the grid displacement field at future times. Based on this, quickly estimate the deformation process of the computational domain. Update the computational grid according to the predicted grid displacement field, and input the updated deformation computational grid into generator two to predict the spatial distribution and temporal evolution characteristics of fluid field or multiphysics state variables at future times. Output at least one multiphysics prediction result at a time for engineering structure response analysis, operation status assessment or rapid prediction and early warning under extreme conditions.

2. The prediction method according to claim 1, characterized in that: In step (1), the input dataset includes multiphysics-related data of the target structure at multiple historical moments, including but not limited to one or more of structural surface displacement, velocity and pressure; monitoring points are arranged on the surface of the solid structure to naturally record the moving boundary position of the deformable structure. For the output target value, the flow field variables are projected onto the deformed mesh defined by the moving boundary displacement, and the mesh deformation displacement is controlled by the weighted Laplace equation. Dirichlet boundary conditions with zero computational domain boundary displacement are applied to the mesh deformation boundary, and the displacement at the moving boundary is equal to the actual motion displacement.

3. The prediction method according to claim 1, characterized in that: In step (2), the dual generators adopt the same encoder-decoder architecture. The parameters of each layer of the decoder are related to the parameters of the corresponding layer in the encoder, and a recurrent neural network is integrated. The historical sparse monitoring point data is expanded in dimension through three stacked fully connected layers and converted into a higher-dimensional feature representation. This feature is reshaped into a two-dimensional matrix consistent with the physical field dimension. By introducing a domain mask, the amplified historical sparse monitoring point data is combined with the domain mask of the physical field to form a fused feature matrix. As the spatial dimension decreases, the number of channels increases accordingly, thereby deepening the neural network hierarchy and enhancing its ability to capture features. A ReLU activation function is connected after each convolution operation. For the first convolutional layer, its kernel size is set to 1×1, and the remaining convolutional layers use a 3×3 kernel. The convolutional long short-term memory neural network unit receives the multi-time step features extracted by the encoder in sequence, generates n future time step predictions in an autoregressive manner, and then provides the predictions to the decoder. The decoder path reconstructs the flow field through a series of upsampling operations and skip connections to ensure the restoration of fine-scale details and accurate field prediction.

4. The prediction method according to claim 1, characterized in that: In step (2), the loss function of the dual generator is implemented as follows: (3a) Mean square error loss term, which is used to measure the difference between the reconstructed physical field and the real physical field based on the mean square error of the variable values ​​at each grid point; (3b) Adversarial loss term: use a discriminator to evaluate the probability that the generated result comes from real data, so as to improve the overall fidelity of the physical field reconstructed by the generator; (3c) The residual loss term of the continuity equation is defined as a quantitative index of the deviation between the reconstructed flow field and the incompressible condition; (3d) Mesh smoothness constraint error term, which aims to limit abrupt displacement changes between adjacent nodes by applying Laplace or weighted Laplace constraints to the mesh displacement field, thereby ensuring the smoothness of mesh deformation.

5. The prediction method according to claim 4, characterized in that: In step (3c), the residual loss term of the continuity equation is discretized using the finite difference method. During this calculation, the effects of mesh deformation can be offset by obtaining the inverse of the Jacobian matrix.

6. The prediction method according to claim 1, characterized in that: In step (3), the prediction model is trained in the PyTorch environment. The Adam optimizer is used to minimize the generator loss function. After introducing the optimal transmission distance, the discriminator effectively quantifies the difference between the real data distribution and the data distribution generated by the generator. To ensure that the Lipshitz constraint is met, a gradient penalty term is introduced into the loss function of the discriminator. This gradient penalty term is used to apply a penalty when the gradient norm of the interpolated sample deviates from the target value of 1, thereby ensuring the stability of the training process.

7. The prediction method according to claim 1, characterized in that: In step (4), the generator system uses a small sample transfer training method to perform highly flexible prediction and evaluation of fluid-structure interaction problems of moving boundaries under different conditions, and is further used to calculate the hydrodynamic load on the wall of the structure.

Images