A power transmission line real-time wind field reconstruction method based on physical information neural operator

Abstract

This invention discloses a real-time wind field reconstruction method for transmission lines based on a physical information neural operator, relating to the fields of digital power grid and artificial intelligence. It aims to address the pain point of existing wind field reconstruction technologies, which cannot simultaneously achieve reconstruction accuracy, real-time inference efficiency, and physical consistency. This invention first delineates the wind field computational domain of the transmission line and performs adaptive grid discretization, collects multi-source sparse observation data, constructs and standardizes the initial sparse input tensor, builds a PINO model containing a lifting module, a core processing module, and a projection module, and constructs a composite loss function that adaptively weights data and physical losses. After model training and optimal weight selection, real-time wind field reconstruction and physical residual correction of the transmission line are achieved based on the optimal model. This invention balances millisecond-level inference efficiency with high reconstruction accuracy, ensures wind field physical consistency, and meets the engineering requirements for power grid disaster prevention and early warning.

Description

Technical Field

[0001] This invention relates to the fields of digital power grid and artificial intelligence technology, specifically to a method for real-time wind farm reconstruction of transmission lines based on physical information neural operators. Background Technology

[0002] As the core artery of national energy transmission, the safe and stable operation of power transmission lines is the core foundation for the construction of new power systems and the prevention and mitigation of power grid disasters. my country's power transmission line corridors often traverse mountainous and valley terrains with complex and varied topography, and are exposed to harsh and variable micro-meteorological environments for extended periods. Among numerous meteorological disasters, strong winds and wind shear caused by strong winds and complex terrain are the main causes of wind-induced flashovers, line galloping, and even tower collapses in power transmission lines, seriously threatening the safe and stable operation of the power grid. Therefore, to achieve full-area, high-precision perception of wind fields along power transmission lines, and to support the construction of digital twin systems for power transmission lines and power grid disaster prevention, early warning, and dispatching, this field urgently needs a technology capable of real-time, high-precision, and highly physically reliable reconstruction of the micro-meteorological wind fields along power transmission line corridors.

[0003] However, existing technologies have shown significant limitations in addressing the engineering requirements of wind farm reconstruction along transmission lines, failing to simultaneously balance reconstruction accuracy, real-time performance, and physical reliability.

[0004] First, traditional spatial interpolation methods are limited by the deployment conditions along the transmission line. The sensor layout is extremely sparse. Simple mathematical interpolation methods cannot capture the fluid dynamic evolution characteristics of the sensor blank areas. They can only achieve numerical fitting of discrete sampling points. The reconstruction results lack physical credibility and are difficult to truly restore the spatial distribution of wind field under complex terrain. The error is huge in complex scenarios such as mountainous valleys, which cannot meet the accuracy requirements of line disaster prevention.

[0005] Second, traditional computational fluid dynamics (CFD) models; these methods have complete physical mechanisms and can accurately simulate the fluid evolution of wind fields, but their computational cost is extremely high. They require fine meshing and long-cycle iterative solutions for complex terrains. Simulation of a single scenario usually requires high-performance computing clusters to run for hours or even days, which cannot meet the urgent need for real-time response in power grid disaster prevention and dispatch and safety early warning, and is difficult to implement on a large scale.

[0006] Third, the Physical Information Neural Network (PINN) method, which has emerged in recent years; although this type of method introduces physical equation constraints and improves the physical consistency of the results to a certain extent, most existing models adopt point-to-point multilayer perceptron skeletons, which cannot efficiently capture the multi-scale spatial correlation of large-scale wind fields. When facing the scenario of reconstructing wind fields across the entire transmission line, the training process converges slowly, the generalization ability is weak, and the inference speed is still difficult to meet the standards of real-time power grid applications.

[0007] Fourth, data-driven methods, represented by Fourier Neural Operators (FNO), possess the speed advantage of millisecond-level inference due to the characteristics of global convolution in the frequency domain. They can efficiently capture the spatial evolution characteristics of the flow field. However, they are essentially pure data-driven black-box models. In the vast blank areas along the transmission line where sensor observations are extremely sparse, if there is a lack of clear physical law constraints, the flow field generated by them often exhibits phenomena such as non-conservation of mass and sudden changes in wind speed and pressure, which do not conform to the basic laws of fluid mechanics. This results in extremely low physical reliability of the reconstruction results, which can easily lead to misjudgment and omission in actual power engineering disaster prevention applications, and cannot guarantee the safe operation of the transmission line.

[0008] Therefore, it is necessary to design a real-time wind farm reconstruction method for transmission lines based on physical information neural operators. Summary of the Invention

[0009] The purpose of this invention is to provide a real-time wind farm reconstruction method for transmission lines based on physical information neural operators, in order to solve the problems mentioned in the background art that the existing transmission line wind farm reconstruction technology cannot simultaneously take into account reconstruction accuracy, real-time inference efficiency and physical consistency, is difficult to adapt to the sparse sensor layout scenario along the line, and cannot meet the needs of power grid disaster prevention early warning and safe dispatch engineering.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time wind farm reconstruction of transmission lines based on physical information neural operators, comprising the following steps:

[0011] S1: Divide the wind field calculation domain, obtain meteorological grid data, terrain elevation data, and conductor vibration monitoring data along the transmission line within the wind field calculation domain; extract specific sampling point data from the multi-source data according to the actual spatial distribution of micro-meteorological sensors along the transmission line, construct an initial sparse input tensor, and perform preprocessing.

[0012] S2: Construct a physical information neural operator PINO model containing Fourier neural operators, and use the weighted sum of the model data loss term and the physical loss term as the model composite loss function. The weights used to balance the two losses in the composite loss function are dynamically adjusted adaptively by training batches. The physical loss term is constructed based on the residual constraints of the fluid dynamics control equation.

[0013] S3: The PINO model is trained iteratively with the preprocessed initial sparse input tensor as the input and the high-resolution continuous wind field data as the output. The model is trained with the goal of minimizing the composite loss function. The parameters of the PINO model are corrected by the backpropagation algorithm. The optimal model weights are selected and saved based on the generalization performance of the validation set.

[0014] S4: Deploy the optimal PINO model saved after training at the inference end, input sparse observation data along the transmission line in real time and perform forward inference, perform physical residual constraint post-processing correction on the wind field data output by inference, and finally obtain the real-time reconstructed wind field of the entire transmission line through post-processing calculation.

[0015] As a further technical solution of the present invention, S1 specifically includes the following:

[0016] (1) Wind field computational domain partitioning and adaptive grid discretization:

[0017] A rectangular computational domain covering the transmission line and the surrounding wind field influence area is delineated based on the geographical coordinates of the transmission line. The computational domain is discretized into a two-dimensional global grid. According to the adaptive densification rules of the transmission line corridor, the minimum spatial step size is adopted for the grid within the preset range of the line centerline, and the spatial step size is gradually increased by a fixed coefficient for the grid outside the preset range. Simultaneously, digital elevation model data within the computational domain is acquired, and the registration of the terrain elevation data with the global grid is completed to obtain a terrain elevation matrix with the same dimensions as the global grid.

[0018] (2) Acquisition of multi-source basic data and feature decomposition:

[0019] High-resolution meteorological grid data, sparse observation data from micro-meteorological sensors along the transmission line, and conductor vibration monitoring data were collected within the computational domain to achieve time synchronization and spatial registration of multi-source data. Wind speed, wind direction, air pressure, and ambient temperature data were extracted from the meteorological data to calculate zonal and meridional wind speeds. Based on the amplitude and frequency of conductor vibration monitoring, the wind speed components of the corresponding sampling points were obtained through inversion using wind-induced vibration theory. The original meteorological data were corrected by combining topographic elevation data to obtain wind speed and air pressure data adapted to the effects of terrain shading and acceleration.

[0020] (3) Sparse sampling point data extraction and initial sparse input tensor construction:

[0021] Based on the actual spatial distribution of micro-meteorological sensors and vibration monitoring sensors, sampling point data is extracted from the registered multi-source grid data, coordinate registration and grid mapping are completed, and an initial sparse input tensor containing multiple channels including zonal wind speed, meridional wind speed, air pressure, terrain elevation, and conductor vibration characteristics is constructed.

[0022] (4) Standardization preprocessing of initial sparse input tensors:

[0023] The original feature data of the initial sparse input tensor is standardized to unify the value range of each feature data to a preset range. The processing is adapted and adjusted according to the extreme value range of the corresponding feature channel, and finally encapsulated into a standardized initial sparse input tensor that matches the spatial distribution dimension of the sensor.

[0024] As a further technical solution of the present invention, S2 specifically includes the following:

[0025] (1) Overall architecture construction of the PINO model:

[0026] Construct a physical information neural operator model consisting of a cascaded lifting module, a core processing module, and a projection module;

[0027] The boosting module maps the normalized initial sparse input tensor from the sparse sampling dimension to the global grid dimension through a fully connected layer, and then projects it to a high-dimensional latent feature space through a convolutional layer.

[0028] The core processing module works through multi-level Fourier neural operator units. Each unit performs frequency domain processing on the input features by fast Fourier transform, learnable linear transform in the spectral domain, and inverse fast Fourier transform, as well as spatial domain processing by one-dimensional convolution. The two processing results are then fused with the input features and output after nonlinear activation.

[0029] The projection module maps high-dimensional features back to physical space through two fully connected layers, outputting high-resolution continuous wind field data with full-domain dimensions;

[0030] (2) Construction of the composite loss function:

[0031] The weighted sum of the data loss term and the physical loss term is used as the composite loss function, where the weights of the two loss terms are dynamically adjusted according to the ratio of the mean of the full-grid physical residuals of the current training batch to the mean of the fitting residuals of the sampled data.

[0032] Data loss term: The degree of fitting deviation between the global grid zonal wind speed, meridional wind speed, and air pressure output by the model and the multi-source fusion ground truth data is used as the data loss; the multi-source fusion ground truth data is obtained by weighted fusion of meteorological sensor measured data, conductor vibration inversion data, and terrain correction data according to the confidence weight of each data source;

[0033] Physical loss term: The residual deviation of the two-dimensional steady incompressible Navier-Stokes fluid dynamics governing equations is used as the physical loss constraint. The residual of the continuity equation is calculated based on the combined deviation of the zonal wind speed change rate along the zonal direction and the meridional wind speed change rate along the meridional direction; the residual of the zonal momentum equation is calculated based on the combined deviation of the convection term, pressure gradient term, and viscous diffusion term; the residual of the meridional momentum equation is calculated based on the combined deviation of the convection term, meridional pressure gradient term, and viscous diffusion term; air density is dynamically corrected based on real-time air pressure, ambient temperature, and terrain elevation; and air kinematic viscosity is dynamically corrected based on the deviation of ambient temperature from standard temperature. Adaptive weights are used for the residuals of each equation in the physical loss. The weight of the residual of the continuity equation is adjusted based on the deviation of the grid elevation from the average elevation of the line, and the weight of the residual of the momentum equation is adjusted based on the distance between the grid and the centerline of the line.

[0034] As a further technical solution of the present invention, S3 specifically includes the following:

[0035] (1) Temporal partitioning of the dataset:

[0036] Multi-source meteorological grid data is divided into training set, validation set and test set according to time dimension in a fixed proportion. The maximum and minimum values ​​of each feature channel in the training set are fixed as fixed parameters for the whole process normalization.

[0037] (2) Model iterative training optimization:

[0038] An adaptive moment estimation optimizer combined with a cosine annealing learning rate strategy is used to train the model. The learning rate is dynamically adjusted between the maximum and minimum values ​​according to the preset cosine variation law based on the number of iterations. The optimization objective is to minimize the composite loss function, and all parameters of the model are corrected through the backpropagation algorithm. After each training round, the validation set data is inferred, and the inference results are corrected by physical residuals and used as pseudo-labels to supplement the data loss calculation to achieve weakly supervised learning.

[0039] (3) Model generalization evaluation and optimal weight preservation:

[0040] The generalization ability of the model is evaluated by the validation set loss, and the training is terminated by an early stopping mechanism. The model weights with the lowest validation set loss are selected as the optimal weights and saved.

[0041] As a further technical solution of the present invention, S4 specifically includes the following:

[0042] (1) Deployment of the inference model and real-time forward inference:

[0043] After the training is completed, the model is loaded to the inference end, sparse sensor data is collected in real time, and the standardized initial sparse input tensor is constructed by reusing the training set with fixed normalization parameters. The high-resolution initial reconstructed wind field is obtained through one forward propagation of the model.

[0044] (2) Post-processing correction of physical residual constraints in the inference results:

[0045] Calculate the residuals of the full-grid hydrodynamic equations of the initial reconstructed wind field. For grid regions where the residuals exceed the preset threshold, wind speed and air pressure are successively corrected according to the preset correction step size until the full-grid residuals meet the preset convergence conditions or reach the maximum number of corrections. Output the corrected full-domain high-resolution wind field data.

[0046] (3) Wind farm post-processing and engineering output:

[0047] The corrected zonal and meridional wind speeds are vectorized to obtain the wind speed scalar and meteorological standard wind direction of the entire grid. Based on the geographical coordinates of the transmission line, the wind field data specific to the transmission line corridor is extracted from the wind field of the entire area and output to the transmission line wind deflection and galloping early warning system to complete the real-time wind field reconstruction.

[0048] As a further technical solution of the present invention, the preprocessing of the initial sparse input tensor includes normalization:

[0049]

[0050] in: : The original feature data of the initial sparse input tensor; The maximum and minimum values ​​of the corresponding feature channels in the training set are fixed during the training phase and directly reused during the inference phase to ensure consistent distribution of training and inference data. : Standardized feature data after normalization;

[0051] Then, based on the sensor arrangement order, it is packaged into a dimension of The standardized initial sparse input tensor, : These represent the latitudinal and longitudinal dimensions of the sparse sampling points, corresponding to the spatial distribution of the sensors, and the global grid dimension, respectively. To establish a mapping relationship, the model enhancement module completes the mapping from sparse dimensions to global dimensions.

[0052] As a further technical solution of the present invention, the composite loss function is:

[0053]

[0054] in: Data loss function; Physical loss function; : The hyperparameter weights that balance the two losses;

[0055] Data loss items:

[0056]

[0057] in: The total number of pixels in the global grid. ; : Represents the width and height of the global grid, respectively; PINO model in global mesh The output shows the zonal wind speed, meridional wind speed, and air pressure values. Multi-source fusion of truth data;

[0058] Physical loss term: Constructed based on the residual constraints of the two-dimensional steady incompressible Navier-Stkes fluid dynamics governing equations, and embedded in the loss function in the form of physical residuals of the equations:

[0059] Continuity equation residuals:

[0060]

[0061] Residual of momentum equation in the x-direction:

[0062]

[0063] y-direction momentum equation residual:

[0064]

[0065] Physical loss:

[0066]

[0067] in: This refers to the zonal wind speed, i.e., the x-direction. This represents the meridional wind speed, i.e., the y-direction. Air density; : Air kinematic viscosity; : These are the adaptive weights of the residuals of the continuity equation, the latitudinal momentum equation, and the meridional momentum equation, respectively.

[0068] As a further technical solution of the present invention, the hyperparameter weights for balancing the two losses are... :

[0069]

[0070] in: : Preset fixed base weight values; The mean absolute value of the physical residuals across the entire grid in the current training batch; : The mean absolute value of the fitting residuals between the model output and the ground truth data at all sampling points within the current training batch.

[0071] As a further technical solution of the present invention, the air density Employing elevation-temperature adaptive dynamic correction:

[0072]

[0073] air kinematic viscosity Temperature adaptive dynamic correction is adopted:

[0074]

[0075] in: : The standard value of air density under standard conditions; : Reference value for air kinematic viscosity under standard conditions; Standard atmosphere; Standard temperature; Real-time air pressure and ambient temperature measured by micro-meteorological sensors; Global Grid The corresponding terrain elevation value; : Gravitational acceleration; : Gas constant for dry air; Temperature correction factor.

[0076] As a further technical solution of the present invention, the adaptive weights of the continuity equation residual, the latitudinal momentum equation residual, and the meridional momentum equation residual are... :

[0077]

[0078]

[0079] in: The average elevation of the entire transmission line; : The maximum elevation difference within the computational domain; : Preset weight amplification factor; Global Grid The corresponding terrain elevation value; Global Grid The vertical distance to the centerline of the transmission line corridor; The pre-set half-width of the transmission line corridor; : The preset minimum value.

[0080] Compared with existing technologies, the advantages of this method for real-time wind farm reconstruction of transmission lines based on physical information neural operators are:

[0081] By utilizing the PINO model, a physical information neural operator, and combining the high computational efficiency of deep learning with the physical constraints of fluid mechanics, compared to purely data-driven black-box models, this approach significantly improves the model's generalization ability and physical consistency under sparse observation data conditions. This is achieved by constructing a composite loss function that adaptively weights data loss and physical loss, combined with weakly supervised pseudo-label optimization during training and iterative correction of physical residuals during inference. This avoids abnormal abrupt changes in the reconstructed wind field that violate the laws of fluid mechanics. Furthermore, by integrating multi-source features from transmission line conductor vibration inversion data, terrain correction data, and micro-meteorological observation data, the scientific validity and reliability of the wind field reconstruction results are further ensured.

[0082] The PINO model's boosting module, through a cascaded structure of grid mapping units and channel boosting units, completes accurate mapping from sparse sampling dimensions to global regular grid dimensions and feature channel expansion based on a fully connected layer. Combined with the transmission line corridor adaptive densification grid discretization mechanism, it effectively solves the adaptation problem between sparse unstructured sensor observation data along the transmission line and the regular grid input required by the model, realizing efficient and high-precision information transmission from discrete sensor sampling points to the global continuous wind field.

[0083] The PINO model, through the parallel processing characteristics of frequency domain global spectral convolution and spatial domain linear transformation of the core Fourier neural operator unit, combined with the gradient stabilization mechanism of skip connections, can efficiently capture the multi-scale spatial correlation of wind fields. It can not only accurately reconstruct large-scale background wind fields, but also restore the micro-scale local wind field characteristics along the transmission line affected by terrain and line structure. At the same time, the full-link embedding of terrain elevation data and dynamic adaptive fluid parameter correction further enhances the fitting ability of local micro-meteorological characteristics and significantly improves the reconstruction accuracy of wind fields in transmission line corridors.

[0084] The trained PINO model has extremely high end-to-end inference efficiency, requiring only one forward propagation to complete the reconstruction of the entire wind field. With the addition of lightweight residual iterative correction, it can achieve millisecond-level wind field inference output. Compared with traditional computational fluid dynamics (CFD) simulation methods, the inference speed is improved by several orders of magnitude. It does not require complex computational domain meshing and long-cycle iterative solutions, and can directly complete real-time wind field reconstruction based on sparse sensor data along the transmission line. It fully meets the stringent real-time requirements of power grid dispatching and transmission line wind deflection and galloping safety early warning, and has strong engineering feasibility and important engineering application value. Attached Figure Description

[0085] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0086] Figure 2 This is a diagram showing the wind field reconstruction results based on the PINO model of this invention. Detailed Implementation

[0087] 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.

[0088] Please see the appendix Figure 1 The present invention provides an embodiment of a method for real-time wind farm reconstruction of transmission lines based on physical information neural operators, comprising the following steps:

[0089] S1: Divide the wind field calculation domain, obtain meteorological grid data, terrain elevation data, and conductor vibration monitoring data along the transmission line within the wind field calculation domain; extract specific sampling point data from the multi-source data according to the actual spatial distribution of micro-meteorological sensors along the transmission line, construct an initial sparse input tensor, and perform preprocessing.

[0090] S2: Construct a physical information neural operator PINO model containing Fourier neural operators, and use the weighted sum of the model data loss term and the physical loss term as the model composite loss function. The weights used to balance the two losses in the composite loss function are dynamically adjusted adaptively by training batches. The physical loss term is constructed based on the residual constraints of the fluid dynamics control equation.

[0091] S3: The PINO model is trained iteratively with the preprocessed initial sparse input tensor as the input and the high-resolution continuous wind field data as the output. The model is trained with the goal of minimizing the composite loss function. The parameters of the PINO model are corrected by the backpropagation algorithm. The optimal model weights are selected and saved based on the generalization performance of the validation set.

[0092] S4: Deploy the optimal PINO model saved after training at the inference end, input sparse observation data along the transmission line in real time and perform forward inference, perform physical residual constraint post-processing correction on the wind field data output by inference, and finally obtain the real-time reconstructed wind field of the entire transmission line through post-processing calculation.

[0093] Furthermore, in another embodiment, S1 specifically includes the following:

[0094] (1) Wind field computational domain partitioning and adaptive grid discretization:

[0095] A rectangular computational domain covering the transmission line and the surrounding wind field influence area is delineated based on the geographical coordinates of the transmission line. The computational domain is discretized into a two-dimensional global grid. According to the adaptive densification rules of the transmission line corridor, the minimum spatial step size is adopted for the grid within the preset range of the line centerline, and the spatial step size is gradually increased by a fixed coefficient for the grid outside the preset range. Simultaneously, digital elevation model data within the computational domain is acquired, and the registration of the terrain elevation data with the global grid is completed to obtain a terrain elevation matrix with the same dimensions as the global grid.

[0096] (2) Acquisition of multi-source basic data and feature decomposition:

[0097] High-resolution meteorological grid data, sparse observation data from micro-meteorological sensors along the transmission line, and conductor vibration monitoring data were collected within the computational domain to achieve time synchronization and spatial registration of multi-source data. Wind speed, wind direction, air pressure, and ambient temperature data were extracted from the meteorological data to calculate zonal and meridional wind speeds. Based on the amplitude and frequency of conductor vibration monitoring, the wind speed components of the corresponding sampling points were obtained through inversion using wind-induced vibration theory. The original meteorological data were corrected by combining topographic elevation data to obtain wind speed and air pressure data adapted to the effects of terrain shading and acceleration.

[0098] (3) Sparse sampling point data extraction and initial sparse input tensor construction:

[0099] Based on the actual spatial distribution of micro-meteorological sensors and vibration monitoring sensors, sampling point data is extracted from the registered multi-source grid data, coordinate registration and grid mapping are completed, and an initial sparse input tensor containing multiple channels including zonal wind speed, meridional wind speed, air pressure, terrain elevation, and conductor vibration characteristics is constructed.

[0100] (4) Standardization preprocessing of initial sparse input tensors:

[0101] The original feature data of the initial sparse input tensor is standardized to unify the value range of each feature data to a preset range. The processing is adapted and adjusted according to the extreme value range of the corresponding feature channel, and finally encapsulated into a standardized initial sparse input tensor that matches the spatial distribution dimension of the sensor.

[0102] Furthermore, in another embodiment, S2 specifically includes the following:

[0103] (1) Overall architecture construction of the PINO model:

[0104] Construct a physical information neural operator model consisting of a cascaded lifting module, a core processing module, and a projection module;

[0105] The boosting module maps the normalized initial sparse input tensor from the sparse sampling dimension to the global grid dimension through a fully connected layer, and then projects it to a high-dimensional latent feature space through a convolutional layer.

[0106] The core processing module works through multi-level Fourier neural operator units. Each unit performs frequency domain processing on the input features by fast Fourier transform, learnable linear transform in the spectral domain, and inverse fast Fourier transform, as well as spatial domain processing by one-dimensional convolution. The two processing results are then fused with the input features and output after nonlinear activation.

[0107] The projection module maps high-dimensional features back to physical space through two fully connected layers, outputting high-resolution continuous wind field data with full-domain dimensions;

[0108] (2) Construction of the composite loss function:

[0109] The weighted sum of the data loss term and the physical loss term is used as the composite loss function, where the weights of the two loss terms are dynamically adjusted according to the ratio of the mean of the full-grid physical residuals of the current training batch to the mean of the fitting residuals of the sampled data.

[0110] Data loss term: The degree of fitting deviation between the global grid zonal wind speed, meridional wind speed, and air pressure output by the model and the multi-source fusion ground truth data is used as the data loss; the multi-source fusion ground truth data is obtained by weighted fusion of meteorological sensor measured data, conductor vibration inversion data, and terrain correction data according to the confidence weight of each data source;

[0111] Physical loss term: The residual deviation of the two-dimensional steady incompressible Navier-Stokes fluid dynamics governing equations is used as the physical loss constraint. The residual of the continuity equation is calculated based on the combined deviation of the rate of change of zonal wind speed along the zonal direction and the rate of change of meridional wind speed along the meridional direction; the residual of the zonal momentum equation is calculated based on the combined deviation of the convection term, pressure gradient term, and viscous diffusion term; the residual of the meridional momentum equation is calculated based on the combined deviation of the convection term, meridional pressure gradient term, and viscous diffusion term; air density is dynamically corrected based on real-time air pressure, ambient temperature, and terrain elevation; and air kinematic viscosity is dynamically corrected based on the deviation of ambient temperature from standard temperature. Adaptive weights are used for the residuals of each equation in the physical loss. The weight of the residual of the continuity equation is adjusted based on the deviation of the grid elevation from the average elevation of the line, and the weight of the residual of the momentum equation is adjusted based on the distance between the grid and the centerline of the line.

[0112] Furthermore, in another embodiment, S3 specifically includes the following:

[0113] (1) Temporal partitioning of the dataset:

[0114] Multi-source meteorological grid data is divided into training set, validation set and test set according to time dimension in a fixed proportion. The maximum and minimum values ​​of each feature channel in the training set are fixed as fixed parameters for the whole process normalization.

[0115] (2) Model iterative training optimization:

[0116] An adaptive moment estimation optimizer combined with a cosine annealing learning rate strategy is used to train the model. The learning rate is dynamically adjusted between the maximum and minimum values ​​according to the preset cosine variation law based on the number of iterations. The optimization objective is to minimize the composite loss function, and all parameters of the model are corrected through the backpropagation algorithm. After each training round, the validation set data is inferred, and the inference results are corrected by physical residuals and used as pseudo-labels to supplement the data loss calculation to achieve weakly supervised learning.

[0117] (3) Model generalization evaluation and optimal weight preservation:

[0118] The generalization ability of the model is evaluated by the validation set loss, and the training is terminated by an early stopping mechanism. The model weights with the lowest validation set loss are selected as the optimal weights and saved.

[0119] Furthermore, in another embodiment, S4 specifically includes the following:

[0120] (1) Deployment of the inference model and real-time forward inference:

[0121] After the training is completed, the model is loaded to the inference end, sparse sensor data is collected in real time, and the standardized initial sparse input tensor is constructed by reusing the training set with fixed normalization parameters. The high-resolution initial reconstructed wind field is obtained through one forward propagation of the model.

[0122] (2) Post-processing correction of physical residual constraints in the inference results:

[0123] Calculate the residuals of the full-grid hydrodynamic equations of the initial reconstructed wind field. For grid regions where the residuals exceed the preset threshold, wind speed and air pressure are successively corrected according to the preset correction step size until the full-grid residuals meet the preset convergence conditions or reach the maximum number of corrections. Output the corrected full-domain high-resolution wind field data.

[0124] (3) Wind farm post-processing and engineering output:

[0125] The corrected zonal and meridional wind speeds are vectorized to obtain the wind speed scalar and meteorological standard wind direction of the entire grid. Based on the geographical coordinates of the transmission line, the wind field data specific to the transmission line corridor is extracted from the wind field of the entire area and output to the transmission line wind deflection and galloping early warning system to complete the real-time wind field reconstruction.

[0126] An example application is provided:

[0127] S1: Delineating the computational domain for reconstructing the wind field, acquiring multi-source data, constructing and preprocessing the initial sparse input tensor:

[0128] S1.1 Wind field computational domain partitioning and adaptive mesh discretization:

[0129] Based on the geographical coordinates of the transmission line, a rectangular computational domain covering the entire transmission line and the surrounding wind field influence area is delineated. This computational domain is discretized into a two-dimensional grid to determine the analysis scope of the model. Simultaneously, an adaptive grid refinement mechanism for the transmission line corridor is adopted to optimize the accuracy of the core areas of interest. The grid division rules are as follows:

[0130]

[0131] in: : grid Spatial step size; Minimum grid step size, preset according to the accuracy requirements of wind field reconstruction, with a value range of 5m~50m; : grid The vertical distance to the centerline of the transmission line corridor; The preset half-width of the transmission line corridor is preset according to the voltage level of the transmission line, and the value ranges from 50m to 200m. : Grid step size growth factor, with a fixed value of 0.05~0.2;

[0132] Simultaneously acquire digital elevation model (DEM) data within the computational domain, complete the registration of terrain elevation data with the wind field computational grid, and obtain a terrain elevation matrix of the same dimension as the computational grid. ,in These are the width and height of the global grid, respectively. For grid The corresponding terrain elevation values, this matrix will be used simultaneously in subsequent stages of initial sparse input tensor construction, dynamic correction of fluid parameters, and adaptive adjustment of physical loss weights;

[0133] S1.2 Multi-source basic data acquisition and feature decomposition:

[0134] High-resolution meteorological grid data, sparse observation data from micro-meteorological sensors along the transmission line, and conductor vibration monitoring data corresponding to the computational domain are acquired. Time synchronization and spatial registration of multi-source data are completed. The specific solution rules are as follows:

[0135] Meteorological basic data processing: Extract wind speed, wind direction, air pressure, and ambient temperature data from meteorological data, and calculate the zonal wind speed based on the wind speed and wind direction data. Meridional wind speed This forms a basic meteorological grid dataset;

[0136] Conductor vibration inversion wind speed calculation: Extracting conductor vibration amplitude based on transmission line conductor micro-wind vibration / galloping monitoring data. Vibration frequency By inverting the wind speed components at the corresponding sampling points using the wind-induced vibration theoretical model, the solution formula is as follows:

[0137]

[0138] in: : Zonal and meridional wind speed components obtained from conductor vibration inversion; The outer diameter of the conductor is determined by the type of conductor used in the transmission line. The Strocha number is fixed at 0.2 for flow around a cylinder. The conductor vibration yaw angle was obtained from actual measurements using sensors.

[0139] Terrain Adaptation Data Correction: Based on terrain elevation data, the original meteorological data is corrected for terrain occlusion or acceleration effects to obtain terrain-adapted wind speed and air pressure data.

[0140]

[0141] in: : Zonal wind speed and air pressure data after terrain adaptation correction; Raw zonal wind speed and air pressure data measured by micro-meteorological sensors; Grid point elevation and reference elevation The difference; : Terrain slope correction coefficient, preset according to the terrain undulation of the calculation domain, with a value range of 0.1~0.5; Gravitational acceleration, with a fixed value. ; The gas constant for dry air has a fixed value. ; : Ambient temperature measured by the micro-weather sensor, in °C;

[0142] S1.3 Initial sparse input tensor construction:

[0143] Based on the actual spatial distribution of micro-meteorological sensors and vibration monitoring sensors along the transmission line, corresponding sampling point data are extracted from the registered multi-source grid data to simulate the true observation data of sparse sensors; coordinate system registration and grid mapping of sampling point data are completed to construct an initial sparse input tensor that meets the requirements of the model input channel;

[0144] The tensor channel dimension is It includes five channels: zonal wind speed, meridional wind speed, air pressure, terrain elevation, and conductor vibration characteristics, providing raw input data for subsequent preprocessing stages;

[0145] S1.4 Initial sparse input tensor normalization preprocessing:

[0146] The initial sparse input tensor constructed in S1.3 is normalized and preprocessed, and then encapsulated into a dimensional tensor according to the sensor arrangement order. The normalized initial sparse input tensor is given by the following formula:

[0147]

[0148] in: : The original feature data of the initial sparse input tensor; The maximum and minimum values ​​of the corresponding feature channels in the training set. This parameter will be fixed during the training phase and directly reused during the inference phase. : Standardized feature data after normalization; : These represent the latitudinal and longitudinal dimensions of the sparse sampling points, corresponding to the spatial distribution of the sensors, and the global grid dimension, respectively. To establish the mapping relationship, the model enhancement module completes the mapping from sparse dimensions to global dimensions;

[0149] S2: Construction of the PINO model of the physical information neural operator and the composite loss function:

[0150] Overall architecture of the S2.1PINO model:

[0151] The constructed PINO model consists of a cascaded lifting module, a core processing module, and a projection module:

[0152] Lifting module: includes mesh mapping unit and channel lifting unit:

[0153] Mesh mapping cells: preprocessed with S1 The initial sparse input tensor is used as the input, and a fully connected layer maps the sparsely sampled tensor to... A complete high-resolution tensor of dimensions, supplementing the global grid features outside of sparse sampling points;

[0154] Channel boosting unit: Projects a high-resolution tensor from the physical space onto a high-dimensional latent feature space through a convolutional layer, resulting in a feature space with dimension 1. High-dimensional feature tensors;

[0155] in: The number of channels in the high-dimensional latent feature space is preset to 64 / 128, and can be dynamically adjusted according to the size of the computational domain grid and the reconstruction accuracy requirements.

[0156] The core processing module consists of several cascaded Fourier neural operator units. Each unit performs frequency domain global spectral convolution and spatial domain linear transformation in parallel. The complete computation flow of a single operator unit is as follows:

[0157]

[0158]

[0159]

[0160] in: The Fourier neural operator unit takes a feature tensor as input and the output of the previous module or unit. The final output feature of the operator unit is input to the next level unit or projection module; : Modal truncation matrix, retaining only the first part of the spectral space Set one low-order modal component and the rest of the high-frequency components to 0. Cut off hyperparameters for learnable modes; The spectral domain can learn linear transformation weight matrices, and the integral operator for wind field evolution can be learned through end-to-end training. Hadamard product operation; : Non-linear activation function, using GELU activation function;

[0161] Projection Module: Employs a multilayer perceptron (MLP) structure consisting of two cascaded fully connected layers. Taking the high-dimensional features output from the core processing module as input, it maps the deeply processed high-dimensional feature tensor back to physical space, outputting... A high-resolution continuous wind field with dimensions including zonal wind speed, meridional wind speed, and air pressure;

[0162] S2.2 Construction of the composite loss function:

[0163] The composite loss function of the model is the weighted sum of the model's data loss term and physical loss term.

[0164] (1) General formula for composite loss:

[0165]

[0166] in: Data loss function; Physical loss function; The hyperparameter weights of the two losses are balanced and dynamically adjusted adaptively using training batches.

[0167]

[0168] in: The preset base weight is a fixed value, ranging from 0.2 to 0.8; The mean absolute value of the physical residuals across the entire grid in the current training batch; The mean absolute value of the fitting residuals between the model output and the ground truth data at all sampling points within the current training batch. The residual calculation rule is consistent with the data loss formula. The residual for a single sampling point is... ;

[0169] (2) Data loss function:

[0170]

[0171] in: The total number of pixels in the global grid. ; : Represents the width and height of the global grid, respectively; PINO model in global mesh The output shows the zonal wind speed, meridional wind speed, and air pressure values. Multi-source fused truth data (referred to as truth data):

[0172]

[0173]

[0174]

[0175] in: : Zonal wind speed, meridional wind speed, and air pressure data measured by micro-meteorological sensors; The wind speed component obtained by conductor vibration inversion in S1.2; S1.2: Wind speed and air pressure data after terrain adaptation correction; : The confidence weight corresponding to the data source, the weight sum is 1, and it is dynamically assigned by the data sampling accuracy and time synchronization.

[0176] (3) Physical loss function:

[0177] The physical loss term is constructed based on the residual constraints of the two-dimensional steady incompressible Navier-Stkes fluid dynamics governing equations and is embedded in the loss function in the form of the physical residuals of the equations:

[0178] Continuity equation residuals:

[0179]

[0180] Residual of momentum equation in the x-direction (zonal):

[0181]

[0182] Residual of momentum equation in the y-direction (meridian):

[0183]

[0184] Air density, using elevation-temperature adaptive dynamic correction:

[0185]

[0186] Air kinematic viscosity, with temperature-adaptive dynamic correction:

[0187]

[0188] in: The standard value for air density is a fixed value. ; The reference value for air kinematic viscosity under standard conditions is fixed at a certain value. ; Standard atmospheric pressure, with a fixed value. ; Standard temperature, with a fixed value. ; Real-time air pressure and ambient temperature measured by micro-meteorological sensors; The global mesh obtained in S1.1 The corresponding terrain elevation value; Temperature correction factor, with a fixed value. ;

[0189] Physical loss function:

[0190]

[0191]

[0192]

[0193] in: : These are the adaptive weights of the residuals of the continuity equation, the zonal momentum equation, and the meridional momentum equation, respectively; The average elevation of the entire transmission line; : The maximum elevation difference within the computational domain; The preset weight amplification factor has a fixed value of 0.5 to 2. Global Grid The vertical distance to the centerline of the transmission line corridor; The pre-set half-width of the transmission line corridor; The preset minimum value has a fixed value. This is used to avoid calculation errors caused by a denominator of 0;

[0194] S3: Iterative Training and Optimal Weight Selection of the PINO Model

[0195] Time-series partitioning of dataset S3.1:

[0196] A time series segmentation strategy was adopted, and the multi-source meteorological grid data acquired by S1 was divided into training set, validation set and test set in a ratio of 7:2:1 according to the time dimension. The time sequence was not disrupted during the segmentation process to avoid data leakage and ensure the model's ability to generalize and predict future wind fields.

[0197] At the same time, fix the maximum value of each feature channel in the training set. Minimum value As the only normalized parameter for S1 preprocessing and S4 inference preprocessing, it is not dynamically updated throughout the process;

[0198] S3.2 Model Iterative Training Optimization:

[0199] Using the training set data, the PINO model is iteratively trained using the Adam optimizer and a cosine annealing learning rate strategy to drive the model to learn from... Initial sparse input tensor to dimension The specific training rules for the nonlinear mapping operator with dense dimension output are as follows:

[0200] Optimizer hyperparameter settings: First-order moment estimation of the Adam optimizer, exponential decay rate Second-order moment estimation of exponential decay rate The weight decay coefficient is ;

[0201] Learning rate update strategy: A cosine annealing learning rate strategy is adopted, and the learning rate update formula is:

[0202]

[0203] in: : No. The learning rate for each iteration; These represent the maximum and minimum values ​​of the learning rate, with default values ​​of [value 1]. , ; : Maximum number of iterations, preset value is 100~2000;

[0204] Iterative training rule: Each training round uses the composite loss function defined by S2. The optimization objective is to minimize all weight parameters of the PINO model using backpropagation. After each training round, forward inference is performed on the validation set data, and the inference results, after physical residual correction, are used as pseudo-labels to supplement the data loss function of the next training round. In the calculation, the weight of the fitting residual in the non-sampling region is set to 0.1 times the weight of the sampling point to achieve weakly supervised learning and enhance the model's ability to fit the wind field across the entire domain and its physical consistency.

[0205] S3.3 Model Generalization Evaluation and Optimal Weight Preservation:

[0206] Using validation set data as input, the generalization ability is evaluated by monitoring the full validation loss of the model at unseen time steps. An early stopping mechanism is set: training is terminated when the validation set loss does not decrease for 20 consecutive iterations. Based on the validation set performance, the model weights with the lowest validation loss are selected as the optimal model weights and saved.

[0207] S4: Model Deployment and Real-time Wind Farm Reconstruction for Transmission Lines

[0208] S4.1 Inference Edge Model Deployment and Real-time Forward Inference:

[0209] The optimal PINO model, trained and saved by S3, is deployed at the inference end to acquire sparse observation data from micro-meteorological sensors and vibration sensors along the transmission line in real time. Following the preprocessing procedure of S1, the normalization parameters of the training set fixed by S3 are reused. , constructed as The dimensionality-normalized initial sparse input tensor is fed into the PINO model and a forward propagation is performed to quickly generate... Full-domain high-resolution initial reconstruction of the wind field;

[0210] S4.2 Post-processing correction of physical residual constraints in inference results:

[0211] For the initial reconstructed wind field output by the model's forward inference, perform physical residual iterative correction. The correction rule is completely consistent with the NS equation residuals defined in S2. The specific steps are as follows:

[0212] Calculate the continuity equation residuals for the entire grid of the initial wind field. 2. Residual of zonal momentum equation Residuals of the meridional momentum equation The physical residual distribution of the entire mesh is obtained;

[0213] Based on the physical residual distribution, for residuals exceeding a preset threshold For the grid region, an iterative correction algorithm based on residual weighting is used. The correction formula is as follows:

[0214]

[0215] in: : The number of iterations for correction, initially 0, incremented by 1 after each correction iteration; : No. Zonal wind speed, meridional wind speed, and air pressure values ​​for each iteration; The physical residual value of the current iteration step; The residual correction step size is fixed at 0.01 to 0.1. The preset physical residual convergence threshold is set according to the accuracy requirements of wind field reconstruction. ; : Maximum number of iterations for correction, default value is 20;

[0216] Iteration termination condition: The physical residual of the entire mesh is lower than a preset threshold. or reaching the maximum number of iterations. Outputs high-resolution wind field data across the entire domain after physical residual correction;

[0217] S4.3 Wind Farm Post-processing and Engineering Output:

[0218] The corrected zonal and meridional wind speeds are vector-synthesized to obtain the wind speed and direction field of the entire grid. The formula is as follows:

[0219]

[0220] in: : The synthesized full-grid wind speed scalar value; Composite wind direction, defined according to meteorological standards, is the direction from which the wind is coming, with due north as 0° and clockwise rotation as the positive direction; The arctangent function in the four quadrants has an output range of... When converting to meteorological wind direction, a corresponding angle mapping is required;

[0221] Finally, based on the geographical coordinates of the transmission line, wind field data specific to the transmission line corridor is extracted from the wind field of the entire region and output to the transmission line wind deviation and galloping early warning system, thus completing the full-process engineering implementation of real-time wind field reconstruction of the transmission line.

[0222] like Figure 2 As shown, the comparison between the model's wind field reconstruction results at a certain time point and the true value is presented, testing the model's ability to accurately reconstruct complex flow field structures under sparse input. The evaluation results show that the model exhibits extremely high reconstruction accuracy on all test sets: the root mean square error (RMSE) of zonal wind speed is 0.200970 m / s, and the mean absolute error (MAE) is 0.157966 m / s; the RMSE of meridional wind speed is 0.261515 m / s, and the MAE is 0.204126 m / s.

[0223] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for real-time wind farm reconstruction of transmission lines based on physical information neural operators, characterized in that: Includes the following steps: S1: Divide the wind field calculation domain, obtain meteorological grid data, terrain elevation data, and conductor vibration monitoring data along the transmission line within the wind field calculation domain; extract specific sampling point data from the multi-source data according to the actual spatial distribution of micro-meteorological sensors along the transmission line, construct an initial sparse input tensor, and perform preprocessing. S2: Construct a physical information neural operator PINO model containing Fourier neural operators, and use the weighted sum of the model data loss term and the physical loss term as the model composite loss function. The weights used to balance the two losses in the composite loss function are dynamically adjusted adaptively by training batches. The physical loss term is constructed based on the residual constraints of the fluid dynamics control equation. S3: The PINO model is trained iteratively using the preprocessed initial sparse input tensor as the input and the high-resolution continuous wind field data as the output. The model is trained iteratively with the goal of minimizing the composite loss function. The parameters of the PINO model are corrected by the backpropagation algorithm. The optimal model weights are selected and saved based on the generalization performance of the validation set. S4: Deploy the optimal PINO model saved after training at the inference end, input sparse observation data along the transmission line in real time and perform forward inference, perform physical residual constraint post-processing correction on the wind field data output by inference, and finally obtain the real-time reconstructed wind field of the entire transmission line through post-processing calculation.

2. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 1, characterized in that: S1 specifically includes the following: (1) Wind field computational domain partitioning and adaptive grid discretization: A rectangular computational domain covering the transmission line and the surrounding wind field influence area is delineated based on the geographical coordinates of the transmission line. The computational domain is discretized into a two-dimensional global grid. According to the adaptive densification rules of the transmission line corridor, the minimum spatial step size is adopted for the grid within the preset range of the line centerline, and the spatial step size is gradually increased by a fixed coefficient for the grid outside the preset range. Simultaneously, digital elevation model data within the computational domain is acquired, and the registration of the terrain elevation data with the global grid is completed to obtain a terrain elevation matrix with the same dimensions as the global grid. (2) Acquisition of multi-source basic data and feature decomposition: High-resolution meteorological grid data, sparse observation data from micro-meteorological sensors along the transmission line, and conductor vibration monitoring data are collected within the computational domain to achieve time synchronization and spatial registration of multi-source data; wind speed, wind direction, air pressure, and ambient temperature data are extracted from the meteorological data to calculate zonal and meridional wind speeds. Based on the amplitude and frequency of conductor vibration monitoring, the wind speed components of the corresponding sampling points are obtained by inversion through wind-induced vibration theory; the original meteorological data are corrected by combining topographic elevation data to obtain wind speed and air pressure data that are adapted to the effects of terrain shading and acceleration. (3) Sparse sampling point data extraction and initial sparse input tensor construction: Based on the actual spatial distribution of micro-meteorological sensors and vibration monitoring sensors, sampling point data is extracted from the registered multi-source grid data, coordinate registration and grid mapping are completed, and an initial sparse input tensor containing multiple channels including zonal wind speed, meridional wind speed, air pressure, terrain elevation, and conductor vibration characteristics is constructed. (4) Standardization preprocessing of initial sparse input tensors: The original feature data of the initial sparse input tensor is standardized to unify the value range of each feature data to a preset range. The processing is adapted and adjusted according to the extreme value range of the corresponding feature channel, and finally encapsulated into a standardized initial sparse input tensor that matches the spatial distribution dimension of the sensor.

3. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 1, characterized in that: S2 specifically includes the following: (1) Overall architecture construction of the PINO model: Construct a physical information neural operator model consisting of a cascaded lifting module, a core processing module, and a projection module; The boosting module maps the normalized initial sparse input tensor from the sparse sampling dimension to the global grid dimension through a fully connected layer, and then projects it to a high-dimensional latent feature space through a convolutional layer. The core processing module works through multi-level Fourier neural operator units. Each unit performs frequency domain processing on the input features by fast Fourier transform, learnable linear transform in the spectral domain, and inverse fast Fourier transform, as well as spatial domain processing by one-dimensional convolution. The two processing results are then fused with the input features and output after nonlinear activation. The projection module maps high-dimensional features back to physical space through two fully connected layers, outputting high-resolution continuous wind field data with full-domain dimensions; (2) Construction of the composite loss function: The weighted sum of the data loss term and the physical loss term is used as the composite loss function, where the weights of the two loss terms are dynamically adjusted according to the ratio of the mean of the full-grid physical residuals of the current training batch to the mean of the fitting residuals of the sampled data. Data loss term: The degree of fitting deviation between the global grid zonal wind speed, meridional wind speed, and air pressure output by the model and the multi-source fused true data is used as the data loss. The multi-source fusion true data is obtained by weighted fusion of meteorological sensor measured data, conductor vibration inversion data, and topographic correction data according to the confidence weight of each data source; Physical loss term: The degree of deviation of the residuals of the two-dimensional steady incompressible Navier-Stokes fluid dynamics governing equations is used as the physical loss constraint. The residuals of the continuity equation are calculated based on the combined deviation of the zonal wind speed along the zonal direction and the meridional wind speed along the meridional direction; the residuals of the zonal momentum equation are calculated based on the combined deviation of the convection term, pressure gradient term, and viscous diffusion term; the residuals of the meridional momentum equation are calculated based on the combined deviation of the convection term, meridional pressure gradient term, and viscous diffusion term. Air density is dynamically corrected based on real-time air pressure, ambient temperature, and terrain elevation; air kinematic viscosity is dynamically corrected based on the deviation of ambient temperature from standard temperature; adaptive weights are used for the residuals of each equation in physical loss; the weights of the residuals of the continuity equation are adjusted based on the deviation of grid elevation from the average elevation of the line; and the weights of the residuals of the momentum equation are adjusted based on the distance between the grid and the centerline of the line.

4. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 1, characterized in that: S3 specifically includes the following: (1) Temporal partitioning of the dataset: Multi-source meteorological grid data is divided into training set, validation set and test set according to time dimension in a fixed proportion. The maximum and minimum values ​​of each feature channel in the training set are fixed as fixed parameters for the whole process normalization. (2) Model iterative training optimization: An adaptive moment estimation optimizer combined with a cosine annealing learning rate strategy is used to train the model. The learning rate is dynamically adjusted between the maximum and minimum values ​​according to the preset cosine variation law based on the number of iterations. The optimization objective is to minimize the composite loss function, and all parameters of the model are corrected through the backpropagation algorithm. After each training round, the validation set data is inferred, and the inference results are corrected by physical residuals and used as pseudo-labels to supplement the data loss calculation to achieve weakly supervised learning. (3) Model generalization evaluation and optimal weight preservation: The generalization ability of the model is evaluated by the validation set loss, and the training is terminated by an early stopping mechanism. The model weights with the lowest validation set loss are selected as the optimal weights and saved.

5. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 1, characterized in that: S4 specifically includes the following: (1) Deployment of the inference model and real-time forward inference: After the training is completed, the model is loaded to the inference end, sparse sensor data is collected in real time, and the standardized initial sparse input tensor is constructed by reusing the training set with fixed normalization parameters. The high-resolution initial reconstructed wind field is obtained through one forward propagation of the model. (2) Post-processing correction of physical residual constraints in the inference results: Calculate the residuals of the full-grid hydrodynamic equations of the initial reconstructed wind field. For grid regions where the residuals exceed the preset threshold, wind speed and air pressure are successively corrected according to the preset correction step size until the full-grid residuals meet the preset convergence conditions or reach the maximum number of corrections. Output the corrected full-domain high-resolution wind field data. (3) Wind farm post-processing and engineering output: The corrected zonal and meridional wind speeds are vectorized to obtain the wind speed scalar and meteorological standard wind direction of the entire grid. Based on the geographical coordinates of the transmission line, the wind field data specific to the transmission line corridor is extracted from the wind field of the entire area and output to the transmission line wind deflection and galloping early warning system to complete the real-time wind field reconstruction.

6. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 2, characterized in that: The preprocessing of the initial sparse input tensor includes normalization: in: : The original feature data of the initial sparse input tensor; The maximum and minimum values ​​of the corresponding feature channels in the training set are fixed during the training phase and directly reused during the inference phase to ensure consistent distribution of training and inference data. : Standardized feature data after normalization; Then, based on the sensor arrangement order, it is packaged into a dimension of The standardized initial sparse input tensor, : These represent the latitudinal and longitudinal arrangement dimensions of the sparse sampling points, corresponding to the spatial distribution number of the sensors, and the global grid dimension, respectively. To establish a mapping relationship, the model enhancement module completes the mapping from sparse dimensions to global dimensions.

7. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 3, characterized in that: The composite loss function: in: Data loss function; Physical loss function; : The hyperparameter weights that balance the two losses; Data loss items: in: The total number of pixels in the global grid. ; : Represents the width and height of the global grid, respectively; PINO model in global mesh The output shows the zonal wind speed, meridional wind speed, and air pressure values. Multi-source fusion of truth data; Physical loss term: Constructed based on the residual constraints of the two-dimensional steady incompressible Navier-Stkes fluid dynamics governing equations, and embedded in the loss function in the form of physical residuals of the equations: Continuity equation residuals: Residual of momentum equation in the x-direction: y-direction momentum equation residual: Physical loss: in: This refers to the zonal wind speed, i.e., the x-direction. This represents the meridional wind speed, i.e., the y-direction. Air density; Air kinematic viscosity; : These are the adaptive weights of the residuals of the continuity equation, the latitudinal momentum equation, and the meridional momentum equation, respectively.

8. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 7, characterized in that: The hyperparameter weights that balance the two losses : in: : Preset fixed base weight values; The mean absolute value of the physical residuals across the entire grid in the current training batch; : The mean absolute value of the fitting residuals between the model output and the ground truth data at all sampling points within the current training batch.

9. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 7, characterized in that: air density Employing elevation-temperature adaptive dynamic correction: air kinematic viscosity Temperature adaptive dynamic correction is adopted: in: : The standard value of air density under standard conditions; : Reference value for air kinematic viscosity under standard conditions; Standard atmosphere; Standard temperature; Real-time air pressure and ambient temperature measured by micro-meteorological sensors; Global Grid The corresponding terrain elevation value; Gravitational acceleration; : Gas constant for dry air; Temperature correction factor.

10. The method for real-time wind farm reconstruction of transmission lines based on physical information neural operators according to claim 7, characterized in that: The adaptive weights of the continuity equation residuals, the zonal momentum equation residuals, and the meridional momentum equation residuals. : in: The average elevation of the entire transmission line; : The maximum elevation difference within the computational domain; : Preset weight amplification factor; Global Grid The corresponding terrain elevation value; Global Grid The vertical distance to the centerline of the transmission line corridor; The pre-set half-width of the transmission line corridor; : The preset minimum value.

Images