Numerical calculation method for multi-field coupling of photovoltaic support based on digital twinning

Abstract

The application discloses a kind of photovoltaic support multi-field coupling numerical calculation method based on digital twinning, it is related to digital twinning and computer numerical simulation technical field.The method extracts the initial three-dimensional geometric model of photovoltaic support pre-set in digital twinning space and is discretized to generate three-dimensional geometric topology grid, and the multi-field coupling boundary condition of real-time meteorological and sensor data fusion is analyzed;Boundary conditions and node geometric properties are spliced to build node initial features, and the cross-field topological connection relationship is constructed to calculate the physical field coupling interaction weight between adjacent nodes;Based on the weight, cross-physical field implicit feature interaction and iterative update are carried out, and the node interaction feature vector is extracted;Based on the vector and in combination with physical equation constraint, the multi-field coupling numerical calculation result is output.The application realizes the accurate mapping and numerical calculation of photovoltaic support high nonlinear two-way fluid-solid coupling effect by constructing the cross-field topological connection and multi-physical field implicit feature interaction network across the fluid-solid interface.

Description

Technical Field

[0001] This invention relates to the field of digital twin and computer numerical simulation technology, specifically to a multi-field coupling numerical calculation method for photovoltaic brackets based on digital twins. Background Technology

[0002] Photovoltaic (PV) brackets are widely used in large-scale PV power plants. During their outdoor service, they are subjected to complex meteorological environments with the coupling effects of multiple physical fields, including airflow, wind load, thermal stress, and structural mechanical fields. With the development of digital twin technology, using virtual digital models to dynamically map the stress state of physical PV brackets and assess their structural health has become an important means to ensure the safety of PV power plant systems and improve operation and maintenance efficiency.

[0003] However, in actual service, the deformation of photovoltaic (PV) supports caused by wind loads reacts to the surrounding flow field, altering the aerodynamic distribution on the support surface and creating a highly nonlinear two-way fluid-structure interaction effect. Existing digital twin numerical calculation methods typically employ sequential solution strategies involving unidirectional static equivalent loading or decoupling, failing to fully consider the dynamic interactive feedback mechanism between fluid mechanics (wind field) and solid mechanics (structural field). This lack of physical coupling mechanism leads to significant errors in the stress analysis of virtual models under complex and extreme weather conditions, making it impossible to accurately reproduce the true dynamic stress state of the physical entity. Consequently, this severely impacts the reliability of PV support structural health assessments and the accuracy of active defense control.

[0004] To address this, a numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins. By introducing a cross-field topology connection mechanism and a multi-physics implicit interaction network, and combining physical constraint decoding and time series recursive model, this invention solves the problem that in the stress analysis of photovoltaic supports, the use of unidirectional static equivalent loading or decoupling solution strategies fails to fully consider the fluid-structure bidirectional dynamic interaction feedback mechanism, thus resulting in the inability to accurately restore the true dynamic stress state of the physical entity.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins includes:

[0008] The initial three-dimensional geometric model of the photovoltaic support is extracted from the digital twin space. The initial three-dimensional geometric model is spatially discretized to generate a three-dimensional geometric topology mesh composed of multiple discrete nodes. Real-time environmental meteorological sequence and physical sensor data of the photovoltaic support are acquired and analyzed. Fluid dynamics and solid mechanics features are extracted respectively to generate multi-field coupled boundary conditions.

[0009] The multi-field coupling boundary conditions are mapped to the corresponding discrete nodes of the three-dimensional geometric topology mesh, the three-dimensional geometric properties of the discrete nodes are extracted, and the initial features of the nodes are constructed by splicing them together; the cross-field topological connection relationship between the discrete nodes in the fluid domain and the solid domain is constructed, and the physical field coupling interaction weights between adjacent nodes are calculated based on the cross-field topological connection relationship and the three-dimensional geometric properties.

[0010] Based on the physical field coupling interaction weights, the initial features of the nodes are subjected to implicit feature interaction and iterative update across physical fields, and node interaction representation vectors containing global multi-field coupling effects are extracted; based on the node interaction representation vectors, the multi-field coupling numerical calculation results of the photovoltaic support at the current timestamp are output.

[0011] Preferably, the process of generating a three-dimensional geometric topology mesh includes:

[0012] Calculate the principal curvature and Gaussian curvature of the surface of the initial three-dimensional geometric model, and identify regions where the curvature gradient exceeds a preset threshold as structural feature points and regions where curvature changes abruptly.

[0013] Based on the structural feature points and the curvature-changing regions, an adaptive background mesh size control strategy is adopted to generate high-density discrete nodes with high spatial resolution in stress concentration regions and flow field boundary layer regions, and low-density discrete nodes with low spatial resolution in gentle, non-critical regions.

[0014] By using the Delaunay space triangulation algorithm, the high-density discrete nodes and the low-density discrete nodes are connected, and the node connection edges in the transition region are smoothed using Laplacian smoothing, thus constructing a three-dimensional geometric topology mesh containing multi-scale non-uniform nodes.

[0015] Preferably, the process of generating multi-field coupled boundary conditions includes:

[0016] Temporal characteristic parameters of wind speed profile distribution, wind direction deflection angle and ambient temperature are extracted from the real-time environmental meteorological sequence and used as fluid dynamic boundary features;

[0017] From the physical sensing data, the strain tensor and local spatial displacement vector on the surface of the key load-bearing components of the photovoltaic bracket are extracted as solid mechanics boundary features.

[0018] A time-window-based spline interpolation algorithm is used to align the fluid dynamics boundary features and the solid mechanical boundary features with time-series frequencies. The aligned features are then mapped to a unified dimensionless feature space for normalization and fused to generate multi-field coupled boundary conditions.

[0019] Preferably, the process of constructing the initial features of a node includes:

[0020] The global absolute spatial coordinates, surface normal vector, and local neighborhood topological features of each discrete node are extracted and combined to form three-dimensional geometric properties;

[0021] The multi-field coupling boundary conditions mapped to the corresponding discrete nodes are input into a high-dimensional feature embedding layer composed of a multilayer perceptron, which converts the low-dimensional physical boundary parameters into high-dimensional dense latent feature vectors.

[0022] The high-dimensional dense latent feature vector and the three-dimensional geometric attributes are concatenated and spliced ​​in the channel dimension to generate initial node features that combine spatial geometric position awareness and physical external load constraint information.

[0023] Preferably, the process of calculating the physical field coupling interaction weights between adjacent nodes includes:

[0024] Based on the Euclidean distance between the fluid domain nodes and the solid domain nodes, a dynamic cutoff radius is set, and within the dynamic cutoff radius, a neighborhood search algorithm is used to establish directed connection edges at the cross-fluid-solid interface, forming the cross-field topological connection relationship.

[0025] Extract the three-dimensional geometric attributes of the nodes connected at both ends of the directed edge, and perform linear transformation to generate query vector and key vector;

[0026] The dot product of the query vector and the key vector is calculated, divided by a scaling factor, and then activated by a normalized exponential function to obtain the attention distribution coefficient characterizing the intensity of the interaction between fluid kinetic energy transfer and solid deformation. This coefficient is then used as the weight of the physical field coupling interaction.

[0027] Preferably, the process of extracting node interaction representation vectors includes:

[0028] Along the cross-field topological connection relationship, the feature vectors of neighboring nodes are weighted and summed with the physical field coupling interaction weights to aggregate cross-field physical information within the local receptive field;

[0029] The aggregated features are concatenated with the features of the target node itself and input into the graph neural network transit unit containing a non-linear activation function and a normalization layer, and the updated node hidden layer state is output.

[0030] The cross-field physical information aggregation and update steps are executed repeatedly until the preset message transmission depth is reached;

[0031] The feature vector of the last layer of the network at the message passing depth is extracted and used as the node interaction representation vector.

[0032] Preferably, the process of calculating the multi-field coupling numerical results includes:

[0033] The node interaction representation vectors are input in parallel to the solid mechanics decoder and the fluid mechanics decoder, respectively. Both the solid mechanics decoder and the fluid mechanics decoder contain multi-branch feedforward neural networks.

[0034] The solid mechanics decoder generates deviatoric stress tensor features and displacement vector features through multi-branch network mapping, and performs physical consistency correction on the stress tensor and displacement vector through a linear elastic physical constraint layer to reconstruct and output dynamic stress field and strain displacement field.

[0035] The fluid dynamics decoder generates three-dimensional velocity vector field features through a multi-branch network, introduces a divergence penalty mapping layer to ensure that the velocity vector field satisfies the fluid incompressibility condition, and reconstructs the pressure distribution and vortex structure features of the output flow field.

[0036] The dynamic stress field, the strain displacement field, the flow field pressure distribution, and the eddy structure characteristics are combined to form the multi-field coupled numerical calculation results.

[0037] Preferably, the calculation method further includes a time-series recursion mechanism to ensure the continuity of physical states, the time-series recursion mechanism including:

[0038] Set up a hidden state unit with a gated cyclic memory structure; after the node interaction representation vector under the current timestamp is reduced in dimensionality by a fully connected layer, it is saved and updated to the context matrix of the hidden state unit;

[0039] When extracting the initial features of the node at the next time stamp, the context matrix is ​​read as the historical prior physical features and fused with the features at the next time stamp through residual connection to perform multi-field coupled dynamic temporal evolution calculation.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. This invention, when generating a three-dimensional geometric mesh, employs an adaptive size control strategy by calculating the curvature gradient. This generates high-density nodes in stress concentration and flow field boundary layer regions, and low-density nodes in flat regions. Valuable computational resources are precisely allocated to key stress areas, avoiding redundant computations caused by global mesh subdivision. It also solves the problem that conventional mesh equivalence processing cannot accurately reproduce the true dynamic stress state of physical entities. Furthermore, time window spline interpolation is used to align and normalize meteorological and sensor data with different sampling frequencies along the time axis, providing a high-quality spatiotemporal data foundation for subsequent multi-field coupling.

[0042] 2. In handling fluid-solid interactions, this invention establishes a topological connection across the fluid-solid interface using a dynamic cutoff radius and introduces an attention mechanism to calculate the interaction weights across physical fields. It also performs feature weighting aggregation and nonlinear updates within the local receptive field using a graph neural network. This method implicitly quantifies and extracts the highly nonlinear bidirectional fluid-structure interaction effect, overcoming the shortcomings of traditional unidirectional static equivalent loading or decoupling sequential solution strategies that fail to fully consider the dynamic physical interaction feedback.

[0043] 3. In the final output of numerical calculation results, this invention decodes and forcibly introduces physical equation constraints through a multi-branch network; it ensures that the fluid field satisfies the incompressibility condition through divergence penalty, and uses Hooke's law to perform linear elastic physical consistency correction on the solid field, ensuring that the decoded and reconstructed field quantities strictly conform to the mechanical laws of real continuous media; combined with a gated cyclic memory structure and residual connection mechanism, it effectively retains and transmits the fluid separation vortex and solid vibration hysteresis effect of the previous moment, realizing continuous dynamic temporal evolution tracking of the stress state of photovoltaic support under complex extreme weather conditions. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins.

[0045] Figure 2 A schematic diagram of the logical structure of the interaction process between the cross-field topology connection of photovoltaic brackets and the implicit features of the graph network;

[0046] Figure 3 This is a schematic diagram of the architecture of a multi-branch network decoding and time series recursion mechanism. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0048] Please see Figures 1 to 3 This invention provides a numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins, and the technical solution is as follows:

[0049] A numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins includes:

[0050] The initial three-dimensional geometric model of the photovoltaic support is extracted from the digital twin space. The initial three-dimensional geometric model is spatially discretized to generate a three-dimensional geometric topology mesh composed of multiple discrete nodes. Real-time environmental meteorological sequence and physical sensor data of the photovoltaic support are acquired and analyzed. Fluid dynamics and solid mechanics features are extracted respectively to generate multi-field coupled boundary conditions.

[0051] The multi-field coupling boundary conditions are mapped to the corresponding discrete nodes of the three-dimensional geometric topology mesh, the three-dimensional geometric properties of the discrete nodes are extracted, and the initial features of the nodes are constructed by splicing them together; the cross-field topological connection relationship between the discrete nodes in the fluid domain and the solid domain is constructed, and the physical field coupling interaction weights between adjacent nodes are calculated based on the cross-field topological connection relationship and the three-dimensional geometric properties.

[0052] Based on the physical field coupling interaction weights, the initial features of the nodes are subjected to implicit feature interaction and iterative update across physical fields, and node interaction representation vectors containing global multi-field coupling effects are extracted; based on the node interaction representation vectors, the multi-field coupling numerical calculation results of the photovoltaic support at the current timestamp are output.

[0053] The process of generating a three-dimensional geometric topology mesh includes: calculating the principal curvature and Gaussian curvature of the surface of the initial three-dimensional geometric model, identifying regions where the curvature gradient exceeds a preset threshold as structural feature points and regions with abrupt curvature changes; based on the structural feature points and the regions with abrupt curvature changes, adopting an adaptive background mesh size control strategy to generate high-density discrete nodes with high spatial resolution in stress concentration regions and flow field boundary layer regions, and low-density discrete nodes with low spatial resolution in smooth, non-critical regions; connecting the high-density discrete nodes and the low-density discrete nodes using the Delaunay spatial triangulation algorithm, and performing Laplacian smoothing on the node connection edges in transition regions to construct a three-dimensional geometric topology mesh containing multi-scale non-uniform nodes.

[0054] The digital twin space is a three-dimensional digital environment built on a virtual simulation engine or digital twin platform. It has a pre-built three-dimensional geometric model asset library of photovoltaic brackets and is equipped with a data interface for receiving data from physical power station sensors and meteorological stations.

[0055] The system reads the device identifier corresponding to the current physical photovoltaic support, calls the data interface of the digital twin space based on the device identifier, and retrieves the matching three-dimensional geometric model file from the pre-set model asset library; it parses the geometric topology data structure of the model file, extracts the vertex coordinate matrix, normal vector data, and assembly constraint parameters between geometric components of all topological patches; it reads the three-dimensional world coordinate system uniformly constructed in the digital twin space, establishes a homogeneous transformation matrix from the model's local coordinate system to the three-dimensional world coordinate system, and performs rotation, translation, and scaling transformations on the vertex coordinate matrix and normal vector data through the homogeneous transformation matrix to complete spatial coordinate alignment and geometric reconstruction, generating an initial three-dimensional geometric model with complete geometric topology information;

[0056] The vertex data of the facets on the surface of the initial 3D geometric model are extracted. The principal curvature equation is solved to calculate the maximum and minimum principal curvature of each vertex, and the two are multiplied to obtain the Gaussian curvature. The spatial derivative of the Gaussian curvature between adjacent vertices is calculated to obtain the curvature gradient. The mean and standard deviation of the global curvature gradient are statistically analyzed to calculate a dynamic preset threshold. The dynamic preset threshold is set as the sum of the mean of the global curvature gradient and the standard deviation of a set multiple. The set multiple is preferably a real number between 1.5 and 3.0 to effectively filter out normal surface undulations and accurately locate sharp geometric abrupt change regions. Regions with curvature gradients exceeding the preset threshold are marked with coordinates and identified as structural feature points and regions with abrupt curvature changes.

[0057] A spatial grid size function is constructed, and an adaptive background grid size control strategy is introduced. An octree spatial partitioning algorithm is used to recursively discretize the model bounding box. The spatial grid size function constructs a grid size field that increases exponentially with spatial distance, centered on the identified structural feature points and the regions with abrupt curvature changes. This maps the geometric regions with abrupt curvature changes to stress concentration regions of the solid structure, and maps the outwardly radiating normal neighboring space to the flow field boundary layer region. For the stress concentration regions and flow field boundary layer regions corresponding to the structural feature points and the regions with abrupt curvature changes, the octree is driven to subdivide into deeper layers, generating high-density discrete nodes with high spatial resolution. For unmarked, gentle, non-critical regions, the octree subdivision depth is limited, generating low-density discrete nodes with low spatial resolution.

[0058] In the process of adopting the adaptive background grid size control strategy, a local densification bias mechanism based on historical fatigue prior knowledge is introduced. Specifically, long-term high-frequency stress monitoring data of the same model of photovoltaic support physical entity during its past life cycle are extracted. Through statistical analysis, high-risk structural connection points and weld areas prone to fatigue damage are identified, forming a spatial fatigue susceptibility probability field. This probability field is superimposed as an additional penalty factor into the spatial grid size function, so that even in areas with gentle curvature but high fatigue risk, the octree spatial partitioning algorithm can be forced to perform deep subdivision, thereby breaking through the limitation of relying solely on pure geometric curvature features to determine grid density. In this way, the digital twin grid not only has geometric fidelity, but also integrates the fatigue decay prior laws of physical entities, effectively improving the sensitivity of photovoltaic support structural health assessment under long service cycles.

[0059] The generated discrete node coordinates are extracted and input into the Delaunay spatial triangulation algorithm. A three-dimensional mesh cell is constructed by testing the empty sphere rule, connecting the high-density discrete nodes and the low-density discrete nodes. The transition region at the boundary between high and low density is extracted, and the node connection edges with distorted adjacent side length ratios are traversed and detected. The specific judgment criterion is whether the ratio of the maximum side length to the minimum side length of adjacent shared mesh cells is greater than a preset aspect ratio threshold. For connection edges that are determined to be distorted, Laplacian smoothing is performed on them. The node coordinates of the transition region are iteratively moved to the spatial weighted average position of the first-order adjacent nodes to eliminate distortion, thus constructing a three-dimensional geometric topology mesh containing multi-scale non-uniform nodes.

[0060] By adopting an adaptive node placement and smooth meshing strategy based on curvature gradient, a multi-scale non-uniform three-dimensional geometric topology mesh was obtained, which enabled the accurate allocation of computing power to the key stress zone and fluid boundary layer region of the structure. This solved the problem that the actual dynamic stress state of the physical entity could not be accurately reproduced due to the mesh equivalence processing during digital twin numerical calculation.

[0061] The process of generating multi-field coupled boundary conditions includes:

[0062] Temporal characteristic parameters of wind speed profile distribution, wind direction deflection angle, and ambient temperature are extracted from the real-time environmental meteorological sequence as fluid dynamics boundary features; strain tensors and local spatial displacement vectors on the surface of key load-bearing components of the photovoltaic support are extracted from the physical sensing data as solid mechanics boundary features; a spline interpolation algorithm based on time windows is used to align the temporal frequency of the fluid dynamics boundary features and the solid mechanics boundary features, and the aligned features are mapped to a unified dimensionless feature space for normalization processing, and multi-field coupled boundary conditions are generated by fusion.

[0063] The system receives real-time environmental meteorological sequence data streams collected by meteorological monitoring equipment, analyzes and extracts wind speed numerical sequences at different altitude levels to construct wind speed profile distribution; it simultaneously acquires the time-varying sequence of incoming flow angles fed back by wind direction sensors as wind direction deflection angles, and extracts the time-series characteristic parameters of ambient temperature collected by temperature sensors. The wind speed profile distribution, the wind direction deflection angle, and the time-series characteristic parameters of ambient temperature are used together as fluid dynamics boundary features.

[0064] The physical sensing data transmitted in real time by strain gauges and displacement sensors arranged on the surface of key load-bearing components of photovoltaic brackets is obtained, and the sensor electrical signals are decoded and converted into physical parameters. By combining multi-directional strain channel data, the strain tensor of each measuring point is calculated based on strain rosette theory, and the local spatial displacement vector is obtained by spatial coordinate system transformation. The strain tensor and the local spatial displacement vector are used as solid mechanical boundary features.

[0065] A fixed-length sliding time window is set, the length of which can be preset to a specific time span based on the characteristic period of real-time environmental and meteorological changes. The fluid dynamics boundary features and the solid mechanical boundary features within the time window are extracted. To address the issue of inconsistent sampling rates between fluid and solid sensors, a spline interpolation algorithm based on the time window is adopted, preferably a cubic spline interpolation algorithm. Using a fixed high-frequency sampling rate as a benchmark, a spline function that continuously changes with time for each feature is constructed. Through resampling at equal time steps, the temporal frequency alignment of the fluid dynamics boundary features and the solid mechanical boundary features is achieved.

[0066] The statistical extreme values ​​of each dimension of the frequency-aligned features within the current time window are extracted, and the aligned features are mapped to a unified dimensionless feature space for normalization. The specific normalization calculation formula is expressed as follows: ,in As input features, and These represent the minimum and maximum values ​​of the feature within the time window, respectively. The normalized dimensionless features are represented by the normalized fluid dynamics boundary feature matrix and the solid mechanics boundary feature matrix, which are then cascaded and combined along the feature channel dimension to generate multi-field coupled boundary conditions.

[0067] By using time window spline interpolation and dimensionless normalization mapping, multi-field coupled boundary conditions for spatiotemporal fusion of multi-source heterogeneous data were obtained, realizing the time-series data alignment and dimensional unification of fluid meteorological loads and solid structure responses.

[0068] The process of constructing the initial features of the nodes includes:

[0069] The global absolute spatial coordinates, surface normal vector, and local neighborhood topological features of each discrete node are extracted and combined to form three-dimensional geometric attributes. The multi-field coupling boundary conditions mapped to the corresponding discrete nodes are input into a high-dimensional feature embedding layer composed of a multilayer perceptron, and the low-dimensional physical boundary parameters are converted into high-dimensional dense latent feature vectors. The high-dimensional dense latent feature vectors and the three-dimensional geometric attributes are concatenated and spliced ​​in the channel dimension to generate initial node features that combine spatial geometric position perception and physical external load constraint information.

[0070] The process involves traversing each discrete node in the 3D geometric topology mesh, reading its 3D coordinate coefficients in the global coordinate system to obtain the global absolute spatial coordinates; using the planar patches formed by surrounding neighboring nodes, the surface normal vectors of each node are obtained by calculating the cross product of coplanar vectors and performing normalization; using the adjacency matrix in graph theory, the connectivity and distribution of first-order neighboring nodes around each node are statistically analyzed to extract local neighborhood topological features that characterize the local connectivity of the mesh; the obtained global absolute spatial coordinates, surface normal vectors, and local neighborhood topological features are encapsulated in a multidimensional array, and flattened and aligned to the one-dimensional feature vector corresponding to each node, thus combining them to form 3D geometric properties.

[0071] A multilayer perceptron network comprising an input layer, multiple hidden layers, and an output layer is constructed and used as a high-dimensional feature embedding layer. The multi-field coupling boundary conditions mapped to the corresponding discrete nodes are used as low-dimensional physical boundary parameters and input into the high-dimensional feature embedding layer composed of the multilayer perceptron. The input tensor is linearly multiplied and biased by the neuron weight matrix in each hidden layer, and a nonlinear activation function is used to perform layer-by-layer feature space mapping to expand the dimension of the low-dimensional physical boundary parameters and convert them into high-dimensional dense latent feature vectors.

[0072] Using tensor splicing operations, the expanded high-dimensional dense latent feature vector and the encapsulated three-dimensional geometric attributes are tensor aligned and concatenated along the preset feature channel dimension. Through channel combination calculation of multimodal data, the physical external load conditions in numerical simulation are deeply bound to the three-dimensional spatial structural features of the nodes, generating initial node features that combine spatial geometric position awareness and physical external load constraint information.

[0073] By extracting node spatial geometric parameters and combining high-dimensional mapping and channel splicing fusion techniques of multilayer perceptron, initial node features with both spatial geometric position perception and physical external load constraint information were obtained, realizing deep feature binding and enhanced expressive power of multi-source heterogeneous cross-field data.

[0074] The process of calculating the physical field coupling interaction weights between adjacent nodes includes:

[0075] Based on the Euclidean distance between the fluid domain nodes and the solid domain nodes, a dynamic cutoff radius is set. Within the dynamic cutoff radius, a neighborhood search algorithm is used to establish directed connecting edges across the fluid-solid interface, forming the cross-field topological connection relationship. The three-dimensional geometric attributes of the nodes connected at both ends of the directed connecting edge are extracted, and a linear transformation is performed to generate a query vector and a key vector. The dot product of the query vector and the key vector is calculated, divided by a scaling factor, and activated by a normalized exponential function to obtain the attention distribution coefficient characterizing the intensity of the interaction between fluid kinetic energy transfer and solid deformation, which is used as the weight of the physical field coupling interaction.

[0076] The sum of squares of the coordinate differences between fluid domain nodes and solid domain nodes in the three-dimensional spatial coordinate system is calculated and the square root is taken to obtain the Euclidean distance between them; the average edge length of the local mesh at the fluid-solid interface is extracted and multiplied by a preset proportional expansion coefficient to calculate the dynamic cutoff radius; using a multidimensional spatial index tree as a neighborhood search algorithm, adjacent nodes on the opposite side of the fluid domain or solid domain are quickly retrieved within the spherical space of the dynamic cutoff radius, and directed connection edges across the fluid-solid interface are established along the direction of transmission of physical forces to form a cross-field topological connection relationship that characterizes the spatial adjacency state;

[0077] Traverse each directed connection edge in the cross-field topology connection relationship, and read the three-dimensional geometric attributes of its source node and target node from the computer memory respectively; set a first weight matrix and a second weight matrix containing learnable network parameters in the model architecture, perform matrix multiplication operation on the three-dimensional geometric attributes of the source node with the first weight matrix, and generate the query vector through linear transformation; perform matrix multiplication operation on the three-dimensional geometric attributes of the target node with the second weight matrix, and generate the key vector through linear transformation.

[0078] The query vector and the key vector are extracted and subjected to a tensor inner product operation to calculate their dot product, thereby quantifying the similarity and correlation of features of adjacent nodes in the latent space. The feature channel dimension value of the key vector is read, and a square root operation is performed on this dimension value to obtain a scaling factor. The dot product is divided by the scaling factor to control the variance of the inner product value and prevent gradient vanishing during backpropagation. The scaled value is then input into a normalized exponential function for nonlinear activation, the mathematical expression of which can be expressed as: ,in For query vector, For key vectors, The dimension value corresponding to the scaling factor. The normalized probability distribution value is mapped to the interval between zero and one. The attention distribution coefficient, which characterizes the intensity of the interaction between fluid kinetic energy transfer and solid deformation, is obtained by calculating this formula and is assigned as the weight of the physical field coupling interaction.

[0079] Before performing nonlinear activation operations, a dynamic physical mask matrix based on the fluid boundary layer thickness is introduced for truncation. The fluid boundary layer thickness at the current interface is estimated in real-time using the fluid dynamics wall law, based on the incoming wind speed and the surface roughness of the photovoltaic support. A dynamic physical mask matrix with the same dimension as the dot product tensor is generated based on this boundary layer thickness. The dot product values ​​of the connecting edges whose spatial distance between the source and target nodes is greater than the boundary layer thickness are truncated by extreme negative values, while the connecting edges located inside the boundary layer thickness retain their original dot product values. The dot product tensor superimposed with the physical mask matrix is ​​input into a normalized exponential function. This attention calculation mechanism, which introduces a physical boundary layer mask, can use prior knowledge of the physical flow field to forcibly cut off unreasonable long-distance false information transmission, accelerating the convergence of the graph neural network and enhancing the physical interpretability of multi-field coupling.

[0080] By using cross-field spatial neighborhood topology search based on dynamic truncation radius and vector inner product activation technique with attention mechanism, we obtained the physical field coupling interaction weights that characterize the interaction strength of cross-field nodes, and realized efficient implicit feature quantification of fluid kinetic energy and solid deformation interaction.

[0081] The process of extracting node interaction representation vectors includes:

[0082] Along the cross-field topological connection, the feature vectors of neighboring nodes are weighted and summed with the physical field coupling interaction weights to aggregate cross-field physical information within the local receptive field; the aggregated features are concatenated with the features of the target node itself and input into the graph neural network transmission unit containing a nonlinear activation function and a normalization layer, and the updated node hidden layer state is output; the cross-field physical information aggregation and update steps are executed cyclically until a preset message passing depth is reached; the feature vector output by the last layer of the network at the message passing depth is extracted as the node interaction representation vector.

[0083] Read the neighbor node index list in the cross-field topology connection relationship, and extract the hidden layer feature vector corresponding to each neighbor node; perform element-wise scalar multiplication operation on the feature vector of the neighbor node and its corresponding physical field coupling interaction weight, traverse all weighted neighbor feature vectors in the local receptive field of the target node and perform tensor weighted summation operation to complete the feature convergence between adjacent nodes in the fluid domain and the solid domain, and realize the cross-field physical information aggregation in the local receptive field;

[0084] Extract the feature vector of the target node's current state. On the preset feature channel dimension, perform tensor concatenation and splicing operation with the target node's own features to form a joint feature matrix. Input the joint feature matrix into a pre-constructed graph neural network transmission unit. The feature dimension is mapped sequentially through the learnable linear weight matrix inside the unit. The mean and variance of the features are calculated by the normalization layer for standardized distribution correction. Then, it is input into a non-linear activation function for activation processing, thereby reconstructing and outputting the updated node hidden layer state.

[0085] An iteration counter is set in computer memory to control the number of forward propagation layers of the network. The updated hidden layer state of the nodes output each time is used as the initial input for the next graph network layer transmission. The cross-field physical information aggregation and update steps are executed cyclically, and the value of the counter is incremented synchronously. The preset message passing depth is a hyperparameter used to control the receptive field size and the number of iterations of the graph neural network. It is preset according to the spatial scale of the three-dimensional geometric topology mesh to fully capture the global fluid-structure interaction effect while avoiding the phenomenon of over-smoothing of node features due to excessive depth of the graph neural network. The features of each node are updated sequentially according to the preset message passing depth. After completing the message passing of the preset number of layers, the features of the last layer node are used as the node interaction representation vector and input into the subsequent decoding branch. The preset number of layers is used to limit the structural depth of the network forward propagation.

[0086] By employing techniques based on weighted summation of local information aggregation and implicit nonlinear updating of graph network layers, a node interaction representation vector integrating multi-physics attributes and spatial connectivity was obtained, realizing implicit dimensionality reduction and computation of the highly nonlinear dynamic interaction mechanism between hydrodynamics and solid mechanics of photovoltaic supports.

[0087] The process of calculating the results of the multi-field coupling numerical calculation includes:

[0088] The node interaction representation vectors are input in parallel into the solid mechanics decoder and the fluid mechanics decoder, both of which contain multi-branch feedforward neural networks. The solid mechanics decoder generates deviatoric stress tensor features and displacement vector features through multi-branch network mapping, and performs physical consistency correction on the stress tensor and displacement vector through a linear elastic physical constraint layer to reconstruct the output dynamic stress field and strain-displacement field. The fluid mechanics decoder generates three-dimensional velocity vector field features through multi-branch network, introduces a divergence penalty mapping layer to ensure that the velocity vector field satisfies the fluid incompressibility condition, and reconstructs the output flow field pressure distribution and vortex structure features. The dynamic stress field, the strain-displacement field, the flow field pressure distribution, and the vortex structure features are summarized to form the multi-field coupled numerical calculation results.

[0089] Extract the node interaction representation vector output by implicit feature interaction, open two parallel computing paths in computer memory, and input the node interaction representation vector into the solid mechanics decoder and the fluid mechanics decoder in parallel; load the pre-built network architecture parameters of the solid mechanics decoder and the fluid mechanics decoder. The underlying architecture of both contains multi-branch feedforward neural networks with different neuron cascade depths. Use their weight matrices to perform feature dimensionality reduction and decoupling calculations on the high-dimensional hidden states.

[0090] Inside the solid mechanics decoder, the input features are distributed to independent network branches through a fully connected layer. Through forward propagation and activation mapping of the multi-branch network, a deviatoric stress tensor feature containing six independent components and a displacement vector feature containing three spatial components are calculated and generated. The deviatoric stress tensor feature and the displacement vector feature are extracted and input into the built-in linear elastic physical constraint layer. Geometric equations and generalized Hooke's law are used to establish a partial differential constraint calculation graph. The numerical deviations of the stress tensor and the displacement vector are physically consistent and corrected. Then, a dynamic stress field and strain-displacement field that conform to the laws of continuous medium mechanics are reconstructed and output.

[0091] Within the fluid dynamics decoder, the input features are processed through a multi-branch network using forward regression mapping to generate a three-dimensional velocity vector field feature in a spatial coordinate system. The fluid decoding branch outputs the velocity field based on the node interaction representation vector and calculates the divergence residual of the velocity field. The square norm of the divergence residual is multiplied by a weighting coefficient and added as a divergence-free penalty term to the loss function of the network model during training, reducing the divergence residual of the velocity field during training. Based on the constraint correction of the fluid incompressibility condition, the pressure distribution and vortex structure features of the output flow field are reconstructed by solving the discretized pressure Poisson equation and extracting the curl of the velocity field.

[0092] In determining the penalty weight coefficient In this process, an adaptive loss weight adjustment mechanism based on the intensity of local vortex flow is introduced. Specifically, after the multi-branch network outputs the three-dimensional velocity vector field characteristics, the flow field curl amplitude in the spatial neighborhood of the current node is calculated in parallel to characterize the intensity of the local vortex structure. This curl amplitude is mapped to a dynamic adjustment factor for the weight of the divergence-free penalty term in the network loss function, so that the penalty weight coefficient is automatically amplified in the vortex separation region with large curl, while the penalty weight coefficient is relatively reduced in the laminar flow region where the flow field is translated. This achieves a dynamic adaptive balance between the data regression error and the partial differential error of the physical equation in the decoding model. Through this adaptive weight allocation based on local fluid physics characteristics, gradient conflicts caused by a globally fixed penalty coefficient are avoided, effectively improving the flow field reconstruction accuracy of the complex vortex region in the wake of the photovoltaic support.

[0093] From the output ports of the two parallel decoders mentioned above, the corrected dynamic stress field, strain displacement field, flow pressure distribution, and eddy structure features are extracted synchronously. In a unified tensor space, the above field features are dimensionally aligned and tensor-stitched based on the spatial grid node index to complete the aggregation of heterogeneous physical field data, and finally form a multi-field coupled numerical calculation result that characterizes the stress state of the photovoltaic support at the current timestamp.

[0094] By introducing a multi-branch feedforward neural network decoding technique combined with fluid-structure interaction control equation constraint verification, multi-field coupled numerical calculation results that satisfy the consistency principle of fluid incompressibility and solid linear elasticity were obtained, realizing the accurate reconstruction and decoding of high-dimensional implicit features into macroscopic spatial physical field quantities.

[0095] The calculation method further includes a time-series recursion mechanism to ensure the continuity of physical states. The time-series recursion mechanism includes: setting a hidden state unit with a gated cyclic memory structure; reducing the dimensionality of the node interaction representation vector at the current timestamp through a fully connected layer, saving and updating it in the context matrix of the hidden state unit; when extracting the initial features of the node at the next timestamp, reading the context matrix as historical prior physical features, performing residual connection fusion with the features of the next timestamp, and performing multi-field coupled dynamic temporal evolution calculation.

[0096] A time-series recursive network framework is initialized in computer memory, and a network structure containing forget gates, input gates, and output gates for activation computation logic is constructed to set up hidden state units with gated recurrent memory structures. Inside the hidden state units, an initial weight matrix is ​​allocated to initialize the internal memory cell states, and a low-level tensor space is established to store and track the flow of features at continuous time steps.

[0097] The node interaction representation vector at the current timestamp is extracted from the graph neural network output and input into a dimensionality reduction module consisting of linear transformation layers. By multiplying the feature tensor with a learnable projection matrix and adding a bias term, redundant information in the channel is eliminated, so that the node interaction representation vector at the current timestamp is output after dimensionality reduction by a fully connected layer. The dimensionality-reduced feature vector is input into the hidden state unit, and the gating state is controlled by the internal activation function Sigmoid function. The amplitude of the feature vector is adjusted by the Tanh function to filter out high-frequency oscillation noise and retain key features such as low-frequency structural deformation. This is saved as time-series state data and updated in the context matrix of the hidden state unit.

[0098] As the simulation time step moves forward, when extracting the initial features of the node at the next time stamp, the context matrix is ​​read through memory addressing instructions; the read matrix data is directly mapped as the historical prior physical features characterizing the hysteresis effect of the fluid-solid interaction at the previous moment; in the corresponding dimension of the feature tensor, the historical prior physical features and the extracted features of the next time stamp are added element by element, and the residual connection is fused with the features of the next time stamp through a shortcut feedback link;

[0099] Using the joint features after residual connection fusion as the initial input benchmark for cross-field interaction of the graph network in the next physical moment, the numerical solution model is driven to evolve frame by frame at a fixed time step to complete the multi-field coupled dynamic temporal evolution calculation.

[0100] By introducing time-series recursion with gated memory structure and cross-time-step residual fusion techniques, we obtained the context matrix and residual fusion features that characterize the continuous evolution of physical states, and realized efficient time-series transmission of fluid separation vortices and solid vibration hysteresis effects between discrete time nodes.

[0101] Example 2

[0102] This embodiment applies a digital twin-based multi-field coupling numerical calculation method for photovoltaic supports to a large-scale photovoltaic power station in a windy and sandy area, focusing on the bidirectional fluid-structure interaction dynamic stress analysis and safety assessment process of a tracking photovoltaic support under extreme gust conditions.

[0103] In the actual service environment of this large-scale photovoltaic power station, the tracking photovoltaic support system, due to its large span and high flexibility, is prone to highly nonlinear fluid-structure interaction phenomena such as torsional flutter when facing complex wind conditions. Traditional unidirectional static equivalent calculation methods often neglect the reaction effect of support deformation on the surrounding wind field distribution, leading to predicted failure under extreme conditions.

[0104] The specific implementation process and effects of this invention are as follows:

[0105] Reference Figures 1 to 3 The initial 3D geometric model of the target photovoltaic support structure (including the main shaft, purlins, rotary reducer, columns, and photovoltaic modules) at the array edge of the photovoltaic power station, which is most severely affected by wind speed, was extracted from the 3D asset library of the digital twin platform. When generating the 3D geometric topology mesh, the algorithm calculated the principal curvature and Gaussian curvature of the support structure model surface. For structural feature points where the curvature gradient exceeds a preset threshold, such as the flange connecting the main shaft and rotary reducer, the ends of the purlins, and the edges of the photovoltaic modules, as well as areas with abrupt curvature changes, an adaptive background mesh size control strategy was adopted.

[0106] When the meteorological station detected a sudden severe convective weather event, the system received real-time data on the wind speed profile of a sudden gust of 25 m / s and the wind direction deflection angle data at a 45-degree angle, while the ambient temperature dropped sharply. Simultaneously, strain gauges located at the bottom of the main beam of the photovoltaic support and displacement sensors at the rotation drive also transmitted high-frequency data on the strain tensor and local spatial displacement vector of the structural surface.

[0107] Because there is a significant difference between the sampling frequency of meteorological data (e.g., 10Hz) and physical sensor data (e.g., 100Hz), a cubic spline interpolation algorithm based on a 100ms time window was used to accurately align the fluid dynamics boundary features and solid mechanics boundary features along the time axis. Subsequently, all features were mapped to a dimensionless feature space for normalization, and the multi-field coupled boundary conditions under the current gust wind condition were fused and accurately mapped onto the previously generated three-dimensional mesh discrete nodes. Through high-dimensional feature embedding of a multilayer perceptron, the low-dimensional sensor physical parameters were transformed into high-dimensional dense latent feature vectors, which were then concatenated with the three-dimensional geometric attributes of the nodes to complete the construction of the initial features.

[0108] During the impact of gusts on photovoltaic modules, the kinetic energy of the wind is converted into surface wind pressure on the modules, causing torsional deformation of the support structure. This torsion alters the angle of attack, thus changing the separation vortex structure of the module's wake. This invention establishes a directed connection edge across the fluid-solid interface by setting a dynamic cutoff radius between the fluid domain (air) nodes and the solid domain (photovoltaic panel) nodes. The system extracts features from connected nodes to generate query vectors and key vectors, and uses a normalized exponential function to activate and derive an attention distribution coefficient. This attention coefficient physically quantifies the intensity of gust kinetic energy transferred to the support structure, as well as the interference weight of the reaction of the support structure's minute torsional deformation on the surrounding flow field.

[0109] The system follows the aforementioned cross-field topological connections, performing weighted aggregation and nonlinear updates of local receptive field features at a preset depth within the graph neural network's transmission unit. Through iterative computation by the graph network, the highly nonlinear bidirectional fluid-structure interaction effect is implicitly quantized and extracted as a global node interaction representation vector. This node interaction representation vector, containing the global multi-field coupling effect, is input in parallel to both the solid mechanics decoder and the fluid mechanics decoder. To ensure that the AI ​​calculation results conform to real-world physical laws, the solid mechanics decoder, after outputting the deviatoric stress tensor and displacement vector features, invokes the generalized Hooke's law for linear elastic physical constraint correction; while the fluid mechanics decoder, when outputting the three-dimensional velocity vector field, introduces a divergence-free penalty term to ensure that the reconstructed wind field strictly satisfies the fluid incompressibility condition.

[0110] This invention preserves the evolution trend of the flow field separation vortices and the hysteresis effect of solid support vibration at the previous moment through the hidden state units of the built-in LSTM-gated cyclic memory structure and the residual connection mechanism. Continuous simulation results show that during the period from 1200ms to 1500ms after the gust impact, the photovoltaic support generated high-frequency micro-amplitude flutter due to aeroelastic instability. The deformation caused by wind load significantly changed the vortex shedding frequency at the trailing edge of the module. This dynamic cumulative effect caused the peak local dynamic stress at the main shaft connection flange to surge by approximately 43.5% compared to the initial static load prediction.

[0111] The beneficial engineering effects achieved by applying this invention to the photovoltaic support structure are as follows: Traditional methods fail to consider the bidirectional feedback mechanism of wind load and deformation. This invention captures the stress surge phenomenon caused by fluid-structure interaction during the period of 1200ms to 1500ms, restoring the true stress state of the physical entity under complex weather conditions. Based on the high-confidence multi-field coupling numerical calculation results output by this method, the photovoltaic power station's main control system can issue millisecond-level response defense and avoidance commands before structural failure, preventing damage to core mechanical components. Within the non-limiting fluid-structure interaction stress threshold, the support structure can still maintain a certain tracking angle and continue generating electricity. This invention provides a solid data foundation for this refined wind and solar control strategy, significantly improving the overall power generation efficiency of large-scale photovoltaic power stations in harsh environments.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins, characterized in that, include: The initial three-dimensional geometric model of the photovoltaic support is extracted from the digital twin space. The initial three-dimensional geometric model is spatially discretized to generate a three-dimensional geometric topology mesh composed of multiple discrete nodes. Real-time environmental meteorological sequence and physical sensor data of the photovoltaic support are acquired and analyzed. Fluid dynamics and solid mechanics features are extracted respectively to generate multi-field coupled boundary conditions. The multi-field coupling boundary conditions are mapped to the corresponding discrete nodes of the three-dimensional geometric topology mesh, the three-dimensional geometric properties of the discrete nodes are extracted, and the initial features of the nodes are constructed by splicing them together. Construct the cross-field topological connection relationship between the discrete nodes in the fluid domain and the solid domain, and calculate the physical field coupling interaction weight between adjacent nodes based on the cross-field topological connection relationship and the three-dimensional geometric properties; Based on the physical field coupling interaction weights, the initial features of the nodes are subjected to implicit feature interaction and iterative update across physical fields, and node interaction representation vectors containing global multi-field coupling effects are extracted; based on the node interaction representation vectors, the multi-field coupling numerical calculation results of the photovoltaic support at the current timestamp are output.

2. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The process of generating a three-dimensional geometric topology mesh includes: The principal curvature and Gaussian curvature of the initial 3D geometric model surface are calculated, and regions where the curvature gradient exceeds a preset threshold are identified as structural feature points and regions with abrupt curvature changes. Based on the structural feature points and the regions with abrupt curvature changes, an adaptive background mesh size control strategy is adopted to generate high-density discrete nodes with high spatial resolution in stress concentration regions and flow field boundary layer regions, and low-density discrete nodes with low spatial resolution in smooth, non-critical regions. The high-density discrete nodes and the low-density discrete nodes are connected using the Delaunay spatial triangulation algorithm, and the node connection edges in the transition region are smoothed using Laplacian smoothing to construct a 3D geometric topology mesh containing multi-scale non-uniform nodes.

3. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The process of generating multi-field coupled boundary conditions includes: Temporal characteristic parameters of wind speed profile distribution, wind direction deflection angle, and ambient temperature are extracted from the real-time environmental meteorological sequence as fluid dynamics boundary features; strain tensors and local spatial displacement vectors on the surface of key load-bearing components of the photovoltaic support are extracted from the physical sensing data as solid mechanics boundary features; a spline interpolation algorithm based on time windows is used to align the temporal frequency of the fluid dynamics boundary features and the solid mechanics boundary features, and the aligned features are mapped to a unified dimensionless feature space for normalization processing, and multi-field coupled boundary conditions are generated by fusion.

4. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The process of constructing initial node features includes: extracting the global absolute spatial coordinates, surface normal vector, and local neighborhood topological features of each discrete node, and combining them to form three-dimensional geometric attributes; inputting the multi-field coupling boundary conditions mapped onto the corresponding discrete node into a high-dimensional feature embedding layer composed of a multilayer perceptron, and converting the low-dimensional physical boundary parameters into a high-dimensional dense latent feature vector; and concatenating the high-dimensional dense latent feature vector with the three-dimensional geometric attributes in the channel dimension to generate initial node features that combine spatial geometric position awareness and physical external load constraint information.

5. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The process of calculating the physical field coupling interaction weights between adjacent nodes includes: Based on the Euclidean distance between the fluid domain nodes and the solid domain nodes, a dynamic cutoff radius is set. Within the dynamic cutoff radius, a neighborhood search algorithm is used to establish directed connecting edges across the fluid-solid interface, forming the cross-field topological connection relationship. The three-dimensional geometric attributes of the nodes connected at both ends of the directed connecting edge are extracted, and a linear transformation is performed to generate a query vector and a key vector. The dot product of the query vector and the key vector is calculated, divided by a scaling factor, and activated by a normalized exponential function to obtain the attention distribution coefficient characterizing the intensity of the interaction between fluid kinetic energy transfer and solid deformation, which is used as the weight of the physical field coupling interaction.

6. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The process of extracting node interaction representation vectors includes: Along the cross-field topological connection, the feature vectors of neighboring nodes are weighted and summed with the physical field coupling interaction weights to aggregate cross-field physical information within the local receptive field; the aggregated features are concatenated with the features of the target node itself and input into the graph neural network transmission unit containing a nonlinear activation function and a normalization layer, and the updated node hidden layer state is output; the cross-field physical information aggregation and update steps are executed cyclically until a preset message passing depth is reached; the feature vector output by the last layer of the network at the message passing depth is extracted as the node interaction representation vector.

7. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The multi-field coupled numerical calculation process includes: inputting the node interaction representation vectors into the solid mechanics decoder and the fluid mechanics decoder in parallel, respectively, wherein both the solid mechanics decoder and the fluid mechanics decoder contain a multi-branch feedforward neural network; The solid mechanics decoder generates deviatoric stress tensor features and displacement vector features through a multi-branch network mapping, and performs physical consistency correction on the stress tensor and displacement vector through a linear elastic physical constraint layer to reconstruct the output dynamic stress field and strain-displacement field. The fluid mechanics decoder generates three-dimensional velocity vector field features through a multi-branch network, introduces a divergence penalty mapping layer to ensure that the velocity vector field satisfies the fluid incompressibility condition, and reconstructs the output flow field pressure distribution and eddy structure features. The dynamic stress field, strain-displacement field, flow field pressure distribution, and eddy structure features are summarized to form the multi-field coupled numerical calculation results.

8. The numerical calculation method for multi-field coupling of photovoltaic supports based on digital twins according to claim 1, characterized in that, The calculation method further includes a time-series recursion mechanism to ensure the continuity of physical states. The time-series recursion mechanism includes: setting a hidden state unit with a gated cyclic memory structure; reducing the dimensionality of the node interaction representation vector at the current timestamp through a fully connected layer, saving and updating it in the context matrix of the hidden state unit; when extracting the initial features of the node at the next timestamp, reading the context matrix as historical prior physical features, performing residual connection fusion with the features of the next timestamp, and performing multi-field coupled dynamic temporal evolution calculation.

Images